fix(cli): 6 cascading test regressions hidden behind client_integration gate

- compact flag: was parsed then discarded (`compact: _`) instead of passed to `run_turn_with_output` — hardcoded `false` meant --compact never took effect - piped stdin vs permission prompter: `read_piped_stdin()` consumed all stdin before `CliPermissionPrompter::decide()` could read interactive approval answers; now only consumes stdin as prompt context when permission mode is `DangerFullAccess` (fully unattended) - session resolver: `resolve_managed_session_path` and `list_managed_sessions` now fall back to the pre-isolation flat `.claw/sessions/` layout so legacy sessions remain accessible - help assertion: match on stable prefix after `/session delete` was added in batch 5 - prompt shorthand: fix copy-paste that changed expected prompt from "help me debug" to "$help overview" - mock parity harness: filter captured requests to `/v1/messages` path only, excluding count_tokens preflight calls added by `be561bf` All 6 failures were pre-existing but masked because `client_integration` always failed first (fixed in 8c6dfe5). Workspace: 810+ tests passing, 0 failing.
fix(api): restore local preflight guard ahead of count_tokens round-trip
2026-04-10 18:14:50 +08:00 · 2026-04-08 14:54:10 +09:00 · 2026-04-08 14:34:38 +09:00 · 2026-04-08 14:11:12 +09:00 · 2026-04-08 14:06:26 +09:00 · 2026-04-08 12:46:04 +09:00
34 changed files with 8740 additions and 403 deletions
--- a/ROADMAP.md
+++ b/ROADMAP.md
@@ -308,7 +308,7 @@ Priority order: P0 = blocks CI/green state, P1 = blocks integration wiring, P2 =
 19. **Subcommand help falls through into runtime/API path** — **done**: `claw doctor --help`, `claw status --help`, `claw sandbox --help`, and nested `mcp`/`skills` help are now intercepted locally without runtime/provider startup, with regression tests covering the direct CLI paths.
 20. **Session state classification gap (working vs blocked vs finished vs truly stale)** — **done**: agent manifests now derive machine states such as `working`, `blocked_background_job`, `blocked_merge_conflict`, `degraded_mcp`, `interrupted_transport`, `finished_pending_report`, and `finished_cleanable`, and terminal-state persistence records commit provenance plus derived state so downstream monitoring can distinguish quiet progress from truly idle sessions.
 21. **Resumed `/status` JSON parity gap** — dogfooding shows fresh `claw status --output-format json` now emits structured JSON, but resumed slash-command status still leaks through a text-shaped path in at least one dispatch path. Local CI-equivalent repro fails `rust/crates/rusty-claude-cli/tests/resume_slash_commands.rs::resumed_status_command_emits_structured_json_when_requested` with `expected value at line 1 column 1`, so resumed automation can receive text where JSON was explicitly requested. **Action:** unify fresh vs resumed `/status` rendering through one output-format contract and add regression coverage so resumed JSON output is guaranteed valid.
-22. **Opaque failure surface for session/runtime crashes** — repeated dogfood-facing failures can currently collapse to generic wrappers like `Something went wrong while processing your request. Please try again, or use /new to start a fresh session.` without exposing whether the fault was provider auth, session corruption, slash-command dispatch, render failure, or transport/runtime panic. This blocks fast self-recovery and turns actionable clawability bugs into blind retries. **Action:** preserve a short user-safe failure class (`provider_auth`, `session_load`, `command_dispatch`, `render`, `runtime_panic`, etc.), attach a local trace/session id, and ensure operators can jump from the chat-visible error to the exact failure log quickly.
+22. **Opaque failure surface for session/runtime crashes** — **done**: `safe_failure_class()` in `error.rs` classifies all API errors into 8 user-safe classes (`provider_auth`, `provider_internal`, `provider_retry_exhausted`, `provider_rate_limit`, `provider_transport`, `provider_error`, `context_window`, `runtime_io`). `format_user_visible_api_error` in `main.rs` attaches session ID + request trace ID to every user-visible error. Coverage in `opaque_provider_wrapper_surfaces_failure_class_session_and_trace` and 3 related tests.
 23. **`doctor --output-format json` check-level structure gap** — **done**: `claw doctor --output-format json` now keeps the human-readable `message`/`report` while also emitting structured per-check diagnostics (`name`, `status`, `summary`, `details`, plus typed fields like workspace paths and sandbox fallback data), with regression coverage in `output_format_contract.rs`.
 24. **Plugin lifecycle init/shutdown test flakes under workspace-parallel execution** — dogfooding surfaced that `build_runtime_runs_plugin_lifecycle_init_and_shutdown` can fail under `cargo test --workspace` while passing in isolation because sibling tests race on tempdir-backed shell init script paths. This is test brittleness rather than a code-path regression, but it still destabilizes CI confidence and wastes diagnosis cycles. **Action:** isolate temp resources per test robustly (unique dirs + no shared cwd assumptions), audit cleanup timing, and add a regression guard so the plugin lifecycle test remains stable under parallel workspace execution.
 26. **Resumed local-command JSON parity gap** — **done**: direct `claw --output-format json` already had structured renderers for `sandbox`, `mcp`, `skills`, `version`, and `init`, but resumed `claw --output-format json --resume <session> /…` paths still fell back to prose because resumed slash dispatch only emitted JSON for `/status`. Resumed `/sandbox`, `/mcp`, `/skills`, `/version`, and `/init` now reuse the same JSON envelopes as their direct CLI counterparts, with regression coverage in `rust/crates/rusty-claude-cli/tests/resume_slash_commands.rs` and `rust/crates/rusty-claude-cli/tests/output_format_contract.rs`.
@@ -385,3 +385,77 @@ to:
 - a **claw-native execution runtime**
 - an **event-native orchestration substrate**
 - a **plugin/hook-first autonomous coding harness**
+
+## Deployment Architecture Gap (filed from dogfood 2026-04-08)
+
+### WorkerState is in the runtime; /state is NOT in opencode serve
+
+**Root cause discovered during batch 8 dogfood.**
+
+`worker_boot.rs` has a solid `WorkerStatus` state machine (`Spawning → TrustRequired → ReadyForPrompt → Running → Finished/Failed`). It is exported from `runtime/src/lib.rs` as a public API. But claw-code is a **plugin** loaded inside the `opencode` binary — it cannot add HTTP routes to `opencode serve`. The HTTP server is 100% owned by the upstream opencode process (v1.3.15).
+
+**Impact:** There is no way to `curl localhost:4710/state` and get back a JSON `WorkerStatus`. Any such endpoint would require either:
+1. Upstreaming a `/state` route into opencode's HTTP server (requires a PR to sst/opencode), or
+2. Writing a sidecar HTTP process that queries the `WorkerRegistry` in-process (possible but fragile), or
+3. Writing `WorkerStatus` to a well-known file path (`.claw/worker-state.json`) that an external observer can poll.
+
+**Recommended path:** Option 3 — emit `WorkerStatus` transitions to `.claw/worker-state.json` on every state change. This is purely within claw-code's plugin scope, requires no upstream changes, and gives clawhip a file it can poll to distinguish a truly stalled worker from a quiet-but-progressing one.
+
+**Action item:** Wire `WorkerRegistry::transition()` to atomically write `.claw/worker-state.json` on every state transition. Add a `claw state` CLI subcommand that reads and prints this file. Add regression test.
+
+**Prior session note:** A previous session summary claimed commit `0984cca` landed a `/state` HTTP endpoint via axum. This was incorrect — no such commit exists on main, axum is not a dependency, and the HTTP server is not ours. The actual work that exists: `worker_boot.rs` with `WorkerStatus` enum + `WorkerRegistry`, fully wired into `runtime/src/lib.rs` as public exports.
+
+## Startup Friction Gap: No Default trusted_roots in Settings (filed 2026-04-08)
+
+### Every lane starts with manual trust babysitting unless caller explicitly passes roots
+
+**Root cause discovered during direct dogfood of WorkerCreate tool.**
+
+`WorkerCreate` accepts a `trusted_roots: Vec<String>` parameter. If the caller omits it (or passes `[]`), every new worker immediately enters `TrustRequired` and stalls — requiring manual intervention to advance to `ReadyForPrompt`. There is no mechanism to configure a default allowlist in `settings.json` or `.claw/settings.json`.
+
+**Impact:** Batch tooling (clawhip, lane orchestrators) must pass `trusted_roots` explicitly on every `WorkerCreate` call. If a batch script forgets the field, all workers in that batch stall silently at `trust_required`. This was the root cause of several "batch 8 lanes not advancing" incidents.
+
+**Recommended fix:**
+1. Add a `trusted_roots` field to `RuntimeConfig` (or a nested `[trust]` table), loaded via `ConfigLoader`.
+2. In `WorkerRegistry::spawn_worker()`, merge config-level `trusted_roots` with any per-call overrides.
+3. Default: empty list (safest). Users opt in by adding their repo paths to settings.
+4. Update `config_validate` schema with the new field.
+
+**Action item:** Wire `RuntimeConfig::trusted_roots()` → `WorkerRegistry::spawn_worker()` default. Cover with test: config with `trusted_roots = ["/tmp"]` → spawning worker in `/tmp/x` auto-resolves trust without caller passing the field.
+
+## Observability Transport Decision (filed 2026-04-08)
+
+### Canonical state surface: CLI/file-based. HTTP endpoint deferred.
+
+**Decision:** `claw state` reading `.claw/worker-state.json` is the **blessed observability contract** for clawhip and downstream tooling. This is not a stepping-stone — it is the supported surface. Build against it.
+
+**Rationale:**
+- claw-code is a plugin running inside the opencode binary. It cannot add HTTP routes to `opencode serve` — that server belongs to upstream sst/opencode.
+- The file-based surface is fully within plugin scope: `emit_state_file()` in `worker_boot.rs` writes atomically on every `WorkerStatus` transition.
+- `claw state --output-format json` gives clawhip everything it needs: `status`, `is_ready`, `seconds_since_update`, `trust_gate_cleared`, `last_event`, `updated_at`.
+- Polling a local file has lower latency and fewer failure modes than an HTTP round-trip to a sidecar.
+- An HTTP state endpoint would require either (a) upstreaming a route to sst/opencode — a multi-week PR cycle with no guarantee of acceptance — or (b) a sidecar process that queries `WorkerRegistry` in-process, which is fragile and adds an extra failure domain.
+
+**What downstream tooling (clawhip) should do:**
+1. After `WorkerCreate`, poll `.claw/worker-state.json` (or run `claw state --output-format json`) in the worker's CWD at whatever interval makes sense (e.g. 5s).
+2. Trust `seconds_since_update > 60` in `trust_required` status as the stall signal.
+3. Call `WorkerResolveTrust` tool to unblock, or `WorkerRestart` to reset.
+
+**HTTP endpoint tracking:** Not scheduled. If a concrete use case emerges that file polling cannot serve (e.g. remote workers over a network boundary), open a new issue to upstream a `/worker/state` route to sst/opencode at that time. Until then: file/CLI is canonical.
+
+## Provider Routing: Model-Name Prefix Must Win Over Env-Var Presence (fixed 2026-04-08, `0530c50`)
+
+### `openai/gpt-4.1-mini` was silently misrouted to Anthropic when ANTHROPIC_API_KEY was set
+
+**Root cause:** `metadata_for_model` returned `None` for any model not matching `claude` or `grok` prefix.
+`detect_provider_kind` then fell through to auth-sniffer order: first `has_auth_from_env_or_saved()` (Anthropic), then `OPENAI_API_KEY`, then `XAI_API_KEY`.
+
+If `ANTHROPIC_API_KEY` was present in the environment (e.g. user has both Anthropic and OpenRouter configured), any unknown model — including explicitly namespaced ones like `openai/gpt-4.1-mini` — was silently routed to the Anthropic client, which then failed with `missing Anthropic credentials` or a confusing 402/auth error rather than routing to OpenAI-compatible.
+
+**Fix:** Added explicit prefix checks in `metadata_for_model`:
+- `openai/` prefix → `ProviderKind::OpenAi`
+- `gpt-` prefix → `ProviderKind::OpenAi`
+
+Model name prefix now wins unconditionally over env-var presence. Regression test locked in: `providers::tests::openai_namespaced_model_routes_to_openai_not_anthropic`.
+
+**Lesson:** Auth-sniffer fallback order is fragile. Any new provider added in the future should be registered in `metadata_for_model` via a model-name prefix, not left to env-var order. This is the canonical extension point.
--- a/USAGE.md
+++ b/USAGE.md
@@ -153,6 +153,133 @@ cd rust
 ./target/debug/claw --model "openai/gpt-4.1-mini" prompt "summarize this repository in one sentence"
 ```

+### Alibaba DashScope (Qwen)
+
+For Qwen models via Alibaba's native DashScope API (higher rate limits than OpenRouter):
+
+```bash
+export DASHSCOPE_API_KEY="sk-..."
+
+cd rust
+./target/debug/claw --model "qwen/qwen-max" prompt "hello"
+# or bare:
+./target/debug/claw --model "qwen-plus" prompt "hello"
+```
+
+Model names starting with `qwen/` or `qwen-` are automatically routed to the DashScope compatible-mode endpoint (`https://dashscope.aliyuncs.com/compatible-mode/v1`). You do **not** need to set `OPENAI_BASE_URL` or unset `ANTHROPIC_API_KEY` — the model prefix wins over the ambient credential sniffer.
+
+Reasoning variants (`qwen-qwq-*`, `qwq-*`, `*-thinking`) automatically strip `temperature`/`top_p`/`frequency_penalty`/`presence_penalty` before the request hits the wire (these params are rejected by reasoning models).
+
+## Supported Providers & Models
+
+`claw` has three built-in provider backends. The provider is selected automatically based on the model name, falling back to whichever credential is present in the environment.
+
+### Provider matrix
+
+| Provider | Protocol | Auth env var(s) | Base URL env var | Default base URL |
+|---|---|---|---|---|
+| **Anthropic** (direct) | Anthropic Messages API | `ANTHROPIC_API_KEY` or `ANTHROPIC_AUTH_TOKEN` or OAuth (`claw login`) | `ANTHROPIC_BASE_URL` | `https://api.anthropic.com` |
+| **xAI** | OpenAI-compatible | `XAI_API_KEY` | `XAI_BASE_URL` | `https://api.x.ai/v1` |
+| **OpenAI-compatible** | OpenAI Chat Completions | `OPENAI_API_KEY` | `OPENAI_BASE_URL` | `https://api.openai.com/v1` |
+| **DashScope** (Alibaba) | OpenAI-compatible | `DASHSCOPE_API_KEY` | `DASHSCOPE_BASE_URL` | `https://dashscope.aliyuncs.com/compatible-mode/v1` |
+
+The OpenAI-compatible backend also serves as the gateway for **OpenRouter**, **Ollama**, and any other service that speaks the OpenAI `/v1/chat/completions` wire format — just point `OPENAI_BASE_URL` at the service.
+
+**Model-name prefix routing:** If a model name starts with `openai/`, `gpt-`, `qwen/`, or `qwen-`, the provider is selected by the prefix regardless of which env vars are set. This prevents accidental misrouting to Anthropic when multiple credentials exist in the environment.
+
+### Tested models and aliases
+
+These are the models registered in the built-in alias table with known token limits:
+
+| Alias | Resolved model name | Provider | Max output tokens | Context window |
+|---|---|---|---|---|
+| `opus` | `claude-opus-4-6` | Anthropic | 32 000 | 200 000 |
+| `sonnet` | `claude-sonnet-4-6` | Anthropic | 64 000 | 200 000 |
+| `haiku` | `claude-haiku-4-5-20251213` | Anthropic | 64 000 | 200 000 |
+| `grok` / `grok-3` | `grok-3` | xAI | 64 000 | 131 072 |
+| `grok-mini` / `grok-3-mini` | `grok-3-mini` | xAI | 64 000 | 131 072 |
+| `grok-2` | `grok-2` | xAI | — | — |
+
+Any model name that does not match an alias is passed through verbatim. This is how you use OpenRouter model slugs (`openai/gpt-4.1-mini`), Ollama tags (`llama3.2`), or full Anthropic model IDs (`claude-sonnet-4-20250514`).
+
+### User-defined aliases
+
+You can add custom aliases in any settings file (`~/.claw/settings.json`, `.claw/settings.json`, or `.claw/settings.local.json`):
+
+```json
+{
+  "aliases": {
+    "fast": "claude-haiku-4-5-20251213",
+    "smart": "claude-opus-4-6",
+    "cheap": "grok-3-mini"
+  }
+}
+```
+
+Local project settings override user-level settings. Aliases resolve through the built-in table, so `"fast": "haiku"` also works.
+
+### How provider detection works
+
+1. If the resolved model name starts with `claude` → Anthropic.
+2. If it starts with `grok` → xAI.
+3. Otherwise, `claw` checks which credential is set: `ANTHROPIC_API_KEY`/`ANTHROPIC_AUTH_TOKEN` first, then `OPENAI_API_KEY`, then `XAI_API_KEY`.
+4. If nothing matches, it defaults to Anthropic.
+
+## FAQ
+
+### What about Codex?
+
+The name "codex" appears in the Claw Code ecosystem but it does **not** refer to OpenAI Codex (the code-generation model). Here is what it means in this project:
+
+- **`oh-my-codex` (OmX)** is the workflow and plugin layer that sits on top of `claw`. It provides planning modes, parallel multi-agent execution, notification routing, and other automation features. See [PHILOSOPHY.md](./PHILOSOPHY.md) and the [oh-my-codex repo](https://github.com/Yeachan-Heo/oh-my-codex).
+- **`.codex/` directories** (e.g. `.codex/skills`, `.codex/agents`, `.codex/commands`) are legacy lookup paths that `claw` still scans alongside the primary `.claw/` directories.
+- **`CODEX_HOME`** is an optional environment variable that points to a custom root for user-level skill and command lookups.
+
+`claw` does **not** support OpenAI Codex sessions, the Codex CLI, or Codex session import/export. If you need to use OpenAI models (like GPT-4.1), configure the OpenAI-compatible provider as shown above in the [OpenAI-compatible endpoint](#openai-compatible-endpoint) and [OpenRouter](#openrouter) sections.
+
+## HTTP proxy support
+
+`claw` honours the standard `HTTP_PROXY`, `HTTPS_PROXY`, and `NO_PROXY` environment variables (both upper- and lower-case spellings are accepted) when issuing outbound requests to Anthropic, OpenAI-, and xAI-compatible endpoints. Set them before launching the CLI and the underlying `reqwest` client will be configured automatically.
+
+### Environment variables
+
+```bash
+export HTTPS_PROXY="http://proxy.corp.example:3128"
+export HTTP_PROXY="http://proxy.corp.example:3128"
+export NO_PROXY="localhost,127.0.0.1,.corp.example"
+
+cd rust
+./target/debug/claw prompt "hello via the corporate proxy"
+```
+
+### Programmatic `proxy_url` config option
+
+As an alternative to per-scheme environment variables, the `ProxyConfig` type exposes a `proxy_url` field that acts as a single catch-all proxy for both HTTP and HTTPS traffic. When `proxy_url` is set it takes precedence over the separate `http_proxy` and `https_proxy` fields.
+
+```rust
+use api::{build_http_client_with, ProxyConfig};
+
+// From a single unified URL (config file, CLI flag, etc.)
+let config = ProxyConfig::from_proxy_url("http://proxy.corp.example:3128");
+let client = build_http_client_with(&config).expect("proxy client");
+
+// Or set the field directly alongside NO_PROXY
+let config = ProxyConfig {
+    proxy_url: Some("http://proxy.corp.example:3128".to_string()),
+    no_proxy: Some("localhost,127.0.0.1".to_string()),
+    ..ProxyConfig::default()
+};
+let client = build_http_client_with(&config).expect("proxy client");
+```
+
+### Notes
+
+- When both `HTTPS_PROXY` and `HTTP_PROXY` are set, the secure proxy applies to `https://` URLs and the plain proxy applies to `http://` URLs.
+- `proxy_url` is a unified alternative: when set, it applies to both `http://` and `https://` destinations, overriding the per-scheme fields.
+- `NO_PROXY` accepts a comma-separated list of host suffixes (for example `.corp.example`) and IP literals.
+- Empty values are treated as unset, so leaving `HTTPS_PROXY=""` in your shell will not enable a proxy.
+- If a proxy URL cannot be parsed, `claw` falls back to a direct (no-proxy) client so existing workflows keep working; double-check the URL if you expected the request to be tunnelled.
+
 ## Common operational commands

 ```bash
--- a/install.sh
+++ b/install.sh
@@ -0,0 +1,394 @@
+#!/usr/bin/env bash
+# Claw Code installer
+#
+# Detects the host OS, verifies the Rust toolchain (rustc + cargo),
+# builds the `claw` binary from the `rust/` workspace, and runs a
+# post-install verification step. Supports Linux, macOS, and WSL.
+#
+# Usage:
+#   ./install.sh                # debug build (fast, default)
+#   ./install.sh --release      # optimized release build
+#   ./install.sh --no-verify    # skip post-install verification
+#   ./install.sh --help         # print usage
+#
+# Environment overrides:
+#   CLAW_BUILD_PROFILE=debug|release   same as --release toggle
+#   CLAW_SKIP_VERIFY=1                 same as --no-verify
+
+set -euo pipefail
+
+# ---------------------------------------------------------------------------
+# Pretty printing
+# ---------------------------------------------------------------------------
+
+if [ -t 1 ] && command -v tput >/dev/null 2>&1 && [ "$(tput colors 2>/dev/null || echo 0)" -ge 8 ]; then
+    COLOR_RESET="$(tput sgr0)"
+    COLOR_BOLD="$(tput bold)"
+    COLOR_DIM="$(tput dim)"
+    COLOR_RED="$(tput setaf 1)"
+    COLOR_GREEN="$(tput setaf 2)"
+    COLOR_YELLOW="$(tput setaf 3)"
+    COLOR_BLUE="$(tput setaf 4)"
+    COLOR_CYAN="$(tput setaf 6)"
+else
+    COLOR_RESET=""
+    COLOR_BOLD=""
+    COLOR_DIM=""
+    COLOR_RED=""
+    COLOR_GREEN=""
+    COLOR_YELLOW=""
+    COLOR_BLUE=""
+    COLOR_CYAN=""
+fi
+
+CURRENT_STEP=0
+TOTAL_STEPS=6
+
+step() {
+    CURRENT_STEP=$((CURRENT_STEP + 1))
+    printf '\n%s[%d/%d]%s %s%s%s\n' \
+        "${COLOR_BLUE}" "${CURRENT_STEP}" "${TOTAL_STEPS}" "${COLOR_RESET}" \
+        "${COLOR_BOLD}" "$1" "${COLOR_RESET}"
+}
+
+info()  { printf '%s  ->%s %s\n' "${COLOR_CYAN}" "${COLOR_RESET}" "$1"; }
+ok()    { printf '%s  ok%s %s\n' "${COLOR_GREEN}" "${COLOR_RESET}" "$1"; }
+warn()  { printf '%s  warn%s %s\n' "${COLOR_YELLOW}" "${COLOR_RESET}" "$1"; }
+error() { printf '%s  error%s %s\n' "${COLOR_RED}" "${COLOR_RESET}" "$1" 1>&2; }
+
+print_banner() {
+    printf '%s' "${COLOR_BOLD}"
+    cat <<'EOF'
+   ____  _                   ____          _
+  / ___|| |  __ _ __      __ / ___|___   __| | ___
+ | |    | | / _` |\ \ /\ / /| |   / _ \ / _` |/ _ \
+ | |___ | || (_| | \ V  V / | |__| (_) | (_| |  __/
+  \____||_| \__,_|  \_/\_/   \____\___/ \__,_|\___|
+EOF
+    printf '%s\n' "${COLOR_RESET}"
+    printf '%sClaw Code installer%s\n' "${COLOR_DIM}" "${COLOR_RESET}"
+}
+
+print_usage() {
+    cat <<'EOF'
+Usage: ./install.sh [options]
+
+Options:
+  --release       Build the optimized release profile (slower, smaller binary).
+  --debug         Build the debug profile (default, faster compile).
+  --no-verify     Skip the post-install verification step.
+  -h, --help      Show this help text and exit.
+
+Environment overrides:
+  CLAW_BUILD_PROFILE   debug | release
+  CLAW_SKIP_VERIFY     set to 1 to skip verification
+EOF
+}
+
+# ---------------------------------------------------------------------------
+# Argument parsing
+# ---------------------------------------------------------------------------
+
+BUILD_PROFILE="${CLAW_BUILD_PROFILE:-debug}"
+SKIP_VERIFY="${CLAW_SKIP_VERIFY:-0}"
+
+while [ "$#" -gt 0 ]; do
+    case "$1" in
+        --release)
+            BUILD_PROFILE="release"
+            ;;
+        --debug)
+            BUILD_PROFILE="debug"
+            ;;
+        --no-verify)
+            SKIP_VERIFY="1"
+            ;;
+        -h|--help)
+            print_usage
+            exit 0
+            ;;
+        *)
+            error "unknown argument: $1"
+            print_usage
+            exit 2
+            ;;
+    esac
+    shift
+done
+
+case "${BUILD_PROFILE}" in
+    debug|release) ;;
+    *)
+        error "invalid build profile: ${BUILD_PROFILE} (expected debug or release)"
+        exit 2
+        ;;
+esac
+
+# ---------------------------------------------------------------------------
+# Troubleshooting hints
+# ---------------------------------------------------------------------------
+
+print_troubleshooting() {
+    cat <<EOF
+
+${COLOR_BOLD}Troubleshooting${COLOR_RESET}
+${COLOR_DIM}---------------${COLOR_RESET}
+
+  ${COLOR_BOLD}1. Rust toolchain missing${COLOR_RESET}
+     Install Rust via rustup:
+       curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
+     Then reload your shell or run:
+       source "\$HOME/.cargo/env"
+
+  ${COLOR_BOLD}2. Linux: missing system packages${COLOR_RESET}
+     The build needs git, pkg-config, and OpenSSL headers.
+     Debian/Ubuntu:
+       sudo apt-get update && sudo apt-get install -y \\
+         git pkg-config libssl-dev ca-certificates build-essential
+     Fedora/RHEL:
+       sudo dnf install -y git pkgconf-pkg-config openssl-devel gcc
+     Arch:
+       sudo pacman -S --needed git pkgconf openssl base-devel
+
+  ${COLOR_BOLD}3. macOS: missing Xcode CLT${COLOR_RESET}
+     Install the command line tools:
+       xcode-select --install
+
+  ${COLOR_BOLD}4. Windows users${COLOR_RESET}
+     Run this script from inside a WSL distro (Ubuntu/Debian recommended).
+     Native Windows builds are not supported by this installer.
+
+  ${COLOR_BOLD}5. Build fails partway through${COLOR_RESET}
+     Try a clean build:
+       cd rust && cargo clean && cargo build --workspace
+     If the failure mentions ring/openssl, double check step 2.
+
+  ${COLOR_BOLD}6. 'claw' not found after install${COLOR_RESET}
+     The binary lives at:
+       rust/target/${BUILD_PROFILE}/claw
+     Add it to your PATH or invoke it with the full path.
+
+EOF
+}
+
+trap 'rc=$?; if [ "$rc" -ne 0 ]; then error "installation failed (exit ${rc})"; print_troubleshooting; fi' EXIT
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+require_cmd() {
+    command -v "$1" >/dev/null 2>&1
+}
+
+# ---------------------------------------------------------------------------
+# Step 1: detect OS / arch / WSL
+# ---------------------------------------------------------------------------
+
+print_banner
+step "Detecting host environment"
+
+UNAME_S="$(uname -s 2>/dev/null || echo unknown)"
+UNAME_M="$(uname -m 2>/dev/null || echo unknown)"
+OS_FAMILY="unknown"
+IS_WSL="0"
+
+case "${UNAME_S}" in
+    Linux*)
+        OS_FAMILY="linux"
+        if grep -qiE 'microsoft|wsl' /proc/version 2>/dev/null; then
+            IS_WSL="1"
+        fi
+        ;;
+    Darwin*)
+        OS_FAMILY="macos"
+        ;;
+    MINGW*|MSYS*|CYGWIN*)
+        OS_FAMILY="windows-shell"
+        ;;
+esac
+
+info "uname:        ${UNAME_S} ${UNAME_M}"
+info "os family:    ${OS_FAMILY}"
+if [ "${IS_WSL}" = "1" ]; then
+    info "wsl:          yes"
+fi
+
+case "${OS_FAMILY}" in
+    linux|macos)
+        ok "supported platform detected"
+        ;;
+    windows-shell)
+        error "Detected a native Windows shell (MSYS/Cygwin/MinGW)."
+        error "Please re-run this script from inside a WSL distribution."
+        exit 1
+        ;;
+    *)
+        error "Unsupported or unknown OS: ${UNAME_S}"
+        error "Supported: Linux, macOS, and Windows via WSL."
+        exit 1
+        ;;
+esac
+
+# ---------------------------------------------------------------------------
+# Step 2: locate the Rust workspace
+# ---------------------------------------------------------------------------
+
+step "Locating the Rust workspace"
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+RUST_DIR="${SCRIPT_DIR}/rust"
+
+if [ ! -d "${RUST_DIR}" ]; then
+    error "Could not find rust/ workspace next to install.sh"
+    error "Expected: ${RUST_DIR}"
+    exit 1
+fi
+
+if [ ! -f "${RUST_DIR}/Cargo.toml" ]; then
+    error "Missing ${RUST_DIR}/Cargo.toml — repository layout looks unexpected."
+    exit 1
+fi
+
+ok "workspace at ${RUST_DIR}"
+
+# ---------------------------------------------------------------------------
+# Step 3: prerequisite checks
+# ---------------------------------------------------------------------------
+
+step "Checking prerequisites"
+
+MISSING_PREREQS=0
+
+if require_cmd rustc; then
+    RUSTC_VERSION="$(rustc --version 2>/dev/null || echo 'unknown')"
+    ok "rustc found: ${RUSTC_VERSION}"
+else
+    error "rustc not found in PATH"
+    MISSING_PREREQS=1
+fi
+
+if require_cmd cargo; then
+    CARGO_VERSION="$(cargo --version 2>/dev/null || echo 'unknown')"
+    ok "cargo found: ${CARGO_VERSION}"
+else
+    error "cargo not found in PATH"
+    MISSING_PREREQS=1
+fi
+
+if require_cmd git; then
+    ok "git found:  $(git --version 2>/dev/null || echo 'unknown')"
+else
+    warn "git not found — some workflows (login, session export) may degrade"
+fi
+
+if [ "${OS_FAMILY}" = "linux" ]; then
+    if require_cmd pkg-config; then
+        ok "pkg-config found"
+    else
+        warn "pkg-config not found — may be required for OpenSSL-linked crates"
+    fi
+fi
+
+if [ "${OS_FAMILY}" = "macos" ]; then
+    if ! require_cmd cc && ! xcode-select -p >/dev/null 2>&1; then
+        warn "Xcode command line tools not detected — run: xcode-select --install"
+    fi
+fi
+
+if [ "${MISSING_PREREQS}" -ne 0 ]; then
+    error "Missing required tools. See troubleshooting below."
+    exit 1
+fi
+
+# ---------------------------------------------------------------------------
+# Step 4: build the workspace
+# ---------------------------------------------------------------------------
+
+step "Building the claw workspace (${BUILD_PROFILE})"
+
+CARGO_FLAGS=("build" "--workspace")
+if [ "${BUILD_PROFILE}" = "release" ]; then
+    CARGO_FLAGS+=("--release")
+fi
+
+info "running: cargo ${CARGO_FLAGS[*]}"
+info "this may take a few minutes on the first build"
+
+(
+    cd "${RUST_DIR}"
+    CARGO_TERM_COLOR="${CARGO_TERM_COLOR:-always}" cargo "${CARGO_FLAGS[@]}"
+)
+
+CLAW_BIN="${RUST_DIR}/target/${BUILD_PROFILE}/claw"
+
+if [ ! -x "${CLAW_BIN}" ]; then
+    error "Expected binary not found at ${CLAW_BIN}"
+    error "The build reported success but the binary is missing — check cargo output above."
+    exit 1
+fi
+
+ok "built ${CLAW_BIN}"
+
+# ---------------------------------------------------------------------------
+# Step 5: post-install verification
+# ---------------------------------------------------------------------------
+
+step "Verifying the installed binary"
+
+if [ "${SKIP_VERIFY}" = "1" ]; then
+    warn "verification skipped (--no-verify or CLAW_SKIP_VERIFY=1)"
+else
+    info "running: claw --version"
+    if VERSION_OUT="$("${CLAW_BIN}" --version 2>&1)"; then
+        ok "claw --version -> ${VERSION_OUT}"
+    else
+        error "claw --version failed:"
+        printf '%s\n' "${VERSION_OUT}" 1>&2
+        exit 1
+    fi
+
+    info "running: claw --help (smoke test)"
+    if "${CLAW_BIN}" --help >/dev/null 2>&1; then
+        ok "claw --help responded"
+    else
+        error "claw --help failed"
+        exit 1
+    fi
+fi
+
+# ---------------------------------------------------------------------------
+# Step 6: next steps
+# ---------------------------------------------------------------------------
+
+step "Next steps"
+
+cat <<EOF
+${COLOR_GREEN}Claw Code is built and ready.${COLOR_RESET}
+
+  Binary:  ${COLOR_BOLD}${CLAW_BIN}${COLOR_RESET}
+  Profile: ${BUILD_PROFILE}
+
+Try it out:
+
+  ${COLOR_DIM}# interactive REPL${COLOR_RESET}
+  ${CLAW_BIN}
+
+  ${COLOR_DIM}# one-shot prompt${COLOR_RESET}
+  ${CLAW_BIN} prompt "summarize this repository"
+
+  ${COLOR_DIM}# health check (run /doctor inside the REPL)${COLOR_RESET}
+  ${CLAW_BIN}
+  /doctor
+
+Authentication:
+
+  export ANTHROPIC_API_KEY="sk-ant-..."
+  ${COLOR_DIM}# or use OAuth:${COLOR_RESET}
+  ${CLAW_BIN} login
+
+For deeper docs, see USAGE.md and rust/README.md.
+EOF
+
+# clear the failure trap on clean exit
+trap - EXIT
--- a/rust/Cargo.lock
+++ b/rust/Cargo.lock
@@ -1580,6 +1580,7 @@ version = "0.1.0"
 dependencies = [
 "api",
 "commands",
+ "flate2",
 "plugins",
 "reqwest",
 "runtime",
--- a/rust/crates/api/src/http_client.rs
+++ b/rust/crates/api/src/http_client.rs
@@ -0,0 +1,344 @@
+use crate::error::ApiError;
+
+const HTTP_PROXY_KEYS: [&str; 2] = ["HTTP_PROXY", "http_proxy"];
+const HTTPS_PROXY_KEYS: [&str; 2] = ["HTTPS_PROXY", "https_proxy"];
+const NO_PROXY_KEYS: [&str; 2] = ["NO_PROXY", "no_proxy"];
+
+/// Snapshot of the proxy-related environment variables that influence the
+/// outbound HTTP client. Captured up front so callers can inspect, log, and
+/// test the resolved configuration without re-reading the process environment.
+///
+/// When `proxy_url` is set it acts as a single catch-all proxy for both
+/// HTTP and HTTPS traffic, taking precedence over the per-scheme fields.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct ProxyConfig {
+    pub http_proxy: Option<String>,
+    pub https_proxy: Option<String>,
+    pub no_proxy: Option<String>,
+    /// Optional unified proxy URL that applies to both HTTP and HTTPS.
+    /// When set, this takes precedence over `http_proxy` and `https_proxy`.
+    pub proxy_url: Option<String>,
+}
+
+impl ProxyConfig {
+    /// Read proxy settings from the live process environment, honouring both
+    /// the upper- and lower-case spellings used by curl, git, and friends.
+    #[must_use]
+    pub fn from_env() -> Self {
+        Self::from_lookup(|key| std::env::var(key).ok())
+    }
+
+    /// Create a proxy configuration from a single URL that applies to both
+    /// HTTP and HTTPS traffic. This is the config-file alternative to setting
+    /// `HTTP_PROXY` and `HTTPS_PROXY` environment variables separately.
+    #[must_use]
+    pub fn from_proxy_url(url: impl Into<String>) -> Self {
+        Self {
+            proxy_url: Some(url.into()),
+            ..Self::default()
+        }
+    }
+
+    fn from_lookup<F>(mut lookup: F) -> Self
+    where
+        F: FnMut(&str) -> Option<String>,
+    {
+        Self {
+            http_proxy: first_non_empty(&HTTP_PROXY_KEYS, &mut lookup),
+            https_proxy: first_non_empty(&HTTPS_PROXY_KEYS, &mut lookup),
+            no_proxy: first_non_empty(&NO_PROXY_KEYS, &mut lookup),
+            proxy_url: None,
+        }
+    }
+
+    #[must_use]
+    pub fn is_empty(&self) -> bool {
+        self.proxy_url.is_none() && self.http_proxy.is_none() && self.https_proxy.is_none()
+    }
+}
+
+/// Build a `reqwest::Client` that honours the standard `HTTP_PROXY`,
+/// `HTTPS_PROXY`, and `NO_PROXY` environment variables. When no proxy is
+/// configured the client behaves identically to `reqwest::Client::new()`.
+pub fn build_http_client() -> Result<reqwest::Client, ApiError> {
+    build_http_client_with(&ProxyConfig::from_env())
+}
+
+/// Infallible counterpart to [`build_http_client`] for constructors that
+/// historically returned `Self` rather than `Result<Self, _>`. When the proxy
+/// configuration is malformed we fall back to a default client so that
+/// callers retain the previous behaviour and the failure surfaces on the
+/// first outbound request instead of at construction time.
+#[must_use]
+pub fn build_http_client_or_default() -> reqwest::Client {
+    build_http_client().unwrap_or_else(|_| reqwest::Client::new())
+}
+
+/// Build a `reqwest::Client` from an explicit [`ProxyConfig`]. Used by tests
+/// and by callers that want to override process-level environment lookups.
+///
+/// When `config.proxy_url` is set it overrides the per-scheme `http_proxy`
+/// and `https_proxy` fields and is registered as both an HTTP and HTTPS
+/// proxy so a single value can route every outbound request.
+pub fn build_http_client_with(config: &ProxyConfig) -> Result<reqwest::Client, ApiError> {
+    let mut builder = reqwest::Client::builder().no_proxy();
+
+    let no_proxy = config
+        .no_proxy
+        .as_deref()
+        .and_then(reqwest::NoProxy::from_string);
+
+    let (http_proxy_url, https_proxy_url) = match config.proxy_url.as_deref() {
+        Some(unified) => (Some(unified), Some(unified)),
+        None => (config.http_proxy.as_deref(), config.https_proxy.as_deref()),
+    };
+
+    if let Some(url) = https_proxy_url {
+        let mut proxy = reqwest::Proxy::https(url)?;
+        if let Some(filter) = no_proxy.clone() {
+            proxy = proxy.no_proxy(Some(filter));
+        }
+        builder = builder.proxy(proxy);
+    }
+
+    if let Some(url) = http_proxy_url {
+        let mut proxy = reqwest::Proxy::http(url)?;
+        if let Some(filter) = no_proxy.clone() {
+            proxy = proxy.no_proxy(Some(filter));
+        }
+        builder = builder.proxy(proxy);
+    }
+
+    Ok(builder.build()?)
+}
+
+fn first_non_empty<F>(keys: &[&str], lookup: &mut F) -> Option<String>
+where
+    F: FnMut(&str) -> Option<String>,
+{
+    keys.iter()
+        .find_map(|key| lookup(key).filter(|value| !value.is_empty()))
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::HashMap;
+
+    use super::{build_http_client_with, ProxyConfig};
+
+    fn config_from_map(pairs: &[(&str, &str)]) -> ProxyConfig {
+        let map: HashMap<String, String> = pairs
+            .iter()
+            .map(|(key, value)| ((*key).to_string(), (*value).to_string()))
+            .collect();
+        ProxyConfig::from_lookup(|key| map.get(key).cloned())
+    }
+
+    #[test]
+    fn proxy_config_is_empty_when_no_env_vars_are_set() {
+        // given
+        let config = config_from_map(&[]);
+
+        // when
+        let empty = config.is_empty();
+
+        // then
+        assert!(empty);
+        assert_eq!(config, ProxyConfig::default());
+    }
+
+    #[test]
+    fn proxy_config_reads_uppercase_http_https_and_no_proxy() {
+        // given
+        let pairs = [
+            ("HTTP_PROXY", "http://proxy.internal:3128"),
+            ("HTTPS_PROXY", "http://secure.internal:3129"),
+            ("NO_PROXY", "localhost,127.0.0.1,.corp"),
+        ];
+
+        // when
+        let config = config_from_map(&pairs);
+
+        // then
+        assert_eq!(
+            config.http_proxy.as_deref(),
+            Some("http://proxy.internal:3128")
+        );
+        assert_eq!(
+            config.https_proxy.as_deref(),
+            Some("http://secure.internal:3129")
+        );
+        assert_eq!(
+            config.no_proxy.as_deref(),
+            Some("localhost,127.0.0.1,.corp")
+        );
+        assert!(!config.is_empty());
+    }
+
+    #[test]
+    fn proxy_config_falls_back_to_lowercase_keys() {
+        // given
+        let pairs = [
+            ("http_proxy", "http://lower.internal:3128"),
+            ("https_proxy", "http://lower-secure.internal:3129"),
+            ("no_proxy", ".lower"),
+        ];
+
+        // when
+        let config = config_from_map(&pairs);
+
+        // then
+        assert_eq!(
+            config.http_proxy.as_deref(),
+            Some("http://lower.internal:3128")
+        );
+        assert_eq!(
+            config.https_proxy.as_deref(),
+            Some("http://lower-secure.internal:3129")
+        );
+        assert_eq!(config.no_proxy.as_deref(), Some(".lower"));
+    }
+
+    #[test]
+    fn proxy_config_prefers_uppercase_over_lowercase_when_both_set() {
+        // given
+        let pairs = [
+            ("HTTP_PROXY", "http://upper.internal:3128"),
+            ("http_proxy", "http://lower.internal:3128"),
+        ];
+
+        // when
+        let config = config_from_map(&pairs);
+
+        // then
+        assert_eq!(
+            config.http_proxy.as_deref(),
+            Some("http://upper.internal:3128")
+        );
+    }
+
+    #[test]
+    fn proxy_config_treats_empty_strings_as_unset() {
+        // given
+        let pairs = [("HTTP_PROXY", ""), ("http_proxy", "")];
+
+        // when
+        let config = config_from_map(&pairs);
+
+        // then
+        assert!(config.http_proxy.is_none());
+    }
+
+    #[test]
+    fn build_http_client_succeeds_when_no_proxy_is_configured() {
+        // given
+        let config = ProxyConfig::default();
+
+        // when
+        let result = build_http_client_with(&config);
+
+        // then
+        assert!(result.is_ok());
+    }
+
+    #[test]
+    fn build_http_client_succeeds_with_valid_http_and_https_proxies() {
+        // given
+        let config = ProxyConfig {
+            http_proxy: Some("http://proxy.internal:3128".to_string()),
+            https_proxy: Some("http://secure.internal:3129".to_string()),
+            no_proxy: Some("localhost,127.0.0.1".to_string()),
+            proxy_url: None,
+        };
+
+        // when
+        let result = build_http_client_with(&config);
+
+        // then
+        assert!(result.is_ok());
+    }
+
+    #[test]
+    fn build_http_client_returns_http_error_for_invalid_proxy_url() {
+        // given
+        let config = ProxyConfig {
+            http_proxy: None,
+            https_proxy: Some("not a url".to_string()),
+            no_proxy: None,
+            proxy_url: None,
+        };
+
+        // when
+        let result = build_http_client_with(&config);
+
+        // then
+        let error = result.expect_err("invalid proxy URL must be reported as a build failure");
+        assert!(
+            matches!(error, crate::error::ApiError::Http(_)),
+            "expected ApiError::Http for invalid proxy URL, got: {error:?}"
+        );
+    }
+
+    #[test]
+    fn from_proxy_url_sets_unified_field_and_leaves_per_scheme_empty() {
+        // given / when
+        let config = ProxyConfig::from_proxy_url("http://unified.internal:3128");
+
+        // then
+        assert_eq!(
+            config.proxy_url.as_deref(),
+            Some("http://unified.internal:3128")
+        );
+        assert!(config.http_proxy.is_none());
+        assert!(config.https_proxy.is_none());
+        assert!(!config.is_empty());
+    }
+
+    #[test]
+    fn build_http_client_succeeds_with_unified_proxy_url() {
+        // given
+        let config = ProxyConfig {
+            proxy_url: Some("http://unified.internal:3128".to_string()),
+            no_proxy: Some("localhost".to_string()),
+            ..ProxyConfig::default()
+        };
+
+        // when
+        let result = build_http_client_with(&config);
+
+        // then
+        assert!(result.is_ok());
+    }
+
+    #[test]
+    fn proxy_url_takes_precedence_over_per_scheme_fields() {
+        // given – both per-scheme and unified are set
+        let config = ProxyConfig {
+            http_proxy: Some("http://per-scheme.internal:1111".to_string()),
+            https_proxy: Some("http://per-scheme.internal:2222".to_string()),
+            no_proxy: None,
+            proxy_url: Some("http://unified.internal:3128".to_string()),
+        };
+
+        // when – building succeeds (the unified URL is valid)
+        let result = build_http_client_with(&config);
+
+        // then
+        assert!(result.is_ok());
+    }
+
+    #[test]
+    fn build_http_client_returns_error_for_invalid_unified_proxy_url() {
+        // given
+        let config = ProxyConfig::from_proxy_url("not a url");
+
+        // when
+        let result = build_http_client_with(&config);
+
+        // then
+        assert!(
+            matches!(result, Err(crate::error::ApiError::Http(_))),
+            "invalid unified proxy URL should fail: {result:?}"
+        );
+    }
+}
--- a/rust/crates/api/src/lib.rs
+++ b/rust/crates/api/src/lib.rs
@@ -1,5 +1,6 @@
 mod client;
 mod error;
+mod http_client;
 mod prompt_cache;
 mod providers;
 mod sse;
@@ -10,6 +11,9 @@ pub use client::{
    resolve_startup_auth_source, MessageStream, OAuthTokenSet, ProviderClient,
 };
 pub use error::ApiError;
+pub use http_client::{
+    build_http_client, build_http_client_or_default, build_http_client_with, ProxyConfig,
+};
 pub use prompt_cache::{
    CacheBreakEvent, PromptCache, PromptCacheConfig, PromptCachePaths, PromptCacheRecord,
    PromptCacheStats,
@@ -17,7 +21,8 @@ pub use prompt_cache::{
 pub use providers::anthropic::{AnthropicClient, AnthropicClient as ApiClient, AuthSource};
 pub use providers::openai_compat::{OpenAiCompatClient, OpenAiCompatConfig};
 pub use providers::{
-    detect_provider_kind, max_tokens_for_model, resolve_model_alias, ProviderKind,
+    detect_provider_kind, max_tokens_for_model, max_tokens_for_model_with_override,
+    resolve_model_alias, ProviderKind,
 };
 pub use sse::{parse_frame, SseParser};
 pub use types::{
--- a/rust/crates/api/src/prompt_cache.rs
+++ b/rust/crates/api/src/prompt_cache.rs
@@ -704,6 +704,7 @@ mod tests {
            tools: None,
            tool_choice: None,
            stream: false,
+            ..Default::default()
        }
    }

--- a/rust/crates/api/src/providers/anthropic.rs
+++ b/rust/crates/api/src/providers/anthropic.rs
@@ -1,4 +1,5 @@
 use std::collections::VecDeque;
+use std::sync::atomic::{AtomicU64, Ordering};
 use std::sync::{Arc, Mutex};
 use std::time::{Duration, SystemTime, UNIX_EPOCH};

@@ -12,6 +13,7 @@ use serde_json::{Map, Value};
 use telemetry::{AnalyticsEvent, AnthropicRequestProfile, ClientIdentity, SessionTracer};

 use crate::error::ApiError;
+use crate::http_client::build_http_client_or_default;
 use crate::prompt_cache::{PromptCache, PromptCacheRecord, PromptCacheStats};

 use super::{model_token_limit, resolve_model_alias, Provider, ProviderFuture};
@@ -21,9 +23,9 @@ use crate::types::{MessageDeltaEvent, MessageRequest, MessageResponse, StreamEve
 pub const DEFAULT_BASE_URL: &str = "https://api.anthropic.com";
 const REQUEST_ID_HEADER: &str = "request-id";
 const ALT_REQUEST_ID_HEADER: &str = "x-request-id";
-const DEFAULT_INITIAL_BACKOFF: Duration = Duration::from_millis(200);
-const DEFAULT_MAX_BACKOFF: Duration = Duration::from_secs(2);
-const DEFAULT_MAX_RETRIES: u32 = 2;
+const DEFAULT_INITIAL_BACKOFF: Duration = Duration::from_secs(1);
+const DEFAULT_MAX_BACKOFF: Duration = Duration::from_secs(128);
+const DEFAULT_MAX_RETRIES: u32 = 8;

 #[derive(Debug, Clone, PartialEq, Eq)]
 pub enum AuthSource {
@@ -127,7 +129,7 @@ impl AnthropicClient {
    #[must_use]
    pub fn new(api_key: impl Into<String>) -> Self {
        Self {
-            http: reqwest::Client::new(),
+            http: build_http_client_or_default(),
            auth: AuthSource::ApiKey(api_key.into()),
            base_url: DEFAULT_BASE_URL.to_string(),
            max_retries: DEFAULT_MAX_RETRIES,
@@ -143,7 +145,7 @@ impl AnthropicClient {
    #[must_use]
    pub fn from_auth(auth: AuthSource) -> Self {
        Self {
-            http: reqwest::Client::new(),
+            http: build_http_client_or_default(),
            auth,
            base_url: DEFAULT_BASE_URL.to_string(),
            max_retries: DEFAULT_MAX_RETRIES,
@@ -452,7 +454,7 @@ impl AnthropicClient {
                break;
            }

-            tokio::time::sleep(self.backoff_for_attempt(attempts)?).await;
+            tokio::time::sleep(self.jittered_backoff_for_attempt(attempts)?).await;
        }

        Err(ApiError::RetriesExhausted {
@@ -485,10 +487,21 @@ impl AnthropicClient {
    }

    async fn preflight_message_request(&self, request: &MessageRequest) -> Result<(), ApiError> {
+        // Always run the local byte-estimate guard first. This catches
+        // oversized requests even if the remote count_tokens endpoint is
+        // unreachable, misconfigured, or unimplemented (e.g., third-party
+        // Anthropic-compatible gateways). If byte estimation already flags
+        // the request as oversized, reject immediately without a network
+        // round trip.
+        super::preflight_message_request(request)?;
+
        let Some(limit) = model_token_limit(&request.model) else {
            return Ok(());
        };

+        // Best-effort refinement using the Anthropic count_tokens endpoint.
+        // On any failure (network, parse, auth), fall back to the local
+        // byte-estimate result which already passed above.
        let counted_input_tokens = match self.count_tokens(request).await {
            Ok(count) => count,
            Err(_) => return Ok(()),
@@ -513,7 +526,10 @@ impl AnthropicClient {
            input_tokens: u32,
        }

-        let request_url = format!("{}/v1/messages/count_tokens", self.base_url.trim_end_matches('/'));
+        let request_url = format!(
+            "{}/v1/messages/count_tokens",
+            self.base_url.trim_end_matches('/')
+        );
        let mut request_body = self.request_profile.render_json_body(request)?;
        strip_unsupported_beta_body_fields(&mut request_body);
        let response = self
@@ -526,12 +542,7 @@ impl AnthropicClient {
        let response = expect_success(response).await?;
        let body = response.text().await.map_err(ApiError::from)?;
        let parsed = serde_json::from_str::<CountTokensResponse>(&body).map_err(|error| {
-            ApiError::json_deserialize(
-                "Anthropic count_tokens",
-                &request.model,
-                &body,
-                error,
-            )
+            ApiError::json_deserialize("Anthropic count_tokens", &request.model, &body, error)
        })?;
        Ok(parsed.input_tokens)
    }
@@ -568,6 +579,42 @@ impl AnthropicClient {
            .checked_mul(multiplier)
            .map_or(self.max_backoff, |delay| delay.min(self.max_backoff)))
    }
+
+    fn jittered_backoff_for_attempt(&self, attempt: u32) -> Result<Duration, ApiError> {
+        let base = self.backoff_for_attempt(attempt)?;
+        Ok(base + jitter_for_base(base))
+    }
+}
+
+/// Process-wide counter that guarantees distinct jitter samples even when
+/// the system clock resolution is coarser than consecutive retry sleeps.
+static JITTER_COUNTER: AtomicU64 = AtomicU64::new(0);
+
+/// Returns a random additive jitter in `[0, base]` to decorrelate retries
+/// from multiple concurrent clients. Entropy is drawn from the nanosecond
+/// wall clock mixed with a monotonic counter and run through a splitmix64
+/// finalizer; adequate for retry jitter (no cryptographic requirement).
+fn jitter_for_base(base: Duration) -> Duration {
+    let base_nanos = u64::try_from(base.as_nanos()).unwrap_or(u64::MAX);
+    if base_nanos == 0 {
+        return Duration::ZERO;
+    }
+    let raw_nanos = SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map(|elapsed| u64::try_from(elapsed.as_nanos()).unwrap_or(u64::MAX))
+        .unwrap_or(0);
+    let tick = JITTER_COUNTER.fetch_add(1, Ordering::Relaxed);
+    // splitmix64 finalizer — mixes the low bits so large bases still see
+    // jitter across their full range instead of being clamped to subsec nanos.
+    let mut mixed = raw_nanos
+        .wrapping_add(tick)
+        .wrapping_add(0x9E37_79B9_7F4A_7C15);
+    mixed = (mixed ^ (mixed >> 30)).wrapping_mul(0xBF58_476D_1CE4_E5B9);
+    mixed = (mixed ^ (mixed >> 27)).wrapping_mul(0x94D0_49BB_1331_11EB);
+    mixed ^= mixed >> 31;
+    // Inclusive upper bound: jitter may equal `base`, matching "up to base".
+    let jitter_nanos = mixed % base_nanos.saturating_add(1);
+    Duration::from_nanos(jitter_nanos)
 }

 impl AuthSource {
@@ -894,6 +941,15 @@ const fn is_retryable_status(status: reqwest::StatusCode) -> bool {
 fn strip_unsupported_beta_body_fields(body: &mut Value) {
    if let Some(object) = body.as_object_mut() {
        object.remove("betas");
+        // These fields are OpenAI-compatible only; Anthropic rejects them.
+        object.remove("frequency_penalty");
+        object.remove("presence_penalty");
+        // Anthropic uses "stop_sequences" not "stop". Convert if present.
+        if let Some(stop_val) = object.remove("stop") {
+            if stop_val.as_array().map_or(false, |a| !a.is_empty()) {
+                object.insert("stop_sequences".to_string(), stop_val);
+            }
+        }
    }
 }

@@ -1223,6 +1279,7 @@ mod tests {
            tools: None,
            tool_choice: None,
            stream: false,
+            ..Default::default()
        };

        assert!(request.with_streaming().stream);
@@ -1249,6 +1306,58 @@ mod tests {
        );
    }

+    #[test]
+    fn jittered_backoff_stays_within_additive_bounds_and_varies() {
+        let client = AnthropicClient::new("test-key").with_retry_policy(
+            8,
+            Duration::from_secs(1),
+            Duration::from_secs(128),
+        );
+        let mut samples = Vec::with_capacity(64);
+        for _ in 0..64 {
+            let base = client.backoff_for_attempt(3).expect("base attempt 3");
+            let jittered = client
+                .jittered_backoff_for_attempt(3)
+                .expect("jittered attempt 3");
+            assert!(
+                jittered >= base,
+                "jittered delay {jittered:?} must be at least the base {base:?}"
+            );
+            assert!(
+                jittered <= base * 2,
+                "jittered delay {jittered:?} must not exceed base*2 {:?}",
+                base * 2
+            );
+            samples.push(jittered);
+        }
+        let distinct: std::collections::HashSet<_> = samples.iter().collect();
+        assert!(
+            distinct.len() > 1,
+            "jitter should produce varied delays across samples, got {samples:?}"
+        );
+    }
+
+    #[test]
+    fn default_retry_policy_matches_exponential_schedule() {
+        let client = AnthropicClient::new("test-key");
+        assert_eq!(
+            client.backoff_for_attempt(1).expect("attempt 1"),
+            Duration::from_secs(1)
+        );
+        assert_eq!(
+            client.backoff_for_attempt(2).expect("attempt 2"),
+            Duration::from_secs(2)
+        );
+        assert_eq!(
+            client.backoff_for_attempt(3).expect("attempt 3"),
+            Duration::from_secs(4)
+        );
+        assert_eq!(
+            client.backoff_for_attempt(8).expect("attempt 8"),
+            Duration::from_secs(128)
+        );
+    }
+
    #[test]
    fn retryable_statuses_are_detected() {
        assert!(super::is_retryable_status(
@@ -1350,6 +1459,52 @@ mod tests {
        assert_eq!(body, original);
    }

+    #[test]
+    fn strip_removes_openai_only_fields_and_converts_stop() {
+        let mut body = serde_json::json!({
+            "model": "claude-sonnet-4-6",
+            "max_tokens": 1024,
+            "temperature": 0.7,
+            "frequency_penalty": 0.5,
+            "presence_penalty": 0.3,
+            "stop": ["\n"],
+        });
+
+        super::strip_unsupported_beta_body_fields(&mut body);
+
+        // temperature is kept (Anthropic supports it)
+        assert_eq!(body["temperature"], serde_json::json!(0.7));
+        // frequency_penalty and presence_penalty are removed
+        assert!(
+            body.get("frequency_penalty").is_none(),
+            "frequency_penalty must be stripped for Anthropic"
+        );
+        assert!(
+            body.get("presence_penalty").is_none(),
+            "presence_penalty must be stripped for Anthropic"
+        );
+        // stop is renamed to stop_sequences
+        assert!(body.get("stop").is_none(), "stop must be renamed");
+        assert_eq!(body["stop_sequences"], serde_json::json!(["\n"]));
+    }
+
+    #[test]
+    fn strip_does_not_add_empty_stop_sequences() {
+        let mut body = serde_json::json!({
+            "model": "claude-sonnet-4-6",
+            "max_tokens": 1024,
+            "stop": [],
+        });
+
+        super::strip_unsupported_beta_body_fields(&mut body);
+
+        assert!(body.get("stop").is_none());
+        assert!(
+            body.get("stop_sequences").is_none(),
+            "empty stop should not produce stop_sequences"
+        );
+    }
+
    #[test]
    fn rendered_request_body_strips_betas_for_standard_messages_endpoint() {
        let client = AnthropicClient::new("test-key").with_beta("tools-2026-04-01");
@@ -1361,6 +1516,7 @@ mod tests {
            tools: None,
            tool_choice: None,
            stream: false,
+            ..Default::default()
        };

        let mut rendered = client
--- a/rust/crates/api/src/providers/mod.rs
+++ b/rust/crates/api/src/providers/mod.rs
@@ -169,6 +169,31 @@ pub fn metadata_for_model(model: &str) -> Option<ProviderMetadata> {
            default_base_url: openai_compat::DEFAULT_XAI_BASE_URL,
        });
    }
+    // Explicit provider-namespaced models (e.g. "openai/gpt-4.1-mini") must
+    // route to the correct provider regardless of which auth env vars are set.
+    // Without this, detect_provider_kind falls through to the auth-sniffer
+    // order and misroutes to Anthropic if ANTHROPIC_API_KEY is present.
+    if canonical.starts_with("openai/") || canonical.starts_with("gpt-") {
+        return Some(ProviderMetadata {
+            provider: ProviderKind::OpenAi,
+            auth_env: "OPENAI_API_KEY",
+            base_url_env: "OPENAI_BASE_URL",
+            default_base_url: openai_compat::DEFAULT_OPENAI_BASE_URL,
+        });
+    }
+    // Alibaba DashScope compatible-mode endpoint. Routes qwen/* and bare
+    // qwen-* model names (qwen-max, qwen-plus, qwen-turbo, qwen-qwq, etc.)
+    // to the OpenAI-compat client pointed at DashScope's /compatible-mode/v1.
+    // Uses the OpenAi provider kind because DashScope speaks the OpenAI REST
+    // shape — only the base URL and auth env var differ.
+    if canonical.starts_with("qwen/") || canonical.starts_with("qwen-") {
+        return Some(ProviderMetadata {
+            provider: ProviderKind::OpenAi,
+            auth_env: "DASHSCOPE_API_KEY",
+            base_url_env: "DASHSCOPE_BASE_URL",
+            default_base_url: openai_compat::DEFAULT_DASHSCOPE_BASE_URL,
+        });
+    }
    None
 }

@@ -204,6 +229,14 @@ pub fn max_tokens_for_model(model: &str) -> u32 {
    )
 }

+/// Returns the effective max output tokens for a model, preferring a plugin
+/// override when present. Falls back to [`max_tokens_for_model`] when the
+/// override is `None`.
+#[must_use]
+pub fn max_tokens_for_model_with_override(model: &str, plugin_override: Option<u32>) -> u32 {
+    plugin_override.unwrap_or_else(|| max_tokens_for_model(model))
+}
+
 #[must_use]
 pub fn model_token_limit(model: &str) -> Option<ModelTokenLimit> {
    let canonical = resolve_model_alias(model);
@@ -323,8 +356,9 @@ mod tests {
    };

    use super::{
-        detect_provider_kind, load_dotenv_file, max_tokens_for_model, model_token_limit,
-        parse_dotenv, preflight_message_request, resolve_model_alias, ProviderKind,
+        detect_provider_kind, load_dotenv_file, max_tokens_for_model,
+        max_tokens_for_model_with_override, model_token_limit, parse_dotenv,
+        preflight_message_request, resolve_model_alias, ProviderKind,
    };

    #[test]
@@ -343,12 +377,114 @@ mod tests {
        );
    }

+    #[test]
+    fn openai_namespaced_model_routes_to_openai_not_anthropic() {
+        // Regression: "openai/gpt-4.1-mini" was misrouted to Anthropic when
+        // ANTHROPIC_API_KEY was set because metadata_for_model returned None
+        // and detect_provider_kind fell through to auth-sniffer order.
+        // The model prefix must win over env-var presence.
+        let kind = super::metadata_for_model("openai/gpt-4.1-mini")
+            .map(|m| m.provider)
+            .unwrap_or_else(|| detect_provider_kind("openai/gpt-4.1-mini"));
+        assert_eq!(
+            kind,
+            ProviderKind::OpenAi,
+            "openai/ prefix must route to OpenAi regardless of ANTHROPIC_API_KEY"
+        );
+
+        // Also cover bare gpt- prefix
+        let kind2 = super::metadata_for_model("gpt-4o")
+            .map(|m| m.provider)
+            .unwrap_or_else(|| detect_provider_kind("gpt-4o"));
+        assert_eq!(kind2, ProviderKind::OpenAi);
+    }
+
+    #[test]
+    fn qwen_prefix_routes_to_dashscope_not_anthropic() {
+        // User request from Discord #clawcode-get-help: web3g wants to use
+        // Qwen 3.6 Plus via native Alibaba DashScope API (not OpenRouter,
+        // which has lower rate limits). metadata_for_model must route
+        // qwen/* and bare qwen-* to the OpenAi provider kind pointed at
+        // the DashScope compatible-mode endpoint, regardless of whether
+        // ANTHROPIC_API_KEY is present in the environment.
+        let meta = super::metadata_for_model("qwen/qwen-max")
+            .expect("qwen/ prefix must resolve to DashScope metadata");
+        assert_eq!(meta.provider, ProviderKind::OpenAi);
+        assert_eq!(meta.auth_env, "DASHSCOPE_API_KEY");
+        assert_eq!(meta.base_url_env, "DASHSCOPE_BASE_URL");
+        assert!(meta.default_base_url.contains("dashscope.aliyuncs.com"));
+
+        // Bare qwen- prefix also routes
+        let meta2 = super::metadata_for_model("qwen-plus")
+            .expect("qwen- prefix must resolve to DashScope metadata");
+        assert_eq!(meta2.provider, ProviderKind::OpenAi);
+        assert_eq!(meta2.auth_env, "DASHSCOPE_API_KEY");
+
+        // detect_provider_kind must agree even if ANTHROPIC_API_KEY is set
+        let kind = detect_provider_kind("qwen/qwen3-coder");
+        assert_eq!(
+            kind,
+            ProviderKind::OpenAi,
+            "qwen/ prefix must win over auth-sniffer order"
+        );
+    }
+
    #[test]
    fn keeps_existing_max_token_heuristic() {
        assert_eq!(max_tokens_for_model("opus"), 32_000);
        assert_eq!(max_tokens_for_model("grok-3"), 64_000);
    }

+    #[test]
+    fn plugin_config_max_output_tokens_overrides_model_default() {
+        // given
+        let nanos = std::time::SystemTime::now()
+            .duration_since(std::time::UNIX_EPOCH)
+            .expect("time should be after epoch")
+            .as_nanos();
+        let root = std::env::temp_dir().join(format!("api-plugin-max-tokens-{nanos}"));
+        let cwd = root.join("project");
+        let home = root.join("home").join(".claw");
+        std::fs::create_dir_all(cwd.join(".claw")).expect("project config dir");
+        std::fs::create_dir_all(&home).expect("home config dir");
+        std::fs::write(
+            home.join("settings.json"),
+            r#"{
+              "plugins": {
+                "maxOutputTokens": 12345
+              }
+            }"#,
+        )
+        .expect("write plugin settings");
+
+        // when
+        let loaded = runtime::ConfigLoader::new(&cwd, &home)
+            .load()
+            .expect("config should load");
+        let plugin_override = loaded.plugins().max_output_tokens();
+        let effective = max_tokens_for_model_with_override("claude-opus-4-6", plugin_override);
+
+        // then
+        assert_eq!(plugin_override, Some(12345));
+        assert_eq!(effective, 12345);
+        assert_ne!(effective, max_tokens_for_model("claude-opus-4-6"));
+
+        std::fs::remove_dir_all(root).expect("cleanup temp dir");
+    }
+
+    #[test]
+    fn max_tokens_for_model_with_override_falls_back_when_plugin_unset() {
+        // given
+        let plugin_override: Option<u32> = None;
+
+        // when
+        let effective = max_tokens_for_model_with_override("claude-opus-4-6", plugin_override);
+
+        // then
+        assert_eq!(effective, max_tokens_for_model("claude-opus-4-6"));
+        assert_eq!(effective, 32_000);
+    }
+
    #[test]
    fn returns_context_window_metadata_for_supported_models() {
        assert_eq!(
@@ -387,6 +523,7 @@ mod tests {
            }]),
            tool_choice: Some(ToolChoice::Auto),
            stream: true,
+            ..Default::default()
        };

        let error = preflight_message_request(&request)
@@ -425,6 +562,7 @@ mod tests {
            tools: None,
            tool_choice: None,
            stream: false,
+            ..Default::default()
        };

        preflight_message_request(&request)
--- a/rust/crates/api/src/providers/openai_compat.rs
+++ b/rust/crates/api/src/providers/openai_compat.rs
@@ -1,10 +1,12 @@
 use std::collections::{BTreeMap, VecDeque};
-use std::time::Duration;
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::time::{Duration, SystemTime, UNIX_EPOCH};

 use serde::Deserialize;
 use serde_json::{json, Value};

 use crate::error::ApiError;
+use crate::http_client::build_http_client_or_default;
 use crate::types::{
    ContentBlockDelta, ContentBlockDeltaEvent, ContentBlockStartEvent, ContentBlockStopEvent,
    InputContentBlock, InputMessage, MessageDelta, MessageDeltaEvent, MessageRequest,
@@ -16,11 +18,12 @@ use super::{preflight_message_request, Provider, ProviderFuture};

 pub const DEFAULT_XAI_BASE_URL: &str = "https://api.x.ai/v1";
 pub const DEFAULT_OPENAI_BASE_URL: &str = "https://api.openai.com/v1";
+pub const DEFAULT_DASHSCOPE_BASE_URL: &str = "https://dashscope.aliyuncs.com/compatible-mode/v1";
 const REQUEST_ID_HEADER: &str = "request-id";
 const ALT_REQUEST_ID_HEADER: &str = "x-request-id";
-const DEFAULT_INITIAL_BACKOFF: Duration = Duration::from_millis(200);
-const DEFAULT_MAX_BACKOFF: Duration = Duration::from_secs(2);
-const DEFAULT_MAX_RETRIES: u32 = 2;
+const DEFAULT_INITIAL_BACKOFF: Duration = Duration::from_secs(1);
+const DEFAULT_MAX_BACKOFF: Duration = Duration::from_secs(128);
+const DEFAULT_MAX_RETRIES: u32 = 8;

 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub struct OpenAiCompatConfig {
@@ -32,6 +35,7 @@ pub struct OpenAiCompatConfig {

 const XAI_ENV_VARS: &[&str] = &["XAI_API_KEY"];
 const OPENAI_ENV_VARS: &[&str] = &["OPENAI_API_KEY"];
+const DASHSCOPE_ENV_VARS: &[&str] = &["DASHSCOPE_API_KEY"];

 impl OpenAiCompatConfig {
    #[must_use]
@@ -53,11 +57,27 @@ impl OpenAiCompatConfig {
            default_base_url: DEFAULT_OPENAI_BASE_URL,
        }
    }
+
+    /// Alibaba DashScope compatible-mode endpoint (Qwen family models).
+    /// Uses the OpenAI-compatible REST shape at /compatible-mode/v1.
+    /// Requested via Discord #clawcode-get-help: native Alibaba API for
+    /// higher rate limits than going through OpenRouter.
+    #[must_use]
+    pub const fn dashscope() -> Self {
+        Self {
+            provider_name: "DashScope",
+            api_key_env: "DASHSCOPE_API_KEY",
+            base_url_env: "DASHSCOPE_BASE_URL",
+            default_base_url: DEFAULT_DASHSCOPE_BASE_URL,
+        }
+    }
+
    #[must_use]
    pub fn credential_env_vars(self) -> &'static [&'static str] {
        match self.provider_name {
            "xAI" => XAI_ENV_VARS,
            "OpenAI" => OPENAI_ENV_VARS,
+            "DashScope" => DASHSCOPE_ENV_VARS,
            _ => &[],
        }
    }
@@ -81,7 +101,7 @@ impl OpenAiCompatClient {
    #[must_use]
    pub fn new(api_key: impl Into<String>, config: OpenAiCompatConfig) -> Self {
        Self {
-            http: reqwest::Client::new(),
+            http: build_http_client_or_default(),
            api_key: api_key.into(),
            config,
            base_url: read_base_url(config),
@@ -133,12 +153,7 @@ impl OpenAiCompatClient {
        let request_id = request_id_from_headers(response.headers());
        let body = response.text().await.map_err(ApiError::from)?;
        let payload = serde_json::from_str::<ChatCompletionResponse>(&body).map_err(|error| {
-            ApiError::json_deserialize(
-                self.config.provider_name,
-                &request.model,
-                &body,
-                error,
-            )
+            ApiError::json_deserialize(self.config.provider_name, &request.model, &body, error)
        })?;
        let mut normalized = normalize_response(&request.model, payload)?;
        if normalized.request_id.is_none() {
@@ -158,10 +173,7 @@ impl OpenAiCompatClient {
        Ok(MessageStream {
            request_id: request_id_from_headers(response.headers()),
            response,
-            parser: OpenAiSseParser::with_context(
-                self.config.provider_name,
-                request.model.clone(),
-            ),
+            parser: OpenAiSseParser::with_context(self.config.provider_name, request.model.clone()),
            pending: VecDeque::new(),
            done: false,
            state: StreamState::new(request.model.clone()),
@@ -190,7 +202,7 @@ impl OpenAiCompatClient {
                break retryable_error;
            }

-            tokio::time::sleep(self.backoff_for_attempt(attempts)?).await;
+            tokio::time::sleep(self.jittered_backoff_for_attempt(attempts)?).await;
        };

        Err(ApiError::RetriesExhausted {
@@ -226,6 +238,39 @@ impl OpenAiCompatClient {
            .checked_mul(multiplier)
            .map_or(self.max_backoff, |delay| delay.min(self.max_backoff)))
    }
+
+    fn jittered_backoff_for_attempt(&self, attempt: u32) -> Result<Duration, ApiError> {
+        let base = self.backoff_for_attempt(attempt)?;
+        Ok(base + jitter_for_base(base))
+    }
+}
+
+/// Process-wide counter that guarantees distinct jitter samples even when
+/// the system clock resolution is coarser than consecutive retry sleeps.
+static JITTER_COUNTER: AtomicU64 = AtomicU64::new(0);
+
+/// Returns a random additive jitter in `[0, base]` to decorrelate retries
+/// from multiple concurrent clients. Entropy is drawn from the nanosecond
+/// wall clock mixed with a monotonic counter and run through a splitmix64
+/// finalizer; adequate for retry jitter (no cryptographic requirement).
+fn jitter_for_base(base: Duration) -> Duration {
+    let base_nanos = u64::try_from(base.as_nanos()).unwrap_or(u64::MAX);
+    if base_nanos == 0 {
+        return Duration::ZERO;
+    }
+    let raw_nanos = SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map(|elapsed| u64::try_from(elapsed.as_nanos()).unwrap_or(u64::MAX))
+        .unwrap_or(0);
+    let tick = JITTER_COUNTER.fetch_add(1, Ordering::Relaxed);
+    let mut mixed = raw_nanos
+        .wrapping_add(tick)
+        .wrapping_add(0x9E37_79B9_7F4A_7C15);
+    mixed = (mixed ^ (mixed >> 30)).wrapping_mul(0xBF58_476D_1CE4_E5B9);
+    mixed = (mixed ^ (mixed >> 27)).wrapping_mul(0x94D0_49BB_1331_11EB);
+    mixed ^= mixed >> 31;
+    let jitter_nanos = mixed % base_nanos.saturating_add(1);
+    Duration::from_nanos(jitter_nanos)
 }

 impl Provider for OpenAiCompatClient {
@@ -657,6 +702,25 @@ struct ErrorBody {
    message: Option<String>,
 }

+/// Returns true for models known to reject tuning parameters like temperature,
+/// top_p, frequency_penalty, and presence_penalty. These are typically
+/// reasoning/chain-of-thought models with fixed sampling.
+fn is_reasoning_model(model: &str) -> bool {
+    let lowered = model.to_ascii_lowercase();
+    // Strip any provider/ prefix for the check (e.g. qwen/qwen-qwq -> qwen-qwq)
+    let canonical = lowered.rsplit('/').next().unwrap_or(lowered.as_str());
+    // OpenAI reasoning models
+    canonical.starts_with("o1")
+        || canonical.starts_with("o3")
+        || canonical.starts_with("o4")
+        // xAI reasoning: grok-3-mini always uses reasoning mode
+        || canonical == "grok-3-mini"
+        // Alibaba DashScope reasoning variants (QwQ + Qwen3-Thinking family)
+        || canonical.starts_with("qwen-qwq")
+        || canonical.starts_with("qwq")
+        || canonical.contains("thinking")
+}
+
 fn build_chat_completion_request(request: &MessageRequest, config: OpenAiCompatConfig) -> Value {
    let mut messages = Vec::new();
    if let Some(system) = request.system.as_ref().filter(|value| !value.is_empty()) {
@@ -688,6 +752,30 @@ fn build_chat_completion_request(request: &MessageRequest, config: OpenAiCompatC
        payload["tool_choice"] = openai_tool_choice(tool_choice);
    }

+    // OpenAI-compatible tuning parameters — only included when explicitly set.
+    // Reasoning models (o1/o3/o4/grok-3-mini) reject these params with 400;
+    // silently strip them to avoid cryptic provider errors.
+    if !is_reasoning_model(&request.model) {
+        if let Some(temperature) = request.temperature {
+            payload["temperature"] = json!(temperature);
+        }
+        if let Some(top_p) = request.top_p {
+            payload["top_p"] = json!(top_p);
+        }
+        if let Some(frequency_penalty) = request.frequency_penalty {
+            payload["frequency_penalty"] = json!(frequency_penalty);
+        }
+        if let Some(presence_penalty) = request.presence_penalty {
+            payload["presence_penalty"] = json!(presence_penalty);
+        }
+    }
+    // stop is generally safe for all providers
+    if let Some(stop) = &request.stop {
+        if !stop.is_empty() {
+            payload["stop"] = json!(stop);
+        }
+    }
+
    payload
 }

@@ -976,8 +1064,9 @@ impl StringExt for String {
 #[cfg(test)]
 mod tests {
    use super::{
-        build_chat_completion_request, chat_completions_endpoint, normalize_finish_reason,
-        openai_tool_choice, parse_tool_arguments, OpenAiCompatClient, OpenAiCompatConfig,
+        build_chat_completion_request, chat_completions_endpoint, is_reasoning_model,
+        normalize_finish_reason, openai_tool_choice, parse_tool_arguments, OpenAiCompatClient,
+        OpenAiCompatConfig,
    };
    use crate::error::ApiError;
    use crate::types::{
@@ -1016,6 +1105,7 @@ mod tests {
                }]),
                tool_choice: Some(ToolChoice::Auto),
                stream: false,
+                ..Default::default()
            },
            OpenAiCompatConfig::xai(),
        );
@@ -1038,6 +1128,7 @@ mod tests {
                tools: None,
                tool_choice: None,
                stream: true,
+                ..Default::default()
            },
            OpenAiCompatConfig::openai(),
        );
@@ -1056,6 +1147,7 @@ mod tests {
                tools: None,
                tool_choice: None,
                stream: true,
+                ..Default::default()
            },
            OpenAiCompatConfig::xai(),
        );
@@ -1126,4 +1218,104 @@ mod tests {
        assert_eq!(normalize_finish_reason("stop"), "end_turn");
        assert_eq!(normalize_finish_reason("tool_calls"), "tool_use");
    }
+
+    #[test]
+    fn tuning_params_included_in_payload_when_set() {
+        let request = MessageRequest {
+            model: "gpt-4o".to_string(),
+            max_tokens: 1024,
+            messages: vec![],
+            system: None,
+            tools: None,
+            tool_choice: None,
+            stream: false,
+            temperature: Some(0.7),
+            top_p: Some(0.9),
+            frequency_penalty: Some(0.5),
+            presence_penalty: Some(0.3),
+            stop: Some(vec!["\n".to_string()]),
+        };
+        let payload = build_chat_completion_request(&request, OpenAiCompatConfig::openai());
+        assert_eq!(payload["temperature"], 0.7);
+        assert_eq!(payload["top_p"], 0.9);
+        assert_eq!(payload["frequency_penalty"], 0.5);
+        assert_eq!(payload["presence_penalty"], 0.3);
+        assert_eq!(payload["stop"], json!(["\n"]));
+    }
+
+    #[test]
+    fn reasoning_model_strips_tuning_params() {
+        let request = MessageRequest {
+            model: "o1-mini".to_string(),
+            max_tokens: 1024,
+            messages: vec![],
+            stream: false,
+            temperature: Some(0.7),
+            top_p: Some(0.9),
+            frequency_penalty: Some(0.5),
+            presence_penalty: Some(0.3),
+            stop: Some(vec!["\n".to_string()]),
+            ..Default::default()
+        };
+        let payload = build_chat_completion_request(&request, OpenAiCompatConfig::openai());
+        assert!(
+            payload.get("temperature").is_none(),
+            "reasoning model should strip temperature"
+        );
+        assert!(
+            payload.get("top_p").is_none(),
+            "reasoning model should strip top_p"
+        );
+        assert!(payload.get("frequency_penalty").is_none());
+        assert!(payload.get("presence_penalty").is_none());
+        // stop is safe for all providers
+        assert_eq!(payload["stop"], json!(["\n"]));
+    }
+
+    #[test]
+    fn grok_3_mini_is_reasoning_model() {
+        assert!(is_reasoning_model("grok-3-mini"));
+        assert!(is_reasoning_model("o1"));
+        assert!(is_reasoning_model("o1-mini"));
+        assert!(is_reasoning_model("o3-mini"));
+        assert!(!is_reasoning_model("gpt-4o"));
+        assert!(!is_reasoning_model("grok-3"));
+        assert!(!is_reasoning_model("claude-sonnet-4-6"));
+    }
+
+    #[test]
+    fn qwen_reasoning_variants_are_detected() {
+        // QwQ reasoning model
+        assert!(is_reasoning_model("qwen-qwq-32b"));
+        assert!(is_reasoning_model("qwen/qwen-qwq-32b"));
+        // Qwen3 thinking family
+        assert!(is_reasoning_model("qwen3-30b-a3b-thinking"));
+        assert!(is_reasoning_model("qwen/qwen3-30b-a3b-thinking"));
+        // Bare qwq
+        assert!(is_reasoning_model("qwq-plus"));
+        // Regular Qwen models must NOT be classified as reasoning
+        assert!(!is_reasoning_model("qwen-max"));
+        assert!(!is_reasoning_model("qwen/qwen-plus"));
+        assert!(!is_reasoning_model("qwen-turbo"));
+    }
+
+    #[test]
+    fn tuning_params_omitted_from_payload_when_none() {
+        let request = MessageRequest {
+            model: "gpt-4o".to_string(),
+            max_tokens: 1024,
+            messages: vec![],
+            stream: false,
+            ..Default::default()
+        };
+        let payload = build_chat_completion_request(&request, OpenAiCompatConfig::openai());
+        assert!(
+            payload.get("temperature").is_none(),
+            "temperature should be absent"
+        );
+        assert!(payload.get("top_p").is_none(), "top_p should be absent");
+        assert!(payload.get("frequency_penalty").is_none());
+        assert!(payload.get("presence_penalty").is_none());
+        assert!(payload.get("stop").is_none());
+    }
 }
--- a/rust/crates/api/src/types.rs
+++ b/rust/crates/api/src/types.rs
@@ -2,7 +2,7 @@ use runtime::{pricing_for_model, TokenUsage, UsageCostEstimate};
 use serde::{Deserialize, Serialize};
 use serde_json::Value;

-#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, Default)]
 pub struct MessageRequest {
    pub model: String,
    pub max_tokens: u32,
@@ -15,6 +15,17 @@ pub struct MessageRequest {
    pub tool_choice: Option<ToolChoice>,
    #[serde(default, skip_serializing_if = "std::ops::Not::not")]
    pub stream: bool,
+    /// OpenAI-compatible tuning parameters. Optional — omitted from payload when None.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub temperature: Option<f64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub top_p: Option<f64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub frequency_penalty: Option<f64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub presence_penalty: Option<f64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub stop: Option<Vec<String>>,
 }

 impl MessageRequest {
--- a/rust/crates/api/tests/client_integration.rs
+++ b/rust/crates/api/tests/client_integration.rs
@@ -127,6 +127,7 @@ async fn send_message_blocks_oversized_requests_before_the_http_call() {
            tools: None,
            tool_choice: None,
            stream: false,
+            ..Default::default()
        })
        .await
        .expect_err("oversized request should fail local context-window preflight");
@@ -545,6 +546,71 @@ async fn surfaces_retry_exhaustion_for_persistent_retryable_errors() {
    }
 }

+#[tokio::test]
+async fn retries_multiple_retryable_failures_with_exponential_backoff_and_jitter() {
+    let state = Arc::new(Mutex::new(Vec::<CapturedRequest>::new()));
+    let server = spawn_server(
+        state.clone(),
+        vec![
+            http_response(
+                "429 Too Many Requests",
+                "application/json",
+                "{\"type\":\"error\",\"error\":{\"type\":\"rate_limit_error\",\"message\":\"slow down\"}}",
+            ),
+            http_response(
+                "500 Internal Server Error",
+                "application/json",
+                "{\"type\":\"error\",\"error\":{\"type\":\"api_error\",\"message\":\"boom\"}}",
+            ),
+            http_response(
+                "503 Service Unavailable",
+                "application/json",
+                "{\"type\":\"error\",\"error\":{\"type\":\"overloaded_error\",\"message\":\"busy\"}}",
+            ),
+            http_response(
+                "429 Too Many Requests",
+                "application/json",
+                "{\"type\":\"error\",\"error\":{\"type\":\"rate_limit_error\",\"message\":\"slow down again\"}}",
+            ),
+            http_response(
+                "503 Service Unavailable",
+                "application/json",
+                "{\"type\":\"error\",\"error\":{\"type\":\"overloaded_error\",\"message\":\"still busy\"}}",
+            ),
+            http_response(
+                "200 OK",
+                "application/json",
+                "{\"id\":\"msg_exp_retry\",\"type\":\"message\",\"role\":\"assistant\",\"content\":[{\"type\":\"text\",\"text\":\"Recovered after 5\"}],\"model\":\"claude-3-7-sonnet-latest\",\"stop_reason\":\"end_turn\",\"stop_sequence\":null,\"usage\":{\"input_tokens\":3,\"output_tokens\":2}}",
+            ),
+        ],
+    )
+    .await;
+
+    let client = ApiClient::new("test-key")
+        .with_base_url(server.base_url())
+        .with_retry_policy(8, Duration::from_millis(1), Duration::from_millis(4));
+    let started_at = std::time::Instant::now();
+
+    let response = client
+        .send_message(&sample_request(false))
+        .await
+        .expect("8-retry policy should absorb 5 retryable failures");
+
+    let elapsed = started_at.elapsed();
+    assert_eq!(response.total_tokens(), 5);
+    assert_eq!(
+        state.lock().await.len(),
+        6,
+        "client should issue 1 original + 5 retry requests before the 200"
+    );
+    // Jittered sleeps are bounded by 2 * max_backoff per retry (base + jitter),
+    // so 5 sleeps fit comfortably below this upper bound with generous slack.
+    assert!(
+        elapsed < Duration::from_secs(5),
+        "retries should complete promptly, took {elapsed:?}"
+    );
+}
+
 #[tokio::test]
 #[allow(clippy::await_holding_lock)]
 async fn send_message_reuses_recent_completion_cache_entries() {
@@ -676,6 +742,7 @@ async fn live_stream_smoke_test() {
            tools: None,
            tool_choice: None,
            stream: false,
+            ..Default::default()
        })
        .await
        .expect("live stream should start");
@@ -856,5 +923,6 @@ fn sample_request(stream: bool) -> MessageRequest {
        }]),
        tool_choice: Some(ToolChoice::Auto),
        stream,
+        ..Default::default()
    }
 }
--- a/rust/crates/api/tests/openai_compat_integration.rs
+++ b/rust/crates/api/tests/openai_compat_integration.rs
@@ -88,6 +88,7 @@ async fn send_message_blocks_oversized_xai_requests_before_the_http_call() {
            tools: None,
            tool_choice: None,
            stream: false,
+            ..Default::default()
        })
        .await
        .expect_err("oversized request should fail local context-window preflight");
@@ -496,6 +497,7 @@ fn sample_request(stream: bool) -> MessageRequest {
        }]),
        tool_choice: Some(ToolChoice::Auto),
        stream,
+        ..Default::default()
    }
 }

--- a/rust/crates/api/tests/proxy_integration.rs
+++ b/rust/crates/api/tests/proxy_integration.rs
@@ -0,0 +1,173 @@
+use std::ffi::OsString;
+use std::sync::{Mutex, OnceLock};
+
+use api::{build_http_client_with, ProxyConfig};
+
+fn env_lock() -> std::sync::MutexGuard<'static, ()> {
+    static LOCK: OnceLock<Mutex<()>> = OnceLock::new();
+    LOCK.get_or_init(|| Mutex::new(()))
+        .lock()
+        .unwrap_or_else(std::sync::PoisonError::into_inner)
+}
+
+struct EnvVarGuard {
+    key: &'static str,
+    original: Option<OsString>,
+}
+
+impl EnvVarGuard {
+    fn set(key: &'static str, value: Option<&str>) -> Self {
+        let original = std::env::var_os(key);
+        match value {
+            Some(value) => std::env::set_var(key, value),
+            None => std::env::remove_var(key),
+        }
+        Self { key, original }
+    }
+}
+
+impl Drop for EnvVarGuard {
+    fn drop(&mut self) {
+        match &self.original {
+            Some(value) => std::env::set_var(self.key, value),
+            None => std::env::remove_var(self.key),
+        }
+    }
+}
+
+#[test]
+fn proxy_config_from_env_reads_uppercase_proxy_vars() {
+    // given
+    let _lock = env_lock();
+    let _http = EnvVarGuard::set("HTTP_PROXY", Some("http://proxy.corp:3128"));
+    let _https = EnvVarGuard::set("HTTPS_PROXY", Some("http://secure.corp:3129"));
+    let _no = EnvVarGuard::set("NO_PROXY", Some("localhost,127.0.0.1"));
+    let _http_lower = EnvVarGuard::set("http_proxy", None);
+    let _https_lower = EnvVarGuard::set("https_proxy", None);
+    let _no_lower = EnvVarGuard::set("no_proxy", None);
+
+    // when
+    let config = ProxyConfig::from_env();
+
+    // then
+    assert_eq!(config.http_proxy.as_deref(), Some("http://proxy.corp:3128"));
+    assert_eq!(
+        config.https_proxy.as_deref(),
+        Some("http://secure.corp:3129")
+    );
+    assert_eq!(config.no_proxy.as_deref(), Some("localhost,127.0.0.1"));
+    assert!(config.proxy_url.is_none());
+    assert!(!config.is_empty());
+}
+
+#[test]
+fn proxy_config_from_env_reads_lowercase_proxy_vars() {
+    // given
+    let _lock = env_lock();
+    let _http = EnvVarGuard::set("HTTP_PROXY", None);
+    let _https = EnvVarGuard::set("HTTPS_PROXY", None);
+    let _no = EnvVarGuard::set("NO_PROXY", None);
+    let _http_lower = EnvVarGuard::set("http_proxy", Some("http://lower.corp:3128"));
+    let _https_lower = EnvVarGuard::set("https_proxy", Some("http://lower-secure.corp:3129"));
+    let _no_lower = EnvVarGuard::set("no_proxy", Some(".internal"));
+
+    // when
+    let config = ProxyConfig::from_env();
+
+    // then
+    assert_eq!(config.http_proxy.as_deref(), Some("http://lower.corp:3128"));
+    assert_eq!(
+        config.https_proxy.as_deref(),
+        Some("http://lower-secure.corp:3129")
+    );
+    assert_eq!(config.no_proxy.as_deref(), Some(".internal"));
+    assert!(!config.is_empty());
+}
+
+#[test]
+fn proxy_config_from_env_is_empty_when_no_vars_set() {
+    // given
+    let _lock = env_lock();
+    let _http = EnvVarGuard::set("HTTP_PROXY", None);
+    let _https = EnvVarGuard::set("HTTPS_PROXY", None);
+    let _no = EnvVarGuard::set("NO_PROXY", None);
+    let _http_lower = EnvVarGuard::set("http_proxy", None);
+    let _https_lower = EnvVarGuard::set("https_proxy", None);
+    let _no_lower = EnvVarGuard::set("no_proxy", None);
+
+    // when
+    let config = ProxyConfig::from_env();
+
+    // then
+    assert!(config.is_empty());
+    assert!(config.http_proxy.is_none());
+    assert!(config.https_proxy.is_none());
+    assert!(config.no_proxy.is_none());
+}
+
+#[test]
+fn proxy_config_from_env_treats_empty_values_as_unset() {
+    // given
+    let _lock = env_lock();
+    let _http = EnvVarGuard::set("HTTP_PROXY", Some(""));
+    let _https = EnvVarGuard::set("HTTPS_PROXY", Some(""));
+    let _http_lower = EnvVarGuard::set("http_proxy", Some(""));
+    let _https_lower = EnvVarGuard::set("https_proxy", Some(""));
+    let _no = EnvVarGuard::set("NO_PROXY", Some(""));
+    let _no_lower = EnvVarGuard::set("no_proxy", Some(""));
+
+    // when
+    let config = ProxyConfig::from_env();
+
+    // then
+    assert!(config.is_empty());
+}
+
+#[test]
+fn build_client_with_env_proxy_config_succeeds() {
+    // given
+    let _lock = env_lock();
+    let _http = EnvVarGuard::set("HTTP_PROXY", Some("http://proxy.corp:3128"));
+    let _https = EnvVarGuard::set("HTTPS_PROXY", Some("http://secure.corp:3129"));
+    let _no = EnvVarGuard::set("NO_PROXY", Some("localhost"));
+    let _http_lower = EnvVarGuard::set("http_proxy", None);
+    let _https_lower = EnvVarGuard::set("https_proxy", None);
+    let _no_lower = EnvVarGuard::set("no_proxy", None);
+    let config = ProxyConfig::from_env();
+
+    // when
+    let result = build_http_client_with(&config);
+
+    // then
+    assert!(result.is_ok());
+}
+
+#[test]
+fn build_client_with_proxy_url_config_succeeds() {
+    // given
+    let config = ProxyConfig::from_proxy_url("http://unified.corp:3128");
+
+    // when
+    let result = build_http_client_with(&config);
+
+    // then
+    assert!(result.is_ok());
+}
+
+#[test]
+fn proxy_config_from_env_prefers_uppercase_over_lowercase() {
+    // given
+    let _lock = env_lock();
+    let _http_upper = EnvVarGuard::set("HTTP_PROXY", Some("http://upper.corp:3128"));
+    let _http_lower = EnvVarGuard::set("http_proxy", Some("http://lower.corp:3128"));
+    let _https = EnvVarGuard::set("HTTPS_PROXY", None);
+    let _https_lower = EnvVarGuard::set("https_proxy", None);
+    let _no = EnvVarGuard::set("NO_PROXY", None);
+    let _no_lower = EnvVarGuard::set("no_proxy", None);
+
+    // when
+    let config = ProxyConfig::from_env();
+
+    // then
+    assert_eq!(config.http_proxy.as_deref(), Some("http://upper.corp:3128"));
+}
--- a/rust/crates/commands/src/lib.rs
+++ b/rust/crates/commands/src/lib.rs
@@ -221,8 +221,10 @@ const SLASH_COMMAND_SPECS: &[SlashCommandSpec] = &[
    SlashCommandSpec {
        name: "session",
        aliases: &[],
-        summary: "List, switch, or fork managed local sessions",
-        argument_hint: Some("[list|switch <session-id>|fork [branch-name]]"),
+        summary: "List, switch, fork, or delete managed local sessions",
+        argument_hint: Some(
+            "[list|switch <session-id>|fork [branch-name]|delete <session-id> [--force]]",
+        ),
        resume_supported: false,
    },
    SlashCommandSpec {
@@ -1188,6 +1190,9 @@ pub enum SlashCommand {
    AddDir {
        path: Option<String>,
    },
+    History {
+        count: Option<String>,
+    },
    Unknown(String),
 }

@@ -1421,6 +1426,9 @@ pub fn validate_slash_command_input(
        "tag" => SlashCommand::Tag { label: remainder },
        "output-style" => SlashCommand::OutputStyle { style: remainder },
        "add-dir" => SlashCommand::AddDir { path: remainder },
+        "history" => SlashCommand::History {
+            count: optional_single_arg(command, &args, "[count]")?,
+        },
        other => SlashCommand::Unknown(other.to_string()),
    }))
 }
@@ -1520,7 +1528,7 @@ fn parse_session_command(args: &[&str]) -> Result<SlashCommand, SlashCommandPars
            action: Some("list".to_string()),
            target: None,
        }),
-        ["list", ..] => Err(usage_error("session", "[list|switch <session-id>|fork [branch-name]]")),
+        ["list", ..] => Err(usage_error("session", "[list|switch <session-id>|fork [branch-name]|delete <session-id> [--force]]")),
        ["switch"] => Err(usage_error("session switch", "<session-id>")),
        ["switch", target] => Ok(SlashCommand::Session {
            action: Some("switch".to_string()),
@@ -1544,12 +1552,33 @@ fn parse_session_command(args: &[&str]) -> Result<SlashCommand, SlashCommandPars
            "session",
            "/session fork [branch-name]",
        )),
-        [action, ..] => Err(command_error(
+        ["delete"] => Err(usage_error("session delete", "<session-id> [--force]")),
+        ["delete", target] => Ok(SlashCommand::Session {
+            action: Some("delete".to_string()),
+            target: Some((*target).to_string()),
+        }),
+        ["delete", target, "--force"] => Ok(SlashCommand::Session {
+            action: Some("delete-force".to_string()),
+            target: Some((*target).to_string()),
+        }),
+        ["delete", _target, unexpected] => Err(command_error(
            &format!(
-                "Unknown /session action '{action}'. Use list, switch <session-id>, or fork [branch-name]."
+                "Unsupported /session delete flag '{unexpected}'. Use --force to skip confirmation."
            ),
            "session",
-            "/session [list|switch <session-id>|fork [branch-name]]",
+            "/session delete <session-id> [--force]",
+        )),
+        ["delete", ..] => Err(command_error(
+            "Unexpected arguments for /session delete.",
+            "session",
+            "/session delete <session-id> [--force]",
+        )),
+        [action, ..] => Err(command_error(
+            &format!(
+                "Unknown /session action '{action}'. Use list, switch <session-id>, fork [branch-name], or delete <session-id> [--force]."
+            ),
+            "session",
+            "/session [list|switch <session-id>|fork [branch-name]|delete <session-id> [--force]]",
        )),
    }
 }
@@ -1786,24 +1815,29 @@ pub fn resume_supported_slash_commands() -> Vec<&'static SlashCommandSpec> {

 fn slash_command_category(name: &str) -> &'static str {
    match name {
-        "help" | "status" | "sandbox" | "model" | "permissions" | "cost" | "resume" | "session"
-        | "version" | "login" | "logout" | "usage" | "stats" | "rename" | "privacy-settings" => {
-            "Session & visibility"
-        }
-        "compact" | "clear" | "config" | "memory" | "init" | "diff" | "commit" | "pr" | "issue"
-        | "export" | "plugin" | "branch" | "add-dir" | "files" | "hooks" | "release-notes" => {
-            "Workspace & git"
-        }
-        "agents" | "skills" | "teleport" | "debug-tool-call" | "mcp" | "context" | "tasks"
-        | "doctor" | "ide" | "desktop" => "Discovery & debugging",
-        "bughunter" | "ultraplan" | "review" | "security-review" | "advisor" | "insights" => {
-            "Analysis & automation"
-        }
-        "theme" | "vim" | "voice" | "color" | "effort" | "fast" | "brief" | "output-style"
-        | "keybindings" | "stickers" => "Appearance & input",
-        "copy" | "share" | "feedback" | "summary" | "tag" | "thinkback" | "plan" | "exit"
-        | "upgrade" | "rewind" => "Communication & control",
-        _ => "Other",
+        "help" | "status" | "cost" | "resume" | "session" | "version" | "login" | "logout"
+        | "usage" | "stats" | "rename" | "clear" | "compact" | "history" | "tokens" | "cache"
+        | "exit" | "summary" | "tag" | "thinkback" | "copy" | "share" | "feedback" | "rewind"
+        | "pin" | "unpin" | "bookmarks" | "context" | "files" | "focus" | "unfocus" | "retry"
+        | "stop" | "undo" => "Session",
+        "diff" | "commit" | "pr" | "issue" | "branch" | "blame" | "log" | "git" | "stash"
+        | "init" | "export" | "plan" | "review" | "security-review" | "bughunter" | "ultraplan"
+        | "teleport" | "refactor" | "fix" | "autofix" | "explain" | "docs" | "perf" | "search"
+        | "references" | "definition" | "hover" | "symbols" | "map" | "web" | "image"
+        | "screenshot" | "paste" | "listen" | "speak" | "test" | "lint" | "build" | "run"
+        | "format" | "parallel" | "multi" | "macro" | "alias" | "templates" | "migrate"
+        | "benchmark" | "cron" | "agent" | "subagent" | "agents" | "skills" | "team" | "plugin"
+        | "mcp" | "hooks" | "tasks" | "advisor" | "insights" | "release-notes" | "chat"
+        | "approve" | "deny" | "allowed-tools" | "add-dir" => "Tools",
+        "model" | "permissions" | "config" | "memory" | "theme" | "vim" | "voice" | "color"
+        | "effort" | "fast" | "brief" | "output-style" | "keybindings" | "privacy-settings"
+        | "stickers" | "language" | "profile" | "max-tokens" | "temperature" | "system-prompt"
+        | "api-key" | "terminal-setup" | "notifications" | "telemetry" | "providers" | "env"
+        | "project" | "reasoning" | "budget" | "rate-limit" | "workspace" | "reset" | "ide"
+        | "desktop" | "upgrade" => "Config",
+        "debug-tool-call" | "doctor" | "sandbox" | "diagnostics" | "tool-details" | "changelog"
+        | "metrics" => "Debug",
+        _ => "Tools",
    }
 }

@@ -1912,12 +1946,7 @@ pub fn render_slash_command_help() -> String {
        String::new(),
    ];

-    let categories = [
-        "Session & visibility",
-        "Workspace & git",
-        "Discovery & debugging",
-        "Analysis & automation",
-    ];
+    let categories = ["Session", "Tools", "Config", "Debug"];

    for category in categories {
        lines.push(category.to_string());
@@ -1930,6 +1959,12 @@ pub fn render_slash_command_help() -> String {
        lines.push(String::new());
    }

+    lines.push("Keyboard shortcuts".to_string());
+    lines.push("  Up/Down              Navigate prompt history".to_string());
+    lines.push("  Tab                  Complete commands, modes, and recent sessions".to_string());
+    lines.push("  Ctrl-C               Clear input (or exit on empty prompt)".to_string());
+    lines.push("  Shift+Enter/Ctrl+J   Insert a newline".to_string());
+
    lines
        .into_iter()
        .rev()
@@ -2314,8 +2349,7 @@ pub fn resolve_skill_invocation(
            .unwrap_or_default();
        if !skill_token.is_empty() {
            if let Err(error) = resolve_skill_path(cwd, skill_token) {
-                let mut message =
-                    format!("Unknown skill: {skill_token} ({error})");
+                let mut message = format!("Unknown skill: {skill_token} ({error})");
                let roots = discover_skill_roots(cwd);
                if let Ok(available) = load_skills_from_roots(&roots) {
                    let names: Vec<String> = available
@@ -2324,15 +2358,10 @@ pub fn resolve_skill_invocation(
                        .map(|s| s.name.clone())
                        .collect();
                    if !names.is_empty() {
-                        message.push_str(&format!(
-                            "\n  Available skills: {}",
-                            names.join(", ")
-                        ));
+                        message.push_str(&format!("\n  Available skills: {}", names.join(", ")));
                    }
                }
-                message.push_str(
-                    "\n  Usage: /skills [list|install <path>|help|<skill> [args]]",
-                );
+                message.push_str("\n  Usage: /skills [list|install <path>|help|<skill> [args]]");
                return Err(message);
            }
        }
@@ -3942,6 +3971,7 @@ pub fn handle_slash_command(
        | SlashCommand::Tag { .. }
        | SlashCommand::OutputStyle { .. }
        | SlashCommand::AddDir { .. }
+        | SlashCommand::History { .. }
        | SlashCommand::Unknown(_) => None,
    }
 }
@@ -4256,6 +4286,47 @@ mod tests {
        );
    }

+    #[test]
+    fn parses_history_command_without_count() {
+        // given
+        let input = "/history";
+
+        // when
+        let parsed = SlashCommand::parse(input);
+
+        // then
+        assert_eq!(parsed, Ok(Some(SlashCommand::History { count: None })));
+    }
+
+    #[test]
+    fn parses_history_command_with_numeric_count() {
+        // given
+        let input = "/history 25";
+
+        // when
+        let parsed = SlashCommand::parse(input);
+
+        // then
+        assert_eq!(
+            parsed,
+            Ok(Some(SlashCommand::History {
+                count: Some("25".to_string())
+            }))
+        );
+    }
+
+    #[test]
+    fn rejects_history_with_extra_arguments() {
+        // given
+        let input = "/history 25 extra";
+
+        // when
+        let error = parse_error_message(input);
+
+        // then
+        assert!(error.contains("Usage: /history [count]"));
+    }
+
    #[test]
    fn rejects_unexpected_arguments_for_no_arg_commands() {
        // given
@@ -4297,7 +4368,7 @@ mod tests {

        // then
        assert!(error.contains("Usage: /teleport <symbol-or-path>"));
-        assert!(error.contains("  Category         Discovery & debugging"));
+        assert!(error.contains("  Category         Tools"));
    }

    #[test]
@@ -4371,10 +4442,10 @@ mod tests {
        let help = render_slash_command_help();
        assert!(help.contains("Start here        /status, /diff, /agents, /skills, /commit"));
        assert!(help.contains("[resume]          also works with --resume SESSION.jsonl"));
-        assert!(help.contains("Session & visibility"));
-        assert!(help.contains("Workspace & git"));
-        assert!(help.contains("Discovery & debugging"));
-        assert!(help.contains("Analysis & automation"));
+        assert!(help.contains("Session"));
+        assert!(help.contains("Tools"));
+        assert!(help.contains("Config"));
+        assert!(help.contains("Debug"));
        assert!(help.contains("/help"));
        assert!(help.contains("/status"));
        assert!(help.contains("/sandbox"));
@@ -4398,7 +4469,7 @@ mod tests {
        assert!(help.contains("/diff"));
        assert!(help.contains("/version"));
        assert!(help.contains("/export [file]"));
-        assert!(help.contains("/session [list|switch <session-id>|fork [branch-name]]"));
+        assert!(help.contains("/session"), "help must mention /session");
        assert!(help.contains("/sandbox"));
        assert!(help.contains(
            "/plugin [list|install <path>|enable <name>|disable <name>|uninstall <id>|update <id>]"
@@ -4411,6 +4482,53 @@ mod tests {
        assert!(resume_supported_slash_commands().len() >= 39);
    }

+    #[test]
+    fn renders_help_with_grouped_categories_and_keyboard_shortcuts() {
+        // given
+        let categories = ["Session", "Tools", "Config", "Debug"];
+
+        // when
+        let help = render_slash_command_help();
+
+        // then
+        for category in categories {
+            assert!(
+                help.contains(category),
+                "expected help to contain category {category}"
+            );
+        }
+        let session_index = help.find("Session").expect("Session header should exist");
+        let tools_index = help.find("Tools").expect("Tools header should exist");
+        let config_index = help.find("Config").expect("Config header should exist");
+        let debug_index = help.find("Debug").expect("Debug header should exist");
+        assert!(session_index < tools_index);
+        assert!(tools_index < config_index);
+        assert!(config_index < debug_index);
+
+        assert!(help.contains("Keyboard shortcuts"));
+        assert!(help.contains("Up/Down              Navigate prompt history"));
+        assert!(help.contains("Tab                  Complete commands, modes, and recent sessions"));
+        assert!(help.contains("Ctrl-C               Clear input (or exit on empty prompt)"));
+        assert!(help.contains("Shift+Enter/Ctrl+J   Insert a newline"));
+
+        // every command should still render with a summary line
+        for spec in slash_command_specs() {
+            let usage = match spec.argument_hint {
+                Some(hint) => format!("/{} {hint}", spec.name),
+                None => format!("/{}", spec.name),
+            };
+            assert!(
+                help.contains(&usage),
+                "expected help to contain command {usage}"
+            );
+            assert!(
+                help.contains(spec.summary),
+                "expected help to contain summary for /{}",
+                spec.name
+            );
+        }
+    }
+
    #[test]
    fn renders_per_command_help_detail() {
        // given
@@ -4423,7 +4541,7 @@ mod tests {
        assert!(help.contains("/plugin"));
        assert!(help.contains("Summary          Manage Claw Code plugins"));
        assert!(help.contains("Aliases          /plugins, /marketplace"));
-        assert!(help.contains("Category         Workspace & git"));
+        assert!(help.contains("Category         Tools"));
    }

    #[test]
@@ -4431,7 +4549,7 @@ mod tests {
        let help = render_slash_command_help_detail("mcp").expect("detail help should exist");
        assert!(help.contains("/mcp"));
        assert!(help.contains("Summary          Inspect configured MCP servers"));
-        assert!(help.contains("Category         Discovery & debugging"));
+        assert!(help.contains("Category         Tools"));
        assert!(help.contains("Resume           Supported with --resume SESSION.jsonl"));
    }

--- a/rust/crates/runtime/Cargo.toml
+++ b/rust/crates/runtime/Cargo.toml
@@ -13,7 +13,7 @@ regex = "1"
 serde = { version = "1", features = ["derive"] }
 serde_json.workspace = true
 telemetry = { path = "../telemetry" }
-tokio = { version = "1", features = ["io-util", "macros", "process", "rt", "rt-multi-thread", "time"] }
+tokio = { version = "1", features = ["io-std", "io-util", "macros", "process", "rt", "rt-multi-thread", "time"] }
 walkdir = "2"

 [lints]
--- a/rust/crates/runtime/src/config.rs
+++ b/rust/crates/runtime/src/config.rs
@@ -48,6 +48,7 @@ pub struct RuntimePluginConfig {
    install_root: Option<String>,
    registry_path: Option<String>,
    bundled_root: Option<String>,
+    max_output_tokens: Option<u32>,
 }

 /// Structured feature configuration consumed by runtime subsystems.
@@ -58,9 +59,21 @@ pub struct RuntimeFeatureConfig {
    mcp: McpConfigCollection,
    oauth: Option<OAuthConfig>,
    model: Option<String>,
+    aliases: BTreeMap<String, String>,
    permission_mode: Option<ResolvedPermissionMode>,
    permission_rules: RuntimePermissionRuleConfig,
    sandbox: SandboxConfig,
+    provider_fallbacks: ProviderFallbackConfig,
+    trusted_roots: Vec<String>,
+}
+
+/// Ordered chain of fallback model identifiers used when the primary
+/// provider returns a retryable failure (429/500/503/etc.). The chain is
+/// strict: each entry is tried in order until one succeeds.
+#[derive(Debug, Clone, PartialEq, Eq, Default)]
+pub struct ProviderFallbackConfig {
+    primary: Option<String>,
+    fallbacks: Vec<String>,
 }

 /// Hook command lists grouped by lifecycle stage.
@@ -259,17 +272,33 @@ impl ConfigLoader {
        let mut merged = BTreeMap::new();
        let mut loaded_entries = Vec::new();
        let mut mcp_servers = BTreeMap::new();
+        let mut all_warnings = Vec::new();

        for entry in self.discover() {
-            let Some(value) = read_optional_json_object(&entry.path)? else {
+            crate::config_validate::check_unsupported_format(&entry.path)?;
+            let Some(parsed) = read_optional_json_object(&entry.path)? else {
                continue;
            };
-            validate_optional_hooks_config(&value, &entry.path)?;
-            merge_mcp_servers(&mut mcp_servers, entry.source, &value, &entry.path)?;
-            deep_merge_objects(&mut merged, &value);
+            let validation = crate::config_validate::validate_config_file(
+                &parsed.object,
+                &parsed.source,
+                &entry.path,
+            );
+            if !validation.is_ok() {
+                let first_error = &validation.errors[0];
+                return Err(ConfigError::Parse(first_error.to_string()));
+            }
+            all_warnings.extend(validation.warnings);
+            validate_optional_hooks_config(&parsed.object, &entry.path)?;
+            merge_mcp_servers(&mut mcp_servers, entry.source, &parsed.object, &entry.path)?;
+            deep_merge_objects(&mut merged, &parsed.object);
            loaded_entries.push(entry);
        }

+        for warning in &all_warnings {
+            eprintln!("warning: {warning}");
+        }
+
        let merged_value = JsonValue::Object(merged.clone());

        let feature_config = RuntimeFeatureConfig {
@@ -280,9 +309,12 @@ impl ConfigLoader {
            },
            oauth: parse_optional_oauth_config(&merged_value, "merged settings.oauth")?,
            model: parse_optional_model(&merged_value),
+            aliases: parse_optional_aliases(&merged_value)?,
            permission_mode: parse_optional_permission_mode(&merged_value)?,
            permission_rules: parse_optional_permission_rules(&merged_value)?,
            sandbox: parse_optional_sandbox_config(&merged_value)?,
+            provider_fallbacks: parse_optional_provider_fallbacks(&merged_value)?,
+            trusted_roots: parse_optional_trusted_roots(&merged_value)?,
        };

        Ok(RuntimeConfig {
@@ -353,6 +385,11 @@ impl RuntimeConfig {
        self.feature_config.model.as_deref()
    }

+    #[must_use]
+    pub fn aliases(&self) -> &BTreeMap<String, String> {
+        &self.feature_config.aliases
+    }
+
    #[must_use]
    pub fn permission_mode(&self) -> Option<ResolvedPermissionMode> {
        self.feature_config.permission_mode
@@ -367,6 +404,16 @@ impl RuntimeConfig {
    pub fn sandbox(&self) -> &SandboxConfig {
        &self.feature_config.sandbox
    }
+
+    #[must_use]
+    pub fn provider_fallbacks(&self) -> &ProviderFallbackConfig {
+        &self.feature_config.provider_fallbacks
+    }
+
+    #[must_use]
+    pub fn trusted_roots(&self) -> &[String] {
+        &self.feature_config.trusted_roots
+    }
 }

 impl RuntimeFeatureConfig {
@@ -407,6 +454,11 @@ impl RuntimeFeatureConfig {
        self.model.as_deref()
    }

+    #[must_use]
+    pub fn aliases(&self) -> &BTreeMap<String, String> {
+        &self.aliases
+    }
+
    #[must_use]
    pub fn permission_mode(&self) -> Option<ResolvedPermissionMode> {
        self.permission_mode
@@ -421,6 +473,38 @@ impl RuntimeFeatureConfig {
    pub fn sandbox(&self) -> &SandboxConfig {
        &self.sandbox
    }
+
+    #[must_use]
+    pub fn provider_fallbacks(&self) -> &ProviderFallbackConfig {
+        &self.provider_fallbacks
+    }
+
+    #[must_use]
+    pub fn trusted_roots(&self) -> &[String] {
+        &self.trusted_roots
+    }
+}
+
+impl ProviderFallbackConfig {
+    #[must_use]
+    pub fn new(primary: Option<String>, fallbacks: Vec<String>) -> Self {
+        Self { primary, fallbacks }
+    }
+
+    #[must_use]
+    pub fn primary(&self) -> Option<&str> {
+        self.primary.as_deref()
+    }
+
+    #[must_use]
+    pub fn fallbacks(&self) -> &[String] {
+        &self.fallbacks
+    }
+
+    #[must_use]
+    pub fn is_empty(&self) -> bool {
+        self.fallbacks.is_empty()
+    }
 }

 impl RuntimePluginConfig {
@@ -449,6 +533,15 @@ impl RuntimePluginConfig {
        self.bundled_root.as_deref()
    }

+    #[must_use]
+    pub fn max_output_tokens(&self) -> Option<u32> {
+        self.max_output_tokens
+    }
+
+    pub fn set_max_output_tokens(&mut self, max_output_tokens: Option<u32>) {
+        self.max_output_tokens = max_output_tokens;
+    }
+
    pub fn set_plugin_state(&mut self, plugin_id: String, enabled: bool) {
        self.enabled_plugins.insert(plugin_id, enabled);
    }
@@ -572,9 +665,13 @@ impl McpServerConfig {
    }
 }

-fn read_optional_json_object(
-    path: &Path,
-) -> Result<Option<BTreeMap<String, JsonValue>>, ConfigError> {
+/// Parsed JSON object paired with its raw source text for validation.
+struct ParsedConfigFile {
+    object: BTreeMap<String, JsonValue>,
+    source: String,
+}
+
+fn read_optional_json_object(path: &Path) -> Result<Option<ParsedConfigFile>, ConfigError> {
    let is_legacy_config = path.file_name().and_then(|name| name.to_str()) == Some(".claw.json");
    let contents = match fs::read_to_string(path) {
        Ok(contents) => contents,
@@ -583,7 +680,10 @@ fn read_optional_json_object(
    };

    if contents.trim().is_empty() {
-        return Ok(Some(BTreeMap::new()));
+        return Ok(Some(ParsedConfigFile {
+            object: BTreeMap::new(),
+            source: contents,
+        }));
    }

    let parsed = match JsonValue::parse(&contents) {
@@ -600,7 +700,10 @@ fn read_optional_json_object(
            path.display()
        )));
    };
-    Ok(Some(object.clone()))
+    Ok(Some(ParsedConfigFile {
+        object: object.clone(),
+        source: contents,
+    }))
 }

 fn merge_mcp_servers(
@@ -637,6 +740,13 @@ fn parse_optional_model(root: &JsonValue) -> Option<String> {
        .map(ToOwned::to_owned)
 }

+fn parse_optional_aliases(root: &JsonValue) -> Result<BTreeMap<String, String>, ConfigError> {
+    let Some(object) = root.as_object() else {
+        return Ok(BTreeMap::new());
+    };
+    Ok(optional_string_map(object, "aliases", "merged settings")?.unwrap_or_default())
+}
+
 fn parse_optional_hooks_config(root: &JsonValue) -> Result<RuntimeHookConfig, ConfigError> {
    let Some(object) = root.as_object() else {
        return Ok(RuntimeHookConfig::default());
@@ -714,6 +824,7 @@ fn parse_optional_plugin_config(root: &JsonValue) -> Result<RuntimePluginConfig,
        optional_string(plugins, "registryPath", "merged settings.plugins")?.map(str::to_string);
    config.bundled_root =
        optional_string(plugins, "bundledRoot", "merged settings.plugins")?.map(str::to_string);
+    config.max_output_tokens = optional_u32(plugins, "maxOutputTokens", "merged settings.plugins")?;
    Ok(config)
 }

@@ -776,6 +887,33 @@ fn parse_optional_sandbox_config(root: &JsonValue) -> Result<SandboxConfig, Conf
    })
 }

+fn parse_optional_provider_fallbacks(
+    root: &JsonValue,
+) -> Result<ProviderFallbackConfig, ConfigError> {
+    let Some(object) = root.as_object() else {
+        return Ok(ProviderFallbackConfig::default());
+    };
+    let Some(value) = object.get("providerFallbacks") else {
+        return Ok(ProviderFallbackConfig::default());
+    };
+    let entry = expect_object(value, "merged settings.providerFallbacks")?;
+    let primary =
+        optional_string(entry, "primary", "merged settings.providerFallbacks")?.map(str::to_string);
+    let fallbacks = optional_string_array(entry, "fallbacks", "merged settings.providerFallbacks")?
+        .unwrap_or_default();
+    Ok(ProviderFallbackConfig { primary, fallbacks })
+}
+
+fn parse_optional_trusted_roots(root: &JsonValue) -> Result<Vec<String>, ConfigError> {
+    let Some(object) = root.as_object() else {
+        return Ok(Vec::new());
+    };
+    Ok(
+        optional_string_array(object, "trustedRoots", "merged settings.trustedRoots")?
+            .unwrap_or_default(),
+    )
+}
+
 fn parse_filesystem_mode_label(value: &str) -> Result<FilesystemIsolationMode, ConfigError> {
    match value {
        "off" => Ok(FilesystemIsolationMode::Off),
@@ -957,6 +1095,27 @@ fn optional_u16(
    }
 }

+fn optional_u32(
+    object: &BTreeMap<String, JsonValue>,
+    key: &str,
+    context: &str,
+) -> Result<Option<u32>, ConfigError> {
+    match object.get(key) {
+        Some(value) => {
+            let Some(number) = value.as_i64() else {
+                return Err(ConfigError::Parse(format!(
+                    "{context}: field {key} must be a non-negative integer"
+                )));
+            };
+            let number = u32::try_from(number).map_err(|_| {
+                ConfigError::Parse(format!("{context}: field {key} is out of range"))
+            })?;
+            Ok(Some(number))
+        }
+        None => Ok(None),
+    }
+}
+
 fn optional_u64(
    object: &BTreeMap<String, JsonValue>,
    key: &str,
@@ -1247,6 +1406,113 @@ mod tests {
        fs::remove_dir_all(root).expect("cleanup temp dir");
    }

+    #[test]
+    fn parses_provider_fallbacks_chain_with_primary_and_ordered_fallbacks() {
+        // given
+        let root = temp_dir();
+        let cwd = root.join("project");
+        let home = root.join("home").join(".claw");
+        fs::create_dir_all(cwd.join(".claw")).expect("project config dir");
+        fs::create_dir_all(&home).expect("home config dir");
+        fs::write(
+            home.join("settings.json"),
+            r#"{
+              "providerFallbacks": {
+                "primary": "claude-opus-4-6",
+                "fallbacks": ["grok-3", "grok-3-mini"]
+              }
+            }"#,
+        )
+        .expect("write provider fallback settings");
+
+        // when
+        let loaded = ConfigLoader::new(&cwd, &home)
+            .load()
+            .expect("config should load");
+
+        // then
+        let chain = loaded.provider_fallbacks();
+        assert_eq!(chain.primary(), Some("claude-opus-4-6"));
+        assert_eq!(
+            chain.fallbacks(),
+            &["grok-3".to_string(), "grok-3-mini".to_string()]
+        );
+        assert!(!chain.is_empty());
+
+        fs::remove_dir_all(root).expect("cleanup temp dir");
+    }
+
+    #[test]
+    fn provider_fallbacks_default_is_empty_when_unset() {
+        // given
+        let root = temp_dir();
+        let cwd = root.join("project");
+        let home = root.join("home").join(".claw");
+        fs::create_dir_all(&home).expect("home config dir");
+        fs::create_dir_all(&cwd).expect("project dir");
+        fs::write(home.join("settings.json"), "{}").expect("write empty settings");
+
+        // when
+        let loaded = ConfigLoader::new(&cwd, &home)
+            .load()
+            .expect("config should load");
+
+        // then
+        let chain = loaded.provider_fallbacks();
+        assert_eq!(chain.primary(), None);
+        assert!(chain.fallbacks().is_empty());
+        assert!(chain.is_empty());
+
+        fs::remove_dir_all(root).expect("cleanup temp dir");
+    }
+
+    #[test]
+    fn parses_trusted_roots_from_settings() {
+        // given
+        let root = temp_dir();
+        let cwd = root.join("project");
+        let home = root.join("home").join(".claw");
+        fs::create_dir_all(&home).expect("home config dir");
+        fs::create_dir_all(&cwd).expect("project dir");
+        fs::write(
+            home.join("settings.json"),
+            r#"{"trustedRoots": ["/tmp/worktrees", "/home/user/projects"]}"#,
+        )
+        .expect("write settings");
+
+        // when
+        let loaded = ConfigLoader::new(&cwd, &home)
+            .load()
+            .expect("config should load");
+
+        // then
+        let roots = loaded.trusted_roots();
+        assert_eq!(roots, ["/tmp/worktrees", "/home/user/projects"]);
+
+        fs::remove_dir_all(root).expect("cleanup temp dir");
+    }
+
+    #[test]
+    fn trusted_roots_default_is_empty_when_unset() {
+        // given
+        let root = temp_dir();
+        let cwd = root.join("project");
+        let home = root.join("home").join(".claw");
+        fs::create_dir_all(&home).expect("home config dir");
+        fs::create_dir_all(&cwd).expect("project dir");
+        fs::write(home.join("settings.json"), "{}").expect("write empty settings");
+
+        // when
+        let loaded = ConfigLoader::new(&cwd, &home)
+            .load()
+            .expect("config should load");
+
+        // then
+        assert!(loaded.trusted_roots().is_empty());
+
+        fs::remove_dir_all(root).expect("cleanup temp dir");
+    }
+
    #[test]
    fn parses_typed_mcp_and_oauth_config() {
        let root = temp_dir();
@@ -1493,6 +1759,49 @@ mod tests {
        fs::remove_dir_all(root).expect("cleanup temp dir");
    }

+    #[test]
+    fn parses_user_defined_model_aliases_from_settings() {
+        // given
+        let root = temp_dir();
+        let cwd = root.join("project");
+        let home = root.join("home").join(".claw");
+        fs::create_dir_all(cwd.join(".claw")).expect("project config dir");
+        fs::create_dir_all(&home).expect("home config dir");
+
+        fs::write(
+            home.join("settings.json"),
+            r#"{"aliases":{"fast":"claude-haiku-4-5-20251213","smart":"claude-opus-4-6"}}"#,
+        )
+        .expect("write user settings");
+        fs::write(
+            cwd.join(".claw").join("settings.local.json"),
+            r#"{"aliases":{"smart":"claude-sonnet-4-6","cheap":"grok-3-mini"}}"#,
+        )
+        .expect("write local settings");
+
+        // when
+        let loaded = ConfigLoader::new(&cwd, &home)
+            .load()
+            .expect("config should load");
+
+        // then
+        let aliases = loaded.aliases();
+        assert_eq!(
+            aliases.get("fast").map(String::as_str),
+            Some("claude-haiku-4-5-20251213")
+        );
+        assert_eq!(
+            aliases.get("smart").map(String::as_str),
+            Some("claude-sonnet-4-6")
+        );
+        assert_eq!(
+            aliases.get("cheap").map(String::as_str),
+            Some("grok-3-mini")
+        );
+
+        fs::remove_dir_all(root).expect("cleanup temp dir");
+    }
+
    #[test]
    fn empty_settings_file_loads_defaults() {
        // given
@@ -1574,12 +1883,13 @@ mod tests {
            .load()
            .expect_err("config should fail");

-        // then
+        // then — config validation now catches the mixed array before the hooks parser
        let rendered = error.to_string();
-        assert!(rendered.contains(&format!(
-            "{}: hooks: field PreToolUse must contain only strings",
-            project_settings.display()
-        )));
+        assert!(
+            rendered.contains("hooks.PreToolUse")
+                && rendered.contains("must be an array of strings"),
+            "expected validation error for hooks.PreToolUse, got: {rendered}"
+        );
        assert!(!rendered.contains("merged settings.hooks"));

        fs::remove_dir_all(root).expect("cleanup temp dir");
@@ -1645,4 +1955,157 @@ mod tests {
        assert!(config.state_for("missing", true));
        assert!(!config.state_for("missing", false));
    }
+
+    #[test]
+    fn validates_unknown_top_level_keys_with_line_and_field_name() {
+        // given
+        let root = temp_dir();
+        let cwd = root.join("project");
+        let home = root.join("home").join(".claw");
+        let user_settings = home.join("settings.json");
+        fs::create_dir_all(&home).expect("home config dir");
+        fs::create_dir_all(&cwd).expect("project dir");
+        fs::write(
+            &user_settings,
+            "{\n  \"model\": \"opus\",\n  \"telemetry\": true\n}\n",
+        )
+        .expect("write user settings");
+
+        // when
+        let error = ConfigLoader::new(&cwd, &home)
+            .load()
+            .expect_err("config should fail");
+
+        // then
+        let rendered = error.to_string();
+        assert!(
+            rendered.contains(&user_settings.display().to_string()),
+            "error should include file path, got: {rendered}"
+        );
+        assert!(
+            rendered.contains("line 3"),
+            "error should include line number, got: {rendered}"
+        );
+        assert!(
+            rendered.contains("telemetry"),
+            "error should name the offending field, got: {rendered}"
+        );
+
+        fs::remove_dir_all(root).expect("cleanup temp dir");
+    }
+
+    #[test]
+    fn validates_deprecated_top_level_keys_with_replacement_guidance() {
+        // given
+        let root = temp_dir();
+        let cwd = root.join("project");
+        let home = root.join("home").join(".claw");
+        let user_settings = home.join("settings.json");
+        fs::create_dir_all(&home).expect("home config dir");
+        fs::create_dir_all(&cwd).expect("project dir");
+        fs::write(
+            &user_settings,
+            "{\n  \"model\": \"opus\",\n  \"allowedTools\": [\"Read\"]\n}\n",
+        )
+        .expect("write user settings");
+
+        // when
+        let error = ConfigLoader::new(&cwd, &home)
+            .load()
+            .expect_err("config should fail");
+
+        // then
+        let rendered = error.to_string();
+        assert!(
+            rendered.contains(&user_settings.display().to_string()),
+            "error should include file path, got: {rendered}"
+        );
+        assert!(
+            rendered.contains("line 3"),
+            "error should include line number, got: {rendered}"
+        );
+        assert!(
+            rendered.contains("allowedTools"),
+            "error should call out the unknown field, got: {rendered}"
+        );
+        // allowedTools is an unknown key; validator should name it in the error
+        assert!(
+            rendered.contains("allowedTools"),
+            "error should name the offending field, got: {rendered}"
+        );
+
+        fs::remove_dir_all(root).expect("cleanup temp dir");
+    }
+
+    #[test]
+    fn validates_wrong_type_for_known_field_with_field_path() {
+        // given
+        let root = temp_dir();
+        let cwd = root.join("project");
+        let home = root.join("home").join(".claw");
+        let user_settings = home.join("settings.json");
+        fs::create_dir_all(&home).expect("home config dir");
+        fs::create_dir_all(&cwd).expect("project dir");
+        fs::write(
+            &user_settings,
+            "{\n  \"hooks\": {\n    \"PreToolUse\": \"not-an-array\"\n  }\n}\n",
+        )
+        .expect("write user settings");
+
+        // when
+        let error = ConfigLoader::new(&cwd, &home)
+            .load()
+            .expect_err("config should fail");
+
+        // then
+        let rendered = error.to_string();
+        assert!(
+            rendered.contains(&user_settings.display().to_string()),
+            "error should include file path, got: {rendered}"
+        );
+        assert!(
+            rendered.contains("hooks"),
+            "error should include field path component 'hooks', got: {rendered}"
+        );
+        assert!(
+            rendered.contains("PreToolUse"),
+            "error should describe the type mismatch, got: {rendered}"
+        );
+        assert!(
+            rendered.contains("array"),
+            "error should describe the expected type, got: {rendered}"
+        );
+
+        fs::remove_dir_all(root).expect("cleanup temp dir");
+    }
+
+    #[test]
+    fn unknown_top_level_key_suggests_closest_match() {
+        // given
+        let root = temp_dir();
+        let cwd = root.join("project");
+        let home = root.join("home").join(".claw");
+        let user_settings = home.join("settings.json");
+        fs::create_dir_all(&home).expect("home config dir");
+        fs::create_dir_all(&cwd).expect("project dir");
+        fs::write(&user_settings, "{\n  \"modle\": \"opus\"\n}\n").expect("write user settings");
+
+        // when
+        let error = ConfigLoader::new(&cwd, &home)
+            .load()
+            .expect_err("config should fail");
+
+        // then
+        let rendered = error.to_string();
+        assert!(
+            rendered.contains("modle"),
+            "error should name the offending field, got: {rendered}"
+        );
+        assert!(
+            rendered.contains("model"),
+            "error should suggest the closest known key, got: {rendered}"
+        );
+
+        fs::remove_dir_all(root).expect("cleanup temp dir");
+    }
 }
--- a/rust/crates/runtime/src/config_validate.rs
+++ b/rust/crates/runtime/src/config_validate.rs
@@ -0,0 +1,901 @@
+use std::collections::BTreeMap;
+use std::path::Path;
+
+use crate::config::ConfigError;
+use crate::json::JsonValue;
+
+/// Diagnostic emitted when a config file contains a suspect field.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct ConfigDiagnostic {
+    pub path: String,
+    pub field: String,
+    pub line: Option<usize>,
+    pub kind: DiagnosticKind,
+}
+
+/// Classification of the diagnostic.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum DiagnosticKind {
+    UnknownKey {
+        suggestion: Option<String>,
+    },
+    WrongType {
+        expected: &'static str,
+        got: &'static str,
+    },
+    Deprecated {
+        replacement: &'static str,
+    },
+}
+
+impl std::fmt::Display for ConfigDiagnostic {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        let location = self
+            .line
+            .map_or_else(String::new, |line| format!(" (line {line})"));
+        match &self.kind {
+            DiagnosticKind::UnknownKey { suggestion: None } => {
+                write!(f, "{}: unknown key \"{}\"{location}", self.path, self.field)
+            }
+            DiagnosticKind::UnknownKey {
+                suggestion: Some(hint),
+            } => {
+                write!(
+                    f,
+                    "{}: unknown key \"{}\"{location}. Did you mean \"{}\"?",
+                    self.path, self.field, hint
+                )
+            }
+            DiagnosticKind::WrongType { expected, got } => {
+                write!(
+                    f,
+                    "{}: field \"{}\" must be {expected}, got {got}{location}",
+                    self.path, self.field
+                )
+            }
+            DiagnosticKind::Deprecated { replacement } => {
+                write!(
+                    f,
+                    "{}: field \"{}\" is deprecated{location}. Use \"{replacement}\" instead",
+                    self.path, self.field
+                )
+            }
+        }
+    }
+}
+
+/// Result of validating a single config file.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct ValidationResult {
+    pub errors: Vec<ConfigDiagnostic>,
+    pub warnings: Vec<ConfigDiagnostic>,
+}
+
+impl ValidationResult {
+    #[must_use]
+    pub fn is_ok(&self) -> bool {
+        self.errors.is_empty()
+    }
+
+    fn merge(&mut self, other: Self) {
+        self.errors.extend(other.errors);
+        self.warnings.extend(other.warnings);
+    }
+}
+
+// ---- known-key schema ----
+
+/// Expected type for a config field.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum FieldType {
+    String,
+    Bool,
+    Object,
+    StringArray,
+    Number,
+}
+
+impl FieldType {
+    fn label(self) -> &'static str {
+        match self {
+            Self::String => "a string",
+            Self::Bool => "a boolean",
+            Self::Object => "an object",
+            Self::StringArray => "an array of strings",
+            Self::Number => "a number",
+        }
+    }
+
+    fn matches(self, value: &JsonValue) -> bool {
+        match self {
+            Self::String => value.as_str().is_some(),
+            Self::Bool => value.as_bool().is_some(),
+            Self::Object => value.as_object().is_some(),
+            Self::StringArray => value
+                .as_array()
+                .is_some_and(|arr| arr.iter().all(|v| v.as_str().is_some())),
+            Self::Number => value.as_i64().is_some(),
+        }
+    }
+}
+
+fn json_type_label(value: &JsonValue) -> &'static str {
+    match value {
+        JsonValue::Null => "null",
+        JsonValue::Bool(_) => "a boolean",
+        JsonValue::Number(_) => "a number",
+        JsonValue::String(_) => "a string",
+        JsonValue::Array(_) => "an array",
+        JsonValue::Object(_) => "an object",
+    }
+}
+
+struct FieldSpec {
+    name: &'static str,
+    expected: FieldType,
+}
+
+struct DeprecatedField {
+    name: &'static str,
+    replacement: &'static str,
+}
+
+const TOP_LEVEL_FIELDS: &[FieldSpec] = &[
+    FieldSpec {
+        name: "$schema",
+        expected: FieldType::String,
+    },
+    FieldSpec {
+        name: "model",
+        expected: FieldType::String,
+    },
+    FieldSpec {
+        name: "hooks",
+        expected: FieldType::Object,
+    },
+    FieldSpec {
+        name: "permissions",
+        expected: FieldType::Object,
+    },
+    FieldSpec {
+        name: "permissionMode",
+        expected: FieldType::String,
+    },
+    FieldSpec {
+        name: "mcpServers",
+        expected: FieldType::Object,
+    },
+    FieldSpec {
+        name: "oauth",
+        expected: FieldType::Object,
+    },
+    FieldSpec {
+        name: "enabledPlugins",
+        expected: FieldType::Object,
+    },
+    FieldSpec {
+        name: "plugins",
+        expected: FieldType::Object,
+    },
+    FieldSpec {
+        name: "sandbox",
+        expected: FieldType::Object,
+    },
+    FieldSpec {
+        name: "env",
+        expected: FieldType::Object,
+    },
+    FieldSpec {
+        name: "aliases",
+        expected: FieldType::Object,
+    },
+    FieldSpec {
+        name: "providerFallbacks",
+        expected: FieldType::Object,
+    },
+    FieldSpec {
+        name: "trustedRoots",
+        expected: FieldType::StringArray,
+    },
+];
+
+const HOOKS_FIELDS: &[FieldSpec] = &[
+    FieldSpec {
+        name: "PreToolUse",
+        expected: FieldType::StringArray,
+    },
+    FieldSpec {
+        name: "PostToolUse",
+        expected: FieldType::StringArray,
+    },
+    FieldSpec {
+        name: "PostToolUseFailure",
+        expected: FieldType::StringArray,
+    },
+];
+
+const PERMISSIONS_FIELDS: &[FieldSpec] = &[
+    FieldSpec {
+        name: "defaultMode",
+        expected: FieldType::String,
+    },
+    FieldSpec {
+        name: "allow",
+        expected: FieldType::StringArray,
+    },
+    FieldSpec {
+        name: "deny",
+        expected: FieldType::StringArray,
+    },
+    FieldSpec {
+        name: "ask",
+        expected: FieldType::StringArray,
+    },
+];
+
+const PLUGINS_FIELDS: &[FieldSpec] = &[
+    FieldSpec {
+        name: "enabled",
+        expected: FieldType::Object,
+    },
+    FieldSpec {
+        name: "externalDirectories",
+        expected: FieldType::StringArray,
+    },
+    FieldSpec {
+        name: "installRoot",
+        expected: FieldType::String,
+    },
+    FieldSpec {
+        name: "registryPath",
+        expected: FieldType::String,
+    },
+    FieldSpec {
+        name: "bundledRoot",
+        expected: FieldType::String,
+    },
+    FieldSpec {
+        name: "maxOutputTokens",
+        expected: FieldType::Number,
+    },
+];
+
+const SANDBOX_FIELDS: &[FieldSpec] = &[
+    FieldSpec {
+        name: "enabled",
+        expected: FieldType::Bool,
+    },
+    FieldSpec {
+        name: "namespaceRestrictions",
+        expected: FieldType::Bool,
+    },
+    FieldSpec {
+        name: "networkIsolation",
+        expected: FieldType::Bool,
+    },
+    FieldSpec {
+        name: "filesystemMode",
+        expected: FieldType::String,
+    },
+    FieldSpec {
+        name: "allowedMounts",
+        expected: FieldType::StringArray,
+    },
+];
+
+const OAUTH_FIELDS: &[FieldSpec] = &[
+    FieldSpec {
+        name: "clientId",
+        expected: FieldType::String,
+    },
+    FieldSpec {
+        name: "authorizeUrl",
+        expected: FieldType::String,
+    },
+    FieldSpec {
+        name: "tokenUrl",
+        expected: FieldType::String,
+    },
+    FieldSpec {
+        name: "callbackPort",
+        expected: FieldType::Number,
+    },
+    FieldSpec {
+        name: "manualRedirectUrl",
+        expected: FieldType::String,
+    },
+    FieldSpec {
+        name: "scopes",
+        expected: FieldType::StringArray,
+    },
+];
+
+const DEPRECATED_FIELDS: &[DeprecatedField] = &[
+    DeprecatedField {
+        name: "permissionMode",
+        replacement: "permissions.defaultMode",
+    },
+    DeprecatedField {
+        name: "enabledPlugins",
+        replacement: "plugins.enabled",
+    },
+];
+
+// ---- line-number resolution ----
+
+/// Find the 1-based line number where a JSON key first appears in the raw source.
+fn find_key_line(source: &str, key: &str) -> Option<usize> {
+    // Search for `"key"` followed by optional whitespace and a colon.
+    let needle = format!("\"{key}\"");
+    let mut search_start = 0;
+    while let Some(offset) = source[search_start..].find(&needle) {
+        let absolute = search_start + offset;
+        let after = absolute + needle.len();
+        // Verify the next non-whitespace char is `:` to confirm this is a key, not a value.
+        if source[after..].chars().find(|ch| !ch.is_ascii_whitespace()) == Some(':') {
+            return Some(source[..absolute].chars().filter(|&ch| ch == '\n').count() + 1);
+        }
+        search_start = after;
+    }
+    None
+}
+
+// ---- core validation ----
+
+fn validate_object_keys(
+    object: &BTreeMap<String, JsonValue>,
+    known_fields: &[FieldSpec],
+    prefix: &str,
+    source: &str,
+    path_display: &str,
+) -> ValidationResult {
+    let mut result = ValidationResult {
+        errors: Vec::new(),
+        warnings: Vec::new(),
+    };
+
+    let known_names: Vec<&str> = known_fields.iter().map(|f| f.name).collect();
+
+    for (key, value) in object {
+        let field_path = if prefix.is_empty() {
+            key.clone()
+        } else {
+            format!("{prefix}.{key}")
+        };
+
+        if let Some(spec) = known_fields.iter().find(|f| f.name == key) {
+            // Type check.
+            if !spec.expected.matches(value) {
+                result.errors.push(ConfigDiagnostic {
+                    path: path_display.to_string(),
+                    field: field_path,
+                    line: find_key_line(source, key),
+                    kind: DiagnosticKind::WrongType {
+                        expected: spec.expected.label(),
+                        got: json_type_label(value),
+                    },
+                });
+            }
+        } else if DEPRECATED_FIELDS.iter().any(|d| d.name == key) {
+            // Deprecated key — handled separately, not an unknown-key error.
+        } else {
+            // Unknown key.
+            let suggestion = suggest_field(key, &known_names);
+            result.errors.push(ConfigDiagnostic {
+                path: path_display.to_string(),
+                field: field_path,
+                line: find_key_line(source, key),
+                kind: DiagnosticKind::UnknownKey { suggestion },
+            });
+        }
+    }
+
+    result
+}
+
+fn suggest_field(input: &str, candidates: &[&str]) -> Option<String> {
+    let input_lower = input.to_ascii_lowercase();
+    candidates
+        .iter()
+        .filter_map(|candidate| {
+            let distance = simple_edit_distance(&input_lower, &candidate.to_ascii_lowercase());
+            (distance <= 3).then_some((distance, *candidate))
+        })
+        .min_by_key(|(distance, _)| *distance)
+        .map(|(_, name)| name.to_string())
+}
+
+fn simple_edit_distance(left: &str, right: &str) -> usize {
+    if left.is_empty() {
+        return right.len();
+    }
+    if right.is_empty() {
+        return left.len();
+    }
+    let right_chars: Vec<char> = right.chars().collect();
+    let mut previous: Vec<usize> = (0..=right_chars.len()).collect();
+    let mut current = vec![0; right_chars.len() + 1];
+
+    for (left_index, left_char) in left.chars().enumerate() {
+        current[0] = left_index + 1;
+        for (right_index, right_char) in right_chars.iter().enumerate() {
+            let cost = usize::from(left_char != *right_char);
+            current[right_index + 1] = (previous[right_index + 1] + 1)
+                .min(current[right_index] + 1)
+                .min(previous[right_index] + cost);
+        }
+        previous.clone_from(&current);
+    }
+
+    previous[right_chars.len()]
+}
+
+/// Validate a parsed config file's keys and types against the known schema.
+///
+/// Returns diagnostics (errors and deprecation warnings) without blocking the load.
+pub fn validate_config_file(
+    object: &BTreeMap<String, JsonValue>,
+    source: &str,
+    file_path: &Path,
+) -> ValidationResult {
+    let path_display = file_path.display().to_string();
+    let mut result = validate_object_keys(object, TOP_LEVEL_FIELDS, "", source, &path_display);
+
+    // Check deprecated fields.
+    for deprecated in DEPRECATED_FIELDS {
+        if object.contains_key(deprecated.name) {
+            result.warnings.push(ConfigDiagnostic {
+                path: path_display.clone(),
+                field: deprecated.name.to_string(),
+                line: find_key_line(source, deprecated.name),
+                kind: DiagnosticKind::Deprecated {
+                    replacement: deprecated.replacement,
+                },
+            });
+        }
+    }
+
+    // Validate known nested objects.
+    if let Some(hooks) = object.get("hooks").and_then(JsonValue::as_object) {
+        result.merge(validate_object_keys(
+            hooks,
+            HOOKS_FIELDS,
+            "hooks",
+            source,
+            &path_display,
+        ));
+    }
+    if let Some(permissions) = object.get("permissions").and_then(JsonValue::as_object) {
+        result.merge(validate_object_keys(
+            permissions,
+            PERMISSIONS_FIELDS,
+            "permissions",
+            source,
+            &path_display,
+        ));
+    }
+    if let Some(plugins) = object.get("plugins").and_then(JsonValue::as_object) {
+        result.merge(validate_object_keys(
+            plugins,
+            PLUGINS_FIELDS,
+            "plugins",
+            source,
+            &path_display,
+        ));
+    }
+    if let Some(sandbox) = object.get("sandbox").and_then(JsonValue::as_object) {
+        result.merge(validate_object_keys(
+            sandbox,
+            SANDBOX_FIELDS,
+            "sandbox",
+            source,
+            &path_display,
+        ));
+    }
+    if let Some(oauth) = object.get("oauth").and_then(JsonValue::as_object) {
+        result.merge(validate_object_keys(
+            oauth,
+            OAUTH_FIELDS,
+            "oauth",
+            source,
+            &path_display,
+        ));
+    }
+
+    result
+}
+
+/// Check whether a file path uses an unsupported config format (e.g. TOML).
+pub fn check_unsupported_format(file_path: &Path) -> Result<(), ConfigError> {
+    if let Some(ext) = file_path.extension().and_then(|e| e.to_str()) {
+        if ext.eq_ignore_ascii_case("toml") {
+            return Err(ConfigError::Parse(format!(
+                "{}: TOML config files are not supported. Use JSON (settings.json) instead",
+                file_path.display()
+            )));
+        }
+    }
+    Ok(())
+}
+
+/// Format all diagnostics into a human-readable report.
+#[must_use]
+pub fn format_diagnostics(result: &ValidationResult) -> String {
+    let mut lines = Vec::new();
+    for warning in &result.warnings {
+        lines.push(format!("warning: {warning}"));
+    }
+    for error in &result.errors {
+        lines.push(format!("error: {error}"));
+    }
+    lines.join("\n")
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::path::PathBuf;
+
+    fn test_path() -> PathBuf {
+        PathBuf::from("/test/settings.json")
+    }
+
+    #[test]
+    fn detects_unknown_top_level_key() {
+        // given
+        let source = r#"{"model": "opus", "unknownField": true}"#;
+        let parsed = JsonValue::parse(source).expect("valid json");
+        let object = parsed.as_object().expect("object");
+
+        // when
+        let result = validate_config_file(object, source, &test_path());
+
+        // then
+        assert_eq!(result.errors.len(), 1);
+        assert_eq!(result.errors[0].field, "unknownField");
+        assert!(matches!(
+            result.errors[0].kind,
+            DiagnosticKind::UnknownKey { .. }
+        ));
+    }
+
+    #[test]
+    fn detects_wrong_type_for_model() {
+        // given
+        let source = r#"{"model": 123}"#;
+        let parsed = JsonValue::parse(source).expect("valid json");
+        let object = parsed.as_object().expect("object");
+
+        // when
+        let result = validate_config_file(object, source, &test_path());
+
+        // then
+        assert_eq!(result.errors.len(), 1);
+        assert_eq!(result.errors[0].field, "model");
+        assert!(matches!(
+            result.errors[0].kind,
+            DiagnosticKind::WrongType {
+                expected: "a string",
+                got: "a number"
+            }
+        ));
+    }
+
+    #[test]
+    fn detects_deprecated_permission_mode() {
+        // given
+        let source = r#"{"permissionMode": "plan"}"#;
+        let parsed = JsonValue::parse(source).expect("valid json");
+        let object = parsed.as_object().expect("object");
+
+        // when
+        let result = validate_config_file(object, source, &test_path());
+
+        // then
+        assert_eq!(result.warnings.len(), 1);
+        assert_eq!(result.warnings[0].field, "permissionMode");
+        assert!(matches!(
+            result.warnings[0].kind,
+            DiagnosticKind::Deprecated {
+                replacement: "permissions.defaultMode"
+            }
+        ));
+    }
+
+    #[test]
+    fn detects_deprecated_enabled_plugins() {
+        // given
+        let source = r#"{"enabledPlugins": {"tool-guard@builtin": true}}"#;
+        let parsed = JsonValue::parse(source).expect("valid json");
+        let object = parsed.as_object().expect("object");
+
+        // when
+        let result = validate_config_file(object, source, &test_path());
+
+        // then
+        assert_eq!(result.warnings.len(), 1);
+        assert_eq!(result.warnings[0].field, "enabledPlugins");
+        assert!(matches!(
+            result.warnings[0].kind,
+            DiagnosticKind::Deprecated {
+                replacement: "plugins.enabled"
+            }
+        ));
+    }
+
+    #[test]
+    fn reports_line_number_for_unknown_key() {
+        // given
+        let source = "{\n  \"model\": \"opus\",\n  \"badKey\": true\n}";
+        let parsed = JsonValue::parse(source).expect("valid json");
+        let object = parsed.as_object().expect("object");
+
+        // when
+        let result = validate_config_file(object, source, &test_path());
+
+        // then
+        assert_eq!(result.errors.len(), 1);
+        assert_eq!(result.errors[0].line, Some(3));
+        assert_eq!(result.errors[0].field, "badKey");
+    }
+
+    #[test]
+    fn reports_line_number_for_wrong_type() {
+        // given
+        let source = "{\n  \"model\": 42\n}";
+        let parsed = JsonValue::parse(source).expect("valid json");
+        let object = parsed.as_object().expect("object");
+
+        // when
+        let result = validate_config_file(object, source, &test_path());
+
+        // then
+        assert_eq!(result.errors.len(), 1);
+        assert_eq!(result.errors[0].line, Some(2));
+    }
+
+    #[test]
+    fn validates_nested_hooks_keys() {
+        // given
+        let source = r#"{"hooks": {"PreToolUse": ["cmd"], "BadHook": ["x"]}}"#;
+        let parsed = JsonValue::parse(source).expect("valid json");
+        let object = parsed.as_object().expect("object");
+
+        // when
+        let result = validate_config_file(object, source, &test_path());
+
+        // then
+        assert_eq!(result.errors.len(), 1);
+        assert_eq!(result.errors[0].field, "hooks.BadHook");
+    }
+
+    #[test]
+    fn validates_nested_permissions_keys() {
+        // given
+        let source = r#"{"permissions": {"allow": ["Read"], "denyAll": true}}"#;
+        let parsed = JsonValue::parse(source).expect("valid json");
+        let object = parsed.as_object().expect("object");
+
+        // when
+        let result = validate_config_file(object, source, &test_path());
+
+        // then
+        assert_eq!(result.errors.len(), 1);
+        assert_eq!(result.errors[0].field, "permissions.denyAll");
+    }
+
+    #[test]
+    fn validates_nested_sandbox_keys() {
+        // given
+        let source = r#"{"sandbox": {"enabled": true, "containerMode": "strict"}}"#;
+        let parsed = JsonValue::parse(source).expect("valid json");
+        let object = parsed.as_object().expect("object");
+
+        // when
+        let result = validate_config_file(object, source, &test_path());
+
+        // then
+        assert_eq!(result.errors.len(), 1);
+        assert_eq!(result.errors[0].field, "sandbox.containerMode");
+    }
+
+    #[test]
+    fn validates_nested_plugins_keys() {
+        // given
+        let source = r#"{"plugins": {"installRoot": "/tmp", "autoUpdate": true}}"#;
+        let parsed = JsonValue::parse(source).expect("valid json");
+        let object = parsed.as_object().expect("object");
+
+        // when
+        let result = validate_config_file(object, source, &test_path());
+
+        // then
+        assert_eq!(result.errors.len(), 1);
+        assert_eq!(result.errors[0].field, "plugins.autoUpdate");
+    }
+
+    #[test]
+    fn validates_nested_oauth_keys() {
+        // given
+        let source = r#"{"oauth": {"clientId": "abc", "secret": "hidden"}}"#;
+        let parsed = JsonValue::parse(source).expect("valid json");
+        let object = parsed.as_object().expect("object");
+
+        // when
+        let result = validate_config_file(object, source, &test_path());
+
+        // then
+        assert_eq!(result.errors.len(), 1);
+        assert_eq!(result.errors[0].field, "oauth.secret");
+    }
+
+    #[test]
+    fn valid_config_produces_no_diagnostics() {
+        // given
+        let source = r#"{
+  "model": "opus",
+  "hooks": {"PreToolUse": ["guard"]},
+  "permissions": {"defaultMode": "plan", "allow": ["Read"]},
+  "mcpServers": {},
+  "sandbox": {"enabled": false}
+}"#;
+        let parsed = JsonValue::parse(source).expect("valid json");
+        let object = parsed.as_object().expect("object");
+
+        // when
+        let result = validate_config_file(object, source, &test_path());
+
+        // then
+        assert!(result.is_ok());
+        assert!(result.warnings.is_empty());
+    }
+
+    #[test]
+    fn suggests_close_field_name() {
+        // given
+        let source = r#"{"modle": "opus"}"#;
+        let parsed = JsonValue::parse(source).expect("valid json");
+        let object = parsed.as_object().expect("object");
+
+        // when
+        let result = validate_config_file(object, source, &test_path());
+
+        // then
+        assert_eq!(result.errors.len(), 1);
+        match &result.errors[0].kind {
+            DiagnosticKind::UnknownKey {
+                suggestion: Some(s),
+            } => assert_eq!(s, "model"),
+            other => panic!("expected suggestion, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn format_diagnostics_includes_all_entries() {
+        // given
+        let source = r#"{"permissionMode": "plan", "badKey": 1}"#;
+        let parsed = JsonValue::parse(source).expect("valid json");
+        let object = parsed.as_object().expect("object");
+        let result = validate_config_file(object, source, &test_path());
+
+        // when
+        let output = format_diagnostics(&result);
+
+        // then
+        assert!(output.contains("warning:"));
+        assert!(output.contains("error:"));
+        assert!(output.contains("badKey"));
+        assert!(output.contains("permissionMode"));
+    }
+
+    #[test]
+    fn check_unsupported_format_rejects_toml() {
+        // given
+        let path = PathBuf::from("/home/.claw/settings.toml");
+
+        // when
+        let result = check_unsupported_format(&path);
+
+        // then
+        assert!(result.is_err());
+        let message = result.unwrap_err().to_string();
+        assert!(message.contains("TOML"));
+        assert!(message.contains("settings.toml"));
+    }
+
+    #[test]
+    fn check_unsupported_format_allows_json() {
+        // given
+        let path = PathBuf::from("/home/.claw/settings.json");
+
+        // when / then
+        assert!(check_unsupported_format(&path).is_ok());
+    }
+
+    #[test]
+    fn wrong_type_in_nested_sandbox_field() {
+        // given
+        let source = r#"{"sandbox": {"enabled": "yes"}}"#;
+        let parsed = JsonValue::parse(source).expect("valid json");
+        let object = parsed.as_object().expect("object");
+
+        // when
+        let result = validate_config_file(object, source, &test_path());
+
+        // then
+        assert_eq!(result.errors.len(), 1);
+        assert_eq!(result.errors[0].field, "sandbox.enabled");
+        assert!(matches!(
+            result.errors[0].kind,
+            DiagnosticKind::WrongType {
+                expected: "a boolean",
+                got: "a string"
+            }
+        ));
+    }
+
+    #[test]
+    fn display_format_unknown_key_with_line() {
+        // given
+        let diag = ConfigDiagnostic {
+            path: "/test/settings.json".to_string(),
+            field: "badKey".to_string(),
+            line: Some(5),
+            kind: DiagnosticKind::UnknownKey { suggestion: None },
+        };
+
+        // when
+        let output = diag.to_string();
+
+        // then
+        assert_eq!(
+            output,
+            r#"/test/settings.json: unknown key "badKey" (line 5)"#
+        );
+    }
+
+    #[test]
+    fn display_format_wrong_type_with_line() {
+        // given
+        let diag = ConfigDiagnostic {
+            path: "/test/settings.json".to_string(),
+            field: "model".to_string(),
+            line: Some(2),
+            kind: DiagnosticKind::WrongType {
+                expected: "a string",
+                got: "a number",
+            },
+        };
+
+        // when
+        let output = diag.to_string();
+
+        // then
+        assert_eq!(
+            output,
+            r#"/test/settings.json: field "model" must be a string, got a number (line 2)"#
+        );
+    }
+
+    #[test]
+    fn display_format_deprecated_with_line() {
+        // given
+        let diag = ConfigDiagnostic {
+            path: "/test/settings.json".to_string(),
+            field: "permissionMode".to_string(),
+            line: Some(3),
+            kind: DiagnosticKind::Deprecated {
+                replacement: "permissions.defaultMode",
+            },
+        };
+
+        // when
+        let output = diag.to_string();
+
+        // then
+        assert_eq!(
+            output,
+            r#"/test/settings.json: field "permissionMode" is deprecated (line 3). Use "permissions.defaultMode" instead"#
+        );
+    }
+}
--- a/rust/crates/runtime/src/conversation.rs
+++ b/rust/crates/runtime/src/conversation.rs
@@ -504,6 +504,10 @@ where
        &self.session
    }

+    pub fn session_mut(&mut self) -> &mut Session {
+        &mut self.session
+    }
+
    #[must_use]
    pub fn fork_session(&self, branch_name: Option<String>) -> Session {
        self.session.fork(branch_name)
@@ -890,6 +894,7 @@ mod tests {
                current_date: "2026-03-31".to_string(),
                git_status: None,
                git_diff: None,
+                git_context: None,
                instruction_files: Vec::new(),
            })
            .with_os("linux", "6.8")
--- a/rust/crates/runtime/src/git_context.rs
+++ b/rust/crates/runtime/src/git_context.rs
@@ -0,0 +1,324 @@
+use std::path::Path;
+use std::process::Command;
+
+/// A single git commit entry from the log.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct GitCommitEntry {
+    pub hash: String,
+    pub subject: String,
+}
+
+/// Git-aware context gathered at startup for injection into the system prompt.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct GitContext {
+    pub branch: Option<String>,
+    pub recent_commits: Vec<GitCommitEntry>,
+    pub staged_files: Vec<String>,
+}
+
+const MAX_RECENT_COMMITS: usize = 5;
+
+impl GitContext {
+    /// Detect the git context from the given working directory.
+    ///
+    /// Returns `None` when the directory is not inside a git repository.
+    #[must_use]
+    pub fn detect(cwd: &Path) -> Option<Self> {
+        // Quick gate: is this a git repo at all?
+        let rev_parse = Command::new("git")
+            .args(["rev-parse", "--is-inside-work-tree"])
+            .current_dir(cwd)
+            .output()
+            .ok()?;
+        if !rev_parse.status.success() {
+            return None;
+        }
+
+        Some(Self {
+            branch: read_branch(cwd),
+            recent_commits: read_recent_commits(cwd),
+            staged_files: read_staged_files(cwd),
+        })
+    }
+
+    /// Render a human-readable summary suitable for system-prompt injection.
+    #[must_use]
+    pub fn render(&self) -> String {
+        let mut lines = Vec::new();
+
+        if let Some(branch) = &self.branch {
+            lines.push(format!("Git branch: {branch}"));
+        }
+
+        if !self.recent_commits.is_empty() {
+            lines.push(String::new());
+            lines.push("Recent commits:".to_string());
+            for entry in &self.recent_commits {
+                lines.push(format!("  {} {}", entry.hash, entry.subject));
+            }
+        }
+
+        if !self.staged_files.is_empty() {
+            lines.push(String::new());
+            lines.push("Staged files:".to_string());
+            for file in &self.staged_files {
+                lines.push(format!("  {file}"));
+            }
+        }
+
+        lines.join("\n")
+    }
+}
+
+fn read_branch(cwd: &Path) -> Option<String> {
+    let output = Command::new("git")
+        .args(["rev-parse", "--abbrev-ref", "HEAD"])
+        .current_dir(cwd)
+        .output()
+        .ok()?;
+    if !output.status.success() {
+        return None;
+    }
+    let branch = String::from_utf8(output.stdout).ok()?;
+    let trimmed = branch.trim();
+    if trimmed.is_empty() || trimmed == "HEAD" {
+        None
+    } else {
+        Some(trimmed.to_string())
+    }
+}
+
+fn read_recent_commits(cwd: &Path) -> Vec<GitCommitEntry> {
+    let output = Command::new("git")
+        .args([
+            "--no-optional-locks",
+            "log",
+            "--oneline",
+            "-n",
+            &MAX_RECENT_COMMITS.to_string(),
+            "--no-decorate",
+        ])
+        .current_dir(cwd)
+        .output()
+        .ok();
+    let Some(output) = output else {
+        return Vec::new();
+    };
+    if !output.status.success() {
+        return Vec::new();
+    }
+    let stdout = String::from_utf8(output.stdout).unwrap_or_default();
+    stdout
+        .lines()
+        .filter_map(|line| {
+            let line = line.trim();
+            if line.is_empty() {
+                return None;
+            }
+            let (hash, subject) = line.split_once(' ')?;
+            Some(GitCommitEntry {
+                hash: hash.to_string(),
+                subject: subject.to_string(),
+            })
+        })
+        .collect()
+}
+
+fn read_staged_files(cwd: &Path) -> Vec<String> {
+    let output = Command::new("git")
+        .args(["--no-optional-locks", "diff", "--cached", "--name-only"])
+        .current_dir(cwd)
+        .output()
+        .ok();
+    let Some(output) = output else {
+        return Vec::new();
+    };
+    if !output.status.success() {
+        return Vec::new();
+    }
+    let stdout = String::from_utf8(output.stdout).unwrap_or_default();
+    stdout
+        .lines()
+        .filter(|line| !line.trim().is_empty())
+        .map(|line| line.trim().to_string())
+        .collect()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{GitCommitEntry, GitContext};
+    use std::fs;
+    use std::process::Command;
+    use std::time::{SystemTime, UNIX_EPOCH};
+
+    fn temp_dir(label: &str) -> std::path::PathBuf {
+        let nanos = SystemTime::now()
+            .duration_since(UNIX_EPOCH)
+            .expect("time should be after epoch")
+            .as_nanos();
+        std::env::temp_dir().join(format!("runtime-git-context-{label}-{nanos}"))
+    }
+
+    fn env_lock() -> std::sync::MutexGuard<'static, ()> {
+        crate::test_env_lock()
+    }
+
+    fn ensure_valid_cwd() {
+        if std::env::current_dir().is_err() {
+            std::env::set_current_dir(env!("CARGO_MANIFEST_DIR"))
+                .expect("test cwd should be recoverable");
+        }
+    }
+
+    #[test]
+    fn returns_none_for_non_git_directory() {
+        // given
+        let _guard = env_lock();
+        ensure_valid_cwd();
+        let root = temp_dir("non-git");
+        fs::create_dir_all(&root).expect("create dir");
+
+        // when
+        let context = GitContext::detect(&root);
+
+        // then
+        assert!(context.is_none());
+        fs::remove_dir_all(root).expect("cleanup");
+    }
+
+    #[test]
+    fn detects_branch_name_and_commits() {
+        // given
+        let _guard = env_lock();
+        ensure_valid_cwd();
+        let root = temp_dir("branch-commits");
+        fs::create_dir_all(&root).expect("create dir");
+        git(&root, &["init", "--quiet", "--initial-branch=main"]);
+        git(&root, &["config", "user.email", "tests@example.com"]);
+        git(&root, &["config", "user.name", "Git Context Tests"]);
+        fs::write(root.join("a.txt"), "a\n").expect("write a");
+        git(&root, &["add", "a.txt"]);
+        git(&root, &["commit", "-m", "first commit", "--quiet"]);
+        fs::write(root.join("b.txt"), "b\n").expect("write b");
+        git(&root, &["add", "b.txt"]);
+        git(&root, &["commit", "-m", "second commit", "--quiet"]);
+
+        // when
+        let context = GitContext::detect(&root).expect("should detect git repo");
+
+        // then
+        assert_eq!(context.branch.as_deref(), Some("main"));
+        assert_eq!(context.recent_commits.len(), 2);
+        assert_eq!(context.recent_commits[0].subject, "second commit");
+        assert_eq!(context.recent_commits[1].subject, "first commit");
+        assert!(context.staged_files.is_empty());
+        fs::remove_dir_all(root).expect("cleanup");
+    }
+
+    #[test]
+    fn detects_staged_files() {
+        // given
+        let _guard = env_lock();
+        ensure_valid_cwd();
+        let root = temp_dir("staged");
+        fs::create_dir_all(&root).expect("create dir");
+        git(&root, &["init", "--quiet", "--initial-branch=main"]);
+        git(&root, &["config", "user.email", "tests@example.com"]);
+        git(&root, &["config", "user.name", "Git Context Tests"]);
+        fs::write(root.join("init.txt"), "init\n").expect("write init");
+        git(&root, &["add", "init.txt"]);
+        git(&root, &["commit", "-m", "initial", "--quiet"]);
+        fs::write(root.join("staged.txt"), "staged\n").expect("write staged");
+        git(&root, &["add", "staged.txt"]);
+
+        // when
+        let context = GitContext::detect(&root).expect("should detect git repo");
+
+        // then
+        assert_eq!(context.staged_files, vec!["staged.txt"]);
+        fs::remove_dir_all(root).expect("cleanup");
+    }
+
+    #[test]
+    fn render_formats_all_sections() {
+        // given
+        let context = GitContext {
+            branch: Some("feat/test".to_string()),
+            recent_commits: vec![
+                GitCommitEntry {
+                    hash: "abc1234".to_string(),
+                    subject: "add feature".to_string(),
+                },
+                GitCommitEntry {
+                    hash: "def5678".to_string(),
+                    subject: "fix bug".to_string(),
+                },
+            ],
+            staged_files: vec!["src/main.rs".to_string()],
+        };
+
+        // when
+        let rendered = context.render();
+
+        // then
+        assert!(rendered.contains("Git branch: feat/test"));
+        assert!(rendered.contains("abc1234 add feature"));
+        assert!(rendered.contains("def5678 fix bug"));
+        assert!(rendered.contains("src/main.rs"));
+    }
+
+    #[test]
+    fn render_omits_empty_sections() {
+        // given
+        let context = GitContext {
+            branch: Some("main".to_string()),
+            recent_commits: Vec::new(),
+            staged_files: Vec::new(),
+        };
+
+        // when
+        let rendered = context.render();
+
+        // then
+        assert!(rendered.contains("Git branch: main"));
+        assert!(!rendered.contains("Recent commits:"));
+        assert!(!rendered.contains("Staged files:"));
+    }
+
+    #[test]
+    fn limits_to_five_recent_commits() {
+        // given
+        let _guard = env_lock();
+        ensure_valid_cwd();
+        let root = temp_dir("five-commits");
+        fs::create_dir_all(&root).expect("create dir");
+        git(&root, &["init", "--quiet", "--initial-branch=main"]);
+        git(&root, &["config", "user.email", "tests@example.com"]);
+        git(&root, &["config", "user.name", "Git Context Tests"]);
+        for i in 1..=8 {
+            let name = format!("file{i}.txt");
+            fs::write(root.join(&name), format!("{i}\n")).expect("write file");
+            git(&root, &["add", &name]);
+            git(&root, &["commit", "-m", &format!("commit {i}"), "--quiet"]);
+        }
+
+        // when
+        let context = GitContext::detect(&root).expect("should detect git repo");
+
+        // then
+        assert_eq!(context.recent_commits.len(), 5);
+        assert_eq!(context.recent_commits[0].subject, "commit 8");
+        assert_eq!(context.recent_commits[4].subject, "commit 4");
+        fs::remove_dir_all(root).expect("cleanup");
+    }
+
+    fn git(cwd: &std::path::Path, args: &[&str]) {
+        let status = Command::new("git")
+            .args(args)
+            .current_dir(cwd)
+            .output()
+            .unwrap_or_else(|_| panic!("git {args:?} should run"))
+            .status;
+        assert!(status.success(), "git {args:?} failed");
+    }
+}
--- a/rust/crates/runtime/src/lib.rs
+++ b/rust/crates/runtime/src/lib.rs
@@ -10,8 +10,10 @@ mod bootstrap;
 pub mod branch_lock;
 mod compact;
 mod config;
+pub mod config_validate;
 mod conversation;
 mod file_ops;
+mod git_context;
 pub mod green_contract;
 mod hooks;
 mod json;
@@ -20,6 +22,7 @@ pub mod lsp_client;
 mod mcp;
 mod mcp_client;
 pub mod mcp_lifecycle_hardened;
+pub mod mcp_server;
 mod mcp_stdio;
 pub mod mcp_tool_bridge;
 mod oauth;
@@ -32,9 +35,10 @@ pub mod recovery_recipes;
 mod remote;
 pub mod sandbox;
 mod session;
-#[cfg(test)]
-mod session_control;
+pub mod session_control;
+pub use session_control::SessionStore;
 mod sse;
+pub mod stale_base;
 pub mod stale_branch;
 pub mod summary_compression;
 pub mod task_packet;
@@ -56,10 +60,14 @@ pub use config::{
    ConfigEntry, ConfigError, ConfigLoader, ConfigSource, McpConfigCollection,
    McpManagedProxyServerConfig, McpOAuthConfig, McpRemoteServerConfig, McpSdkServerConfig,
    McpServerConfig, McpStdioServerConfig, McpTransport, McpWebSocketServerConfig, OAuthConfig,
-    ResolvedPermissionMode, RuntimeConfig, RuntimeFeatureConfig, RuntimeHookConfig,
-    RuntimePermissionRuleConfig, RuntimePluginConfig, ScopedMcpServerConfig,
+    ProviderFallbackConfig, ResolvedPermissionMode, RuntimeConfig, RuntimeFeatureConfig,
+    RuntimeHookConfig, RuntimePermissionRuleConfig, RuntimePluginConfig, ScopedMcpServerConfig,
    CLAW_SETTINGS_SCHEMA_NAME,
 };
+pub use config_validate::{
+    check_unsupported_format, format_diagnostics, validate_config_file, ConfigDiagnostic,
+    DiagnosticKind, ValidationResult,
+};
 pub use conversation::{
    auto_compaction_threshold_from_env, ApiClient, ApiRequest, AssistantEvent, AutoCompactionEvent,
    ConversationRuntime, PromptCacheEvent, RuntimeError, StaticToolExecutor, ToolError,
@@ -70,6 +78,7 @@ pub use file_ops::{
    GrepSearchInput, GrepSearchOutput, ReadFileOutput, StructuredPatchHunk, TextFilePayload,
    WriteFileOutput,
 };
+pub use git_context::{GitCommitEntry, GitContext};
 pub use hooks::{
    HookAbortSignal, HookEvent, HookProgressEvent, HookProgressReporter, HookRunResult, HookRunner,
 };
@@ -89,6 +98,7 @@ pub use mcp_lifecycle_hardened::{
    McpDegradedReport, McpErrorSurface, McpFailedServer, McpLifecyclePhase, McpLifecycleState,
    McpLifecycleValidator, McpPhaseResult,
 };
+pub use mcp_server::{McpServer, McpServerSpec, ToolCallHandler, MCP_SERVER_PROTOCOL_VERSION};
 pub use mcp_stdio::{
    spawn_mcp_stdio_process, JsonRpcError, JsonRpcId, JsonRpcRequest, JsonRpcResponse,
    ManagedMcpTool, McpDiscoveryFailure, McpInitializeClientInfo, McpInitializeParams,
@@ -138,9 +148,13 @@ pub use sandbox::{
 };
 pub use session::{
    ContentBlock, ConversationMessage, MessageRole, Session, SessionCompaction, SessionError,
-    SessionFork,
+    SessionFork, SessionPromptEntry,
 };
 pub use sse::{IncrementalSseParser, SseEvent};
+pub use stale_base::{
+    check_base_commit, format_stale_base_warning, read_claw_base_file, resolve_expected_base,
+    BaseCommitSource, BaseCommitState,
+};
 pub use stale_branch::{
    apply_policy, check_freshness, BranchFreshness, StaleBranchAction, StaleBranchEvent,
    StaleBranchPolicy,
--- a/rust/crates/runtime/src/mcp_server.rs
+++ b/rust/crates/runtime/src/mcp_server.rs
@@ -0,0 +1,440 @@
+//! Minimal Model Context Protocol (MCP) server.
+//!
+//! Implements a newline-safe, LSP-framed JSON-RPC server over stdio that
+//! answers `initialize`, `tools/list`, and `tools/call` requests. The framing
+//! matches the client transport implemented in [`crate::mcp_stdio`] so this
+//! server can be driven by either an external MCP client (e.g. Claude
+//! Desktop) or `claw`'s own [`McpServerManager`](crate::McpServerManager).
+//!
+//! The server is intentionally small: it exposes a list of pre-built
+//! [`McpTool`] descriptors and delegates `tools/call` to a caller-supplied
+//! handler. Tool execution itself lives in the `tools` crate; this module is
+//! purely the transport + dispatch loop.
+//!
+//! [`McpTool`]: crate::mcp_stdio::McpTool
+
+use std::io;
+
+use serde_json::{json, Value as JsonValue};
+use tokio::io::{
+    stdin, stdout, AsyncBufReadExt, AsyncReadExt, AsyncWriteExt, BufReader, Stdin, Stdout,
+};
+
+use crate::mcp_stdio::{
+    JsonRpcError, JsonRpcId, JsonRpcRequest, JsonRpcResponse, McpInitializeResult,
+    McpInitializeServerInfo, McpListToolsResult, McpTool, McpToolCallContent, McpToolCallParams,
+    McpToolCallResult,
+};
+
+/// Protocol version the server advertises during `initialize`.
+///
+/// Matches the version used by the built-in client in
+/// [`crate::mcp_stdio`], so the two stay in lockstep.
+pub const MCP_SERVER_PROTOCOL_VERSION: &str = "2025-03-26";
+
+/// Synchronous handler invoked for every `tools/call` request.
+///
+/// Returning `Ok(text)` yields a single `text` content block and
+/// `isError: false`. Returning `Err(message)` yields a `text` block with the
+/// error and `isError: true`, mirroring the error-surfacing convention used
+/// elsewhere in claw.
+pub type ToolCallHandler =
+    Box<dyn Fn(&str, &JsonValue) -> Result<String, String> + Send + Sync + 'static>;
+
+/// Configuration for an [`McpServer`] instance.
+///
+/// Named `McpServerSpec` rather than `McpServerConfig` to avoid colliding
+/// with the existing client-side [`crate::config::McpServerConfig`] that
+/// describes *remote* MCP servers the runtime connects to.
+pub struct McpServerSpec {
+    /// Name advertised in the `serverInfo` field of the `initialize` response.
+    pub server_name: String,
+    /// Version advertised in the `serverInfo` field of the `initialize`
+    /// response.
+    pub server_version: String,
+    /// Tool descriptors returned for `tools/list`.
+    pub tools: Vec<McpTool>,
+    /// Handler invoked for `tools/call`.
+    pub tool_handler: ToolCallHandler,
+}
+
+/// Minimal MCP stdio server.
+///
+/// The server runs a blocking read/dispatch/write loop over the current
+/// process's stdin/stdout, terminating cleanly when the peer closes the
+/// stream.
+pub struct McpServer {
+    spec: McpServerSpec,
+    stdin: BufReader<Stdin>,
+    stdout: Stdout,
+}
+
+impl McpServer {
+    #[must_use]
+    pub fn new(spec: McpServerSpec) -> Self {
+        Self {
+            spec,
+            stdin: BufReader::new(stdin()),
+            stdout: stdout(),
+        }
+    }
+
+    /// Runs the server until the client closes stdin.
+    ///
+    /// Returns `Ok(())` on clean EOF; any other I/O error is propagated so
+    /// callers can log and exit non-zero.
+    pub async fn run(&mut self) -> io::Result<()> {
+        loop {
+            let Some(payload) = read_frame(&mut self.stdin).await? else {
+                return Ok(());
+            };
+
+            // Requests and notifications share a wire format; the absence of
+            // `id` distinguishes notifications, which must never receive a
+            // response.
+            let message: JsonValue = match serde_json::from_slice(&payload) {
+                Ok(value) => value,
+                Err(error) => {
+                    // Parse error with null id per JSON-RPC 2.0 §4.2.
+                    let response = JsonRpcResponse::<JsonValue> {
+                        jsonrpc: "2.0".to_string(),
+                        id: JsonRpcId::Null,
+                        result: None,
+                        error: Some(JsonRpcError {
+                            code: -32700,
+                            message: format!("parse error: {error}"),
+                            data: None,
+                        }),
+                    };
+                    write_response(&mut self.stdout, &response).await?;
+                    continue;
+                }
+            };
+
+            if message.get("id").is_none() {
+                // Notification: dispatch for side effects only (e.g. log),
+                // but send no reply.
+                continue;
+            }
+
+            let request: JsonRpcRequest<JsonValue> = match serde_json::from_value(message) {
+                Ok(request) => request,
+                Err(error) => {
+                    let response = JsonRpcResponse::<JsonValue> {
+                        jsonrpc: "2.0".to_string(),
+                        id: JsonRpcId::Null,
+                        result: None,
+                        error: Some(JsonRpcError {
+                            code: -32600,
+                            message: format!("invalid request: {error}"),
+                            data: None,
+                        }),
+                    };
+                    write_response(&mut self.stdout, &response).await?;
+                    continue;
+                }
+            };
+
+            let response = self.dispatch(request);
+            write_response(&mut self.stdout, &response).await?;
+        }
+    }
+
+    fn dispatch(&self, request: JsonRpcRequest<JsonValue>) -> JsonRpcResponse<JsonValue> {
+        let id = request.id.clone();
+        match request.method.as_str() {
+            "initialize" => self.handle_initialize(id),
+            "tools/list" => self.handle_tools_list(id),
+            "tools/call" => self.handle_tools_call(id, request.params),
+            other => JsonRpcResponse {
+                jsonrpc: "2.0".to_string(),
+                id,
+                result: None,
+                error: Some(JsonRpcError {
+                    code: -32601,
+                    message: format!("method not found: {other}"),
+                    data: None,
+                }),
+            },
+        }
+    }
+
+    fn handle_initialize(&self, id: JsonRpcId) -> JsonRpcResponse<JsonValue> {
+        let result = McpInitializeResult {
+            protocol_version: MCP_SERVER_PROTOCOL_VERSION.to_string(),
+            capabilities: json!({ "tools": {} }),
+            server_info: McpInitializeServerInfo {
+                name: self.spec.server_name.clone(),
+                version: self.spec.server_version.clone(),
+            },
+        };
+        JsonRpcResponse {
+            jsonrpc: "2.0".to_string(),
+            id,
+            result: serde_json::to_value(result).ok(),
+            error: None,
+        }
+    }
+
+    fn handle_tools_list(&self, id: JsonRpcId) -> JsonRpcResponse<JsonValue> {
+        let result = McpListToolsResult {
+            tools: self.spec.tools.clone(),
+            next_cursor: None,
+        };
+        JsonRpcResponse {
+            jsonrpc: "2.0".to_string(),
+            id,
+            result: serde_json::to_value(result).ok(),
+            error: None,
+        }
+    }
+
+    fn handle_tools_call(
+        &self,
+        id: JsonRpcId,
+        params: Option<JsonValue>,
+    ) -> JsonRpcResponse<JsonValue> {
+        let Some(params) = params else {
+            return invalid_params_response(id, "missing params for tools/call");
+        };
+        let call: McpToolCallParams = match serde_json::from_value(params) {
+            Ok(value) => value,
+            Err(error) => {
+                return invalid_params_response(id, &format!("invalid tools/call params: {error}"));
+            }
+        };
+        let arguments = call.arguments.unwrap_or_else(|| json!({}));
+        let tool_result = (self.spec.tool_handler)(&call.name, &arguments);
+        let (text, is_error) = match tool_result {
+            Ok(text) => (text, false),
+            Err(message) => (message, true),
+        };
+        let mut data = std::collections::BTreeMap::new();
+        data.insert("text".to_string(), JsonValue::String(text));
+        let call_result = McpToolCallResult {
+            content: vec![McpToolCallContent {
+                kind: "text".to_string(),
+                data,
+            }],
+            structured_content: None,
+            is_error: Some(is_error),
+            meta: None,
+        };
+        JsonRpcResponse {
+            jsonrpc: "2.0".to_string(),
+            id,
+            result: serde_json::to_value(call_result).ok(),
+            error: None,
+        }
+    }
+}
+
+fn invalid_params_response(id: JsonRpcId, message: &str) -> JsonRpcResponse<JsonValue> {
+    JsonRpcResponse {
+        jsonrpc: "2.0".to_string(),
+        id,
+        result: None,
+        error: Some(JsonRpcError {
+            code: -32602,
+            message: message.to_string(),
+            data: None,
+        }),
+    }
+}
+
+/// Reads a single LSP-framed JSON-RPC payload from `reader`.
+///
+/// Returns `Ok(None)` on clean EOF before any header bytes have been read,
+/// matching how [`crate::mcp_stdio::McpStdioProcess`] treats stream closure.
+async fn read_frame(reader: &mut BufReader<Stdin>) -> io::Result<Option<Vec<u8>>> {
+    let mut content_length: Option<usize> = None;
+    let mut first_header = true;
+    loop {
+        let mut line = String::new();
+        let bytes_read = reader.read_line(&mut line).await?;
+        if bytes_read == 0 {
+            if first_header {
+                return Ok(None);
+            }
+            return Err(io::Error::new(
+                io::ErrorKind::UnexpectedEof,
+                "MCP stdio stream closed while reading headers",
+            ));
+        }
+        first_header = false;
+        if line == "\r\n" || line == "\n" {
+            break;
+        }
+        let header = line.trim_end_matches(['\r', '\n']);
+        if let Some((name, value)) = header.split_once(':') {
+            if name.trim().eq_ignore_ascii_case("Content-Length") {
+                let parsed = value
+                    .trim()
+                    .parse::<usize>()
+                    .map_err(|error| io::Error::new(io::ErrorKind::InvalidData, error))?;
+                content_length = Some(parsed);
+            }
+        }
+    }
+
+    let content_length = content_length.ok_or_else(|| {
+        io::Error::new(io::ErrorKind::InvalidData, "missing Content-Length header")
+    })?;
+    let mut payload = vec![0_u8; content_length];
+    reader.read_exact(&mut payload).await?;
+    Ok(Some(payload))
+}
+
+async fn write_response(
+    stdout: &mut Stdout,
+    response: &JsonRpcResponse<JsonValue>,
+) -> io::Result<()> {
+    let body = serde_json::to_vec(response)
+        .map_err(|error| io::Error::new(io::ErrorKind::InvalidData, error))?;
+    let header = format!("Content-Length: {}\r\n\r\n", body.len());
+    stdout.write_all(header.as_bytes()).await?;
+    stdout.write_all(&body).await?;
+    stdout.flush().await
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn dispatch_initialize_returns_server_info() {
+        let server = McpServer {
+            spec: McpServerSpec {
+                server_name: "test".to_string(),
+                server_version: "9.9.9".to_string(),
+                tools: Vec::new(),
+                tool_handler: Box::new(|_, _| Ok(String::new())),
+            },
+            stdin: BufReader::new(stdin()),
+            stdout: stdout(),
+        };
+        let request = JsonRpcRequest::<JsonValue> {
+            jsonrpc: "2.0".to_string(),
+            id: JsonRpcId::Number(1),
+            method: "initialize".to_string(),
+            params: None,
+        };
+        let response = server.dispatch(request);
+        assert_eq!(response.id, JsonRpcId::Number(1));
+        assert!(response.error.is_none());
+        let result = response.result.expect("initialize result");
+        assert_eq!(result["protocolVersion"], MCP_SERVER_PROTOCOL_VERSION);
+        assert_eq!(result["serverInfo"]["name"], "test");
+        assert_eq!(result["serverInfo"]["version"], "9.9.9");
+    }
+
+    #[test]
+    fn dispatch_tools_list_returns_registered_tools() {
+        let tool = McpTool {
+            name: "echo".to_string(),
+            description: Some("Echo".to_string()),
+            input_schema: Some(json!({"type": "object"})),
+            annotations: None,
+            meta: None,
+        };
+        let server = McpServer {
+            spec: McpServerSpec {
+                server_name: "test".to_string(),
+                server_version: "0.0.0".to_string(),
+                tools: vec![tool.clone()],
+                tool_handler: Box::new(|_, _| Ok(String::new())),
+            },
+            stdin: BufReader::new(stdin()),
+            stdout: stdout(),
+        };
+        let request = JsonRpcRequest::<JsonValue> {
+            jsonrpc: "2.0".to_string(),
+            id: JsonRpcId::Number(2),
+            method: "tools/list".to_string(),
+            params: None,
+        };
+        let response = server.dispatch(request);
+        assert!(response.error.is_none());
+        let result = response.result.expect("tools/list result");
+        assert_eq!(result["tools"][0]["name"], "echo");
+    }
+
+    #[test]
+    fn dispatch_tools_call_wraps_handler_output() {
+        let server = McpServer {
+            spec: McpServerSpec {
+                server_name: "test".to_string(),
+                server_version: "0.0.0".to_string(),
+                tools: Vec::new(),
+                tool_handler: Box::new(|name, args| Ok(format!("called {name} with {args}"))),
+            },
+            stdin: BufReader::new(stdin()),
+            stdout: stdout(),
+        };
+        let request = JsonRpcRequest::<JsonValue> {
+            jsonrpc: "2.0".to_string(),
+            id: JsonRpcId::Number(3),
+            method: "tools/call".to_string(),
+            params: Some(json!({
+                "name": "echo",
+                "arguments": {"text": "hi"}
+            })),
+        };
+        let response = server.dispatch(request);
+        assert!(response.error.is_none());
+        let result = response.result.expect("tools/call result");
+        assert_eq!(result["isError"], false);
+        assert_eq!(result["content"][0]["type"], "text");
+        assert!(result["content"][0]["text"]
+            .as_str()
+            .unwrap()
+            .starts_with("called echo"));
+    }
+
+    #[test]
+    fn dispatch_tools_call_surfaces_handler_error() {
+        let server = McpServer {
+            spec: McpServerSpec {
+                server_name: "test".to_string(),
+                server_version: "0.0.0".to_string(),
+                tools: Vec::new(),
+                tool_handler: Box::new(|_, _| Err("boom".to_string())),
+            },
+            stdin: BufReader::new(stdin()),
+            stdout: stdout(),
+        };
+        let request = JsonRpcRequest::<JsonValue> {
+            jsonrpc: "2.0".to_string(),
+            id: JsonRpcId::Number(4),
+            method: "tools/call".to_string(),
+            params: Some(json!({"name": "broken"})),
+        };
+        let response = server.dispatch(request);
+        let result = response.result.expect("tools/call result");
+        assert_eq!(result["isError"], true);
+        assert_eq!(result["content"][0]["text"], "boom");
+    }
+
+    #[test]
+    fn dispatch_unknown_method_returns_method_not_found() {
+        let server = McpServer {
+            spec: McpServerSpec {
+                server_name: "test".to_string(),
+                server_version: "0.0.0".to_string(),
+                tools: Vec::new(),
+                tool_handler: Box::new(|_, _| Ok(String::new())),
+            },
+            stdin: BufReader::new(stdin()),
+            stdout: stdout(),
+        };
+        let request = JsonRpcRequest::<JsonValue> {
+            jsonrpc: "2.0".to_string(),
+            id: JsonRpcId::Number(5),
+            method: "nonsense".to_string(),
+            params: None,
+        };
+        let response = server.dispatch(request);
+        let error = response.error.expect("error payload");
+        assert_eq!(error.code, -32601);
+    }
+}
--- a/rust/crates/runtime/src/prompt.rs
+++ b/rust/crates/runtime/src/prompt.rs
@@ -4,6 +4,7 @@ use std::path::{Path, PathBuf};
 use std::process::Command;

 use crate::config::{ConfigError, ConfigLoader, RuntimeConfig};
+use crate::git_context::GitContext;

 /// Errors raised while assembling the final system prompt.
 #[derive(Debug)]
@@ -56,6 +57,7 @@ pub struct ProjectContext {
    pub current_date: String,
    pub git_status: Option<String>,
    pub git_diff: Option<String>,
+    pub git_context: Option<GitContext>,
    pub instruction_files: Vec<ContextFile>,
 }

@@ -71,6 +73,7 @@ impl ProjectContext {
            current_date: current_date.into(),
            git_status: None,
            git_diff: None,
+            git_context: None,
            instruction_files,
        })
    }
@@ -82,6 +85,7 @@ impl ProjectContext {
        let mut context = Self::discover(cwd, current_date)?;
        context.git_status = read_git_status(&context.cwd);
        context.git_diff = read_git_diff(&context.cwd);
+        context.git_context = GitContext::detect(&context.cwd);
        Ok(context)
    }
 }
@@ -299,11 +303,27 @@ fn render_project_context(project_context: &ProjectContext) -> String {
        lines.push("Git status snapshot:".to_string());
        lines.push(status.clone());
    }
+    if let Some(ref gc) = project_context.git_context {
+        if !gc.recent_commits.is_empty() {
+            lines.push(String::new());
+            lines.push("Recent commits (last 5):".to_string());
+            for c in &gc.recent_commits {
+                lines.push(format!("  {} {}", c.hash, c.subject));
+            }
+        }
+    }
    if let Some(diff) = &project_context.git_diff {
        lines.push(String::new());
        lines.push("Git diff snapshot:".to_string());
        lines.push(diff.clone());
    }
+    if let Some(git_context) = &project_context.git_context {
+        let rendered = git_context.render();
+        if !rendered.is_empty() {
+            lines.push(String::new());
+            lines.push(rendered);
+        }
+    }
    lines.join("\n")
 }

@@ -639,6 +659,88 @@ mod tests {
        fs::remove_dir_all(root).expect("cleanup temp dir");
    }

+    #[test]
+    fn discover_with_git_includes_recent_commits_and_renders_them() {
+        // given: a git repo with three commits and a current branch
+        let _guard = env_lock();
+        ensure_valid_cwd();
+        let root = temp_dir();
+        fs::create_dir_all(&root).expect("root dir");
+        std::process::Command::new("git")
+            .args(["init", "--quiet", "-b", "main"])
+            .current_dir(&root)
+            .status()
+            .expect("git init should run");
+        std::process::Command::new("git")
+            .args(["config", "user.email", "tests@example.com"])
+            .current_dir(&root)
+            .status()
+            .expect("git config email should run");
+        std::process::Command::new("git")
+            .args(["config", "user.name", "Runtime Prompt Tests"])
+            .current_dir(&root)
+            .status()
+            .expect("git config name should run");
+        for (file, message) in [
+            ("a.txt", "first commit"),
+            ("b.txt", "second commit"),
+            ("c.txt", "third commit"),
+        ] {
+            fs::write(root.join(file), "x\n").expect("write commit file");
+            std::process::Command::new("git")
+                .args(["add", file])
+                .current_dir(&root)
+                .status()
+                .expect("git add should run");
+            std::process::Command::new("git")
+                .args(["commit", "-m", message, "--quiet"])
+                .current_dir(&root)
+                .status()
+                .expect("git commit should run");
+        }
+        fs::write(root.join("d.txt"), "staged\n").expect("write staged file");
+        std::process::Command::new("git")
+            .args(["add", "d.txt"])
+            .current_dir(&root)
+            .status()
+            .expect("git add staged should run");
+
+        // when: discovering project context with git auto-include
+        let context =
+            ProjectContext::discover_with_git(&root, "2026-03-31").expect("context should load");
+        let rendered = SystemPromptBuilder::new()
+            .with_os("linux", "6.8")
+            .with_project_context(context.clone())
+            .render();
+
+        // then: branch, recent commits and staged files are present in context
+        let gc = context
+            .git_context
+            .as_ref()
+            .expect("git context should be present");
+        let commits: String = gc
+            .recent_commits
+            .iter()
+            .map(|c| c.subject.clone())
+            .collect::<Vec<_>>()
+            .join("\n");
+        assert!(commits.contains("first commit"));
+        assert!(commits.contains("second commit"));
+        assert!(commits.contains("third commit"));
+        assert_eq!(gc.recent_commits.len(), 3);
+
+        let status = context.git_status.as_deref().expect("status snapshot");
+        assert!(status.contains("## main"));
+        assert!(status.contains("A  d.txt"));
+
+        assert!(rendered.contains("Recent commits (last 5):"));
+        assert!(rendered.contains("first commit"));
+        assert!(rendered.contains("Git status snapshot:"));
+        assert!(rendered.contains("## main"));
+
+        fs::remove_dir_all(root).expect("cleanup temp dir");
+    }
+
    #[test]
    fn discover_with_git_includes_diff_snapshot_for_tracked_changes() {
        let _guard = env_lock();
--- a/rust/crates/runtime/src/session.rs
+++ b/rust/crates/runtime/src/session.rs
@@ -65,6 +65,13 @@ pub struct SessionFork {
    pub branch_name: Option<String>,
 }

+/// A single user prompt recorded with a timestamp for history tracking.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct SessionPromptEntry {
+    pub timestamp_ms: u64,
+    pub text: String,
+}
+
 #[derive(Debug, Clone, PartialEq, Eq)]
 struct SessionPersistence {
    path: PathBuf,
@@ -88,6 +95,7 @@ pub struct Session {
    pub compaction: Option<SessionCompaction>,
    pub fork: Option<SessionFork>,
    pub workspace_root: Option<PathBuf>,
+    pub prompt_history: Vec<SessionPromptEntry>,
    persistence: Option<SessionPersistence>,
 }

@@ -101,6 +109,7 @@ impl PartialEq for Session {
            && self.compaction == other.compaction
            && self.fork == other.fork
            && self.workspace_root == other.workspace_root
+            && self.prompt_history == other.prompt_history
    }
 }

@@ -151,6 +160,7 @@ impl Session {
            compaction: None,
            fork: None,
            workspace_root: None,
+            prompt_history: Vec::new(),
            persistence: None,
        }
    }
@@ -252,6 +262,7 @@ impl Session {
                branch_name: normalize_optional_string(branch_name),
            }),
            workspace_root: self.workspace_root.clone(),
+            prompt_history: self.prompt_history.clone(),
            persistence: None,
        }
    }
@@ -295,6 +306,17 @@ impl Session {
                JsonValue::String(workspace_root_to_string(workspace_root)?),
            );
        }
+        if !self.prompt_history.is_empty() {
+            object.insert(
+                "prompt_history".to_string(),
+                JsonValue::Array(
+                    self.prompt_history
+                        .iter()
+                        .map(SessionPromptEntry::to_jsonl_record)
+                        .collect(),
+                ),
+            );
+        }
        Ok(JsonValue::Object(object))
    }

@@ -339,6 +361,16 @@ impl Session {
            .get("workspace_root")
            .and_then(JsonValue::as_str)
            .map(PathBuf::from);
+        let prompt_history = object
+            .get("prompt_history")
+            .and_then(JsonValue::as_array)
+            .map(|entries| {
+                entries
+                    .iter()
+                    .filter_map(SessionPromptEntry::from_json_opt)
+                    .collect()
+            })
+            .unwrap_or_default();
        Ok(Self {
            version,
            session_id,
@@ -348,6 +380,7 @@ impl Session {
            compaction,
            fork,
            workspace_root,
+            prompt_history,
            persistence: None,
        })
    }
@@ -361,6 +394,7 @@ impl Session {
        let mut compaction = None;
        let mut fork = None;
        let mut workspace_root = None;
+        let mut prompt_history = Vec::new();

        for (line_number, raw_line) in contents.lines().enumerate() {
            let line = raw_line.trim();
@@ -414,6 +448,13 @@ impl Session {
                        object.clone(),
                    ))?);
                }
+                "prompt_history" => {
+                    if let Some(entry) =
+                        SessionPromptEntry::from_json_opt(&JsonValue::Object(object.clone()))
+                    {
+                        prompt_history.push(entry);
+                    }
+                }
                other => {
                    return Err(SessionError::Format(format!(
                        "unsupported JSONL record type at line {}: {other}",
@@ -433,15 +474,36 @@ impl Session {
            compaction,
            fork,
            workspace_root,
+            prompt_history,
            persistence: None,
        })
    }

+    /// Record a user prompt with the current wall-clock timestamp.
+    ///
+    /// The entry is appended to the in-memory history and, when a persistence
+    /// path is configured, incrementally written to the JSONL session file.
+    pub fn push_prompt_entry(&mut self, text: impl Into<String>) -> Result<(), SessionError> {
+        let timestamp_ms = current_time_millis();
+        let entry = SessionPromptEntry {
+            timestamp_ms,
+            text: text.into(),
+        };
+        self.prompt_history.push(entry);
+        let entry_ref = self.prompt_history.last().expect("entry was just pushed");
+        self.append_persisted_prompt_entry(entry_ref)
+    }
+
    fn render_jsonl_snapshot(&self) -> Result<String, SessionError> {
        let mut lines = vec![self.meta_record()?.render()];
        if let Some(compaction) = &self.compaction {
            lines.push(compaction.to_jsonl_record()?.render());
        }
+        lines.extend(
+            self.prompt_history
+                .iter()
+                .map(|entry| entry.to_jsonl_record().render()),
+        );
        lines.extend(
            self.messages
                .iter()
@@ -468,6 +530,25 @@ impl Session {
        Ok(())
    }

+    fn append_persisted_prompt_entry(
+        &self,
+        entry: &SessionPromptEntry,
+    ) -> Result<(), SessionError> {
+        let Some(path) = self.persistence_path() else {
+            return Ok(());
+        };
+
+        let needs_bootstrap = !path.exists() || fs::metadata(path)?.len() == 0;
+        if needs_bootstrap {
+            self.save_to_path(path)?;
+            return Ok(());
+        }
+
+        let mut file = OpenOptions::new().append(true).open(path)?;
+        writeln!(file, "{}", entry.to_jsonl_record().render())?;
+        Ok(())
+    }
+
    fn meta_record(&self) -> Result<JsonValue, SessionError> {
        let mut object = BTreeMap::new();
        object.insert(
@@ -784,6 +865,33 @@ impl SessionFork {
    }
 }

+impl SessionPromptEntry {
+    #[must_use]
+    pub fn to_jsonl_record(&self) -> JsonValue {
+        let mut object = BTreeMap::new();
+        object.insert(
+            "type".to_string(),
+            JsonValue::String("prompt_history".to_string()),
+        );
+        object.insert(
+            "timestamp_ms".to_string(),
+            JsonValue::Number(i64::try_from(self.timestamp_ms).unwrap_or(i64::MAX)),
+        );
+        object.insert("text".to_string(), JsonValue::String(self.text.clone()));
+        JsonValue::Object(object)
+    }
+
+    fn from_json_opt(value: &JsonValue) -> Option<Self> {
+        let object = value.as_object()?;
+        let timestamp_ms = object
+            .get("timestamp_ms")
+            .and_then(JsonValue::as_i64)
+            .and_then(|value| u64::try_from(value).ok())?;
+        let text = object.get("text").and_then(JsonValue::as_str)?.to_string();
+        Some(Self { timestamp_ms, text })
+    }
+}
+
 fn message_record(message: &ConversationMessage) -> JsonValue {
    let mut object = BTreeMap::new();
    object.insert("type".to_string(), JsonValue::String("message".to_string()));
@@ -1326,3 +1434,63 @@ mod tests {
            .collect()
    }
 }
+
+/// Per-worktree session isolation: returns a session directory namespaced
+/// by the workspace fingerprint of the given working directory.
+/// This prevents parallel `opencode serve` instances from colliding.
+/// Called by external consumers (e.g. clawhip) to enumerate sessions for a CWD.
+#[allow(dead_code)]
+pub fn workspace_sessions_dir(cwd: &std::path::Path) -> Result<std::path::PathBuf, SessionError> {
+    let store = crate::session_control::SessionStore::from_cwd(cwd).map_err(|e| {
+        SessionError::Io(std::io::Error::new(
+            std::io::ErrorKind::Other,
+            e.to_string(),
+        ))
+    })?;
+    Ok(store.sessions_dir().to_path_buf())
+}
+
+#[cfg(test)]
+mod workspace_sessions_dir_tests {
+    use super::*;
+    use std::fs;
+
+    #[test]
+    fn workspace_sessions_dir_returns_fingerprinted_path_for_valid_cwd() {
+        let tmp = std::env::temp_dir().join("claw-session-dir-test");
+        fs::create_dir_all(&tmp).expect("create temp dir");
+
+        let result = workspace_sessions_dir(&tmp);
+        assert!(
+            result.is_ok(),
+            "workspace_sessions_dir should succeed for a valid CWD, got: {:?}",
+            result
+        );
+        let dir = result.unwrap();
+        // The returned path should be non-empty and end with a hash component
+        assert!(!dir.as_os_str().is_empty());
+        // Two calls with the same CWD should produce identical paths (deterministic)
+        let result2 = workspace_sessions_dir(&tmp).unwrap();
+        assert_eq!(dir, result2, "workspace_sessions_dir must be deterministic");
+
+        fs::remove_dir_all(&tmp).ok();
+    }
+
+    #[test]
+    fn workspace_sessions_dir_differs_for_different_cwds() {
+        let tmp_a = std::env::temp_dir().join("claw-session-dir-a");
+        let tmp_b = std::env::temp_dir().join("claw-session-dir-b");
+        fs::create_dir_all(&tmp_a).expect("create dir a");
+        fs::create_dir_all(&tmp_b).expect("create dir b");
+
+        let dir_a = workspace_sessions_dir(&tmp_a).expect("dir a");
+        let dir_b = workspace_sessions_dir(&tmp_b).expect("dir b");
+        assert_ne!(
+            dir_a, dir_b,
+            "different CWDs must produce different session dirs"
+        );
+
+        fs::remove_dir_all(&tmp_a).ok();
+        fs::remove_dir_all(&tmp_b).ok();
+    }
+}
--- a/rust/crates/runtime/src/session_control.rs
+++ b/rust/crates/runtime/src/session_control.rs
@@ -7,6 +7,252 @@ use std::time::UNIX_EPOCH;

 use crate::session::{Session, SessionError};

+/// Per-worktree session store that namespaces on-disk session files by
+/// workspace fingerprint so that parallel `opencode serve` instances never
+/// collide.
+///
+/// Create via [`SessionStore::from_cwd`] (derives the store path from the
+/// server's working directory) or [`SessionStore::from_data_dir`] (honours an
+/// explicit `--data-dir` flag).  Both constructors produce a directory layout
+/// of `<data_dir>/sessions/<workspace_hash>/` where `<workspace_hash>` is a
+/// stable hex digest of the canonical workspace root.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct SessionStore {
+    /// Resolved root of the session namespace, e.g.
+    /// `/home/user/project/.claw/sessions/a1b2c3d4e5f60718/`.
+    sessions_root: PathBuf,
+    /// The canonical workspace path that was fingerprinted.
+    workspace_root: PathBuf,
+}
+
+impl SessionStore {
+    /// Build a store from the server's current working directory.
+    ///
+    /// The on-disk layout becomes `<cwd>/.claw/sessions/<workspace_hash>/`.
+    pub fn from_cwd(cwd: impl AsRef<Path>) -> Result<Self, SessionControlError> {
+        let cwd = cwd.as_ref();
+        let sessions_root = cwd
+            .join(".claw")
+            .join("sessions")
+            .join(workspace_fingerprint(cwd));
+        fs::create_dir_all(&sessions_root)?;
+        Ok(Self {
+            sessions_root,
+            workspace_root: cwd.to_path_buf(),
+        })
+    }
+
+    /// Build a store from an explicit `--data-dir` flag.
+    ///
+    /// The on-disk layout becomes `<data_dir>/sessions/<workspace_hash>/`
+    /// where `<workspace_hash>` is derived from `workspace_root`.
+    pub fn from_data_dir(
+        data_dir: impl AsRef<Path>,
+        workspace_root: impl AsRef<Path>,
+    ) -> Result<Self, SessionControlError> {
+        let workspace_root = workspace_root.as_ref();
+        let sessions_root = data_dir
+            .as_ref()
+            .join("sessions")
+            .join(workspace_fingerprint(workspace_root));
+        fs::create_dir_all(&sessions_root)?;
+        Ok(Self {
+            sessions_root,
+            workspace_root: workspace_root.to_path_buf(),
+        })
+    }
+
+    /// The fully resolved sessions directory for this namespace.
+    #[must_use]
+    pub fn sessions_dir(&self) -> &Path {
+        &self.sessions_root
+    }
+
+    /// The workspace root this store is bound to.
+    #[must_use]
+    pub fn workspace_root(&self) -> &Path {
+        &self.workspace_root
+    }
+
+    pub fn create_handle(&self, session_id: &str) -> SessionHandle {
+        let id = session_id.to_string();
+        let path = self
+            .sessions_root
+            .join(format!("{id}.{PRIMARY_SESSION_EXTENSION}"));
+        SessionHandle { id, path }
+    }
+
+    pub fn resolve_reference(&self, reference: &str) -> Result<SessionHandle, SessionControlError> {
+        if is_session_reference_alias(reference) {
+            let latest = self.latest_session()?;
+            return Ok(SessionHandle {
+                id: latest.id,
+                path: latest.path,
+            });
+        }
+
+        let direct = PathBuf::from(reference);
+        let candidate = if direct.is_absolute() {
+            direct.clone()
+        } else {
+            self.workspace_root.join(&direct)
+        };
+        let looks_like_path = direct.extension().is_some() || direct.components().count() > 1;
+        let path = if candidate.exists() {
+            candidate
+        } else if looks_like_path {
+            return Err(SessionControlError::Format(
+                format_missing_session_reference(reference),
+            ));
+        } else {
+            self.resolve_managed_path(reference)?
+        };
+
+        Ok(SessionHandle {
+            id: session_id_from_path(&path).unwrap_or_else(|| reference.to_string()),
+            path,
+        })
+    }
+
+    pub fn resolve_managed_path(&self, session_id: &str) -> Result<PathBuf, SessionControlError> {
+        for extension in [PRIMARY_SESSION_EXTENSION, LEGACY_SESSION_EXTENSION] {
+            let path = self.sessions_root.join(format!("{session_id}.{extension}"));
+            if path.exists() {
+                return Ok(path);
+            }
+        }
+        Err(SessionControlError::Format(
+            format_missing_session_reference(session_id),
+        ))
+    }
+
+    pub fn list_sessions(&self) -> Result<Vec<ManagedSessionSummary>, SessionControlError> {
+        let mut sessions = Vec::new();
+        let read_result = fs::read_dir(&self.sessions_root);
+        let entries = match read_result {
+            Ok(entries) => entries,
+            Err(err) if err.kind() == std::io::ErrorKind::NotFound => return Ok(sessions),
+            Err(err) => return Err(err.into()),
+        };
+        for entry in entries {
+            let entry = entry?;
+            let path = entry.path();
+            if !is_managed_session_file(&path) {
+                continue;
+            }
+            let metadata = entry.metadata()?;
+            let modified_epoch_millis = metadata
+                .modified()
+                .ok()
+                .and_then(|time| time.duration_since(UNIX_EPOCH).ok())
+                .map(|duration| duration.as_millis())
+                .unwrap_or_default();
+            let (id, message_count, parent_session_id, branch_name) =
+                match Session::load_from_path(&path) {
+                    Ok(session) => {
+                        let parent_session_id = session
+                            .fork
+                            .as_ref()
+                            .map(|fork| fork.parent_session_id.clone());
+                        let branch_name = session
+                            .fork
+                            .as_ref()
+                            .and_then(|fork| fork.branch_name.clone());
+                        (
+                            session.session_id,
+                            session.messages.len(),
+                            parent_session_id,
+                            branch_name,
+                        )
+                    }
+                    Err(_) => (
+                        path.file_stem()
+                            .and_then(|value| value.to_str())
+                            .unwrap_or("unknown")
+                            .to_string(),
+                        0,
+                        None,
+                        None,
+                    ),
+                };
+            sessions.push(ManagedSessionSummary {
+                id,
+                path,
+                modified_epoch_millis,
+                message_count,
+                parent_session_id,
+                branch_name,
+            });
+        }
+        sessions.sort_by(|left, right| {
+            right
+                .modified_epoch_millis
+                .cmp(&left.modified_epoch_millis)
+                .then_with(|| right.id.cmp(&left.id))
+        });
+        Ok(sessions)
+    }
+
+    pub fn latest_session(&self) -> Result<ManagedSessionSummary, SessionControlError> {
+        self.list_sessions()?
+            .into_iter()
+            .next()
+            .ok_or_else(|| SessionControlError::Format(format_no_managed_sessions()))
+    }
+
+    pub fn load_session(
+        &self,
+        reference: &str,
+    ) -> Result<LoadedManagedSession, SessionControlError> {
+        let handle = self.resolve_reference(reference)?;
+        let session = Session::load_from_path(&handle.path)?;
+        Ok(LoadedManagedSession {
+            handle: SessionHandle {
+                id: session.session_id.clone(),
+                path: handle.path,
+            },
+            session,
+        })
+    }
+
+    pub fn fork_session(
+        &self,
+        session: &Session,
+        branch_name: Option<String>,
+    ) -> Result<ForkedManagedSession, SessionControlError> {
+        let parent_session_id = session.session_id.clone();
+        let forked = session.fork(branch_name);
+        let handle = self.create_handle(&forked.session_id);
+        let branch_name = forked
+            .fork
+            .as_ref()
+            .and_then(|fork| fork.branch_name.clone());
+        let forked = forked.with_persistence_path(handle.path.clone());
+        forked.save_to_path(&handle.path)?;
+        Ok(ForkedManagedSession {
+            parent_session_id,
+            handle,
+            session: forked,
+            branch_name,
+        })
+    }
+}
+
+/// Stable hex fingerprint of a workspace path.
+///
+/// Uses FNV-1a (64-bit) to produce a 16-char hex string that partitions the
+/// on-disk session directory per workspace root.
+#[must_use]
+pub fn workspace_fingerprint(workspace_root: &Path) -> String {
+    let input = workspace_root.to_string_lossy();
+    let mut hash = 0xcbf2_9ce4_8422_2325_u64;
+    for byte in input.as_bytes() {
+        hash ^= u64::from(*byte);
+        hash = hash.wrapping_mul(0x0100_0000_01b3);
+    }
+    format!("{hash:016x}")
+}
+
 pub const PRIMARY_SESSION_EXTENSION: &str = "jsonl";
 pub const LEGACY_SESSION_EXTENSION: &str = "json";
 pub const LATEST_SESSION_REFERENCE: &str = "latest";
@@ -333,7 +579,7 @@ mod tests {
    use super::{
        create_managed_session_handle_for, fork_managed_session_for, is_session_reference_alias,
        list_managed_sessions_for, load_managed_session_for, resolve_session_reference_for,
-        ManagedSessionSummary, LATEST_SESSION_REFERENCE,
+        workspace_fingerprint, ManagedSessionSummary, SessionStore, LATEST_SESSION_REFERENCE,
    };
    use crate::session::Session;
    use std::fs;
@@ -456,4 +702,172 @@ mod tests {
        );
        fs::remove_dir_all(root).expect("temp dir should clean up");
    }
+
+    // ------------------------------------------------------------------
+    // Per-worktree session isolation (SessionStore) tests
+    // ------------------------------------------------------------------
+
+    fn persist_session_via_store(store: &SessionStore, text: &str) -> Session {
+        let mut session = Session::new();
+        session
+            .push_user_text(text)
+            .expect("session message should save");
+        let handle = store.create_handle(&session.session_id);
+        let session = session.with_persistence_path(handle.path.clone());
+        session
+            .save_to_path(&handle.path)
+            .expect("session should persist");
+        session
+    }
+
+    #[test]
+    fn workspace_fingerprint_is_deterministic_and_differs_per_path() {
+        // given
+        let path_a = Path::new("/tmp/worktree-alpha");
+        let path_b = Path::new("/tmp/worktree-beta");
+
+        // when
+        let fp_a1 = workspace_fingerprint(path_a);
+        let fp_a2 = workspace_fingerprint(path_a);
+        let fp_b = workspace_fingerprint(path_b);
+
+        // then
+        assert_eq!(fp_a1, fp_a2, "same path must produce the same fingerprint");
+        assert_ne!(
+            fp_a1, fp_b,
+            "different paths must produce different fingerprints"
+        );
+        assert_eq!(fp_a1.len(), 16, "fingerprint must be a 16-char hex string");
+    }
+
+    #[test]
+    fn session_store_from_cwd_isolates_sessions_by_workspace() {
+        // given
+        let base = temp_dir();
+        let workspace_a = base.join("repo-alpha");
+        let workspace_b = base.join("repo-beta");
+        fs::create_dir_all(&workspace_a).expect("workspace a should exist");
+        fs::create_dir_all(&workspace_b).expect("workspace b should exist");
+
+        let store_a = SessionStore::from_cwd(&workspace_a).expect("store a should build");
+        let store_b = SessionStore::from_cwd(&workspace_b).expect("store b should build");
+
+        // when
+        let session_a = persist_session_via_store(&store_a, "alpha work");
+        let _session_b = persist_session_via_store(&store_b, "beta work");
+
+        // then — each store only sees its own sessions
+        let list_a = store_a.list_sessions().expect("list a");
+        let list_b = store_b.list_sessions().expect("list b");
+        assert_eq!(list_a.len(), 1, "store a should see exactly one session");
+        assert_eq!(list_b.len(), 1, "store b should see exactly one session");
+        assert_eq!(list_a[0].id, session_a.session_id);
+        assert_ne!(
+            store_a.sessions_dir(),
+            store_b.sessions_dir(),
+            "session directories must differ across workspaces"
+        );
+        fs::remove_dir_all(base).expect("temp dir should clean up");
+    }
+
+    #[test]
+    fn session_store_from_data_dir_namespaces_by_workspace() {
+        // given
+        let base = temp_dir();
+        let data_dir = base.join("global-data");
+        let workspace_a = PathBuf::from("/tmp/project-one");
+        let workspace_b = PathBuf::from("/tmp/project-two");
+        fs::create_dir_all(&data_dir).expect("data dir should exist");
+
+        let store_a =
+            SessionStore::from_data_dir(&data_dir, &workspace_a).expect("store a should build");
+        let store_b =
+            SessionStore::from_data_dir(&data_dir, &workspace_b).expect("store b should build");
+
+        // when
+        persist_session_via_store(&store_a, "work in project-one");
+        persist_session_via_store(&store_b, "work in project-two");
+
+        // then
+        assert_ne!(
+            store_a.sessions_dir(),
+            store_b.sessions_dir(),
+            "data-dir stores must namespace by workspace"
+        );
+        assert_eq!(store_a.list_sessions().expect("list a").len(), 1);
+        assert_eq!(store_b.list_sessions().expect("list b").len(), 1);
+        assert_eq!(store_a.workspace_root(), workspace_a.as_path());
+        assert_eq!(store_b.workspace_root(), workspace_b.as_path());
+        fs::remove_dir_all(base).expect("temp dir should clean up");
+    }
+
+    #[test]
+    fn session_store_create_and_load_round_trip() {
+        // given
+        let base = temp_dir();
+        fs::create_dir_all(&base).expect("base dir should exist");
+        let store = SessionStore::from_cwd(&base).expect("store should build");
+        let session = persist_session_via_store(&store, "round-trip message");
+
+        // when
+        let loaded = store
+            .load_session(&session.session_id)
+            .expect("session should load via store");
+
+        // then
+        assert_eq!(loaded.handle.id, session.session_id);
+        assert_eq!(loaded.session.messages.len(), 1);
+        fs::remove_dir_all(base).expect("temp dir should clean up");
+    }
+
+    #[test]
+    fn session_store_latest_and_resolve_reference() {
+        // given
+        let base = temp_dir();
+        fs::create_dir_all(&base).expect("base dir should exist");
+        let store = SessionStore::from_cwd(&base).expect("store should build");
+        let _older = persist_session_via_store(&store, "older");
+        wait_for_next_millisecond();
+        let newer = persist_session_via_store(&store, "newer");
+
+        // when
+        let latest = store.latest_session().expect("latest should resolve");
+        let handle = store
+            .resolve_reference("latest")
+            .expect("latest alias should resolve");
+
+        // then
+        assert_eq!(latest.id, newer.session_id);
+        assert_eq!(handle.id, newer.session_id);
+        fs::remove_dir_all(base).expect("temp dir should clean up");
+    }
+
+    #[test]
+    fn session_store_fork_stays_in_same_namespace() {
+        // given
+        let base = temp_dir();
+        fs::create_dir_all(&base).expect("base dir should exist");
+        let store = SessionStore::from_cwd(&base).expect("store should build");
+        let source = persist_session_via_store(&store, "parent work");
+
+        // when
+        let forked = store
+            .fork_session(&source, Some("bugfix".to_string()))
+            .expect("fork should succeed");
+        let sessions = store.list_sessions().expect("list sessions");
+
+        // then
+        assert_eq!(
+            sessions.len(),
+            2,
+            "forked session must land in the same namespace"
+        );
+        assert_eq!(forked.parent_session_id, source.session_id);
+        assert_eq!(forked.branch_name.as_deref(), Some("bugfix"));
+        assert!(
+            forked.handle.path.starts_with(store.sessions_dir()),
+            "forked session path must be inside the store namespace"
+        );
+        fs::remove_dir_all(base).expect("temp dir should clean up");
+    }
 }
--- a/rust/crates/runtime/src/stale_base.rs
+++ b/rust/crates/runtime/src/stale_base.rs
@@ -0,0 +1,429 @@
+#![allow(clippy::must_use_candidate)]
+use std::path::Path;
+use std::process::Command;
+
+/// Outcome of comparing the worktree HEAD against the expected base commit.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum BaseCommitState {
+    /// HEAD matches the expected base commit.
+    Matches,
+    /// HEAD has diverged from the expected base.
+    Diverged { expected: String, actual: String },
+    /// No expected base was supplied (neither flag nor file).
+    NoExpectedBase,
+    /// The working directory is not inside a git repository.
+    NotAGitRepo,
+}
+
+/// Where the expected base commit originated from.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum BaseCommitSource {
+    Flag(String),
+    File(String),
+}
+
+/// Read the `.claw-base` file from the given directory and return the trimmed
+/// commit hash, or `None` when the file is absent or empty.
+pub fn read_claw_base_file(cwd: &Path) -> Option<String> {
+    let path = cwd.join(".claw-base");
+    let content = std::fs::read_to_string(path).ok()?;
+    let trimmed = content.trim();
+    if trimmed.is_empty() {
+        None
+    } else {
+        Some(trimmed.to_string())
+    }
+}
+
+/// Resolve the expected base commit: prefer the `--base-commit` flag value,
+/// fall back to reading `.claw-base` from `cwd`.
+pub fn resolve_expected_base(flag_value: Option<&str>, cwd: &Path) -> Option<BaseCommitSource> {
+    if let Some(value) = flag_value {
+        let trimmed = value.trim();
+        if !trimmed.is_empty() {
+            return Some(BaseCommitSource::Flag(trimmed.to_string()));
+        }
+    }
+    read_claw_base_file(cwd).map(BaseCommitSource::File)
+}
+
+/// Verify that the worktree HEAD matches `expected_base`.
+///
+/// Returns [`BaseCommitState::NoExpectedBase`] when no expected commit is
+/// provided (the check is effectively a no-op in that case).
+pub fn check_base_commit(cwd: &Path, expected_base: Option<&BaseCommitSource>) -> BaseCommitState {
+    let Some(source) = expected_base else {
+        return BaseCommitState::NoExpectedBase;
+    };
+    let expected_raw = match source {
+        BaseCommitSource::Flag(value) | BaseCommitSource::File(value) => value.as_str(),
+    };
+
+    let Some(head_sha) = resolve_head_sha(cwd) else {
+        return BaseCommitState::NotAGitRepo;
+    };
+
+    let Some(expected_sha) = resolve_rev(cwd, expected_raw) else {
+        // If the expected ref cannot be resolved, compare raw strings as a
+        // best-effort fallback (e.g. partial SHA provided by the caller).
+        return if head_sha.starts_with(expected_raw) || expected_raw.starts_with(&head_sha) {
+            BaseCommitState::Matches
+        } else {
+            BaseCommitState::Diverged {
+                expected: expected_raw.to_string(),
+                actual: head_sha,
+            }
+        };
+    };
+
+    if head_sha == expected_sha {
+        BaseCommitState::Matches
+    } else {
+        BaseCommitState::Diverged {
+            expected: expected_sha,
+            actual: head_sha,
+        }
+    }
+}
+
+/// Format a human-readable warning when the base commit has diverged.
+///
+/// Returns `None` for non-warning states (`Matches`, `NoExpectedBase`).
+pub fn format_stale_base_warning(state: &BaseCommitState) -> Option<String> {
+    match state {
+        BaseCommitState::Diverged { expected, actual } => Some(format!(
+            "warning: worktree HEAD ({actual}) does not match expected base commit ({expected}). \
+             Session may run against a stale codebase."
+        )),
+        BaseCommitState::NotAGitRepo => {
+            Some("warning: stale-base check skipped — not inside a git repository.".to_string())
+        }
+        BaseCommitState::Matches | BaseCommitState::NoExpectedBase => None,
+    }
+}
+
+fn resolve_head_sha(cwd: &Path) -> Option<String> {
+    resolve_rev(cwd, "HEAD")
+}
+
+fn resolve_rev(cwd: &Path, rev: &str) -> Option<String> {
+    let output = Command::new("git")
+        .args(["rev-parse", rev])
+        .current_dir(cwd)
+        .output()
+        .ok()?;
+    if !output.status.success() {
+        return None;
+    }
+    let sha = String::from_utf8(output.stdout).ok()?;
+    let trimmed = sha.trim();
+    if trimmed.is_empty() {
+        None
+    } else {
+        Some(trimmed.to_string())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::fs;
+    use std::process::Command;
+    use std::time::{SystemTime, UNIX_EPOCH};
+
+    fn temp_dir() -> std::path::PathBuf {
+        let nanos = SystemTime::now()
+            .duration_since(UNIX_EPOCH)
+            .expect("time should be after epoch")
+            .as_nanos();
+        std::env::temp_dir().join(format!("runtime-stale-base-{nanos}"))
+    }
+
+    fn init_repo(path: &std::path::Path) {
+        fs::create_dir_all(path).expect("create repo dir");
+        run(path, &["init", "--quiet", "-b", "main"]);
+        run(path, &["config", "user.email", "tests@example.com"]);
+        run(path, &["config", "user.name", "Stale Base Tests"]);
+        fs::write(path.join("init.txt"), "initial\n").expect("write init file");
+        run(path, &["add", "."]);
+        run(path, &["commit", "-m", "initial commit", "--quiet"]);
+    }
+
+    fn run(cwd: &std::path::Path, args: &[&str]) {
+        let status = Command::new("git")
+            .args(args)
+            .current_dir(cwd)
+            .status()
+            .unwrap_or_else(|e| panic!("git {} failed to execute: {e}", args.join(" ")));
+        assert!(
+            status.success(),
+            "git {} exited with {status}",
+            args.join(" ")
+        );
+    }
+
+    fn commit_file(repo: &std::path::Path, name: &str, msg: &str) {
+        fs::write(repo.join(name), format!("{msg}\n")).expect("write file");
+        run(repo, &["add", name]);
+        run(repo, &["commit", "-m", msg, "--quiet"]);
+    }
+
+    fn head_sha(repo: &std::path::Path) -> String {
+        let output = Command::new("git")
+            .args(["rev-parse", "HEAD"])
+            .current_dir(repo)
+            .output()
+            .expect("git rev-parse HEAD");
+        String::from_utf8(output.stdout)
+            .expect("valid utf8")
+            .trim()
+            .to_string()
+    }
+
+    #[test]
+    fn matches_when_head_equals_expected_base() {
+        // given
+        let root = temp_dir();
+        init_repo(&root);
+        let sha = head_sha(&root);
+        let source = BaseCommitSource::Flag(sha);
+
+        // when
+        let state = check_base_commit(&root, Some(&source));
+
+        // then
+        assert_eq!(state, BaseCommitState::Matches);
+        fs::remove_dir_all(&root).expect("cleanup");
+    }
+
+    #[test]
+    fn diverged_when_head_moved_past_expected_base() {
+        // given
+        let root = temp_dir();
+        init_repo(&root);
+        let old_sha = head_sha(&root);
+        commit_file(&root, "extra.txt", "move head forward");
+        let new_sha = head_sha(&root);
+        let source = BaseCommitSource::Flag(old_sha.clone());
+
+        // when
+        let state = check_base_commit(&root, Some(&source));
+
+        // then
+        assert_eq!(
+            state,
+            BaseCommitState::Diverged {
+                expected: old_sha,
+                actual: new_sha,
+            }
+        );
+        fs::remove_dir_all(&root).expect("cleanup");
+    }
+
+    #[test]
+    fn no_expected_base_when_source_is_none() {
+        // given
+        let root = temp_dir();
+        init_repo(&root);
+
+        // when
+        let state = check_base_commit(&root, None);
+
+        // then
+        assert_eq!(state, BaseCommitState::NoExpectedBase);
+        fs::remove_dir_all(&root).expect("cleanup");
+    }
+
+    #[test]
+    fn not_a_git_repo_when_outside_repo() {
+        // given
+        let root = temp_dir();
+        fs::create_dir_all(&root).expect("create dir");
+        let source = BaseCommitSource::Flag("abc1234".to_string());
+
+        // when
+        let state = check_base_commit(&root, Some(&source));
+
+        // then
+        assert_eq!(state, BaseCommitState::NotAGitRepo);
+        fs::remove_dir_all(&root).expect("cleanup");
+    }
+
+    #[test]
+    fn reads_claw_base_file() {
+        // given
+        let root = temp_dir();
+        fs::create_dir_all(&root).expect("create dir");
+        fs::write(root.join(".claw-base"), "abc1234def5678\n").expect("write .claw-base");
+
+        // when
+        let value = read_claw_base_file(&root);
+
+        // then
+        assert_eq!(value, Some("abc1234def5678".to_string()));
+        fs::remove_dir_all(&root).expect("cleanup");
+    }
+
+    #[test]
+    fn returns_none_for_missing_claw_base_file() {
+        // given
+        let root = temp_dir();
+        fs::create_dir_all(&root).expect("create dir");
+
+        // when
+        let value = read_claw_base_file(&root);
+
+        // then
+        assert!(value.is_none());
+        fs::remove_dir_all(&root).expect("cleanup");
+    }
+
+    #[test]
+    fn returns_none_for_empty_claw_base_file() {
+        // given
+        let root = temp_dir();
+        fs::create_dir_all(&root).expect("create dir");
+        fs::write(root.join(".claw-base"), "  \n").expect("write empty .claw-base");
+
+        // when
+        let value = read_claw_base_file(&root);
+
+        // then
+        assert!(value.is_none());
+        fs::remove_dir_all(&root).expect("cleanup");
+    }
+
+    #[test]
+    fn resolve_expected_base_prefers_flag_over_file() {
+        // given
+        let root = temp_dir();
+        fs::create_dir_all(&root).expect("create dir");
+        fs::write(root.join(".claw-base"), "from_file\n").expect("write .claw-base");
+
+        // when
+        let source = resolve_expected_base(Some("from_flag"), &root);
+
+        // then
+        assert_eq!(
+            source,
+            Some(BaseCommitSource::Flag("from_flag".to_string()))
+        );
+        fs::remove_dir_all(&root).expect("cleanup");
+    }
+
+    #[test]
+    fn resolve_expected_base_falls_back_to_file() {
+        // given
+        let root = temp_dir();
+        fs::create_dir_all(&root).expect("create dir");
+        fs::write(root.join(".claw-base"), "from_file\n").expect("write .claw-base");
+
+        // when
+        let source = resolve_expected_base(None, &root);
+
+        // then
+        assert_eq!(
+            source,
+            Some(BaseCommitSource::File("from_file".to_string()))
+        );
+        fs::remove_dir_all(&root).expect("cleanup");
+    }
+
+    #[test]
+    fn resolve_expected_base_returns_none_when_nothing_available() {
+        // given
+        let root = temp_dir();
+        fs::create_dir_all(&root).expect("create dir");
+
+        // when
+        let source = resolve_expected_base(None, &root);
+
+        // then
+        assert!(source.is_none());
+        fs::remove_dir_all(&root).expect("cleanup");
+    }
+
+    #[test]
+    fn format_warning_returns_message_for_diverged() {
+        // given
+        let state = BaseCommitState::Diverged {
+            expected: "abc1234".to_string(),
+            actual: "def5678".to_string(),
+        };
+
+        // when
+        let warning = format_stale_base_warning(&state);
+
+        // then
+        let message = warning.expect("should produce warning");
+        assert!(message.contains("abc1234"));
+        assert!(message.contains("def5678"));
+        assert!(message.contains("stale codebase"));
+    }
+
+    #[test]
+    fn format_warning_returns_none_for_matches() {
+        // given
+        let state = BaseCommitState::Matches;
+
+        // when
+        let warning = format_stale_base_warning(&state);
+
+        // then
+        assert!(warning.is_none());
+    }
+
+    #[test]
+    fn format_warning_returns_none_for_no_expected_base() {
+        // given
+        let state = BaseCommitState::NoExpectedBase;
+
+        // when
+        let warning = format_stale_base_warning(&state);
+
+        // then
+        assert!(warning.is_none());
+    }
+
+    #[test]
+    fn matches_with_claw_base_file_in_real_repo() {
+        // given
+        let root = temp_dir();
+        init_repo(&root);
+        let sha = head_sha(&root);
+        fs::write(root.join(".claw-base"), format!("{sha}\n")).expect("write .claw-base");
+        let source = resolve_expected_base(None, &root);
+
+        // when
+        let state = check_base_commit(&root, source.as_ref());
+
+        // then
+        assert_eq!(state, BaseCommitState::Matches);
+        fs::remove_dir_all(&root).expect("cleanup");
+    }
+
+    #[test]
+    fn diverged_with_claw_base_file_after_new_commit() {
+        // given
+        let root = temp_dir();
+        init_repo(&root);
+        let old_sha = head_sha(&root);
+        fs::write(root.join(".claw-base"), format!("{old_sha}\n")).expect("write .claw-base");
+        commit_file(&root, "new.txt", "advance head");
+        let new_sha = head_sha(&root);
+        let source = resolve_expected_base(None, &root);
+
+        // when
+        let state = check_base_commit(&root, source.as_ref());
+
+        // then
+        assert_eq!(
+            state,
+            BaseCommitState::Diverged {
+                expected: old_sha,
+                actual: new_sha,
+            }
+        );
+        fs::remove_dir_all(&root).expect("cleanup");
+    }
+}
--- a/rust/crates/runtime/src/worker_boot.rs
+++ b/rust/crates/runtime/src/worker_boot.rs
@@ -560,6 +560,7 @@ fn push_event(
    let timestamp = now_secs();
    let seq = worker.events.len() as u64 + 1;
    worker.updated_at = timestamp;
+    worker.status = status;
    worker.events.push(WorkerEvent {
        seq,
        kind,
@@ -568,6 +569,50 @@ fn push_event(
        payload,
        timestamp,
    });
+    emit_state_file(worker);
+}
+
+/// Write current worker state to `.claw/worker-state.json` under the worker's cwd.
+/// This is the file-based observability surface: external observers (clawhip, orchestrators)
+/// poll this file instead of requiring an HTTP route on the opencode binary.
+fn emit_state_file(worker: &Worker) {
+    let state_dir = std::path::Path::new(&worker.cwd).join(".claw");
+    if let Err(_) = std::fs::create_dir_all(&state_dir) {
+        return;
+    }
+    let state_path = state_dir.join("worker-state.json");
+    let tmp_path = state_dir.join("worker-state.json.tmp");
+
+    #[derive(serde::Serialize)]
+    struct StateSnapshot<'a> {
+        worker_id: &'a str,
+        status: WorkerStatus,
+        is_ready: bool,
+        trust_gate_cleared: bool,
+        prompt_in_flight: bool,
+        last_event: Option<&'a WorkerEvent>,
+        updated_at: u64,
+        /// Seconds since last state transition. Clawhip uses this to detect
+        /// stalled workers without computing epoch deltas.
+        seconds_since_update: u64,
+    }
+
+    let now = now_secs();
+    let snapshot = StateSnapshot {
+        worker_id: &worker.worker_id,
+        status: worker.status,
+        is_ready: worker.status == WorkerStatus::ReadyForPrompt,
+        trust_gate_cleared: worker.trust_gate_cleared,
+        prompt_in_flight: worker.prompt_in_flight,
+        last_event: worker.events.last(),
+        updated_at: worker.updated_at,
+        seconds_since_update: now.saturating_sub(worker.updated_at),
+    };
+
+    if let Ok(json) = serde_json::to_string_pretty(&snapshot) {
+        let _ = std::fs::write(&tmp_path, json);
+        let _ = std::fs::rename(&tmp_path, &state_path);
+    }
 }

 fn path_matches_allowlist(cwd: &str, trusted_root: &str) -> bool {
@@ -1058,6 +1103,58 @@ mod tests {
            .any(|event| event.kind == WorkerEventKind::Failed));
    }

+    #[test]
+    fn emit_state_file_writes_worker_status_on_transition() {
+        let cwd_path = std::env::temp_dir().join(format!(
+            "claw-state-test-{}",
+            std::time::SystemTime::now()
+                .duration_since(std::time::UNIX_EPOCH)
+                .unwrap_or_default()
+                .as_nanos()
+        ));
+        std::fs::create_dir_all(&cwd_path).expect("test dir should create");
+        let cwd = cwd_path.to_str().expect("test path should be utf8");
+        let registry = WorkerRegistry::new();
+        let worker = registry.create(cwd, &[], true);
+
+        // After create the worker is Spawning — state file should exist
+        let state_path = cwd_path.join(".claw").join("worker-state.json");
+        assert!(
+            state_path.exists(),
+            "state file should exist after worker creation"
+        );
+
+        let raw = std::fs::read_to_string(&state_path).expect("state file should be readable");
+        let value: serde_json::Value =
+            serde_json::from_str(&raw).expect("state file should be valid JSON");
+        assert_eq!(
+            value["status"].as_str(),
+            Some("spawning"),
+            "initial status should be spawning"
+        );
+        assert_eq!(value["is_ready"].as_bool(), Some(false));
+
+        // Transition to ReadyForPrompt by observing trust-cleared text
+        registry
+            .observe(&worker.worker_id, "Ready for input\n>")
+            .expect("observe ready should succeed");
+
+        let raw = std::fs::read_to_string(&state_path)
+            .expect("state file should be readable after observe");
+        let value: serde_json::Value =
+            serde_json::from_str(&raw).expect("state file should be valid JSON after observe");
+        assert_eq!(
+            value["status"].as_str(),
+            Some("ready_for_prompt"),
+            "status should be ready_for_prompt after observe"
+        );
+        assert_eq!(
+            value["is_ready"].as_bool(),
+            Some(true),
+            "is_ready should be true when ReadyForPrompt"
+        );
+    }
+
    #[test]
    fn observe_completion_accepts_normal_finish_with_tokens() {
        let registry = WorkerRegistry::new();
--- a/rust/crates/rusty-claude-cli/src/main.rs
+++ b/rust/crates/rusty-claude-cli/src/main.rs
--- a/rust/crates/rusty-claude-cli/src/render.rs
+++ b/rust/crates/rusty-claude-cli/src/render.rs
@@ -249,13 +249,14 @@ impl TerminalRenderer {

    #[must_use]
    pub fn render_markdown(&self, markdown: &str) -> String {
+        let normalized = normalize_nested_fences(markdown);
        let mut output = String::new();
        let mut state = RenderState::default();
        let mut code_language = String::new();
        let mut code_buffer = String::new();
        let mut in_code_block = false;

-        for event in Parser::new_ext(markdown, Options::all()) {
+        for event in Parser::new_ext(&normalized, Options::all()) {
            self.render_event(
                event,
                &mut state,
@@ -634,8 +635,180 @@ fn apply_code_block_background(line: &str) -> String {
    format!("\u{1b}[48;5;236m{with_background}\u{1b}[0m{trailing_newline}")
 }

+/// Pre-process raw markdown so that fenced code blocks whose body contains
+/// fence markers of equal or greater length are wrapped with a longer fence.
+///
+/// LLMs frequently emit triple-backtick code blocks that contain triple-backtick
+/// examples.  CommonMark (and pulldown-cmark) treats the inner marker as the
+/// closing fence, breaking the render.  This function detects the situation and
+/// upgrades the outer fence to use enough backticks (or tildes) that the inner
+/// markers become ordinary content.
+fn normalize_nested_fences(markdown: &str) -> String {
+    // A fence line is either "labeled" (has an info string ⇒ always an opener)
+    // or "bare" (no info string ⇒ could be opener or closer).
+    #[derive(Debug, Clone)]
+    struct FenceLine {
+        char: char,
+        len: usize,
+        has_info: bool,
+        indent: usize,
+    }
+
+    fn parse_fence_line(line: &str) -> Option<FenceLine> {
+        let trimmed = line.trim_end_matches('\n').trim_end_matches('\r');
+        let indent = trimmed.chars().take_while(|c| *c == ' ').count();
+        if indent > 3 {
+            return None;
+        }
+        let rest = &trimmed[indent..];
+        let ch = rest.chars().next()?;
+        if ch != '`' && ch != '~' {
+            return None;
+        }
+        let len = rest.chars().take_while(|c| *c == ch).count();
+        if len < 3 {
+            return None;
+        }
+        let after = &rest[len..];
+        if ch == '`' && after.contains('`') {
+            return None;
+        }
+        let has_info = !after.trim().is_empty();
+        Some(FenceLine {
+            char: ch,
+            len,
+            has_info,
+            indent,
+        })
+    }
+
+    let lines: Vec<&str> = markdown.split_inclusive('\n').collect();
+    // Handle final line that may lack trailing newline.
+    // split_inclusive already keeps the original chunks, including a
+    // final chunk without '\n' if the input doesn't end with one.
+
+    // First pass: classify every line.
+    let fence_info: Vec<Option<FenceLine>> = lines.iter().map(|l| parse_fence_line(l)).collect();
+
+    // Second pass: pair openers with closers using a stack, recording
+    // (opener_idx, closer_idx) pairs plus the max fence length found between
+    // them.
+    struct StackEntry {
+        line_idx: usize,
+        fence: FenceLine,
+    }
+
+    let mut stack: Vec<StackEntry> = Vec::new();
+    // Paired blocks: (opener_line, closer_line, max_inner_fence_len)
+    let mut pairs: Vec<(usize, usize, usize)> = Vec::new();
+
+    for (i, fi) in fence_info.iter().enumerate() {
+        let Some(fl) = fi else { continue };
+
+        if fl.has_info {
+            // Labeled fence ⇒ always an opener.
+            stack.push(StackEntry {
+                line_idx: i,
+                fence: fl.clone(),
+            });
+        } else {
+            // Bare fence ⇒ try to close the top of the stack if compatible.
+            let closes_top = stack
+                .last()
+                .is_some_and(|top| top.fence.char == fl.char && fl.len >= top.fence.len);
+            if closes_top {
+                let opener = stack.pop().unwrap();
+                // Find max fence length of any fence line strictly between
+                // opener and closer (these are the nested fences).
+                let inner_max = fence_info[opener.line_idx + 1..i]
+                    .iter()
+                    .filter_map(|fi| fi.as_ref().map(|f| f.len))
+                    .max()
+                    .unwrap_or(0);
+                pairs.push((opener.line_idx, i, inner_max));
+            } else {
+                // Treat as opener.
+                stack.push(StackEntry {
+                    line_idx: i,
+                    fence: fl.clone(),
+                });
+            }
+        }
+    }
+
+    // Determine which lines need rewriting.  A pair needs rewriting when
+    // its opener length <= max inner fence length.
+    struct Rewrite {
+        char: char,
+        new_len: usize,
+        indent: usize,
+    }
+    let mut rewrites: std::collections::HashMap<usize, Rewrite> = std::collections::HashMap::new();
+
+    for (opener_idx, closer_idx, inner_max) in &pairs {
+        let opener_fl = fence_info[*opener_idx].as_ref().unwrap();
+        if opener_fl.len <= *inner_max {
+            let new_len = inner_max + 1;
+            let info_part = {
+                let trimmed = lines[*opener_idx]
+                    .trim_end_matches('\n')
+                    .trim_end_matches('\r');
+                let rest = &trimmed[opener_fl.indent..];
+                rest[opener_fl.len..].to_string()
+            };
+            rewrites.insert(
+                *opener_idx,
+                Rewrite {
+                    char: opener_fl.char,
+                    new_len,
+                    indent: opener_fl.indent,
+                },
+            );
+            let closer_fl = fence_info[*closer_idx].as_ref().unwrap();
+            rewrites.insert(
+                *closer_idx,
+                Rewrite {
+                    char: closer_fl.char,
+                    new_len,
+                    indent: closer_fl.indent,
+                },
+            );
+            // Store info string only in the opener; closer keeps the trailing
+            // portion which is already handled through the original line.
+            // Actually, we rebuild both lines from scratch below, including
+            // the info string for the opener.
+            let _ = info_part; // consumed in rebuild
+        }
+    }
+
+    if rewrites.is_empty() {
+        return markdown.to_string();
+    }
+
+    // Rebuild.
+    let mut out = String::with_capacity(markdown.len() + rewrites.len() * 4);
+    for (i, line) in lines.iter().enumerate() {
+        if let Some(rw) = rewrites.get(&i) {
+            let fence_str: String = std::iter::repeat(rw.char).take(rw.new_len).collect();
+            let indent_str: String = std::iter::repeat(' ').take(rw.indent).collect();
+            // Recover the original info string (if any) and trailing newline.
+            let trimmed = line.trim_end_matches('\n').trim_end_matches('\r');
+            let fi = fence_info[i].as_ref().unwrap();
+            let info = &trimmed[fi.indent + fi.len..];
+            let trailing = &line[trimmed.len()..];
+            out.push_str(&indent_str);
+            out.push_str(&fence_str);
+            out.push_str(info);
+            out.push_str(trailing);
+        } else {
+            out.push_str(line);
+        }
+    }
+    out
+}
+
 fn find_stream_safe_boundary(markdown: &str) -> Option<usize> {
-    let mut in_fence = false;
+    let mut open_fence: Option<FenceMarker> = None;
    let mut last_boundary = None;

    for (offset, line) in markdown.split_inclusive('\n').scan(0usize, |cursor, line| {
@@ -643,20 +816,21 @@ fn find_stream_safe_boundary(markdown: &str) -> Option<usize> {
        *cursor += line.len();
        Some((start, line))
    }) {
-        let trimmed = line.trim_start();
-        if trimmed.starts_with("```") || trimmed.starts_with("~~~") {
-            in_fence = !in_fence;
-            if !in_fence {
+        let line_without_newline = line.trim_end_matches('\n');
+        if let Some(opener) = open_fence {
+            if line_closes_fence(line_without_newline, opener) {
+                open_fence = None;
                last_boundary = Some(offset + line.len());
            }
            continue;
        }

-        if in_fence {
+        if let Some(opener) = parse_fence_opener(line_without_newline) {
+            open_fence = Some(opener);
            continue;
        }

-        if trimmed.is_empty() {
+        if line_without_newline.trim().is_empty() {
            last_boundary = Some(offset + line.len());
        }
    }
@@ -664,6 +838,46 @@ fn find_stream_safe_boundary(markdown: &str) -> Option<usize> {
    last_boundary
 }

+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+struct FenceMarker {
+    character: char,
+    length: usize,
+}
+
+fn parse_fence_opener(line: &str) -> Option<FenceMarker> {
+    let indent = line.chars().take_while(|c| *c == ' ').count();
+    if indent > 3 {
+        return None;
+    }
+    let rest = &line[indent..];
+    let character = rest.chars().next()?;
+    if character != '`' && character != '~' {
+        return None;
+    }
+    let length = rest.chars().take_while(|c| *c == character).count();
+    if length < 3 {
+        return None;
+    }
+    let info_string = &rest[length..];
+    if character == '`' && info_string.contains('`') {
+        return None;
+    }
+    Some(FenceMarker { character, length })
+}
+
+fn line_closes_fence(line: &str, opener: FenceMarker) -> bool {
+    let indent = line.chars().take_while(|c| *c == ' ').count();
+    if indent > 3 {
+        return false;
+    }
+    let rest = &line[indent..];
+    let length = rest.chars().take_while(|c| *c == opener.character).count();
+    if length < opener.length {
+        return false;
+    }
+    rest[length..].chars().all(|c| c == ' ' || c == '\t')
+}
+
 fn visible_width(input: &str) -> usize {
    strip_ansi(input).chars().count()
 }
@@ -778,6 +992,60 @@ mod tests {
        assert!(strip_ansi(&code).contains("fn main()"));
    }

+    #[test]
+    fn streaming_state_holds_outer_fence_with_nested_inner_fence() {
+        let renderer = TerminalRenderer::new();
+        let mut state = MarkdownStreamState::default();
+
+        assert_eq!(
+            state.push(&renderer, "````markdown\n```rust\nfn inner() {}\n"),
+            None,
+            "inner triple backticks must not close the outer four-backtick fence"
+        );
+        assert_eq!(
+            state.push(&renderer, "```\n"),
+            None,
+            "closing the inner fence must not flush the outer fence"
+        );
+        let flushed = state
+            .push(&renderer, "````\n")
+            .expect("closing the outer four-backtick fence flushes the buffered block");
+        let plain_text = strip_ansi(&flushed);
+        assert!(plain_text.contains("fn inner()"));
+        assert!(plain_text.contains("```rust"));
+    }
+
+    #[test]
+    fn streaming_state_distinguishes_backtick_and_tilde_fences() {
+        let renderer = TerminalRenderer::new();
+        let mut state = MarkdownStreamState::default();
+
+        assert_eq!(state.push(&renderer, "~~~text\n"), None);
+        assert_eq!(
+            state.push(&renderer, "```\nstill inside tilde fence\n"),
+            None,
+            "a backtick fence cannot close a tilde-opened fence"
+        );
+        assert_eq!(state.push(&renderer, "```\n"), None);
+        let flushed = state
+            .push(&renderer, "~~~\n")
+            .expect("matching tilde marker closes the fence");
+        let plain_text = strip_ansi(&flushed);
+        assert!(plain_text.contains("still inside tilde fence"));
+    }
+
+    #[test]
+    fn renders_nested_fenced_code_block_preserves_inner_markers() {
+        let terminal_renderer = TerminalRenderer::new();
+        let markdown_output =
+            terminal_renderer.markdown_to_ansi("````markdown\n```rust\nfn nested() {}\n```\n````");
+        let plain_text = strip_ansi(&markdown_output);
+
+        assert!(plain_text.contains("╭─ markdown"));
+        assert!(plain_text.contains("```rust"));
+        assert!(plain_text.contains("fn nested()"));
+    }
+
    #[test]
    fn spinner_advances_frames() {
        let terminal_renderer = TerminalRenderer::new();
--- a/rust/crates/rusty-claude-cli/tests/compact_output.rs
+++ b/rust/crates/rusty-claude-cli/tests/compact_output.rs
@@ -0,0 +1,159 @@
+use std::fs;
+use std::path::PathBuf;
+use std::process::{Command, Output};
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::time::{SystemTime, UNIX_EPOCH};
+
+use mock_anthropic_service::{MockAnthropicService, SCENARIO_PREFIX};
+
+static TEMP_COUNTER: AtomicU64 = AtomicU64::new(0);
+
+#[test]
+fn compact_flag_prints_only_final_assistant_text_without_tool_call_details() {
+    // given a workspace pointed at the mock Anthropic service and a fixture file
+    // that the read_file_roundtrip scenario will fetch through a tool call
+    let runtime = tokio::runtime::Runtime::new().expect("tokio runtime should build");
+    let server = runtime
+        .block_on(MockAnthropicService::spawn())
+        .expect("mock service should start");
+    let base_url = server.base_url();
+
+    let workspace = unique_temp_dir("compact-read-file");
+    let config_home = workspace.join("config-home");
+    let home = workspace.join("home");
+    fs::create_dir_all(&workspace).expect("workspace should exist");
+    fs::create_dir_all(&config_home).expect("config home should exist");
+    fs::create_dir_all(&home).expect("home should exist");
+    fs::write(workspace.join("fixture.txt"), "alpha parity line\n").expect("fixture should write");
+
+    // when we run claw in compact text mode against a tool-using scenario
+    let prompt = format!("{SCENARIO_PREFIX}read_file_roundtrip");
+    let output = run_claw(
+        &workspace,
+        &config_home,
+        &home,
+        &base_url,
+        &[
+            "--model",
+            "sonnet",
+            "--permission-mode",
+            "read-only",
+            "--allowedTools",
+            "read_file",
+            "--compact",
+            &prompt,
+        ],
+    );
+
+    // then the command exits successfully and stdout contains exactly the final
+    // assistant text with no tool call IDs, JSON envelopes, or spinner output
+    assert!(
+        output.status.success(),
+        "compact run should succeed\nstdout:\n{}\n\nstderr:\n{}",
+        String::from_utf8_lossy(&output.stdout),
+        String::from_utf8_lossy(&output.stderr),
+    );
+    let stdout = String::from_utf8(output.stdout).expect("stdout should be utf8");
+    let trimmed = stdout.trim_end_matches('\n');
+    assert_eq!(
+        trimmed, "read_file roundtrip complete: alpha parity line",
+        "compact stdout should contain only the final assistant text"
+    );
+    assert!(
+        !stdout.contains("toolu_"),
+        "compact stdout must not leak tool_use_id ({stdout:?})"
+    );
+    assert!(
+        !stdout.contains("\"tool_uses\""),
+        "compact stdout must not leak json envelopes ({stdout:?})"
+    );
+    assert!(
+        !stdout.contains("Thinking"),
+        "compact stdout must not include the spinner banner ({stdout:?})"
+    );
+
+    fs::remove_dir_all(&workspace).expect("workspace cleanup should succeed");
+}
+
+#[test]
+fn compact_flag_streaming_text_only_emits_final_message_text() {
+    // given a workspace pointed at the mock Anthropic service running the
+    // streaming_text scenario which only emits a single assistant text block
+    let runtime = tokio::runtime::Runtime::new().expect("tokio runtime should build");
+    let server = runtime
+        .block_on(MockAnthropicService::spawn())
+        .expect("mock service should start");
+    let base_url = server.base_url();
+
+    let workspace = unique_temp_dir("compact-streaming-text");
+    let config_home = workspace.join("config-home");
+    let home = workspace.join("home");
+    fs::create_dir_all(&workspace).expect("workspace should exist");
+    fs::create_dir_all(&config_home).expect("config home should exist");
+    fs::create_dir_all(&home).expect("home should exist");
+
+    // when we invoke claw with --compact for the streaming text scenario
+    let prompt = format!("{SCENARIO_PREFIX}streaming_text");
+    let output = run_claw(
+        &workspace,
+        &config_home,
+        &home,
+        &base_url,
+        &[
+            "--model",
+            "sonnet",
+            "--permission-mode",
+            "read-only",
+            "--compact",
+            &prompt,
+        ],
+    );
+
+    // then stdout should be exactly the assistant text followed by a newline
+    assert!(
+        output.status.success(),
+        "compact streaming run should succeed\nstdout:\n{}\n\nstderr:\n{}",
+        String::from_utf8_lossy(&output.stdout),
+        String::from_utf8_lossy(&output.stderr),
+    );
+    let stdout = String::from_utf8(output.stdout).expect("stdout should be utf8");
+    assert_eq!(
+        stdout, "Mock streaming says hello from the parity harness.\n",
+        "compact streaming stdout should contain only the final assistant text"
+    );
+
+    fs::remove_dir_all(&workspace).expect("workspace cleanup should succeed");
+}
+
+fn run_claw(
+    cwd: &std::path::Path,
+    config_home: &std::path::Path,
+    home: &std::path::Path,
+    base_url: &str,
+    args: &[&str],
+) -> Output {
+    let mut command = Command::new(env!("CARGO_BIN_EXE_claw"));
+    command
+        .current_dir(cwd)
+        .env_clear()
+        .env("ANTHROPIC_API_KEY", "test-compact-key")
+        .env("ANTHROPIC_BASE_URL", base_url)
+        .env("CLAW_CONFIG_HOME", config_home)
+        .env("HOME", home)
+        .env("NO_COLOR", "1")
+        .env("PATH", "/usr/bin:/bin")
+        .args(args);
+    command.output().expect("claw should launch")
+}
+
+fn unique_temp_dir(label: &str) -> PathBuf {
+    let millis = SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .expect("clock should be after epoch")
+        .as_millis();
+    let counter = TEMP_COUNTER.fetch_add(1, Ordering::Relaxed);
+    std::env::temp_dir().join(format!(
+        "claw-compact-{label}-{}-{millis}-{counter}",
+        std::process::id()
+    ))
+}
--- a/rust/crates/rusty-claude-cli/tests/mock_parity_harness.rs
+++ b/rust/crates/rusty-claude-cli/tests/mock_parity_harness.rs
@@ -183,17 +183,24 @@ fn clean_env_cli_reaches_mock_anthropic_service_across_scripted_parity_scenarios
    }

    let captured = runtime.block_on(server.captured_requests());
-    assert_eq!(
-        captured.len(),
-        21,
-        "twelve scenarios should produce twenty-one requests"
-    );
-    assert!(captured
+    // After `be561bf` added count_tokens preflight, each turn sends an
+    // extra POST to `/v1/messages/count_tokens` before the messages POST.
+    // The original count (21) assumed messages-only requests.  We now
+    // filter to `/v1/messages` and verify that subset matches the original
+    // scenario expectation.
+    let messages_only: Vec<_> = captured
        .iter()
-        .all(|request| request.path == "/v1/messages"));
-    assert!(captured.iter().all(|request| request.stream));
+        .filter(|r| r.path == "/v1/messages")
+        .collect();
+    assert_eq!(
+        messages_only.len(),
+        21,
+        "twelve scenarios should produce twenty-one /v1/messages requests (total captured: {}, includes count_tokens)",
+        captured.len()
+    );
+    assert!(messages_only.iter().all(|request| request.stream));

-    let scenarios = captured
+    let scenarios = messages_only
        .iter()
        .map(|request| request.scenario.as_str())
        .collect::<Vec<_>>();
--- a/rust/crates/tools/Cargo.toml
+++ b/rust/crates/tools/Cargo.toml
@@ -8,6 +8,7 @@ publish.workspace = true
 [dependencies]
 api = { path = "../api" }
 commands = { path = "../commands" }
+flate2 = "1"
 plugins = { path = "../plugins" }
 runtime = { path = "../runtime" }
 reqwest = { version = "0.12", default-features = false, features = ["blocking", "rustls-tls"] }
--- a/rust/crates/tools/src/lib.rs
+++ b/rust/crates/tools/src/lib.rs
--- a/rust/crates/tools/src/pdf_extract.rs
+++ b/rust/crates/tools/src/pdf_extract.rs
@@ -0,0 +1,548 @@
+//! Minimal PDF text extraction.
+//!
+//! Reads a PDF file, locates `/Contents` stream objects, decompresses with
+//! flate2 when the stream uses `/FlateDecode`, and extracts text operators
+//! found between `BT` / `ET` markers.
+
+use std::io::Read as _;
+use std::path::Path;
+
+/// Extract all readable text from a PDF file.
+///
+/// Returns the concatenated text found inside BT/ET operators across all
+/// content streams.  Non-text pages or encrypted PDFs yield an empty string
+/// rather than an error.
+pub fn extract_text(path: &Path) -> Result<String, String> {
+    let data = std::fs::read(path).map_err(|e| format!("failed to read PDF: {e}"))?;
+    Ok(extract_text_from_bytes(&data))
+}
+
+/// Core extraction from raw PDF bytes — useful for testing without touching the
+/// filesystem.
+pub(crate) fn extract_text_from_bytes(data: &[u8]) -> String {
+    let mut all_text = String::new();
+    let mut offset = 0;
+
+    while offset < data.len() {
+        let Some(stream_start) = find_subsequence(&data[offset..], b"stream") else {
+            break;
+        };
+        let abs_start = offset + stream_start;
+
+        // Determine the byte offset right after "stream\r\n" or "stream\n".
+        let content_start = skip_stream_eol(data, abs_start + b"stream".len());
+
+        let Some(end_rel) = find_subsequence(&data[content_start..], b"endstream") else {
+            break;
+        };
+        let content_end = content_start + end_rel;
+
+        // Look backwards from "stream" for a FlateDecode hint in the object
+        // dictionary.  We scan at most 512 bytes before the stream keyword.
+        let dict_window_start = abs_start.saturating_sub(512);
+        let dict_window = &data[dict_window_start..abs_start];
+        let is_flate = find_subsequence(dict_window, b"FlateDecode").is_some();
+
+        // Only process streams whose parent dictionary references /Contents or
+        // looks like a page content stream (contains /Length).  We intentionally
+        // keep this loose to cover both inline and referenced content streams.
+        let raw = &data[content_start..content_end];
+        let decompressed;
+        let stream_bytes: &[u8] = if is_flate {
+            if let Ok(buf) = inflate(raw) {
+                decompressed = buf;
+                &decompressed
+            } else {
+                offset = content_end;
+                continue;
+            }
+        } else {
+            raw
+        };
+
+        let text = extract_bt_et_text(stream_bytes);
+        if !text.is_empty() {
+            if !all_text.is_empty() {
+                all_text.push('\n');
+            }
+            all_text.push_str(&text);
+        }
+
+        offset = content_end;
+    }
+
+    all_text
+}
+
+/// Inflate (zlib / deflate) compressed data via `flate2`.
+fn inflate(data: &[u8]) -> Result<Vec<u8>, String> {
+    let mut decoder = flate2::read::ZlibDecoder::new(data);
+    let mut buf = Vec::new();
+    decoder
+        .read_to_end(&mut buf)
+        .map_err(|e| format!("flate2 inflate error: {e}"))?;
+    Ok(buf)
+}
+
+/// Extract text from PDF content-stream operators between BT and ET markers.
+///
+/// Handles the common text-showing operators:
+/// - `Tj`  — show a string
+/// - `TJ`  — show an array of strings/numbers
+/// - `'`   — move to next line and show string
+/// - `"`   — set spacing, move to next line and show string
+fn extract_bt_et_text(stream: &[u8]) -> String {
+    let text = String::from_utf8_lossy(stream);
+    let mut result = String::new();
+    let mut in_bt = false;
+
+    for line in text.lines() {
+        let trimmed = line.trim();
+        if trimmed == "BT" {
+            in_bt = true;
+            continue;
+        }
+        if trimmed == "ET" {
+            in_bt = false;
+            continue;
+        }
+        if !in_bt {
+            continue;
+        }
+
+        // Tj operator: (text) Tj
+        if trimmed.ends_with("Tj") {
+            if let Some(s) = extract_parenthesized_string(trimmed) {
+                if !result.is_empty() && !result.ends_with('\n') {
+                    result.push(' ');
+                }
+                result.push_str(&s);
+            }
+        }
+        // TJ operator: [ (text) 123 (text) ] TJ
+        else if trimmed.ends_with("TJ") {
+            let extracted = extract_tj_array(trimmed);
+            if !extracted.is_empty() {
+                if !result.is_empty() && !result.ends_with('\n') {
+                    result.push(' ');
+                }
+                result.push_str(&extracted);
+            }
+        }
+        // ' operator: (text) '   and   " operator: aw ac (text) "
+        else if is_newline_show_operator(trimmed) {
+            if let Some(s) = extract_parenthesized_string(trimmed) {
+                if !result.is_empty() {
+                    result.push('\n');
+                }
+                result.push_str(&s);
+            }
+        }
+    }
+
+    result
+}
+
+/// Returns `true` when `trimmed` looks like a `'` or `"` text-show operator.
+fn is_newline_show_operator(trimmed: &str) -> bool {
+    (trimmed.ends_with('\'') && trimmed.len() > 1)
+        || (trimmed.ends_with('"') && trimmed.contains('('))
+}
+
+/// Pull the text from the first `(…)` group, handling escaped parens and
+/// common PDF escape sequences.
+fn extract_parenthesized_string(input: &str) -> Option<String> {
+    let open = input.find('(')?;
+    let bytes = input.as_bytes();
+    let mut depth = 0;
+    let mut result = String::new();
+    let mut i = open;
+
+    while i < bytes.len() {
+        match bytes[i] {
+            b'(' => {
+                if depth > 0 {
+                    result.push('(');
+                }
+                depth += 1;
+            }
+            b')' => {
+                depth -= 1;
+                if depth == 0 {
+                    return Some(result);
+                }
+                result.push(')');
+            }
+            b'\\' if i + 1 < bytes.len() => {
+                i += 1;
+                match bytes[i] {
+                    b'n' => result.push('\n'),
+                    b'r' => result.push('\r'),
+                    b't' => result.push('\t'),
+                    b'\\' => result.push('\\'),
+                    b'(' => result.push('('),
+                    b')' => result.push(')'),
+                    // Octal sequences — up to 3 digits.
+                    d @ b'0'..=b'7' => {
+                        let mut octal = u32::from(d - b'0');
+                        for _ in 0..2 {
+                            if i + 1 < bytes.len()
+                                && bytes[i + 1].is_ascii_digit()
+                                && bytes[i + 1] <= b'7'
+                            {
+                                i += 1;
+                                octal = octal * 8 + u32::from(bytes[i] - b'0');
+                            } else {
+                                break;
+                            }
+                        }
+                        if let Some(ch) = char::from_u32(octal) {
+                            result.push(ch);
+                        }
+                    }
+                    other => result.push(char::from(other)),
+                }
+            }
+            ch => result.push(char::from(ch)),
+        }
+        i += 1;
+    }
+
+    None // unbalanced
+}
+
+/// Extract concatenated strings from a TJ array like `[ (Hello) -120 (World) ] TJ`.
+fn extract_tj_array(input: &str) -> String {
+    let mut result = String::new();
+    let Some(bracket_start) = input.find('[') else {
+        return result;
+    };
+    let Some(bracket_end) = input.rfind(']') else {
+        return result;
+    };
+    let inner = &input[bracket_start + 1..bracket_end];
+
+    let mut i = 0;
+    let bytes = inner.as_bytes();
+    while i < bytes.len() {
+        if bytes[i] == b'(' {
+            // Reconstruct the parenthesized string and extract it.
+            if let Some(s) = extract_parenthesized_string(&inner[i..]) {
+                result.push_str(&s);
+                // Skip past the closing paren.
+                let mut depth = 0u32;
+                for &b in &bytes[i..] {
+                    i += 1;
+                    if b == b'(' {
+                        depth += 1;
+                    } else if b == b')' {
+                        depth -= 1;
+                        if depth == 0 {
+                            break;
+                        }
+                    }
+                }
+                continue;
+            }
+        }
+        i += 1;
+    }
+
+    result
+}
+
+/// Skip past the end-of-line marker that immediately follows the `stream`
+/// keyword.  Per the PDF spec this is either `\r\n` or `\n`.
+fn skip_stream_eol(data: &[u8], pos: usize) -> usize {
+    if pos < data.len() && data[pos] == b'\r' {
+        if pos + 1 < data.len() && data[pos + 1] == b'\n' {
+            return pos + 2;
+        }
+        return pos + 1;
+    }
+    if pos < data.len() && data[pos] == b'\n' {
+        return pos + 1;
+    }
+    pos
+}
+
+/// Simple byte-subsequence search.
+fn find_subsequence(haystack: &[u8], needle: &[u8]) -> Option<usize> {
+    haystack
+        .windows(needle.len())
+        .position(|window| window == needle)
+}
+
+/// Check if a user-supplied path looks like a PDF file reference.
+#[must_use]
+pub fn looks_like_pdf_path(text: &str) -> Option<&str> {
+    for token in text.split_whitespace() {
+        let cleaned = token.trim_matches(|c: char| c == '\'' || c == '"' || c == '`');
+        if let Some(dot_pos) = cleaned.rfind('.') {
+            if cleaned[dot_pos + 1..].eq_ignore_ascii_case("pdf") && dot_pos > 0 {
+                return Some(cleaned);
+            }
+        }
+    }
+    None
+}
+
+/// Auto-extract text from a PDF path mentioned in a user prompt.
+///
+/// Returns `Some((path, extracted_text))` when a `.pdf` path is detected and
+/// the file exists, otherwise `None`.
+#[must_use]
+pub fn maybe_extract_pdf_from_prompt(prompt: &str) -> Option<(String, String)> {
+    let pdf_path = looks_like_pdf_path(prompt)?;
+    let path = Path::new(pdf_path);
+    if !path.exists() {
+        return None;
+    }
+    let text = extract_text(path).ok()?;
+    if text.is_empty() {
+        return None;
+    }
+    Some((pdf_path.to_string(), text))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Build a minimal valid PDF with a single page containing uncompressed
+    /// text.  This is the smallest PDF structure that exercises the BT/ET
+    /// extraction path.
+    fn build_simple_pdf(text: &str) -> Vec<u8> {
+        let content_stream = format!("BT\n/F1 12 Tf\n({text}) Tj\nET");
+        let stream_bytes = content_stream.as_bytes();
+        let mut pdf = Vec::new();
+
+        // Header
+        pdf.extend_from_slice(b"%PDF-1.4\n");
+
+        // Object 1 — Catalog
+        let obj1_offset = pdf.len();
+        pdf.extend_from_slice(b"1 0 obj\n<< /Type /Catalog /Pages 2 0 R >>\nendobj\n");
+
+        // Object 2 — Pages
+        let obj2_offset = pdf.len();
+        pdf.extend_from_slice(b"2 0 obj\n<< /Type /Pages /Kids [3 0 R] /Count 1 >>\nendobj\n");
+
+        // Object 3 — Page
+        let obj3_offset = pdf.len();
+        pdf.extend_from_slice(
+            b"3 0 obj\n<< /Type /Page /Parent 2 0 R /Contents 4 0 R >>\nendobj\n",
+        );
+
+        // Object 4 — Content stream (uncompressed)
+        let obj4_offset = pdf.len();
+        let length = stream_bytes.len();
+        let header = format!("4 0 obj\n<< /Length {length} >>\nstream\n");
+        pdf.extend_from_slice(header.as_bytes());
+        pdf.extend_from_slice(stream_bytes);
+        pdf.extend_from_slice(b"\nendstream\nendobj\n");
+
+        // Cross-reference table
+        let xref_offset = pdf.len();
+        pdf.extend_from_slice(b"xref\n0 5\n");
+        pdf.extend_from_slice(b"0000000000 65535 f \n");
+        pdf.extend_from_slice(format!("{obj1_offset:010} 00000 n \n").as_bytes());
+        pdf.extend_from_slice(format!("{obj2_offset:010} 00000 n \n").as_bytes());
+        pdf.extend_from_slice(format!("{obj3_offset:010} 00000 n \n").as_bytes());
+        pdf.extend_from_slice(format!("{obj4_offset:010} 00000 n \n").as_bytes());
+
+        // Trailer
+        pdf.extend_from_slice(b"trailer\n<< /Size 5 /Root 1 0 R >>\n");
+        pdf.extend_from_slice(format!("startxref\n{xref_offset}\n%%EOF\n").as_bytes());
+
+        pdf
+    }
+
+    /// Build a minimal PDF with flate-compressed content stream.
+    fn build_flate_pdf(text: &str) -> Vec<u8> {
+        use flate2::write::ZlibEncoder;
+        use flate2::Compression;
+        use std::io::Write as _;
+
+        let content_stream = format!("BT\n/F1 12 Tf\n({text}) Tj\nET");
+        let mut encoder = ZlibEncoder::new(Vec::new(), Compression::default());
+        encoder
+            .write_all(content_stream.as_bytes())
+            .expect("compress");
+        let compressed = encoder.finish().expect("finish");
+
+        let mut pdf = Vec::new();
+        pdf.extend_from_slice(b"%PDF-1.4\n");
+
+        let obj1_offset = pdf.len();
+        pdf.extend_from_slice(b"1 0 obj\n<< /Type /Catalog /Pages 2 0 R >>\nendobj\n");
+
+        let obj2_offset = pdf.len();
+        pdf.extend_from_slice(b"2 0 obj\n<< /Type /Pages /Kids [3 0 R] /Count 1 >>\nendobj\n");
+
+        let obj3_offset = pdf.len();
+        pdf.extend_from_slice(
+            b"3 0 obj\n<< /Type /Page /Parent 2 0 R /Contents 4 0 R >>\nendobj\n",
+        );
+
+        let obj4_offset = pdf.len();
+        let length = compressed.len();
+        let header = format!("4 0 obj\n<< /Length {length} /Filter /FlateDecode >>\nstream\n");
+        pdf.extend_from_slice(header.as_bytes());
+        pdf.extend_from_slice(&compressed);
+        pdf.extend_from_slice(b"\nendstream\nendobj\n");
+
+        let xref_offset = pdf.len();
+        pdf.extend_from_slice(b"xref\n0 5\n");
+        pdf.extend_from_slice(b"0000000000 65535 f \n");
+        pdf.extend_from_slice(format!("{obj1_offset:010} 00000 n \n").as_bytes());
+        pdf.extend_from_slice(format!("{obj2_offset:010} 00000 n \n").as_bytes());
+        pdf.extend_from_slice(format!("{obj3_offset:010} 00000 n \n").as_bytes());
+        pdf.extend_from_slice(format!("{obj4_offset:010} 00000 n \n").as_bytes());
+
+        pdf.extend_from_slice(b"trailer\n<< /Size 5 /Root 1 0 R >>\n");
+        pdf.extend_from_slice(format!("startxref\n{xref_offset}\n%%EOF\n").as_bytes());
+
+        pdf
+    }
+
+    #[test]
+    fn extracts_uncompressed_text_from_minimal_pdf() {
+        // given
+        let pdf_bytes = build_simple_pdf("Hello World");
+
+        // when
+        let text = extract_text_from_bytes(&pdf_bytes);
+
+        // then
+        assert_eq!(text, "Hello World");
+    }
+
+    #[test]
+    fn extracts_text_from_flate_compressed_stream() {
+        // given
+        let pdf_bytes = build_flate_pdf("Compressed PDF Text");
+
+        // when
+        let text = extract_text_from_bytes(&pdf_bytes);
+
+        // then
+        assert_eq!(text, "Compressed PDF Text");
+    }
+
+    #[test]
+    fn handles_tj_array_operator() {
+        // given
+        let stream = b"BT\n/F1 12 Tf\n[ (Hello) -120 ( World) ] TJ\nET";
+        // Build a raw PDF with TJ array operator instead of simple Tj.
+        let content_stream = std::str::from_utf8(stream).unwrap();
+        let raw = format!(
+            "%PDF-1.4\n1 0 obj\n<< /Type /Catalog >>\nendobj\n\
+             2 0 obj\n<< /Length {} >>\nstream\n{}\nendstream\nendobj\n%%EOF\n",
+            content_stream.len(),
+            content_stream
+        );
+        let pdf_bytes = raw.into_bytes();
+
+        // when
+        let text = extract_text_from_bytes(&pdf_bytes);
+
+        // then
+        assert_eq!(text, "Hello World");
+    }
+
+    #[test]
+    fn handles_escaped_parentheses() {
+        // given
+        let content = b"BT\n(Hello \\(World\\)) Tj\nET";
+        let raw = format!(
+            "%PDF-1.4\n1 0 obj\n<< /Length {} >>\nstream\n",
+            content.len()
+        );
+        let mut pdf_bytes = raw.into_bytes();
+        pdf_bytes.extend_from_slice(content);
+        pdf_bytes.extend_from_slice(b"\nendstream\nendobj\n%%EOF\n");
+
+        // when
+        let text = extract_text_from_bytes(&pdf_bytes);
+
+        // then
+        assert_eq!(text, "Hello (World)");
+    }
+
+    #[test]
+    fn returns_empty_for_non_pdf_data() {
+        // given
+        let data = b"This is not a PDF file at all";
+
+        // when
+        let text = extract_text_from_bytes(data);
+
+        // then
+        assert!(text.is_empty());
+    }
+
+    #[test]
+    fn extracts_text_from_file_on_disk() {
+        // given
+        let pdf_bytes = build_simple_pdf("Disk Test");
+        let dir = std::env::temp_dir().join("clawd-pdf-extract-test");
+        std::fs::create_dir_all(&dir).unwrap();
+        let pdf_path = dir.join("test.pdf");
+        std::fs::write(&pdf_path, &pdf_bytes).unwrap();
+
+        // when
+        let text = extract_text(&pdf_path).unwrap();
+
+        // then
+        assert_eq!(text, "Disk Test");
+
+        // cleanup
+        let _ = std::fs::remove_dir_all(&dir);
+    }
+
+    #[test]
+    fn looks_like_pdf_path_detects_pdf_references() {
+        // given / when / then
+        assert_eq!(
+            looks_like_pdf_path("Please read /tmp/report.pdf"),
+            Some("/tmp/report.pdf")
+        );
+        assert_eq!(looks_like_pdf_path("Check file.PDF now"), Some("file.PDF"));
+        assert_eq!(looks_like_pdf_path("no pdf here"), None);
+    }
+
+    #[test]
+    fn maybe_extract_pdf_from_prompt_returns_none_for_missing_file() {
+        // given
+        let prompt = "Read /tmp/nonexistent-abc123.pdf please";
+
+        // when
+        let result = maybe_extract_pdf_from_prompt(prompt);
+
+        // then
+        assert!(result.is_none());
+    }
+
+    #[test]
+    fn maybe_extract_pdf_from_prompt_extracts_existing_file() {
+        // given
+        let pdf_bytes = build_simple_pdf("Auto Extracted");
+        let dir = std::env::temp_dir().join("clawd-pdf-auto-extract-test");
+        std::fs::create_dir_all(&dir).unwrap();
+        let pdf_path = dir.join("auto.pdf");
+        std::fs::write(&pdf_path, &pdf_bytes).unwrap();
+        let prompt = format!("Summarize {}", pdf_path.display());
+
+        // when
+        let result = maybe_extract_pdf_from_prompt(&prompt);
+
+        // then
+        let (path, text) = result.expect("should extract");
+        assert_eq!(path, pdf_path.display().to_string());
+        assert_eq!(text, "Auto Extracted");
+
+        // cleanup
+        let _ = std::fs::remove_dir_all(&dir);
+    }
+}