CI has been red since be561bf ('Use Anthropic count tokens for preflight')
because that commit replaced the free-function preflight_message_request
(byte-estimate guard) with an instance method that silently returns Ok on
any count_tokens failure:
let counted_input_tokens = match self.count_tokens(request).await {
Ok(count) => count,
Err(_) => return Ok(()), // <-- silent bypass
};
Two consequences:
1. client_integration::send_message_blocks_oversized_requests_before_the_http_call
has been FAILING on every CI run since be561bf. The mock server in that
test only has one HTTP response queued (a bare '{}' to satisfy the main
request), so the count_tokens POST parses into an empty body that fails
to deserialize into CountTokensResponse -> Err -> silent bypass -> the
oversized 600k-char request proceeds to the mock instead of being
rejected with ContextWindowExceeded as the test expects.
2. In production, any third-party Anthropic-compatible gateway that doesn't
implement /v1/messages/count_tokens (OpenRouter, Cloudflare AI Gateway,
etc.) would silently disable the preflight guard entirely, letting
oversized requests hit the upstream only to fail there with a provider-
side context-window error. This is exactly the 'opaque failure surface'
ROADMAP #22 asked us to avoid.
Fix: call the free-function super::preflight_message_request(request)? as
the first step in the instance method, before any network round-trip. This
guarantees the byte-estimate guard always fires, whether or not the remote
count_tokens endpoint is reachable. The count_tokens refinement still runs
afterward when available for more precise token counting, but it is now
strictly additive — it can only catch more cases, never silently skip the
guard.
Test results:
- cargo test -p api --lib: 89 passed, 0 failed
- cargo test --release -p api (all test binaries): 118 passed, 0 failed
- cargo test --release -p api --test client_integration \
send_message_blocks_oversized_requests_before_the_http_call: passes
- cargo fmt --check: clean
This unblocks the Rust CI workflow which has been red on every push since
be561bf landed.
Users in Discord #clawcode-get-help (web3g) asked for Qwen 3.6 Plus via
native Alibaba DashScope API instead of OpenRouter, which has stricter
rate limits. This commit adds first-class routing for qwen/ and bare
qwen- prefixed model names.
Changes:
- DEFAULT_DASHSCOPE_BASE_URL constant: /compatible-mode/v1 endpoint
- OpenAiCompatConfig::dashscope() factory mirroring openai()/xai()
- DASHSCOPE_ENV_VARS + credential_env_vars() wiring
- metadata_for_model: qwen/ and qwen- prefix routes to DashScope with
auth_env=DASHSCOPE_API_KEY, reuses ProviderKind::OpenAi because
DashScope speaks the OpenAI REST shape
- is_reasoning_model: detect qwen-qwq, qwq-*, and *-thinking variants
so tuning params (temperature, top_p, etc.) get stripped before
payload assembly (same pattern as o1/o3/grok-3-mini)
Tests added:
- providers::tests::qwen_prefix_routes_to_dashscope_not_anthropic
- openai_compat::tests::qwen_reasoning_variants_are_detected
89 api lib tests passing, 0 failing. cargo fmt --check: clean.
Closes the user-reported gap: 'use Qwen 3.6 Plus via Alibaba API
directly, not OpenRouter' without needing OPENAI_BASE_URL override
or unsetting ANTHROPIC_API_KEY.
build_runtime_runs_plugin_lifecycle_init_and_shutdown was the only test
that set/removed ANTHROPIC_API_KEY without holding the env_lock mutex.
Under parallel workspace execution, other tests racing on the same env
var could wipe the key mid-construction, causing a flaky credential error.
Root cause: process-wide env vars are shared mutable state. All other
tests that touch ANTHROPIC_API_KEY already use env_lock(). This test
was the only holdout.
Fix: add let _guard = env_lock(); at the top of the test.
openai_compat_integration.rs and client_integration.rs had MessageRequest
constructions without the new tuning param fields (temperature, top_p,
frequency_penalty, presence_penalty, stop) added in c667d47.
Added ..Default::default() to all 4 sites. cargo fmt applied.
This was the root cause of CI red on main (E0063 compile error in
integration tests, not caught by --lib tests).
renders_help_from_shared_specs hardcoded the exact /session usage string,
which broke when /session delete was added in batch 5. Relaxed to check
for /session presence instead of exact subcommand list.
Pre-existing test brittleness (not caused by recent commits).
687 workspace lib tests passing, 0 failing.
MessageRequest now carries OpenAI-compatible tuning params (c667d47), but
the Anthropic API does not support frequency_penalty or presence_penalty,
and uses 'stop_sequences' instead of 'stop'. Without this fix, setting
these params with a Claude model would produce 400 errors.
Changes to strip_unsupported_beta_body_fields:
- Remove frequency_penalty and presence_penalty from Anthropic request body
- Convert stop → stop_sequences (only when non-empty)
- temperature and top_p are preserved (Anthropic supports both)
Tests added:
- strip_removes_openai_only_fields_and_converts_stop
- strip_does_not_add_empty_stop_sequences
87 api lib tests passing, 0 failing.
cargo check --workspace: clean.
Reasoning models reject temperature, top_p, frequency_penalty, and
presence_penalty with 400 errors. Instead of letting these flow through
and returning cryptic provider errors, strip them silently at the
request-builder boundary.
is_reasoning_model() classifies: o1*, o3*, o4*, grok-3-mini.
stop sequences are preserved (safe for all providers).
Tests added:
- reasoning_model_strips_tuning_params: o1-mini strips all 4 params, keeps stop
- grok_3_mini_is_reasoning_model: classification coverage for grok-3-mini, o1,
o3-mini, and negative cases (gpt-4o, grok-3, claude)
85 api lib tests passing, 0 failing.
MessageRequest was missing standard OpenAI-compatible generation tuning
parameters. Callers had no way to control temperature, top_p,
frequency_penalty, presence_penalty, or stop sequences.
Changes:
- Added 5 optional fields to MessageRequest (all Option, None by default)
- Wired into build_chat_completion_request: only included in payload when set
- All existing construction sites updated with ..Default::default()
- MessageRequest now derives Default for ergonomic partial construction
Tests added:
- tuning_params_included_in_payload_when_set: all 5 params flow into JSON
- tuning_params_omitted_from_payload_when_none: absent params stay absent
83 api lib tests passing, 0 failing.
cargo check --workspace: 0 warnings.
metadata_for_model returned None for unknown models like openai/gpt-4.1-mini,
causing detect_provider_kind to fall through to auth-sniffer order. If
ANTHROPIC_API_KEY was set, the model was silently misrouted to Anthropic
and the user got a confusing 'missing Anthropic credentials' error.
Fix: add explicit prefix checks for 'openai/' and 'gpt-' in
metadata_for_model so the model name wins over env-var presence.
Regression test added: openai_namespaced_model_routes_to_openai_not_anthropic
- 'openai/gpt-4.1-mini' routes to OpenAi
- 'gpt-4o' routes to OpenAi
Reported and reproduced by gaebal-gajae against current main.
81 api lib tests passing, 0 failing.
WorkerGet had zero test coverage. WorkerAwaitReady and WorkerSendPrompt
had only one happy-path test each with no error paths.
Added 4 tests:
- worker_get_returns_worker_state: WorkerGet fetches correct worker_id/status/cwd
- worker_get_on_unknown_id_returns_error: unknown id -> 'worker not found'
- worker_await_ready_on_spawning_worker_returns_not_ready: ready=false on spawning worker
- worker_send_prompt_on_non_ready_worker_returns_error: sending prompt before ready fails
94 tool tests passing, 0 failing.
recovery_loop_state_file_reflects_transitions reads the actual state
file after each transition to verify the canonical observability surface
reflects the full stall->resolve->ready progression:
spawning (state file exists, seconds_since_update present)
-> trust_required (is_ready=false, trust_gate_cleared=false in file)
-> spawning (trust_gate_cleared=true after WorkerResolveTrust)
-> ready_for_prompt (is_ready=true after ready screen observe)
This is the end-to-end proof gaebal-gajae called for: clawhip polling
.claw/worker-state.json will see truthful state at every step of the
recovery loop, including the seconds_since_update staleness signal.
90 tool tests passing, 0 failing.
Proves the clawhip restart/recover flow that gaebal-gajae flagged:
1. stall_detect_and_resolve_trust_end_to_end
- Worker created without trusted_roots -> trust_auto_resolve=false
- WorkerObserve with trust-prompt text -> status=trust_required, gate cleared=false
- WorkerResolveTrust -> status=spawning, trust_gate_cleared=true
- WorkerObserve with ready text -> status=ready_for_prompt
Full resolve path verified end-to-end.
2. stall_detect_and_restart_recovery_end_to_end
- Worker stalls at trust_required
- WorkerRestart resets to spawning, trust_gate_cleared=false
Documents the restart-then-re-acquire-trust flow.
Note: seconds_since_update is in .claw/worker-state.json (state file),
not in the Worker tool output struct. Staleness detection via state file
is covered by emit_state_file_writes_worker_status_on_transition in
worker_boot.rs tests.
87 tool tests passing, 0 failing.
Three dead-code warnings eliminated from cargo check:
1. KNOWN_TOP_LEVEL_KEYS / DEPRECATED_TOP_LEVEL_KEYS in config.rs
- Superseded by config_validate::TOP_LEVEL_FIELDS and DEPRECATED_FIELDS
- Were out of date (missing aliases, providerFallbacks, trustedRoots)
- Removed
2. read_git_recent_commits in prompt.rs
- Private function, never called anywhere in the codebase
- Removed
3. workspace_sessions_dir in session.rs
- Public API scaffolded for session isolation (#41)
- Genuinely useful for external consumers (clawhip enumerating sessions)
- Added 2 tests: deterministic path for same CWD, different path for different CWDs
- Annotated with #[allow(dead_code)] since it is external-facing API
cargo check --workspace: 0 warnings remaining
430 runtime tests passing, 0 failing
WorkerRestart and WorkerTerminate had zero test coverage despite being
public tools in the tool spec. Also confirms one design decision worth
noting: restart resets trust_gate_cleared=false, so an allowlisted
worker that gets restarted must re-acquire trust via the normal observe
flow (by design — trust is per-session, not per-CWD).
Tests added:
- worker_terminate_sets_finished_status
- worker_restart_resets_to_spawning (verifies status=spawning,
prompt_in_flight=false, trust_gate_cleared=false)
- worker_terminate_on_unknown_id_returns_error
- worker_restart_on_unknown_id_returns_error
85 tool tests passing, 0 failing.
Previously WorkerCreate passed trusted_roots directly to spawn_worker
with no config-level default. Any batch script omitting the field
stalled all workers at TrustRequired with no recovery path.
Now run_worker_create loads RuntimeConfig from the worker CWD before
spawning and merges config.trusted_roots() with per-call overrides.
Per-call overrides still take effect; config provides the default.
Add test: worker_create_merges_config_trusted_roots_without_per_call_override
- writes .claw/settings.json with trustedRoots=[<os-temp-dir>] in a temp worktree
- calls WorkerCreate with no trusted_roots field
- asserts trust_auto_resolve=true (config roots matched the CWD)
81 tool tests passing, 0 failing.
Closes the startup-friction gap filed in ROADMAP (dd97c49).
WorkerCreate required trusted_roots on every call with no config-level
default. Any batch script that omitted the field stalled all workers at
TrustRequired with no auto-recovery path.
Changes:
- RuntimeFeatureConfig: add trusted_roots: Vec<String> field
- ConfigLoader: wire parse_optional_trusted_roots() for 'trustedRoots' key
- RuntimeConfig / RuntimeFeatureConfig: expose trusted_roots() accessor
- config_validate: add trustedRoots to TOP_LEVEL_FIELDS schema (StringArray)
- Tests: parses_trusted_roots_from_settings + trusted_roots_default_is_empty_when_unset
Callers can now set trusted_roots in .claw/settings.json:
{ "trustedRoots": ["/tmp/worktrees"] }
WorkerRegistry::spawn_worker() callers should merge config.trusted_roots()
with any per-call overrides (wiring left for follow-up).
Two real schema gaps found via dogfood (cargo test -p runtime):
1. aliases and providerFallbacks not in TOP_LEVEL_FIELDS
- Both are valid config keys parsed by config.rs
- Validator was rejecting them as unknown keys
- 2 tests failing: parses_user_defined_model_aliases,
parses_provider_fallbacks_chain
2. Deprecated keys were being flagged as unknown before the deprecated
check ran (unknown-key check runs first in validate_object_keys)
- Added early-exit for deprecated keys in unknown-key loop
- Keeps deprecated→warning behavior for permissionMode/enabledPlugins
which still appear in valid legacy configs
3. Config integration tests had assertions on format strings that never
matched the actual validator output (path:3: vs path: ... (line N))
- Updated assertions to check for path + line + field name as
independent substrings instead of a format that was never produced
426 tests passing, 0 failing.
Clawhip needs to distinguish a stalled trust_required worker from one
that just transitioned. Without a pre-computed staleness field it has
to compute epoch delta itself from updated_at.
seconds_since_update = now - updated_at at snapshot write time.
Clawhip threshold: > 60s in trust_required = stalled; act.
WorkerStatus is fully tracked in worker_boot.rs but was invisible to
external observers (clawhip, orchestrators) because opencode serve's
HTTP server is upstream and not ours to extend.
Solution: atomic file-based observability.
- emit_state_file() writes .claw/worker-state.json on every push_event()
call (tmp write + rename for atomicity)
- Snapshot includes: worker_id, status, is_ready, trust_gate_cleared,
prompt_in_flight, last_event, updated_at
- Add 'claw state' CLI subcommand to read and print the file
- Add regression test: emit_state_file_writes_worker_status_on_transition
verifies spawning→ready_for_prompt transition is reflected on disk
This closes the /state dogfood gap without requiring any upstream
opencode changes. Clawhip can now distinguish a truly stalled worker
(status: trust_required or running with no recent updated_at) from a
quiet-but-progressing one.
Add missing imports to test module:
- PromptHistoryEntry, render_prompt_history_report, parse_history_count
- parse_export_args, render_session_markdown
- summarize_tool_payload_for_markdown, short_tool_id
Fixes test compilation errors introduced by new session and export
features from batch 5/6 work.
The CLI was using a flat cwd/.claw/sessions/ path without workspace
fingerprinting, while SessionStore::from_cwd() adds a hash subdirectory.
This mismatch meant the isolation machinery existed but wasn't actually
used by the main session management codepath.
Now sessions_dir() delegates to SessionStore::from_cwd(), ensuring all
session operations use workspace-fingerprinted directories.
Remove #[cfg(test)] gate from session_control module — SessionStore
is now available at runtime, not just in tests. Export SessionStore and
add workspace_sessions_dir() helper that creates fingerprinted session
directories per workspace root.
This is the #41 kill shot: parallel opencode serve instances will use
separate session namespaces based on workspace fingerprint instead of
sharing a global ~/.local/share/opencode/ store.
The CLI already uses cwd/.claw/sessions/ (sessions_dir()), and now
SessionStore::from_cwd() adds workspace hash isolation on top.
