fix(runtime): keep a failed turn's tool results, and size guards from the real context window

Five defects found by auditing the agent stack, all of which let the harness act
confidently on a wrong picture of its own state.

1. A turn that errored discarded the record of tools that had already run. The CLI
   hands run_turn a CLONE of the session; the turn appends the user message, each
   assistant reply, and a tool_result per executed tool to that clone alone. The error
   path dropped it while the side effects of those tools remained on disk, and the
   recovery path then replayed the same user input against a session showing no sign
   they had ever run — inviting a second execution of tools that are not idempotent.
   The clone is now adopted before teardown, and recovery resumes the turn instead of
   replaying it (new ConversationRuntime::resume_turn).

2. Every sizing guard was computed against the model NAME's context window. With a
   local server behind a cloud model alias that is 200k against a much smaller
   backend, which does not make the preflights inaccurate — it makes them unreachable.
   CLAW_CONTEXT_WINDOW now declares the real window; model_token_limit honours it and
   returns a limit for unrecognised models instead of None, max_tokens is capped to a
   quarter of it, and the auto-compaction threshold derives from it (70%) rather than
   defaulting to a 200k-window constant.

3. Whether to recover by discarding conversation history was decided by substring
   matching on rendered error prose, including strings that ordinary transport and
   decode failures also produce — so a dropped connection cost the session its
   transcript. Classification now happens in the API layer, where the error still has
   its type and the request is still in hand: typed overflow errors, plus ambiguous
   transport failures only when the request actually filled the window. The two
   copy-pasted detection blocks are gone.

4. The top-level agent loop had no iteration ceiling (usize::MAX) while spawned
   subagents were bounded at 32. It now gets the same bound, overridable with
   CLAW_MAX_TURN_ITERATIONS.

5. Auto-continue nudges were injected as USER turns, so harness-authored text was
   persisted, re-sent on every later request, and folded into compaction summaries as
   something the user had said. They are recorded as system turns now; the wire
   payload is unchanged.

Also replaces the repository's own self-host section, which described a translator
and a serving stack no longer in the request path. That file is folded into the
system prompt verbatim, so it was briefing the model with an architecture that does
not exist.

(cherry picked from commit c8834cb4d1a3372eb70241429f641162a420297e)

.github/workflows/release.yml

@@ -57,7 +57,7 @@ jobs:
           mkdir -p dist
           cp "target/release/${{ matrix.bin }}" "dist/${{ matrix.artifact_name }}"
           chmod +x "dist/${{ matrix.artifact_name }}"
-          (cd dist && sha256sum "${{ matrix.artifact_name }}" > "${{ matrix.artifact_name }}.sha256")
+          (cd dist && (sha256sum "${{ matrix.artifact_name }}" 2>/dev/null || shasum -a 256 "${{ matrix.artifact_name }}") > "${{ matrix.artifact_name }}.sha256")
 
       - name: Upload workflow artifact
         uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02

.gitignore

@@ -13,3 +13,5 @@ archive/
 status-help.txt
 # Legacy Python port session scratch artifacts
 .port_sessions/
+# Local agent session state
+.gjc/

.gjc/_session-5e2dd1e8-ea4b-4af1-b56b-89ab7cca4299/runtime/runtime-state.json

@@ -1 +0,0 @@
-{"schema_version":1,"session_id":"5e2dd1e8-ea4b-4af1-b56b-89ab7cca4299","state":"running","ready_for_input":false,"updated_at":"2026-10-12T22:09:30.659Z","current_turn_id":null,"last_turn_id":null,"live":true,"reason":null,"source":"agent_session_event","event":"turn_start","cwd":"/home/bellman/Data/Workspace/gc-task-cc-issue-3257-lane","workdir":"/home/bellman/Data/Workspace/gc-task-cc-issue-3257-lane","branch":null,"session_file":"/home/bellman/.gajaeway/broker/agent/sessions/v2-6fji37qf7ld523ivvoiohlcpmz3nu45m2apxyrkusn5nxczjmouq/2026-10-12T22-04-33-362Z_5e2dd1e8-ea4b-4af1-b56b-89ab7cca4299.jsonl","activity":{"seq":115,"last_activity_at":"2026-10-12T22:09:34.144Z","tool":"bash","phase":"started","outcome":null,"elapsed_ms":null,"active_tool_count":1,"active_tools":[{"tool":"bash","started_at":"2026-10-12T22:09:34.144Z"}],"in_flight":[{"digest":"e603a70efb03b6b2b78c94b704197d26fc8a35113543291a21f68b3b08ca11a9","tool":"bash","started_at":"2026-10-12T22:09:34.144Z"}]}}

.gjc/_session-5e2dd1e8-ea4b-4af1-b56b-89ab7cca4299/runtime/runtime-state.json.lock

@@ -1 +0,0 @@
-{"pid":1,"start_time":"unknown","token":"1a2e8a33-76f2-4175-ac23-1521458f12a2","owner_host_id":"7af2fb0e8e37bd543e3d0bda6f205b56f30077b5a9b77e83208e94ae441315ac","released":true}

.gjc/_session-5e2dd1e8-ea4b-4af1-b56b-89ab7cca4299/runtime/runtime-state.json.lock.transition.owner

@@ -1 +0,0 @@
-{"pid":1,"start_time":"unknown","token":"a1342032-af80-4a49-a783-291bf73bf027","owner_host_id":"7af2fb0e8e37bd543e3d0bda6f205b56f30077b5a9b77e83208e94ae441315ac","released":true}

.gjc/state/sdk/5e2dd1e8-ea4b-4af1-b56b-89ab7cca4299.json

@@ -1,12 +0,0 @@
-{
-  "version": 1,
-  "sessionId": "5e2dd1e8-ea4b-4af1-b56b-89ab7cca4299",
-  "pid": 487360,
-  "host": "127.0.0.1",
-  "port": 37541,
-  "url": "ws://127.0.0.1:37541",
-  "token": "LlIfu_aHH57jfys2Ii0ORMrnu56NLAa1",
-  "startedAt": 1791842673967,
-  "updatedAt": 1791842673967,
-  "stale": false
-}

.gjc/state/sdk/5e2dd1e8-ea4b-4af1-b56b-89ab7cca4299.lifecycle.json

@@ -1 +0,0 @@
-{"effectMarker":"f441857a-f591-474b-b34a-6d89b2bf9feb","incarnation":"linux:159099652","pid":487360}

.gjc/state/sdk/5e2dd1e8-ea4b-4af1-b56b-89ab7cca4299.lifecycle.ready.json

@@ -1 +0,0 @@
-{"effectMarker":"f441857a-f591-474b-b34a-6d89b2bf9feb","incarnation":"linux:159099652","pid":487360}

rust/crates/api/src/error.rs

@@ -52,19 +52,13 @@ pub enum ApiError {
         body_snippet: String,
         source: serde_json::Error,
     },
-    Api {
-        status: reqwest::StatusCode,
-        error_type: Option<String>,
-        message: Option<String>,
-        request_id: Option<String>,
-        body: String,
-        retryable: bool,
-        /// Suggested user action based on error type (e.g., "Reduce prompt size" for 413)
-        suggested_action: Option<String>,
-        /// Parsed Retry-After header value (seconds) for 429 responses.
-        /// When present, overrides the exponential backoff delay.
-        retry_after: Option<Duration>,
-    },
+    /// A provider returned an error response.
+    ///
+    /// Boxed: the payload is ~139 bytes, and `ApiError` is the `Err` half of nearly every
+    /// `Result` in this crate, so inlining it inflates the size of every one of those returns
+    /// (clippy::result_large_err). Boxing keeps `ApiError` small and means adding a field here
+    /// cannot silently re-inflate every signature in the crate.
+    Api(Box<ApiErrorDetails>),
     RetriesExhausted {
         attempts: u32,
         last_error: Box<ApiError>,
@@ -81,7 +75,38 @@ pub enum ApiError {
     },
 }
 
+/// Payload of [`ApiError::Api`], held behind a `Box` so the enum stays small.
+#[derive(Debug)]
+pub struct ApiErrorDetails {
+    pub status: reqwest::StatusCode,
+    pub error_type: Option<String>,
+    pub message: Option<String>,
+    pub request_id: Option<String>,
+    pub body: String,
+    pub retryable: bool,
+    /// Suggested user action based on error type (e.g., "Reduce prompt size" for 413)
+    pub suggested_action: Option<String>,
+    /// Parsed Retry-After header value (seconds) for 429 responses.
+    /// When present, overrides the exponential backoff delay.
+    pub retry_after: Option<Duration>,
+}
+
 impl ApiError {
+    /// Convenience constructor so callers do not have to spell out the `Box` at every site.
+    #[must_use]
+    pub fn api(details: ApiErrorDetails) -> Self {
+        Self::Api(Box::new(details))
+    }
+
+    /// Borrow the provider-error payload, if this is one.
+    #[must_use]
+    pub fn api_details(&self) -> Option<&ApiErrorDetails> {
+        match self {
+            Self::Api(details) => Some(details),
+            _ => None,
+        }
+    }
+
     #[must_use]
     pub const fn missing_credentials(
         provider: &'static str,
@@ -137,7 +162,7 @@ impl ApiError {
     /// over the computed backoff delay when it exists.
     pub fn retry_after(&self) -> Option<Duration> {
         match self {
-            Self::Api { retry_after, .. } => *retry_after,
+            Self::Api(details) => details.retry_after,
             Self::RetriesExhausted { last_error, .. } => last_error.retry_after(),
             _ => None,
         }
@@ -146,7 +171,7 @@ impl ApiError {
     pub fn is_retryable(&self) -> bool {
         match self {
             Self::Http(error) => error.is_connect() || error.is_timeout() || error.is_request(),
-            Self::Api { retryable, .. } => *retryable,
+            Self::Api(details) => details.retryable,
             Self::RetriesExhausted { last_error, .. } => last_error.is_retryable(),
             Self::MissingCredentials { .. }
             | Self::ContextWindowExceeded { .. }
@@ -164,7 +189,7 @@ impl ApiError {
     #[must_use]
     pub fn request_id(&self) -> Option<&str> {
         match self {
-            Self::Api { request_id, .. } => request_id.as_deref(),
+            Self::Api(details) => details.request_id.as_deref(),
             Self::RetriesExhausted { last_error, .. } => last_error.request_id(),
             Self::MissingCredentials { .. }
             | Self::ContextWindowExceeded { .. }
@@ -191,12 +216,12 @@ impl ApiError {
             Self::MissingCredentials { .. } | Self::ExpiredOAuthToken | Self::Auth(_) => {
                 "provider_auth"
             }
-            Self::Api { status, .. } if matches!(status.as_u16(), 401 | 403) => "provider_auth",
+            Self::Api(details) if matches!(details.status.as_u16(), 401 | 403) => "provider_auth",
             Self::ContextWindowExceeded { .. } => "context_window",
-            Self::Api { .. } if self.is_context_window_failure() => "context_window",
-            Self::Api { status, .. } if status.as_u16() == 429 => "provider_rate_limit",
-            Self::Api { .. } if self.is_generic_fatal_wrapper() => "provider_internal",
-            Self::Api { .. } => "provider_error",
+            Self::Api(_) if self.is_context_window_failure() => "context_window",
+            Self::Api(details) if details.status.as_u16() == 429 => "provider_rate_limit",
+            Self::Api(_) if self.is_generic_fatal_wrapper() => "provider_internal",
+            Self::Api(_) => "provider_error",
             Self::Http(_) | Self::InvalidSseFrame(_) | Self::BackoffOverflow { .. } => {
                 "provider_transport"
             }
@@ -208,11 +233,12 @@ impl ApiError {
     #[must_use]
     pub fn is_generic_fatal_wrapper(&self) -> bool {
         match self {
-            Self::Api { message, body, .. } => {
-                message
+            Self::Api(details) => {
+                details
+                    .message
                     .as_deref()
                     .is_some_and(looks_like_generic_fatal_wrapper)
-                    || looks_like_generic_fatal_wrapper(body)
+                    || looks_like_generic_fatal_wrapper(&details.body)
             }
             Self::RetriesExhausted { last_error, .. } => last_error.is_generic_fatal_wrapper(),
             Self::MissingCredentials { .. }
@@ -229,21 +255,36 @@ impl ApiError {
         }
     }
 
+    /// True for failures whose own text cannot say what went wrong: transport, decode, and frame
+    /// errors.
+    ///
+    /// These matter because a server that answers an oversized request with a non-SSE HTTP 500
+    /// body lands here rather than in [`Self::is_context_window_failure`] — reqwest reports
+    /// "error decoding response body" and the real cause is lost. Callers that know how large the
+    /// request was can combine the two signals; callers that do not must treat this as unknown,
+    /// never as an overflow.
+    #[must_use]
+    pub fn is_ambiguous_transport_failure(&self) -> bool {
+        match self {
+            Self::Http(_) | Self::Json { .. } | Self::InvalidSseFrame(_) => true,
+            Self::RetriesExhausted { last_error, .. } => {
+                last_error.is_ambiguous_transport_failure()
+            }
+            _ => false,
+        }
+    }
+
     #[must_use]
     pub fn is_context_window_failure(&self) -> bool {
         match self {
             Self::ContextWindowExceeded { .. } => true,
-            Self::Api {
-                status,
-                message,
-                body,
-                ..
-            } => {
-                matches!(status.as_u16(), 400 | 413 | 422)
-                    && (message
+            Self::Api(details) => {
+                matches!(details.status.as_u16(), 400 | 413 | 422)
+                    && (details
+                        .message
                         .as_deref()
                         .is_some_and(looks_like_context_window_error)
-                        || looks_like_context_window_error(body))
+                        || looks_like_context_window_error(&details.body))
             }
             Self::RetriesExhausted { last_error, .. } => last_error.is_context_window_failure(),
             Self::MissingCredentials { .. }
@@ -327,14 +368,14 @@ impl Display for ApiError {
                 "failed to parse {provider} response for model {model}: {source}; first 200 chars of body: {body_snippet}"
             ),
             // #28: enhance 401/403 errors with actionable auth guidance
-            Self::Api {
-                status,
-                error_type,
-                message,
-                request_id,
-                body,
-                ..
-            } if matches!(status.as_u16(), 401 | 403) => {
+            Self::Api(details) if matches!(details.status.as_u16(), 401 | 403) => {
+                let (status, error_type, message, request_id, body) = (
+                    &details.status,
+                    &details.error_type,
+                    &details.message,
+                    &details.request_id,
+                    &details.body,
+                );
                 if let (Some(error_type), Some(message)) = (error_type, message) {
                     write!(f, "api returned {status} ({error_type})")?;
                     if let Some(request_id) = request_id {
@@ -356,14 +397,14 @@ impl Display for ApiError {
                      Run `claw doctor` to verify your credential configuration."
                 )
             }
-            Self::Api {
-                status,
-                error_type,
-                message,
-                request_id,
-                body,
-                ..
-            } => {
+            Self::Api(details) => {
+                let (status, error_type, message, request_id, body) = (
+                    &details.status,
+                    &details.error_type,
+                    &details.message,
+                    &details.request_id,
+                    &details.body,
+                );
                 if let (Some(error_type), Some(message)) = (error_type, message) {
                     write!(f, "api returned {status} ({error_type})")?;
                     if let Some(request_id) = request_id {
@@ -469,6 +510,7 @@ fn truncate_body_snippet(body: &str, max_chars: usize) -> String {
 
 #[cfg(test)]
 mod tests {
+    use super::ApiErrorDetails;
     use super::{truncate_body_snippet, ApiError};
 
     #[test]
@@ -533,7 +575,7 @@ mod tests {
 
     #[test]
     fn detects_generic_fatal_wrapper_and_classifies_it_as_provider_internal() {
-        let error = ApiError::Api {
+        let error = ApiError::Api(Box::new(ApiErrorDetails {
             status: reqwest::StatusCode::INTERNAL_SERVER_ERROR,
             error_type: Some("api_error".to_string()),
             message: Some(
@@ -545,7 +587,7 @@ mod tests {
             retryable: true,
             suggested_action: None,
         retry_after: None,
-        };
+        }));
 
         assert!(error.is_generic_fatal_wrapper());
         assert_eq!(error.safe_failure_class(), "provider_internal");
@@ -557,7 +599,7 @@ mod tests {
     fn retries_exhausted_preserves_nested_request_id_and_failure_class() {
         let error = ApiError::RetriesExhausted {
             attempts: 3,
-            last_error: Box::new(ApiError::Api {
+            last_error: Box::new(ApiError::Api(Box::new(ApiErrorDetails {
                 status: reqwest::StatusCode::BAD_GATEWAY,
                 error_type: Some("api_error".to_string()),
                 message: Some(
@@ -569,7 +611,7 @@ mod tests {
                 retryable: true,
                 suggested_action: None,
             retry_after: None,
-            }),
+            }))),
         };
 
         assert!(error.is_generic_fatal_wrapper());
@@ -579,7 +621,7 @@ mod tests {
 
     #[test]
     fn classifies_provider_context_window_errors() {
-        let error = ApiError::Api {
+        let error = ApiError::Api(Box::new(ApiErrorDetails {
             status: reqwest::StatusCode::BAD_REQUEST,
             error_type: Some("invalid_request_error".to_string()),
             message: Some(
@@ -591,7 +633,7 @@ mod tests {
             retryable: false,
             suggested_action: None,
         retry_after: None,
-        };
+        }));
 
         assert!(error.is_context_window_failure());
         assert_eq!(error.safe_failure_class(), "context_window");
@@ -600,7 +642,7 @@ mod tests {
 
     #[test]
     fn classifies_openai_configured_limit_errors_as_context_window_failures() {
-        let error = ApiError::Api {
+        let error = ApiError::Api(Box::new(ApiErrorDetails {
             status: reqwest::StatusCode::BAD_REQUEST,
             error_type: Some("invalid_request_error".to_string()),
             message: Some(
@@ -612,7 +654,7 @@ mod tests {
             retryable: false,
             suggested_action: None,
             retry_after: None,
-        };
+        }));
 
         assert!(error.is_context_window_failure());
         assert_eq!(error.safe_failure_class(), "context_window");

rust/crates/api/src/lib.rs

@@ -10,7 +10,7 @@ pub use client::{
     oauth_token_is_expired, read_base_url, read_xai_base_url, resolve_saved_oauth_token,
     resolve_startup_auth_source, MessageStream, OAuthTokenSet, ProviderClient,
 };
-pub use error::ApiError;
+pub use error::{ApiError, ApiErrorDetails};
 pub use http_client::{
     build_http_client, build_http_client_or_default, build_http_client_with,
     build_http_client_with_opts, ProxyConfig, TimeoutConfig,
@@ -27,9 +27,10 @@ pub use providers::openai_compat::{
     OpenAiCompatClient, OpenAiCompatConfig,
 };
 pub use providers::{
-    detect_provider_kind, max_tokens_for_model, max_tokens_for_model_with_override,
-    model_family_identity_for, model_family_identity_for_kind, provider_diagnostics_for_model,
-    resolve_model_alias, ProviderDiagnostics, ProviderKind,
+    context_window_override, detect_provider_kind, estimate_message_request_input_tokens,
+    max_tokens_for_model, max_tokens_for_model_with_override, model_family_identity_for,
+    model_family_identity_for_kind, model_token_limit, provider_diagnostics_for_model,
+    resolve_model_alias, ModelTokenLimit, ProviderDiagnostics, ProviderKind,
 };
 pub use sse::{parse_frame, SseParser};
 pub use types::{

rust/crates/api/src/providers/anthropic.rs

@@ -12,7 +12,7 @@ use serde::Deserialize;
 use serde_json::{Map, Value};
 use telemetry::{AnalyticsEvent, AnthropicRequestProfile, ClientIdentity, SessionTracer};
 
-use crate::error::ApiError;
+use crate::error::{ApiError, ApiErrorDetails};
 use crate::http_client::build_http_client_or_default;
 use crate::prompt_cache::{PromptCache, PromptCacheRecord, PromptCacheStats};
 
@@ -892,7 +892,7 @@ async fn expect_success(response: reqwest::Response) -> Result<reqwest::Response
     let retryable = is_retryable_status(status);
     let retry_after = parse_retry_after(&headers, status);
 
-    Err(ApiError::Api {
+    Err(ApiError::Api(Box::new(ApiErrorDetails {
         status,
         error_type: parsed_error
             .as_ref()
@@ -905,7 +905,7 @@ async fn expect_success(response: reqwest::Response) -> Result<reqwest::Response
         retryable,
         suggested_action: None,
         retry_after,
-    })
+    })))
 }
 
 fn parse_retry_after(
@@ -952,86 +952,30 @@ fn is_retryable_400(status: reqwest::StatusCode, body: &str) -> bool {
 const SK_ANT_BEARER_HINT: &str = "sk-ant-* keys go in ANTHROPIC_API_KEY (x-api-key header), not ANTHROPIC_AUTH_TOKEN (Bearer header). Move your key to ANTHROPIC_API_KEY.";
 
 fn enrich_bearer_auth_error(error: ApiError, auth: &AuthSource) -> ApiError {
-    let ApiError::Api {
-        status,
-        error_type,
-        message,
-        request_id,
-        body,
-        retryable,
-        suggested_action,
-        retry_after,
-        ..
-    } = error
-    else {
+    let ApiError::Api(mut details) = error else {
         return error;
     };
-    if status.as_u16() != 401 {
-        return ApiError::Api {
-            status,
-            error_type,
-            message,
-            request_id,
-            body,
-            retryable,
-            suggested_action,
-            retry_after,
-        };
+    if details.status.as_u16() != 401 {
+        return ApiError::Api(details);
     }
     let Some(bearer_token) = auth.bearer_token() else {
-        return ApiError::Api {
-            status,
-            error_type,
-            message,
-            request_id,
-            body,
-            retryable,
-            suggested_action,
-            retry_after,
-        };
+        return ApiError::Api(details);
     };
     if !bearer_token.starts_with("sk-ant-") {
-        return ApiError::Api {
-            status,
-            error_type,
-            message,
-            request_id,
-            body,
-            retryable,
-            suggested_action,
-            retry_after,
-        };
+        return ApiError::Api(details);
     }
     // Only append the hint when the AuthSource is pure BearerToken. If both
     // api_key and bearer_token are present (`ApiKeyAndBearer`), the x-api-key
     // header is already being sent alongside the Bearer header and the 401
     // is coming from a different cause — adding the hint would be misleading.
     if auth.api_key().is_some() {
-        return ApiError::Api {
-            status,
-            error_type,
-            message,
-            request_id,
-            body,
-            retryable,
-            suggested_action,
-            retry_after,
-        };
-    }
-    let enriched_message = match message {
-        Some(existing) => Some(format!("{existing} — hint: {SK_ANT_BEARER_HINT}")),
-        None => Some(format!("hint: {SK_ANT_BEARER_HINT}")),
-    };
-    ApiError::Api {
-        status,
-        error_type,
-        message: enriched_message,
-        request_id,
-        body,
-        retryable,
-        suggested_action,
-        retry_after,
+        return ApiError::Api(details);
     }
+    details.message = Some(match details.message.take() {
+        Some(existing) => format!("{existing} — hint: {SK_ANT_BEARER_HINT}"),
+        None => format!("hint: {SK_ANT_BEARER_HINT}"),
+    });
+    ApiError::Api(details)
 }
 
 fn anthropic_wire_model(model: &str) -> &str {
@@ -1083,6 +1027,7 @@ struct AnthropicErrorBody {
 #[cfg(test)]
 mod tests {
     use super::{ALT_REQUEST_ID_HEADER, REQUEST_ID_HEADER};
+    use crate::error::ApiErrorDetails;
     use std::io::{Read, Write};
     use std::net::TcpListener;
     use std::sync::{Mutex, OnceLock};
@@ -1647,7 +1592,7 @@ mod tests {
     fn enrich_bearer_auth_error_appends_sk_ant_hint_on_401_with_pure_bearer_token() {
         // given
         let auth = AuthSource::BearerToken("sk-ant-api03-deadbeef".to_string());
-        let error = crate::error::ApiError::Api {
+        let error = crate::error::ApiError::Api(Box::new(ApiErrorDetails {
             status: reqwest::StatusCode::UNAUTHORIZED,
             error_type: Some("authentication_error".to_string()),
             message: Some("Invalid bearer token".to_string()),
@@ -1656,7 +1601,7 @@ mod tests {
             retryable: false,
             suggested_action: None,
             retry_after: None,
-        };
+        }));
 
         // when
         let enriched = super::enrich_bearer_auth_error(error, &auth);
@@ -1678,8 +1623,8 @@ mod tests {
             "request id should still flow through the enriched error: {rendered}"
         );
         match enriched {
-            crate::error::ApiError::Api { status, .. } => {
-                assert_eq!(status, reqwest::StatusCode::UNAUTHORIZED);
+            crate::error::ApiError::Api(details) => {
+                assert_eq!(details.status, reqwest::StatusCode::UNAUTHORIZED);
             }
             other => panic!("expected Api variant, got {other:?}"),
         }
@@ -1689,7 +1634,7 @@ mod tests {
     fn enrich_bearer_auth_error_leaves_non_401_errors_unchanged() {
         // given
         let auth = AuthSource::BearerToken("sk-ant-api03-deadbeef".to_string());
-        let error = crate::error::ApiError::Api {
+        let error = crate::error::ApiError::Api(Box::new(ApiErrorDetails {
             status: reqwest::StatusCode::INTERNAL_SERVER_ERROR,
             error_type: Some("api_error".to_string()),
             message: Some("internal server error".to_string()),
@@ -1698,7 +1643,7 @@ mod tests {
             retryable: true,
             suggested_action: None,
             retry_after: None,
-        };
+        }));
 
         // when
         let enriched = super::enrich_bearer_auth_error(error, &auth);
@@ -1719,7 +1664,7 @@ mod tests {
     fn enrich_bearer_auth_error_ignores_401_when_bearer_token_is_not_sk_ant() {
         // given
         let auth = AuthSource::BearerToken("oauth-access-token-opaque".to_string());
-        let error = crate::error::ApiError::Api {
+        let error = crate::error::ApiError::Api(Box::new(ApiErrorDetails {
             status: reqwest::StatusCode::UNAUTHORIZED,
             error_type: Some("authentication_error".to_string()),
             message: Some("Invalid bearer token".to_string()),
@@ -1728,7 +1673,7 @@ mod tests {
             retryable: false,
             suggested_action: None,
             retry_after: None,
-        };
+        }));
 
         // when
         let enriched = super::enrich_bearer_auth_error(error, &auth);
@@ -1748,7 +1693,7 @@ mod tests {
             api_key: "sk-ant-api03-legitimate".to_string(),
             bearer_token: "sk-ant-api03-deadbeef".to_string(),
         };
-        let error = crate::error::ApiError::Api {
+        let error = crate::error::ApiError::Api(Box::new(ApiErrorDetails {
             status: reqwest::StatusCode::UNAUTHORIZED,
             error_type: Some("authentication_error".to_string()),
             message: Some("Invalid bearer token".to_string()),
@@ -1757,7 +1702,7 @@ mod tests {
             retryable: false,
             suggested_action: None,
             retry_after: None,
-        };
+        }));
 
         // when
         let enriched = super::enrich_bearer_auth_error(error, &auth);
@@ -1774,7 +1719,7 @@ mod tests {
     fn enrich_bearer_auth_error_ignores_401_when_auth_source_has_no_bearer() {
         // given
         let auth = AuthSource::ApiKey("sk-ant-api03-legitimate".to_string());
-        let error = crate::error::ApiError::Api {
+        let error = crate::error::ApiError::Api(Box::new(ApiErrorDetails {
             status: reqwest::StatusCode::UNAUTHORIZED,
             error_type: Some("authentication_error".to_string()),
             message: Some("Invalid x-api-key".to_string()),
@@ -1783,7 +1728,7 @@ mod tests {
             retryable: false,
             suggested_action: None,
             retry_after: None,
-        };
+        }));
 
         // when
         let enriched = super::enrich_bearer_auth_error(error, &auth);

rust/crates/api/src/providers/mod.rs

@@ -619,16 +619,43 @@ fn web_passthrough_diagnostic(
     }
 }
 
+/// Overrides the context window every sizing guard is computed against.
+///
+/// The model *name* is not a reliable statement of the window when the endpoint is a local
+/// server: the harness sends its compiled-in default name (a Claude alias) while the backend
+/// serves whatever `-c` it was started with. Every guard downstream of [`model_token_limit`]
+/// then sizes itself against the name's window instead of the real one, which makes the
+/// preflights unable to fire rather than merely inaccurate. This variable is the operator's
+/// statement of the truth, and it wins over the table.
+const CONTEXT_WINDOW_ENV_VAR: &str = "CLAW_CONTEXT_WINDOW";
+
+/// Fraction of the window a single response may be allowed to claim when the window is
+/// declared. Output and input share one budget on a local server, so reserving a quarter for
+/// the reply leaves three quarters for prompt plus transcript.
+const MAX_OUTPUT_SHARE_OF_WINDOW: u32 = 4;
+
+/// The operator-declared context window, when set and parseable as a positive token count.
+#[must_use]
+pub fn context_window_override() -> Option<u32> {
+    parse_context_window_override(std::env::var(CONTEXT_WINDOW_ENV_VAR).ok().as_deref())
+}
+
+#[must_use]
+fn parse_context_window_override(value: Option<&str>) -> Option<u32> {
+    value
+        .and_then(|raw| raw.trim().parse::<u32>().ok())
+        .filter(|window| *window > 0)
+}
+
 #[must_use]
 pub fn max_tokens_for_model(model: &str) -> u32 {
-    let canonical = resolve_model_alias(model);
-    let heuristic = if canonical.contains("opus") {
-        32_000
-    } else {
-        64_000
-    };
-
-    model_token_limit(model).map_or(heuristic, |limit| heuristic.min(limit.max_output_tokens))
+    // A declared window caps the reply too. Asking a 65k server for 32k of output reserves half
+    // the window for a single message, which the server then silently clamps — leaving the
+    // client's own accounting wrong by the difference.
+    model_token_limit(model).map_or_else(
+        || max_tokens_for_model_by_name(model),
+        |limit| max_tokens_for_model_by_name(model).min(limit.max_output_tokens),
+    )
 }
 
 /// Returns the effective max output tokens for a model, preferring a plugin
@@ -641,6 +668,33 @@ pub fn max_tokens_for_model_with_override(model: &str, plugin_override: Option<u
 
 #[must_use]
 pub fn model_token_limit(model: &str) -> Option<ModelTokenLimit> {
+    // A declared window replaces the table's, and supplies one for models the table has never
+    // heard of. Returning `Some` for an unknown model is the point: `None` makes both preflights
+    // return `Ok` without looking at anything, so an unrecognised local model gets no guard at all.
+    if let Some(context_window_tokens) = context_window_override() {
+        return Some(ModelTokenLimit {
+            max_output_tokens: max_tokens_for_model_by_name(model)
+                .min(context_window_tokens / MAX_OUTPUT_SHARE_OF_WINDOW),
+            context_window_tokens,
+        });
+    }
+    model_token_limit_by_name(model)
+}
+
+#[must_use]
+fn max_tokens_for_model_by_name(model: &str) -> u32 {
+    let canonical = resolve_model_alias(model);
+    let heuristic = if canonical.contains("opus") {
+        32_000
+    } else {
+        64_000
+    };
+    model_token_limit_by_name(model)
+        .map_or(heuristic, |limit| heuristic.min(limit.max_output_tokens))
+}
+
+#[must_use]
+fn model_token_limit_by_name(model: &str) -> Option<ModelTokenLimit> {
     let canonical = resolve_model_alias(model);
     let base_model = canonical.rsplit('/').next().unwrap_or(canonical.as_str());
     match base_model {
@@ -723,7 +777,7 @@ pub fn preflight_message_request(request: &MessageRequest) -> Result<(), ApiErro
     Ok(())
 }
 
-fn estimate_message_request_input_tokens(request: &MessageRequest) -> u32 {
+pub fn estimate_message_request_input_tokens(request: &MessageRequest) -> u32 {
     let mut estimate = estimate_serialized_tokens(&request.messages);
     estimate = estimate.saturating_add(estimate_serialized_tokens(&request.system));
     estimate = estimate.saturating_add(estimate_serialized_tokens(&request.tools));

rust/crates/api/src/providers/openai_compat.rs

@@ -7,7 +7,7 @@ use std::time::{Duration, SystemTime, UNIX_EPOCH};
 use serde::Deserialize;
 use serde_json::{json, Value};
 
-use crate::error::ApiError;
+use crate::error::{ApiError, ApiErrorDetails};
 use crate::http_client::build_http_client_or_default;
 use crate::types::{
     ContentBlockDelta, ContentBlockDeltaEvent, ContentBlockStartEvent, ContentBlockStopEvent,
@@ -239,7 +239,7 @@ impl OpenAiCompatClient {
                     .get("code")
                     .and_then(serde_json::Value::as_u64)
                     .map(|c| c as u16);
-                return Err(ApiError::Api {
+                return Err(ApiError::Api(Box::new(ApiErrorDetails {
                     status: reqwest::StatusCode::from_u16(code.unwrap_or(400))
                         .unwrap_or(reqwest::StatusCode::BAD_REQUEST),
                     error_type: err_obj
@@ -255,7 +255,7 @@ impl OpenAiCompatClient {
                             .unwrap_or(reqwest::StatusCode::BAD_REQUEST),
                     ),
                     retry_after: None,
-                });
+                })));
             }
         }
         let payload = serde_json::from_str::<ChatCompletionResponse>(&body).map_err(|error| {
@@ -1613,7 +1613,7 @@ fn parse_sse_frame(
                     .map(|c| c as u16);
                 let status = reqwest::StatusCode::from_u16(code.unwrap_or(500))
                     .unwrap_or(reqwest::StatusCode::INTERNAL_SERVER_ERROR);
-                return Err(ApiError::Api {
+                return Err(ApiError::Api(Box::new(ApiErrorDetails {
                     status,
                     error_type: err_obj
                         .get("type")
@@ -1625,12 +1625,12 @@ fn parse_sse_frame(
                     retryable: false,
                     suggested_action: suggested_action_for_status(status),
                     retry_after: None,
-                });
+                })));
             }
         }
         // Detect HTML responses
         if trimmed.starts_with('<') || trimmed.starts_with("<!") {
-            return Err(ApiError::Api {
+            return Err(ApiError::Api(Box::new(ApiErrorDetails {
                 status: reqwest::StatusCode::BAD_REQUEST,
                 error_type: Some("invalid_response".to_string()),
                 message: Some(
@@ -1641,7 +1641,7 @@ fn parse_sse_frame(
                 retryable: false,
                 suggested_action: Some("verify the API endpoint URL is correct".to_string()),
                 retry_after: None,
-            });
+            })));
         }
         return Ok(None);
     }
@@ -1665,7 +1665,7 @@ fn parse_sse_frame(
                 .map(|c| c as u16);
             let status = reqwest::StatusCode::from_u16(code.unwrap_or(400))
                 .unwrap_or(reqwest::StatusCode::BAD_REQUEST);
-            return Err(ApiError::Api {
+            return Err(ApiError::Api(Box::new(ApiErrorDetails {
                 status,
                 error_type: err_obj
                     .get("type")
@@ -1677,13 +1677,13 @@ fn parse_sse_frame(
                 retryable: false,
                 suggested_action: suggested_action_for_status(status),
                 retry_after: None,
-            });
+            })));
         }
     }
     // Detect HTML or other non-JSON responses early for better error messages
     let trimmed_payload = payload.trim();
     if trimmed_payload.starts_with('<') || trimmed_payload.starts_with("<!") {
-        return Err(ApiError::Api {
+        return Err(ApiError::Api(Box::new(ApiErrorDetails {
             status: reqwest::StatusCode::BAD_REQUEST,
             error_type: Some("invalid_response".to_string()),
             message: Some(
@@ -1694,7 +1694,7 @@ fn parse_sse_frame(
             retryable: false,
             suggested_action: Some("verify the API endpoint URL is correct".to_string()),
             retry_after: None,
-        });
+        })));
     }
     serde_json::from_str::<ChatCompletionChunk>(&payload)
         .map(Some)
@@ -1754,7 +1754,7 @@ async fn expect_success(response: reqwest::Response) -> Result<reqwest::Response
 
     let suggested_action = suggested_action_for_status(status);
 
-    Err(ApiError::Api {
+    Err(ApiError::Api(Box::new(ApiErrorDetails {
         status,
         error_type: parsed_error
             .as_ref()
@@ -1767,7 +1767,7 @@ async fn expect_success(response: reqwest::Response) -> Result<reqwest::Response
         retryable,
         suggested_action,
         retry_after,
-    })
+    })))
 }
 
 fn parse_retry_after(

rust/crates/api/tests/client_integration.rs

@@ -585,14 +585,11 @@ async fn surfaces_retry_exhaustion_for_persistent_retryable_errors() {
             last_error,
         } => {
             assert_eq!(attempts, 2);
-            assert!(matches!(
-                *last_error,
-                ApiError::Api {
-                    status: reqwest::StatusCode::SERVICE_UNAVAILABLE,
-                    retryable: true,
-                    ..
-                }
-            ));
+            let details = last_error
+                .api_details()
+                .expect("last error should be a provider Api error");
+            assert_eq!(details.status, reqwest::StatusCode::SERVICE_UNAVAILABLE);
+            assert!(details.retryable);
         }
         other => panic!("expected retries exhausted, got {other:?}"),
     }

rust/crates/claw-rag-service/src/ingest.rs

@@ -66,7 +66,7 @@ async fn flush_path_batch(
     #[cfg(feature = "qdrant-index")]
     let mut qdrant_points: Vec<ChunkPoint> = Vec::with_capacity(batch.len());
 
-    for ((ord, t), vec) in batch.drain(..).zip(vecs.into_iter()) {
+    for ((ord, t), vec) in batch.drain(..).zip(vecs) {
         let dim = vec.len();
         let cid = insert_chunk(conn, path, ord, &t)?;
         insert_embedding(conn, cid, dim, &vec)?;

rust/crates/mock-anthropic-service/src/lib.rs

@@ -13,6 +13,16 @@ use tokio::task::JoinHandle;
 pub const SCENARIO_PREFIX: &str = "PARITY_SCENARIO:";
 pub const DEFAULT_MODEL: &str = "claude-sonnet-4-6";
 
+/// The assistant text the `context_window_retry` scenario finally streams, after
+/// its first request was refused with a context-overflow error. Exposed so a
+/// test can assert on the exact body instead of restating the literal.
+pub const CONTEXT_WINDOW_RETRY_TEXT: &str = "auto compact retry parity complete.";
+
+/// The partial assistant text the `stream_then_error` scenario streams before it
+/// fails the response mid-stream. Exposed so a test can assert on the exact body
+/// instead of restating the literal.
+pub const STREAM_THEN_ERROR_TEXT: &str = "partial response before the stream failed";
+
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct CapturedRequest {
     pub method: String,
@@ -99,6 +109,8 @@ enum Scenario {
     BashPermissionPromptDenied,
     PluginToolRoundtrip,
     AutoCompactTriggered,
+    ContextWindowRetry,
+    StreamThenError,
     TokenCostReporting,
 }
 
@@ -116,6 +128,8 @@ impl Scenario {
             "bash_permission_prompt_denied" => Some(Self::BashPermissionPromptDenied),
             "plugin_tool_roundtrip" => Some(Self::PluginToolRoundtrip),
             "auto_compact_triggered" => Some(Self::AutoCompactTriggered),
+            "context_window_retry" => Some(Self::ContextWindowRetry),
+            "stream_then_error" => Some(Self::StreamThenError),
             "token_cost_reporting" => Some(Self::TokenCostReporting),
             _ => None,
         }
@@ -134,6 +148,8 @@ impl Scenario {
             Self::BashPermissionPromptDenied => "bash_permission_prompt_denied",
             Self::PluginToolRoundtrip => "plugin_tool_roundtrip",
             Self::AutoCompactTriggered => "auto_compact_triggered",
+            Self::ContextWindowRetry => "context_window_retry",
+            Self::StreamThenError => "stream_then_error",
             Self::TokenCostReporting => "token_cost_reporting",
         }
     }
@@ -149,6 +165,18 @@ async fn handle_connection(
     let scenario = detect_scenario(&request)
         .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidInput, "missing parity scenario"))?;
 
+    // How many times this scenario has already been served *on this endpoint*.
+    // Scenarios that need to behave differently on a retry (reject once, then
+    // succeed) key off this; counting per endpoint keeps the token-counting
+    // preflight on `/v1/messages/count_tokens` from consuming the first attempt.
+    let attempt = {
+        let served = requests.lock().await;
+        served
+            .iter()
+            .filter(|captured| captured.scenario == scenario.name() && captured.path == path)
+            .count()
+    };
+
     requests.lock().await.push(CapturedRequest {
         method,
         path,
@@ -158,7 +186,7 @@ async fn handle_connection(
         raw_body,
     });
 
-    let response = build_http_response(&request, scenario);
+    let response = build_http_response(&request, scenario, attempt);
     socket.write_all(response.as_bytes()).await?;
     Ok(())
 }
@@ -308,7 +336,11 @@ fn flatten_tool_result_content(content: &[api::ToolResultContentBlock]) -> Strin
 }
 
 #[allow(clippy::too_many_lines)]
-fn build_http_response(request: &MessageRequest, scenario: Scenario) -> String {
+fn build_http_response(request: &MessageRequest, scenario: Scenario, attempt: usize) -> String {
+    if let Some(rejection) = context_window_rejection(scenario, attempt) {
+        return rejection;
+    }
+
     let response = if request.stream {
         let body = build_stream_body(request, scenario);
         return http_response(
@@ -329,6 +361,37 @@ fn build_http_response(request: &MessageRequest, scenario: Scenario) -> String {
     )
 }
 
+/// The first request of [`Scenario::ContextWindowRetry`] is refused with a
+/// context-overflow error instead of a streamed answer, so the client's
+/// auto-compact recovery path runs and the *retry* is what produces the
+/// response text. Every later request against the same endpoint succeeds.
+///
+/// HTTP 400 is deliberately chosen: it is not in the client's retryable status
+/// set, so the client surfaces the error rather than silently re-sending it —
+/// which is exactly the situation a real context-overflow rejection creates.
+/// The body carries the `context_window` marker the client greps for.
+fn context_window_rejection(scenario: Scenario, attempt: usize) -> Option<String> {
+    if scenario != Scenario::ContextWindowRetry || attempt > 0 {
+        return None;
+    }
+
+    let body = json!({
+        "type": "error",
+        "error": {
+            "type": "invalid_request_error",
+            "message": "prompt is too long: context_window exceeded for this model",
+        },
+    })
+    .to_string();
+
+    Some(http_response(
+        "400 Bad Request",
+        "application/json",
+        &body,
+        &[("request-id", request_id_for(scenario))],
+    ))
+}
+
 #[allow(clippy::too_many_lines)]
 fn build_stream_body(request: &MessageRequest, scenario: Scenario) -> String {
     match scenario {
@@ -461,6 +524,8 @@ fn build_stream_body(request: &MessageRequest, scenario: Scenario) -> String {
         Scenario::AutoCompactTriggered => {
             final_text_sse_with_usage("auto compact parity complete.", 50_000, 200)
         }
+        Scenario::ContextWindowRetry => final_text_sse(CONTEXT_WINDOW_RETRY_TEXT),
+        Scenario::StreamThenError => stream_then_error_sse(STREAM_THEN_ERROR_TEXT),
         Scenario::TokenCostReporting => {
             final_text_sse_with_usage("token cost reporting parity complete.", 1_000, 500)
         }
@@ -628,6 +693,12 @@ fn build_message_response(request: &MessageRequest, scenario: Scenario) -> Messa
             50_000,
             200,
         ),
+        Scenario::ContextWindowRetry => {
+            text_message_response("msg_context_window_retry", CONTEXT_WINDOW_RETRY_TEXT)
+        }
+        Scenario::StreamThenError => {
+            text_message_response("msg_stream_then_error", STREAM_THEN_ERROR_TEXT)
+        }
         Scenario::TokenCostReporting => text_message_response_with_usage(
             "msg_token_cost_reporting",
             "token cost reporting parity complete.",
@@ -650,6 +721,8 @@ fn request_id_for(scenario: Scenario) -> &'static str {
         Scenario::BashPermissionPromptDenied => "req_bash_permission_prompt_denied",
         Scenario::PluginToolRoundtrip => "req_plugin_tool_roundtrip",
         Scenario::AutoCompactTriggered => "req_auto_compact_triggered",
+        Scenario::ContextWindowRetry => "req_context_window_retry",
+        Scenario::StreamThenError => "req_stream_then_error",
         Scenario::TokenCostReporting => "req_token_cost_reporting",
     }
 }
@@ -971,6 +1044,84 @@ fn final_text_sse(text: &str) -> String {
     body
 }
 
+/// Streams a text block to completion and then fails the response, which is
+/// what a provider does when it dies partway through an answer.
+///
+/// The block is closed before the failure on purpose: the client only flushes
+/// its buffered markdown to the terminal on `content_block_stop`, so without
+/// that frame the partial text would still be sitting in the renderer and the
+/// scenario would not reproduce a turn that streamed text and *then* errored.
+///
+/// The response is cut off in the middle of the failing frame rather than
+/// carrying a complete one. Two things fall out of that, both deliberate.
+///
+/// It is what a provider that dies on the socket actually looks like, and it
+/// keeps the failure in a *later* parser call than the answer text. The whole
+/// response is one `write_all` on the mock's side, so a complete trailing frame
+/// would land in the same read chunk as the deltas — and `SseParser::push`
+/// returns `Err` from inside its frame loop, which discards the events it had
+/// already parsed from that same chunk. The client would then never render the
+/// text at all, and the scenario would be exercising that parser behaviour
+/// instead of the terminal state it is meant to reproduce. A truncated tail
+/// stays in the parser's buffer, so `push` returns the deltas normally and the
+/// failure only surfaces from `finish()`, once the client has consumed every
+/// event ahead of it.
+///
+/// The failure is a frame that does not deserialize — `api::types::StreamEvent`
+/// is `#[serde(tag = "type")]` with no fallback arm, so an unknown frame type or
+/// a truncated payload is an `ApiError::Json`. That is non-retryable, and it is
+/// raised while reading the body rather than while acquiring the response, so the
+/// client reports it on the first attempt rather than retrying.
+fn stream_then_error_sse(text: &str) -> String {
+    let mut body = String::new();
+    append_sse(
+        &mut body,
+        "message_start",
+        json!({
+            "type": "message_start",
+            "message": {
+                "id": unique_message_id(),
+                "type": "message",
+                "role": "assistant",
+                "content": [],
+                "model": DEFAULT_MODEL,
+                "stop_reason": null,
+                "stop_sequence": null,
+                "usage": usage_json(14, 0)
+            }
+        }),
+    );
+    append_sse(
+        &mut body,
+        "content_block_start",
+        json!({
+            "type": "content_block_start",
+            "index": 0,
+            "content_block": {"type": "text", "text": ""}
+        }),
+    );
+    append_sse(
+        &mut body,
+        "content_block_delta",
+        json!({
+            "type": "content_block_delta",
+            "index": 0,
+            "delta": {"type": "text_delta", "text": text}
+        }),
+    );
+    append_sse(
+        &mut body,
+        "content_block_stop",
+        json!({"type": "content_block_stop", "index": 0}),
+    );
+    // No terminating blank line: the frame is still arriving when the provider
+    // drops the connection, so it stays in the client's parser buffer until the
+    // body ends. See the note above for why the failure has to trail the deltas
+    // by a parser call rather than sharing their chunk.
+    body.push_str("event: error\ndata: {\"type\":\"error\",\"error\":{\"type\":\"overloaded_");
+    body
+}
+
 fn final_text_sse_with_usage(text: &str, input_tokens: u32, output_tokens: u32) -> String {
     let mut body = String::new();
     append_sse(

rust/crates/runtime/src/compact.rs

@@ -95,10 +95,21 @@ pub fn get_compact_continuation_message(
 #[must_use]
 pub fn compact_session(session: &Session, config: CompactionConfig) -> CompactionResult {
     if !should_compact(session, config) {
+        // "Nothing to drop" is not the same as "nothing to do". A session can exceed the window
+        // while holding fewer messages than `preserve_recent_messages` — a single read of a large
+        // file does it — and message-granular compaction has no move to make there. Trim the
+        // payloads anyway, so the one lever that still works is not gated behind the one that
+        // does not.
+        let mut trimmed_session = session.clone();
+        trimmed_session.messages = session
+            .messages
+            .iter()
+            .map(truncate_oversized_tool_results)
+            .collect();
         return CompactionResult {
             summary: String::new(),
             formatted_summary: String::new(),
-            compacted_session: session.clone(),
+            compacted_session: trimmed_session,
             removed_message_count: 0,
         };
     }
@@ -165,7 +176,14 @@ pub fn compact_session(session: &Session, config: CompactionConfig) -> Compactio
         k
     };
     let removed = &session.messages[compacted_prefix_len..keep_from];
-    let preserved = session.messages[keep_from..].to_vec();
+    // Dropping whole messages is not enough on its own. An agentic turn is five messages
+    // (user / assistant+tool_use / tool_result / assistant+tool_use / tool_result), so the
+    // preserved tail IS the pair of tool results that overflowed the window, and the only
+    // message compaction may drop is the short user prompt. Trim the payloads that must stay.
+    let preserved = session.messages[keep_from..]
+        .iter()
+        .map(truncate_oversized_tool_results)
+        .collect::<Vec<_>>();
     let summary =
         merge_compact_summaries(existing_summary.as_deref(), &summarize_messages(removed));
     let formatted_summary = format_compact_summary(&summary);
@@ -190,6 +208,78 @@ pub fn compact_session(session: &Session, config: CompactionConfig) -> Compactio
     }
 }
 
+/// Largest tool-result payload compaction will preserve verbatim, in characters.
+///
+/// ~16 KB is ~4 k tokens: big enough that ordinary tool output is never touched, small enough
+/// that several preserved results still fit a small window.
+const MAX_PRESERVED_TOOL_RESULT_CHARS: usize = 16 * 1024;
+
+/// Characters kept from the tail of a trimmed payload. The head carries the structure a model
+/// needs to recognise what it read; the tail is where a file's most recent edits usually are.
+const PRESERVED_TOOL_RESULT_TAIL_CHARS: usize = 2 * 1024;
+
+/// Returns `message` with any oversized `ToolResult` payload replaced by a head/tail excerpt.
+///
+/// Clones only when something is actually trimmed, so the common path stays cheap.
+fn truncate_oversized_tool_results(message: &ConversationMessage) -> ConversationMessage {
+    let needs_trim = message.blocks.iter().any(|block| match block {
+        ContentBlock::ToolResult { output, .. } => {
+            output.chars().count() > MAX_PRESERVED_TOOL_RESULT_CHARS
+        }
+        _ => false,
+    });
+    if !needs_trim {
+        return message.clone();
+    }
+
+    let blocks = message
+        .blocks
+        .iter()
+        .map(|block| match block {
+            ContentBlock::ToolResult {
+                tool_use_id,
+                tool_name,
+                output,
+                is_error,
+            } => ContentBlock::ToolResult {
+                tool_use_id: tool_use_id.clone(),
+                tool_name: tool_name.clone(),
+                output: excerpt_tool_result(output),
+                is_error: *is_error,
+            },
+            other => other.clone(),
+        })
+        .collect();
+
+    ConversationMessage {
+        role: message.role,
+        blocks,
+        usage: message.usage,
+    }
+}
+
+/// Builds a head + marker + tail excerpt of an oversized tool-result payload.
+///
+/// Operates on chars, not bytes, so a multi-byte boundary can never be split.
+fn excerpt_tool_result(output: &str) -> String {
+    let chars: Vec<char> = output.chars().collect();
+    let total = chars.len();
+    if total <= MAX_PRESERVED_TOOL_RESULT_CHARS {
+        return output.to_string();
+    }
+
+    let tail_len = PRESERVED_TOOL_RESULT_TAIL_CHARS.min(total);
+    let head_len = MAX_PRESERVED_TOOL_RESULT_CHARS.saturating_sub(tail_len);
+    let elided = total.saturating_sub(head_len + tail_len);
+    if elided == 0 {
+        return output.to_string();
+    }
+
+    let head: String = chars[..head_len].iter().collect();
+    let tail: String = chars[total - tail_len..].iter().collect();
+    format!("{head}\n\n[... {elided} characters elided by compaction; re-read this file if you need the omitted portion ...]\n\n{tail}")
+}
+
 fn compacted_summary_prefix_len(session: &Session) -> usize {
     usize::from(
         session
@@ -573,7 +663,7 @@ fn extract_summary_timeline(summary: &str) -> Vec<String> {
 #[cfg(test)]
 mod tests {
     use super::{
-        collect_key_files, compact_session, format_compact_summary,
+        collect_key_files, compact_session, estimate_session_tokens, format_compact_summary,
         get_compact_continuation_message, infer_pending_work, should_compact, CompactionConfig,
     };
     use crate::session::{ContentBlock, ConversationMessage, MessageRole, Session};
@@ -763,6 +853,185 @@ mod tests {
     /// user(ToolResult) pair at the boundary. An orphaned tool-result message
     /// without the preceding assistant `tool_calls` causes a 400 on the
     /// OpenAI-compat path (gaebal-gajae repro 2026-04-09).
+    /// The failure this exists for: an agentic turn is 5 messages (user / assistant+tool_use /
+    /// tool_result / assistant+tool_use / tool_result), so `preserve_recent_messages: 4` protects
+    /// exactly the two giant tool results that overflow the window. Message-granular compaction
+    /// can only drop the tiny user prompt, so the payloads themselves must be trimmed.
+    #[test]
+    fn compaction_truncates_oversized_tool_results_it_must_preserve() {
+        let huge = "z".repeat(80_000);
+        let mut session = Session::new();
+        session
+            .push_message(ConversationMessage::user_text("read two files"))
+            .expect("user message");
+        session
+            .push_message(ConversationMessage::assistant(vec![
+                ContentBlock::ToolUse {
+                    id: "call-1".to_string(),
+                    name: "read_file".to_string(),
+                    input: "{\"path\":\"a.rs\"}".to_string(),
+                },
+            ]))
+            .expect("assistant tool use");
+        session
+            .push_message(ConversationMessage::tool_result(
+                "call-1",
+                "read_file",
+                huge.clone(),
+                false,
+            ))
+            .expect("tool result");
+        session
+            .push_message(ConversationMessage::assistant(vec![
+                ContentBlock::ToolUse {
+                    id: "call-2".to_string(),
+                    name: "read_file".to_string(),
+                    input: "{\"path\":\"b.rs\"}".to_string(),
+                },
+            ]))
+            .expect("second assistant tool use");
+        session
+            .push_message(ConversationMessage::tool_result(
+                "call-2",
+                "read_file",
+                huge.clone(),
+                false,
+            ))
+            .expect("second tool result");
+
+        let before = estimate_session_tokens(&session);
+        let result = compact_session(
+            &session,
+            CompactionConfig {
+                preserve_recent_messages: 4,
+                max_estimated_tokens: 0,
+            },
+        );
+        let after = estimate_session_tokens(&result.compacted_session);
+
+        assert!(
+            after < before / 2,
+            "compaction must shrink an oversized preserved tool result: {before} -> {after}"
+        );
+
+        let preserved_output = result
+            .compacted_session
+            .messages
+            .iter()
+            .flat_map(|message| &message.blocks)
+            .find_map(|block| match block {
+                ContentBlock::ToolResult { output, .. } => Some(output.clone()),
+                _ => None,
+            })
+            .expect("the tool result must still be present, not deleted");
+
+        assert!(
+            preserved_output.len() < huge.len(),
+            "the payload must actually be trimmed"
+        );
+        assert!(
+            preserved_output.contains("elided"),
+            "the model must be told the view is partial, got: {}",
+            &preserved_output[..preserved_output.len().min(200)]
+        );
+        assert!(
+            preserved_output.starts_with('z'),
+            "the head of the output must be kept"
+        );
+        assert!(
+            preserved_output.ends_with('z'),
+            "the tail of the output must be kept"
+        );
+    }
+
+    /// A session can exceed the window while holding FEWER messages than
+    /// `preserve_recent_messages` — one read of a large file does it. `should_compact` returns
+    /// false there (nothing is droppable), so trimming must not be gated behind it.
+    #[test]
+    fn oversized_tool_results_are_trimmed_even_when_no_message_can_be_dropped() {
+        let huge = "q".repeat(80_000);
+        let mut session = Session::new();
+        session
+            .push_message(ConversationMessage::user_text("read one enormous file"))
+            .expect("user message");
+        session
+            .push_message(ConversationMessage::assistant(vec![
+                ContentBlock::ToolUse {
+                    id: "call-1".to_string(),
+                    name: "read_file".to_string(),
+                    input: "{\"path\":\"huge.rs\"}".to_string(),
+                },
+            ]))
+            .expect("assistant tool use");
+        session
+            .push_message(ConversationMessage::tool_result(
+                "call-1",
+                "read_file",
+                huge,
+                false,
+            ))
+            .expect("tool result");
+
+        // 3 messages, preserve 4: there is nothing to remove, yet the session is far too big.
+        let before = estimate_session_tokens(&session);
+        let result = compact_session(
+            &session,
+            CompactionConfig {
+                preserve_recent_messages: 4,
+                max_estimated_tokens: 0,
+            },
+        );
+        let after = estimate_session_tokens(&result.compacted_session);
+
+        assert!(
+            after < before / 2,
+            "an untrimmable session must still have its payloads trimmed: {before} -> {after}"
+        );
+    }
+
+    /// Trimming must not touch payloads that already fit — otherwise every ordinary tool result
+    /// would grow an elision marker it does not need.
+    #[test]
+    fn compaction_leaves_small_tool_results_untouched() {
+        let small = "ok".repeat(50);
+        let mut session = Session::new();
+        for index in 0..6 {
+            session
+                .push_message(ConversationMessage::user_text(format!("turn {index}")))
+                .expect("user message");
+            session
+                .push_message(ConversationMessage::tool_result(
+                    format!("call-{index}"),
+                    "read_file",
+                    small.clone(),
+                    false,
+                ))
+                .expect("tool result");
+        }
+
+        let result = compact_session(
+            &session,
+            CompactionConfig {
+                preserve_recent_messages: 4,
+                max_estimated_tokens: 0,
+            },
+        );
+
+        for block in result
+            .compacted_session
+            .messages
+            .iter()
+            .flat_map(|message| &message.blocks)
+        {
+            if let ContentBlock::ToolResult { output, .. } = block {
+                assert_eq!(
+                    output, &small,
+                    "a small tool result must pass through as-is"
+                );
+            }
+        }
+    }
+
     #[test]
     fn compaction_does_not_split_tool_use_tool_result_pair() {
         use crate::session::{ContentBlock, Session};

rust/crates/runtime/src/conversation.rs

@@ -18,6 +18,47 @@ use crate::usage::{TokenUsage, UsageTracker};
 const DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD: u32 = 100_000;
 const AUTO_COMPACTION_THRESHOLD_ENV_VAR: &str = "CLAUDE_CODE_AUTO_COMPACT_INPUT_TOKENS";
 
+const DEFAULT_MAX_STALL_NUDGES: usize = 3;
+const MAX_STALL_NUDGES_ENV_VAR: &str = "CLAW_MAX_STALL_NUDGES";
+/// Placeholder recorded in the trace when a turn is resumed rather than started, so the trace
+/// does not attribute a second user message to a turn that only had one.
+const RESUMED_TURN_MARKER: &str = "[resumed turn]";
+/// Operator-declared context window. Kept as a literal rather than imported because `runtime`
+/// sits below `api` in the dependency graph; `api::context_window_override` reads the same name.
+const CONTEXT_WINDOW_ENV_VAR: &str = "CLAW_CONTEXT_WINDOW";
+/// Share of a declared context window at which compaction is triggered, leaving the remainder
+/// as headroom for the reply plus whatever the next tool result carries.
+const AUTO_COMPACTION_WINDOW_PERCENT: u32 = 70;
+
+/// Injected when the model narrates its next step (or asks for a go-ahead mid-task)
+/// instead of emitting the tool call that would perform it.
+const STALL_NUDGE_PROMPT: &str = "[auto-continue] You ended the turn without calling a tool, \
+     but the work is not finished. Do not describe the next step and do not ask for \
+     confirmation — issue the tool call now and keep going until the task is done. If \
+     everything asked for is genuinely complete, reply with a one-line summary containing \
+     no forward-looking sentence.";
+
+/// Phrases that mark text as an announcement of work the model has not performed yet.
+/// Matched against the final non-empty line of the assistant's reply.
+const PENDING_ACTION_MARKERS: [&str; 16] = [
+    "let me ",
+    "let's ",
+    "lets ",
+    "i'll ",
+    "i will ",
+    "i'm going to",
+    "i am going to",
+    "going to run",
+    "about to ",
+    "now running",
+    "now checking",
+    "now i",
+    "next i",
+    "next, i",
+    "proceeding to",
+    "moving on to",
+];
+
 /// Fully assembled request payload sent to the upstream model client.
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct ApiRequest {
@@ -90,6 +131,7 @@ impl std::error::Error for ToolError {}
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct RuntimeError {
     message: String,
+    context_window: bool,
 }
 
 impl RuntimeError {
@@ -97,8 +139,28 @@ impl RuntimeError {
     pub fn new(message: impl Into<String>) -> Self {
         Self {
             message: message.into(),
+            context_window: false,
         }
     }
+
+    /// A failure the caller may be able to clear by shrinking the session.
+    ///
+    /// This is a decision the API layer makes and hands up, not one the caller reconstructs by
+    /// matching substrings against the rendered message. Recovering from context exhaustion means
+    /// discarding conversation history, so a transport hiccup misread as an overflow destroys the
+    /// session for no reason — the classification has to come from whoever saw the real error.
+    #[must_use]
+    pub fn context_window(message: impl Into<String>) -> Self {
+        Self {
+            message: message.into(),
+            context_window: true,
+        }
+    }
+
+    #[must_use]
+    pub const fn is_context_window_failure(&self) -> bool {
+        self.context_window
+    }
 }
 
 impl Display for RuntimeError {
@@ -140,6 +202,7 @@ pub struct ConversationRuntime<C, T> {
     hook_abort_signal: HookAbortSignal,
     hook_progress_reporter: Option<Box<dyn HookProgressReporter>>,
     session_tracer: Option<SessionTracer>,
+    max_stall_nudges: usize,
 }
 
 impl<C, T> ConversationRuntime<C, T>
@@ -194,9 +257,17 @@ where
             hook_abort_signal: HookAbortSignal::default(),
             hook_progress_reporter: None,
             session_tracer: None,
+            max_stall_nudges: max_stall_nudges_from_env(),
         }
     }
 
+    /// Override how many auto-continue nudges a stalled turn may receive. `0` disables them.
+    #[must_use]
+    pub fn with_max_stall_nudges(mut self, max_stall_nudges: usize) -> Self {
+        self.max_stall_nudges = max_stall_nudges;
+        self
+    }
+
     #[must_use]
     pub fn with_max_iterations(mut self, max_iterations: usize) -> Self {
         self.max_iterations = max_iterations;
@@ -326,14 +397,33 @@ where
         }
     }
 
-    #[allow(clippy::too_many_lines)]
     pub fn run_turn(
         &mut self,
         user_input: impl Into<String>,
+        prompter: Option<&mut dyn PermissionPrompter>,
+    ) -> Result<TurnSummary, RuntimeError> {
+        self.drive_turn(Some(user_input.into()), prompter)
+    }
+
+    /// Continues the turn already recorded in the session, without adding a user message.
+    ///
+    /// Recovery paths retry a turn that failed part-way through. Calling [`Self::run_turn`] again
+    /// with the original input would append that input a second time and invite the model to redo
+    /// tool calls whose side effects have already landed. Resuming reads the transcript as it
+    /// stands — including the tool results that did complete — and carries on from there.
+    pub fn resume_turn(
+        &mut self,
+        prompter: Option<&mut dyn PermissionPrompter>,
+    ) -> Result<TurnSummary, RuntimeError> {
+        self.drive_turn(None, prompter)
+    }
+
+    #[allow(clippy::too_many_lines)]
+    fn drive_turn(
+        &mut self,
+        user_input: Option<String>,
         mut prompter: Option<&mut dyn PermissionPrompter>,
     ) -> Result<TurnSummary, RuntimeError> {
-        let user_input = user_input.into();
-
         // ROADMAP #38: Session-health canary - probe if context was compacted
         if self.session.compaction.is_some() {
             if let Err(error) = self.run_session_health_probe() {
@@ -345,16 +435,24 @@ where
             }
         }
 
-        self.record_turn_started(&user_input);
-        self.session
-            .push_user_text(user_input)
-            .map_err(|error| RuntimeError::new(error.to_string()))?;
+        match user_input {
+            Some(user_input) => {
+                self.record_turn_started(&user_input);
+                self.session
+                    .push_user_text(user_input)
+                    .map_err(|error| RuntimeError::new(error.to_string()))?;
+            }
+            None => self.record_turn_started(RESUMED_TURN_MARKER),
+        }
 
         let mut assistant_messages = Vec::new();
         let mut tool_results = Vec::new();
         let mut prompt_cache_events = Vec::new();
         let mut iterations = 0;
         let mut auto_compaction = None;
+        let max_stall_nudges = self.max_stall_nudges;
+        let mut stall_nudges = 0;
+        let mut tools_ran = false;
 
         loop {
             iterations += 1;
@@ -366,6 +464,23 @@ where
                 return Err(error);
             }
 
+            // Compact BEFORE the request goes out, not only after a reply lands.
+            //
+            // The usage-based check below is post-hoc: it can only see the context once a call
+            // has returned and reported its counters. Between that measurement and the next
+            // request the session can grow enormously — a reasoning model emitting a long
+            // <think> block, then a tool result carrying a large file. Observed live: a call
+            // measured at 17,858 prompt tokens generated 10,842 more, a read_file added ~19k on
+            // top, and the next request hit the provider at 47,755 tokens and was rejected
+            // outright. No amount of tuning the post-hoc threshold prevents that, because
+            // nothing measures the session in between.
+            //
+            // `estimate_session_tokens` is a local heuristic over the transcript, so this costs
+            // no API call and works before any usage has ever been recorded.
+            if let Some(compaction) = self.maybe_auto_compact_before_request() {
+                auto_compaction = Some(compaction);
+            }
+
             let request = ApiRequest {
                 system_prompt: self.system_prompt.clone(),
                 messages: self.session.messages.clone(),
@@ -381,6 +496,7 @@ where
                 match build_assistant_message(events) {
                     Ok(result) => result,
                     Err(error) => {
+                        let error = self.classify_empty_reply(error);
                         self.record_turn_failed(iterations, &error);
                         return Err(error);
                     }
@@ -417,8 +533,35 @@ where
             }
 
             if pending_tool_uses.is_empty() {
+                // A reply with no tool call usually means the turn is done. On small local
+                // models it just as often means the model narrated the next step — or asked
+                // for a go-ahead mid-task — and then emitted EOS, leaving the announced work
+                // undone. Nudge it back into the loop instead of dropping the user at the
+                // prompt, bounded so a model that insists it is finished still gets to stop.
+                let stalled = stall_nudges < max_stall_nudges
+                    && is_stalled_reply(
+                        &assistant_text(
+                            assistant_messages
+                                .last()
+                                .expect("assistant message pushed above"),
+                        ),
+                        tools_ran,
+                    );
+                if stalled {
+                    stall_nudges += 1;
+                    // Pushed as a system turn, not a user one. The model needs to see it, but the
+                    // transcript must not claim the user asked for it: this text is persisted,
+                    // re-sent on every later request, and folded into the next compaction summary,
+                    // where a user-labelled nudge becomes indistinguishable from something the
+                    // user actually said.
+                    self.session
+                        .push_message(ConversationMessage::system_text(STALL_NUDGE_PROMPT))
+                        .map_err(|error| RuntimeError::new(error.to_string()))?;
+                    continue;
+                }
                 break;
             }
+            tools_ran = true;
 
             for (tool_use_id, tool_name, input) in pending_tool_uses {
                 let pre_hook_result = self.run_pre_tool_use_hook(&tool_name, &input);
@@ -573,13 +716,62 @@ where
         self.session
     }
 
+    /// Compaction check run immediately before a request is sent, based on a locally estimated
+    /// transcript size rather than on provider-reported usage.
+    ///
+    /// This is the guard that actually keeps a session inside the context window: it sees growth
+    /// that usage counters cannot, because it runs after tool results and long generations have
+    /// been appended but before the request that would carry them is built.
+    /// Decides whether an empty or unreadable reply was an overflow, by size rather than by text.
+    ///
+    /// A stream that produces no content says nothing about why. Some backends answer an
+    /// oversized request that way instead of with a typed error, so the failure is worth
+    /// recovering from — but only when the session was already large enough for that to be the
+    /// plausible cause. Below the compaction threshold it is a blip, and discarding the
+    /// transcript would cost more than the retry could win.
+    fn classify_empty_reply(&self, error: RuntimeError) -> RuntimeError {
+        if error.is_context_window_failure() {
+            return error;
+        }
+        if estimate_session_tokens(&self.session)
+            >= self.auto_compaction_input_tokens_threshold as usize
+        {
+            return RuntimeError::context_window(error.to_string());
+        }
+        error
+    }
+
+    fn maybe_auto_compact_before_request(&mut self) -> Option<AutoCompactionEvent> {
+        let estimated = estimate_session_tokens(&self.session);
+        if estimated < self.auto_compaction_input_tokens_threshold as usize {
+            return None;
+        }
+        self.compact_now()
+    }
+
     fn maybe_auto_compact(&mut self) -> Option<AutoCompactionEvent> {
-        if self.usage_tracker.cumulative_usage().input_tokens
+        // Gate on the LIVE context size — the most recent call's input counters — not on a
+        // cumulative running total, and count cached input as the context it is.
+        //
+        // The previous gate (`cumulative_usage().input_tokens`) was wrong twice over:
+        //   1. It ignored `cache_read_input_tokens`. With prompt caching a warm call reports
+        //      `input_tokens: 2`, so the counter crawled and compaction never fired however
+        //      large the context grew.
+        //   2. Summing across calls is not a context measurement, and the sum never decreased,
+        //      so once it did trip, every later iteration compacted forever.
+        // Reading the latest turn fixes both: it rises with the real context and falls back
+        // below the threshold as soon as a compaction has done its job.
+        if self.usage_tracker.current_turn_usage().total_input_tokens()
             < self.auto_compaction_input_tokens_threshold
         {
             return None;
         }
 
+        self.compact_now()
+    }
+
+    /// Compacts the session unconditionally, returning `None` when nothing actually changed.
+    fn compact_now(&mut self) -> Option<AutoCompactionEvent> {
         let result = compact_session(
             &self.session,
             CompactionConfig {
@@ -588,7 +780,12 @@ where
             },
         );
 
-        if result.removed_message_count == 0 {
+        // `removed_message_count` alone is the wrong test for "did anything happen": compaction
+        // also trims oversized tool-result payloads in messages it must PRESERVE, and on a
+        // session too short to have anything droppable that trimming is the entire result.
+        // Gating on the count would compute the smaller session and then throw it away.
+        let shrank = result.compacted_session.messages != self.session.messages;
+        if result.removed_message_count == 0 && !shrank {
             return None;
         }
 
@@ -709,19 +906,88 @@ where
 /// Reads the automatic compaction threshold from the environment.
 #[must_use]
 pub fn auto_compaction_threshold_from_env() -> u32 {
-    parse_auto_compaction_threshold(
+    resolve_auto_compaction_threshold(
         std::env::var(AUTO_COMPACTION_THRESHOLD_ENV_VAR)
             .ok()
             .as_deref(),
+        std::env::var(CONTEXT_WINDOW_ENV_VAR).ok().as_deref(),
+    )
+}
+
+/// Resolves the threshold from an explicit setting, else from the declared context window.
+///
+/// `DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD` is a 200k-window number. Against a smaller
+/// server it sits above the window entirely, so the pre-request guard can never fire and the
+/// first sign of trouble is the backend rejecting the request. Deriving from the declared window
+/// keeps the guard meaningful without every caller having to compute a second number by hand.
+#[must_use]
+fn resolve_auto_compaction_threshold(explicit: Option<&str>, context_window: Option<&str>) -> u32 {
+    if let Some(threshold) = parse_positive_u32(explicit) {
+        return threshold;
+    }
+    parse_positive_u32(context_window).map_or(
+        DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD,
+        |window| {
+            (u64::from(window) * u64::from(AUTO_COMPACTION_WINDOW_PERCENT) / 100)
+                .try_into()
+                .unwrap_or(DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD)
+        },
     )
 }
 
 #[must_use]
-fn parse_auto_compaction_threshold(value: Option<&str>) -> u32 {
+fn parse_positive_u32(value: Option<&str>) -> Option<u32> {
     value
         .and_then(|raw| raw.trim().parse::<u32>().ok())
-        .filter(|threshold| *threshold > 0)
-        .unwrap_or(DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD)
+        .filter(|parsed| *parsed > 0)
+}
+
+/// How many auto-continue nudges a single turn may inject. `0` disables the behaviour.
+#[must_use]
+pub fn max_stall_nudges_from_env() -> usize {
+    parse_max_stall_nudges(std::env::var(MAX_STALL_NUDGES_ENV_VAR).ok().as_deref())
+}
+
+#[must_use]
+fn parse_max_stall_nudges(value: Option<&str>) -> usize {
+    value
+        .and_then(|raw| raw.trim().parse::<usize>().ok())
+        .unwrap_or(DEFAULT_MAX_STALL_NUDGES)
+}
+
+/// Concatenated text of an assistant message, ignoring thinking and tool blocks.
+#[must_use]
+fn assistant_text(message: &ConversationMessage) -> String {
+    message
+        .blocks
+        .iter()
+        .filter_map(|block| match block {
+            ContentBlock::Text { text } => Some(text.as_str()),
+            _ => None,
+        })
+        .collect::<Vec<_>>()
+        .join("\n")
+}
+
+/// True when a tool-less reply is a stall rather than a finished turn.
+///
+/// Small local models routinely narrate the next tool call in prose and then emit EOS, which
+/// ends the turn with the announced work never performed. `after_tool_use` additionally treats
+/// a mid-task question as a stall: once tools have run in this turn, stopping to ask for a
+/// go-ahead is the same failure wearing a different hat.
+#[must_use]
+fn is_stalled_reply(text: &str, after_tool_use: bool) -> bool {
+    let Some(last_line) = text.lines().map(str::trim).rfind(|line| !line.is_empty()) else {
+        return false;
+    };
+    let lowered = last_line.to_lowercase();
+
+    if after_tool_use && last_line.ends_with('?') {
+        return true;
+    }
+    PENDING_ACTION_MARKERS
+        .iter()
+        .any(|marker| lowered.contains(marker))
 }
 
 fn build_assistant_message(
@@ -853,11 +1119,13 @@ impl ToolExecutor for StaticToolExecutor {
 #[cfg(test)]
 mod tests {
     use super::{
-        build_assistant_message, parse_auto_compaction_threshold, ApiClient, ApiRequest,
-        AssistantEvent, AutoCompactionEvent, ConversationRuntime, PromptCacheEvent, RuntimeError,
+        assistant_text, build_assistant_message, is_stalled_reply, parse_max_stall_nudges,
+        resolve_auto_compaction_threshold, ApiClient, ApiRequest, AssistantEvent,
+        AutoCompactionEvent, ConversationRuntime, PromptCacheEvent, RuntimeError,
         StaticToolExecutor, ToolExecutor, DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD,
+        DEFAULT_MAX_STALL_NUDGES,
     };
-    use crate::compact::CompactionConfig;
+    use crate::compact::{estimate_session_tokens, CompactionConfig};
     use crate::config::{RuntimeFeatureConfig, RuntimeHookConfig};
     use crate::permissions::{
         PermissionMode, PermissionPolicy, PermissionPromptDecision, PermissionPrompter,
@@ -1701,19 +1969,246 @@ mod tests {
         assert_eq!(runtime.session().messages.len(), 2);
     }
 
+    /// Prompt caching reports the bulk of the context as `cache_read_input_tokens` and leaves
+    /// `input_tokens` at a token or two, so a gate that reads `input_tokens` alone never fires
+    /// on a cached session no matter how large the context has grown.
+    #[test]
+    fn auto_compaction_counts_cached_context_toward_threshold() {
+        struct CachedApi;
+        impl ApiClient for CachedApi {
+            fn stream(
+                &mut self,
+                _request: ApiRequest,
+            ) -> Result<Vec<AssistantEvent>, RuntimeError> {
+                Ok(vec![
+                    AssistantEvent::TextDelta("done".to_string()),
+                    AssistantEvent::Usage(TokenUsage {
+                        input_tokens: 2,
+                        output_tokens: 4,
+                        cache_creation_input_tokens: 1_804,
+                        cache_read_input_tokens: 112_639,
+                    }),
+                    AssistantEvent::MessageStop,
+                ])
+            }
+        }
+
+        let mut session = Session::new();
+        session.messages = vec![
+            crate::session::ConversationMessage::user_text("one"),
+            crate::session::ConversationMessage::assistant(vec![ContentBlock::Text {
+                text: "two".to_string(),
+            }]),
+            crate::session::ConversationMessage::user_text("three"),
+            crate::session::ConversationMessage::assistant(vec![ContentBlock::Text {
+                text: "four".to_string(),
+            }]),
+        ];
+
+        let mut runtime = ConversationRuntime::new(
+            session,
+            CachedApi,
+            StaticToolExecutor::new(),
+            PermissionPolicy::new(PermissionMode::DangerFullAccess),
+            vec!["system".to_string()],
+        )
+        .with_auto_compaction_input_tokens_threshold(100_000);
+
+        let summary = runtime
+            .run_turn("trigger", None)
+            .expect("turn should succeed");
+
+        assert_eq!(
+            summary.auto_compaction,
+            Some(AutoCompactionEvent {
+                removed_message_count: 2,
+            })
+        );
+    }
+
+    /// The threshold describes the LIVE context, not a monotonic running total. Once compaction
+    /// has shrunk the context back down, later small turns must not keep compacting.
+    #[test]
+    fn auto_compaction_stops_once_the_context_shrinks_again() {
+        struct ShrinkingApi {
+            calls: u32,
+        }
+        impl ApiClient for ShrinkingApi {
+            fn stream(
+                &mut self,
+                _request: ApiRequest,
+            ) -> Result<Vec<AssistantEvent>, RuntimeError> {
+                let call = self.calls;
+                self.calls += 1;
+                // The first call is over budget; every call after it is small.
+                let input_tokens = if call == 0 { 150_000 } else { 900 };
+                Ok(vec![
+                    AssistantEvent::TextDelta("done".to_string()),
+                    AssistantEvent::Usage(TokenUsage {
+                        input_tokens,
+                        output_tokens: 4,
+                        cache_creation_input_tokens: 0,
+                        cache_read_input_tokens: 0,
+                    }),
+                    AssistantEvent::MessageStop,
+                ])
+            }
+        }
+
+        let mut session = Session::new();
+        session.messages = vec![
+            crate::session::ConversationMessage::user_text("one"),
+            crate::session::ConversationMessage::assistant(vec![ContentBlock::Text {
+                text: "two".to_string(),
+            }]),
+            crate::session::ConversationMessage::user_text("three"),
+            crate::session::ConversationMessage::assistant(vec![ContentBlock::Text {
+                text: "four".to_string(),
+            }]),
+        ];
+
+        let mut runtime = ConversationRuntime::new(
+            session,
+            ShrinkingApi { calls: 0 },
+            // Every turn after a compaction runs the session health probe, which calls
+            // `glob_search`; without it the probe fails the turn before the assertion is reached.
+            StaticToolExecutor::new().register("glob_search", |_| Ok(String::new())),
+            PermissionPolicy::new(PermissionMode::DangerFullAccess),
+            vec!["system".to_string()],
+        )
+        .with_auto_compaction_input_tokens_threshold(100_000);
+
+        let first = runtime.run_turn("trigger", None).expect("first turn");
+        assert!(
+            first.auto_compaction.is_some(),
+            "the over-budget turn should compact"
+        );
+
+        let second = runtime.run_turn("again", None).expect("second turn");
+        assert_eq!(
+            second.auto_compaction, None,
+            "a small turn after compaction must not compact again"
+        );
+        let third = runtime.run_turn("and again", None).expect("third turn");
+        assert_eq!(
+            third.auto_compaction, None,
+            "compaction must not latch on once the threshold has been crossed"
+        );
+    }
+
+    /// Usage-based compaction is post-hoc: it only sees the context AFTER a call returns. A
+    /// large tool result (or a long reasoning block) appended after that measurement can blow
+    /// the window on the very next request, which the provider rejects outright — so there must
+    /// also be a check BEFORE the request goes out, based on a locally estimated session size.
+    #[test]
+    fn auto_compaction_fires_before_the_first_request_when_the_session_is_already_oversized() {
+        struct RecordingApi {
+            largest_request_messages: std::rc::Rc<std::cell::Cell<usize>>,
+        }
+        impl ApiClient for RecordingApi {
+            fn stream(&mut self, request: ApiRequest) -> Result<Vec<AssistantEvent>, RuntimeError> {
+                self.largest_request_messages.set(
+                    self.largest_request_messages
+                        .get()
+                        .max(request.messages.len()),
+                );
+                Ok(vec![
+                    AssistantEvent::TextDelta("done".to_string()),
+                    // Deliberately no Usage event: nothing has been measured yet, which is
+                    // exactly the situation the pre-send check has to cover.
+                    AssistantEvent::MessageStop,
+                ])
+            }
+        }
+
+        // Six messages that are individually large — the shape of a session that has just
+        // absorbed a couple of big file reads.
+        let bulk = "x".repeat(40_000);
+        let mut session = Session::new();
+        for _ in 0..3 {
+            session
+                .messages
+                .push(crate::session::ConversationMessage::user_text(bulk.clone()));
+            session
+                .messages
+                .push(crate::session::ConversationMessage::assistant(vec![
+                    ContentBlock::Text { text: bulk.clone() },
+                ]));
+        }
+        let before = estimate_session_tokens(&session);
+        assert!(
+            before > 24_000,
+            "test fixture must start over the threshold, got {before}"
+        );
+
+        let largest = std::rc::Rc::new(std::cell::Cell::new(0usize));
+        let mut runtime = ConversationRuntime::new(
+            session,
+            RecordingApi {
+                largest_request_messages: std::rc::Rc::clone(&largest),
+            },
+            StaticToolExecutor::new().register("glob_search", |_| Ok(String::new())),
+            PermissionPolicy::new(PermissionMode::DangerFullAccess),
+            vec!["system".to_string()],
+        )
+        .with_auto_compaction_input_tokens_threshold(24_000);
+
+        let summary = runtime
+            .run_turn("trigger", None)
+            .expect("turn should succeed");
+
+        assert!(
+            summary.auto_compaction.is_some(),
+            "an already-oversized session must compact before the request is sent"
+        );
+        assert!(
+            estimate_session_tokens(runtime.session()) < before,
+            "compaction must actually shrink the session"
+        );
+        assert!(
+            largest.get() < 8,
+            "the request must be sent with the compacted history, not the original 7 messages"
+        );
+    }
+
     #[test]
     fn auto_compaction_threshold_defaults_and_parses_values() {
         assert_eq!(
-            parse_auto_compaction_threshold(None),
+            resolve_auto_compaction_threshold(None, None),
             DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD
         );
-        assert_eq!(parse_auto_compaction_threshold(Some("4321")), 4321);
+        assert_eq!(resolve_auto_compaction_threshold(Some("4321"), None), 4321);
         assert_eq!(
-            parse_auto_compaction_threshold(Some("0")),
+            resolve_auto_compaction_threshold(Some("0"), None),
             DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD
         );
         assert_eq!(
-            parse_auto_compaction_threshold(Some("not-a-number")),
+            resolve_auto_compaction_threshold(Some("not-a-number"), None),
+            DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD
+        );
+    }
+
+    #[test]
+    fn auto_compaction_threshold_derives_from_a_declared_context_window() {
+        // 70% of 65536, so the guard sits inside the window instead of 34k above it.
+        assert_eq!(
+            resolve_auto_compaction_threshold(None, Some("65536")),
+            45875
+        );
+    }
+
+    #[test]
+    fn an_explicit_threshold_still_wins_over_the_declared_window() {
+        assert_eq!(
+            resolve_auto_compaction_threshold(Some("48000"), Some("65536")),
+            48000
+        );
+    }
+
+    #[test]
+    fn an_unparseable_context_window_falls_back_to_the_default_threshold() {
+        assert_eq!(
+            resolve_auto_compaction_threshold(None, Some("plenty")),
             DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD
         );
     }
@@ -1885,6 +2380,146 @@ mod tests {
         assert_eq!(error.to_string(), "unknown tool: missing");
     }
 
+    /// Replays a fixed script of stream outcomes, one per call.
+    struct ScriptedApi {
+        steps: Vec<Result<Vec<AssistantEvent>, String>>,
+        calls: usize,
+    }
+
+    impl ScriptedApi {
+        fn new(steps: Vec<Result<Vec<AssistantEvent>, String>>) -> Self {
+            Self { steps, calls: 0 }
+        }
+    }
+
+    impl ApiClient for ScriptedApi {
+        fn stream(&mut self, _request: ApiRequest) -> Result<Vec<AssistantEvent>, RuntimeError> {
+            let step = self
+                .steps
+                .get(self.calls)
+                .cloned()
+                .unwrap_or_else(|| Err("script exhausted".to_string()));
+            self.calls += 1;
+            step.map_err(RuntimeError::new)
+        }
+    }
+
+    fn tool_call_events(id: &str) -> Vec<AssistantEvent> {
+        vec![
+            AssistantEvent::ToolUse {
+                id: id.to_string(),
+                name: "echo".to_string(),
+                input: "payload".to_string(),
+            },
+            AssistantEvent::MessageStop,
+        ]
+    }
+
+    fn scripted_runtime(
+        steps: Vec<Result<Vec<AssistantEvent>, String>>,
+    ) -> ConversationRuntime<ScriptedApi, StaticToolExecutor> {
+        ConversationRuntime::new(
+            Session::new(),
+            ScriptedApi::new(steps),
+            StaticToolExecutor::new().register("echo", |input| Ok(input.to_string())),
+            PermissionPolicy::new(PermissionMode::DangerFullAccess),
+            vec!["system".to_string()],
+        )
+    }
+
+    #[test]
+    fn a_failed_turn_keeps_the_results_of_tools_that_already_ran() {
+        // given: a turn whose first iteration runs a tool, and whose second call dies
+        let mut runtime = scripted_runtime(vec![
+            Ok(tool_call_events("tool-1")),
+            Err("upstream failed".to_string()),
+        ]);
+
+        // when
+        runtime
+            .run_turn("do the thing", None)
+            .expect_err("the second call fails");
+
+        // then: the executed tool's result is still in the transcript. The side effect happened;
+        // a caller that discards this record will let the model run the same tool again.
+        let tool_results = runtime
+            .session()
+            .messages
+            .iter()
+            .filter(|message| {
+                message
+                    .blocks
+                    .iter()
+                    .any(|block| matches!(block, ContentBlock::ToolResult { .. }))
+            })
+            .count();
+        assert_eq!(tool_results, 1);
+    }
+
+    #[test]
+    fn resume_turn_continues_without_appending_another_user_message() {
+        // given: a session that already carries a user turn and a completed tool result
+        let mut runtime = scripted_runtime(vec![
+            Ok(tool_call_events("tool-1")),
+            Err("upstream failed".to_string()),
+            Ok(vec![
+                AssistantEvent::TextDelta("done".to_string()),
+                AssistantEvent::MessageStop,
+            ]),
+        ]);
+        runtime.run_turn("do the thing", None).expect_err("fails");
+        let user_messages_before = runtime
+            .session()
+            .messages
+            .iter()
+            .filter(|message| message.role == MessageRole::User)
+            .count();
+
+        // when
+        runtime
+            .resume_turn(None)
+            .expect("the resumed turn succeeds");
+
+        // then: the prompt was not repeated
+        let user_messages_after = runtime
+            .session()
+            .messages
+            .iter()
+            .filter(|message| message.role == MessageRole::User)
+            .count();
+        assert_eq!(user_messages_before, user_messages_after);
+    }
+
+    #[test]
+    fn a_stall_nudge_is_recorded_as_a_system_turn_not_a_user_one() {
+        // given: a reply that announces work and stops, so the nudge fires once
+        let mut runtime = nudge_runtime(
+            vec![
+                text_only("Let me check the logs."),
+                text_only("The logs were clean."),
+            ],
+            1,
+        );
+        runtime
+            .run_turn("check the logs", None)
+            .expect("the nudged turn completes");
+
+        // then: exactly one user message — the real one
+        let user_messages = runtime
+            .session()
+            .messages
+            .iter()
+            .filter(|message| message.role == MessageRole::User)
+            .count();
+        assert_eq!(user_messages, 1);
+        assert!(runtime
+            .session()
+            .messages
+            .iter()
+            .any(|message| message.role == MessageRole::System
+                && assistant_text(message).contains("[auto-continue]")));
+    }
+
     #[test]
     fn run_turn_errors_when_max_iterations_is_exceeded() {
         struct LoopingApi;
@@ -1956,4 +2591,136 @@ mod tests {
         // then
         assert_eq!(error.to_string(), "upstream failed");
     }
+
+    fn text_only(text: &str) -> Vec<AssistantEvent> {
+        vec![
+            AssistantEvent::TextDelta(text.to_string()),
+            AssistantEvent::MessageStop,
+        ]
+    }
+
+    struct CountingApi {
+        calls: usize,
+        replies: Vec<Vec<AssistantEvent>>,
+    }
+
+    impl ApiClient for CountingApi {
+        fn stream(&mut self, _request: ApiRequest) -> Result<Vec<AssistantEvent>, RuntimeError> {
+            let reply = self
+                .replies
+                .get(self.calls)
+                .cloned()
+                .unwrap_or_else(|| text_only("Now running the next probe:"));
+            self.calls += 1;
+            Ok(reply)
+        }
+    }
+
+    fn nudge_runtime(
+        replies: Vec<Vec<AssistantEvent>>,
+        max_stall_nudges: usize,
+    ) -> ConversationRuntime<CountingApi, StaticToolExecutor> {
+        ConversationRuntime::new(
+            Session::new(),
+            CountingApi { calls: 0, replies },
+            StaticToolExecutor::new().register("probe", |_| Ok("probe output".to_string())),
+            PermissionPolicy::new(PermissionMode::DangerFullAccess),
+            vec!["system".to_string()],
+        )
+        .with_max_stall_nudges(max_stall_nudges)
+    }
+
+    #[test]
+    fn announced_but_uncalled_tool_is_auto_continued() {
+        // given: the model narrates the next probe instead of calling it
+        let mut runtime = nudge_runtime(
+            vec![
+                text_only("Now running the Ford EEC-V tuner probe (ATSP1):"),
+                vec![
+                    AssistantEvent::ToolUse {
+                        id: "tool-1".to_string(),
+                        name: "probe".to_string(),
+                        input: "atsp1".to_string(),
+                    },
+                    AssistantEvent::MessageStop,
+                ],
+                text_only("No response on ATSP1, which is expected here."),
+            ],
+            3,
+        );
+
+        // when
+        let summary = runtime.run_turn("run the probes", None).expect("turn runs");
+
+        // then: the nudge pulled it back into the loop and the tool actually ran
+        assert_eq!(summary.iterations, 3);
+        assert_eq!(summary.tool_results.len(), 1);
+    }
+
+    #[test]
+    fn nudges_are_bounded_so_a_stuck_model_still_stops() {
+        // given: every reply is an announcement and nothing is ever called
+        let mut runtime = nudge_runtime(Vec::new(), 2);
+
+        // when
+        let summary = runtime.run_turn("run the probes", None).expect("turn runs");
+
+        // then: initial reply plus exactly two nudged retries
+        assert_eq!(summary.iterations, 3);
+        assert!(summary.tool_results.is_empty());
+    }
+
+    #[test]
+    fn a_plain_completion_is_not_nudged() {
+        // given
+        let mut runtime = nudge_runtime(vec![text_only("The DTC scan came back clean.")], 3);
+
+        // when
+        let summary = runtime.run_turn("scan for codes", None).expect("turn runs");
+
+        // then
+        assert_eq!(summary.iterations, 1);
+    }
+
+    #[test]
+    fn zero_budget_disables_auto_continue() {
+        // given
+        let mut runtime = nudge_runtime(vec![text_only("Let me run the DTC scan.")], 0);
+
+        // when
+        let summary = runtime.run_turn("scan for codes", None).expect("turn runs");
+
+        // then
+        assert_eq!(summary.iterations, 1);
+    }
+
+    #[test]
+    fn a_question_stalls_only_after_a_tool_has_run() {
+        assert!(is_stalled_reply("Should I run the DTC scan next?", true));
+        assert!(!is_stalled_reply("Which port is the adapter on?", false));
+    }
+
+    #[test]
+    fn stall_detection_reads_the_last_line_not_the_whole_reply() {
+        // an announcement earlier in the reply does not make a finished turn look stalled
+        assert!(!is_stalled_reply(
+            "Let me check the port.\nThe scan finished with no codes stored.",
+            true
+        ));
+        assert!(is_stalled_reply(
+            "The scan finished.\nNext I'll read the freeze-frame data.",
+            true
+        ));
+    }
+
+    #[test]
+    fn stall_nudge_budget_parses_from_env_value() {
+        // asserted against the literal, not the constant: comparing the parser to the same
+        // constant it falls back to would pass no matter what the default became
+        assert_eq!(DEFAULT_MAX_STALL_NUDGES, 3);
+        assert_eq!(parse_max_stall_nudges(None), 3);
+        assert_eq!(parse_max_stall_nudges(Some("not a number")), 3);
+        assert_eq!(parse_max_stall_nudges(Some(" 5 ")), 5);
+        assert_eq!(parse_max_stall_nudges(Some("0")), 0);
+    }
 }

rust/crates/runtime/src/file_ops.rs

@@ -11,8 +11,48 @@ use serde::{Deserialize, Serialize};
 use walkdir::{DirEntry, WalkDir};
 
 /// Maximum file size that can be read (10 MB).
+///
+/// This is a *process* guard — it stops us slurping a huge or binary blob into memory. It is
+/// not a context guard: 10 MB is on the order of 2.5 M tokens, so it never fires on the files
+/// that actually blow a model's context window. `DEFAULT_READ_LINES` / `MAX_READ_CHARS` below
+/// are the context guard.
 const MAX_READ_SIZE: u64 = 10 * 1024 * 1024;
 
+/// Lines returned by `read_file` when the caller does not pass an explicit `limit`.
+///
+/// This is a *paging* size, not a refusal. An unbounded read is the easiest way for a model to
+/// destroy its own context: one 100 KB source file is ~25 k tokens, which does not fit alongside
+/// a system prompt in a 32 k window. When that happens the backend rejects the whole request and
+/// the session dies — so reading everything at once is what makes it stop. Returning a page plus
+/// `nextOffset` means the read always succeeds and the model can keep going.
+const DEFAULT_READ_LINES: usize = 2000;
+
+/// Hard ceiling on the characters `read_file` returns, applied even when `limit` is explicit.
+///
+/// `DEFAULT_READ_LINES` alone does not bound the payload: minified bundles, JSON blobs and
+/// generated sources routinely put megabytes on a handful of lines. ~64 KB is ~16 k tokens.
+///
+/// That default assumes a large (200 k) context window. Against a small self-hosted backend it
+/// is far too generous: two reads at this size put a request over 47 k tokens, which a 32 k
+/// window rejects outright — and compaction cannot recover it, because the preserved recent
+/// messages ARE the oversized tool results. Override with [`MAX_READ_CHARS_ENV_VAR`].
+const DEFAULT_MAX_READ_CHARS: usize = 64 * 1024;
+
+/// Environment override for [`DEFAULT_MAX_READ_CHARS`], in characters.
+const MAX_READ_CHARS_ENV_VAR: &str = "CLAW_MAX_READ_CHARS";
+
+/// Effective character ceiling for one `read_file` page.
+///
+/// Read per call rather than cached: a long-lived process should pick up a changed budget, and
+/// this runs once per read, not per line.
+fn max_read_chars() -> usize {
+    std::env::var(MAX_READ_CHARS_ENV_VAR)
+        .ok()
+        .and_then(|raw| raw.trim().parse::<usize>().ok())
+        .filter(|chars| *chars > 0)
+        .unwrap_or(DEFAULT_MAX_READ_CHARS)
+}
+
 /// Maximum file size that can be written (10 MB).
 const MAX_WRITE_SIZE: usize = 10 * 1024 * 1024;
 
@@ -80,6 +120,17 @@ pub struct TextFilePayload {
     pub start_line: usize,
     #[serde(rename = "totalLines")]
     pub total_lines: usize,
+    /// True when the file has more lines past this page. Signals the caller that it has NOT seen
+    /// the whole file, so it does not silently reason over a partial view.
+    #[serde(default)]
+    pub truncated: bool,
+    /// Line index to pass back as `offset` to continue reading. Absent once the file is exhausted.
+    #[serde(
+        rename = "nextOffset",
+        default,
+        skip_serializing_if = "Option::is_none"
+    )]
+    pub next_offset: Option<usize>,
 }
 
 /// Output envelope for the `read_file` tool.
@@ -228,10 +279,28 @@ pub fn read_file(
     let content = fs::read_to_string(&absolute_path)?;
     let lines: Vec<&str> = content.lines().collect();
     let start_index = offset.unwrap_or(0).min(lines.len());
-    let end_index = limit.map_or(lines.len(), |limit| {
-        start_index.saturating_add(limit).min(lines.len())
-    });
+    // An absent `limit` pages rather than reading to EOF: see DEFAULT_READ_LINES.
+    let requested_end = start_index
+        .saturating_add(limit.unwrap_or(DEFAULT_READ_LINES))
+        .min(lines.len());
+
+    // Apply the character ceiling on top of the line window, so a file with very long lines
+    // cannot blow the budget the line cap was meant to enforce. Always yield at least one line,
+    // otherwise an oversized single line would return nothing and the caller could not advance.
+    let mut end_index = start_index;
+    let mut chars = 0usize;
+    let char_ceiling = max_read_chars();
+    for line in &lines[start_index..requested_end] {
+        let next = chars + line.chars().count() + 1;
+        if next > char_ceiling && end_index > start_index {
+            break;
+        }
+        chars = next;
+        end_index += 1;
+    }
+
     let selected = lines[start_index..end_index].join("\n");
+    let truncated = end_index < lines.len();
 
     Ok(ReadFileOutput {
         kind: String::from("text"),
@@ -241,6 +310,8 @@ pub fn read_file(
             num_lines: end_index.saturating_sub(start_index),
             start_line: start_index.saturating_add(1),
             total_lines: lines.len(),
+            truncated,
+            next_offset: truncated.then_some(end_index),
         },
     })
 }
@@ -812,11 +883,18 @@ mod tests {
 
     use super::{
         component_contains_glob, derive_glob_walk_root, edit_file, expand_braces, glob_search,
-        grep_search, is_symlink_escape, normalize_for_comparison, read_file,
+        grep_search, is_symlink_escape, max_read_chars, normalize_for_comparison, read_file,
         read_file_in_workspace, write_file, write_file_in_workspace, GrepSearchInput,
-        MAX_WRITE_SIZE,
+        DEFAULT_MAX_READ_CHARS, DEFAULT_READ_LINES, MAX_READ_CHARS_ENV_VAR, MAX_WRITE_SIZE,
     };
 
+    /// Serializes the tests that mutate `CLAW_MAX_READ_CHARS`; env vars are process-global.
+    fn env_lock() -> std::sync::MutexGuard<'static, ()> {
+        static LOCK: std::sync::Mutex<()> = std::sync::Mutex::new(());
+        LOCK.lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner)
+    }
+
     fn temp_path(name: &str) -> std::path::PathBuf {
         let unique = SystemTime::now()
             .duration_since(UNIX_EPOCH)
@@ -837,6 +915,132 @@ mod tests {
         assert_eq!(read_output.file.content, "two");
     }
 
+    #[test]
+    fn unbounded_read_of_a_large_file_pages_instead_of_returning_everything() {
+        // Regression: an unbounded read used to return the whole file, which on a small-context
+        // backend overflows the window and the request is rejected outright — the session dies.
+        let path = temp_path("paged-read.txt");
+        let total = DEFAULT_READ_LINES + 500;
+        let body = (0..total)
+            .map(|i| format!("line {i}"))
+            .collect::<Vec<_>>()
+            .join("\n");
+        write_file(path.to_string_lossy().as_ref(), &body).expect("write should succeed");
+
+        let first = read_file(path.to_string_lossy().as_ref(), None, None).expect("first page");
+        assert_eq!(first.file.num_lines, DEFAULT_READ_LINES);
+        assert_eq!(first.file.total_lines, total);
+        assert!(first.file.truncated, "caller must be told it is partial");
+        assert_eq!(first.file.next_offset, Some(DEFAULT_READ_LINES));
+        assert!(first.file.content.starts_with("line 0\n"));
+
+        // The advertised next_offset must actually resume where the last page stopped.
+        let second = read_file(
+            path.to_string_lossy().as_ref(),
+            first.file.next_offset,
+            None,
+        )
+        .expect("second page");
+        assert_eq!(second.file.num_lines, 500);
+        assert!(!second.file.truncated, "the tail is the final page");
+        assert_eq!(second.file.next_offset, None);
+        assert!(second
+            .file
+            .content
+            .starts_with(&format!("line {DEFAULT_READ_LINES}\n")));
+    }
+
+    #[test]
+    fn char_ceiling_bounds_files_whose_lines_are_enormous() {
+        // A line cap alone does not bound the payload: minified bundles put megabytes on a few
+        // lines, so the byte ceiling has to apply too — and must still yield forward progress.
+        let _guard = env_lock();
+        std::env::remove_var(MAX_READ_CHARS_ENV_VAR);
+        let path = temp_path("long-lines.txt");
+        let huge = "x".repeat(DEFAULT_MAX_READ_CHARS);
+        let body = format!("{huge}\n{huge}\n{huge}");
+        write_file(path.to_string_lossy().as_ref(), &body).expect("write should succeed");
+
+        let page = read_file(path.to_string_lossy().as_ref(), None, None).expect("first page");
+        assert_eq!(page.file.num_lines, 1, "one oversized line still advances");
+        assert!(page.file.truncated);
+        assert_eq!(page.file.next_offset, Some(1));
+
+        // An explicit limit must not be able to defeat the ceiling.
+        let greedy = read_file(path.to_string_lossy().as_ref(), None, Some(3)).expect("greedy");
+        assert_eq!(greedy.file.num_lines, 1);
+        assert!(greedy.file.truncated);
+    }
+
+    /// The 64 KB default is a 200 k-context number. A small self-hosted backend must be able to
+    /// shrink it, or two reads overflow a 32 k window in a single turn.
+    #[test]
+    fn read_char_ceiling_is_configurable_via_the_environment() {
+        let _guard = env_lock();
+        let path = temp_path("configurable-ceiling.txt");
+        // 40 lines of 100 chars = ~4 KB: far under the default ceiling, over a 1 KB one.
+        let body = (0..40)
+            .map(|_| "y".repeat(100))
+            .collect::<Vec<_>>()
+            .join("\n");
+        write_file(path.to_string_lossy().as_ref(), &body).expect("write should succeed");
+
+        std::env::remove_var(MAX_READ_CHARS_ENV_VAR);
+        let whole = read_file(path.to_string_lossy().as_ref(), None, None).expect("default read");
+        assert_eq!(
+            whole.file.num_lines, 40,
+            "default ceiling returns the file whole"
+        );
+        assert!(!whole.file.truncated);
+
+        std::env::set_var(MAX_READ_CHARS_ENV_VAR, "1024");
+        let clipped = read_file(path.to_string_lossy().as_ref(), None, None).expect("clipped read");
+        std::env::remove_var(MAX_READ_CHARS_ENV_VAR);
+        assert!(
+            clipped.file.num_lines < 40,
+            "a smaller ceiling must actually clip, got {} lines",
+            clipped.file.num_lines
+        );
+        assert!(
+            clipped.file.truncated,
+            "a clipped page must report truncated"
+        );
+        assert_eq!(
+            clipped.file.next_offset,
+            Some(clipped.file.num_lines),
+            "the caller must be told where to resume"
+        );
+    }
+
+    #[test]
+    fn read_char_ceiling_ignores_junk_and_zero_values() {
+        let _guard = env_lock();
+        for value in ["", "0", "-5", "not-a-number"] {
+            std::env::set_var(MAX_READ_CHARS_ENV_VAR, value);
+            assert_eq!(
+                max_read_chars(),
+                DEFAULT_MAX_READ_CHARS,
+                "{value:?} must fall back to the default rather than disabling reads"
+            );
+        }
+        std::env::set_var(MAX_READ_CHARS_ENV_VAR, "24000");
+        assert_eq!(max_read_chars(), 24_000);
+        std::env::remove_var(MAX_READ_CHARS_ENV_VAR);
+        assert_eq!(max_read_chars(), DEFAULT_MAX_READ_CHARS);
+    }
+
+    #[test]
+    fn small_files_are_returned_whole_and_not_marked_truncated() {
+        let path = temp_path("small-read.txt");
+        write_file(path.to_string_lossy().as_ref(), "one\ntwo\nthree").expect("write");
+
+        let output = read_file(path.to_string_lossy().as_ref(), None, None).expect("read");
+        assert_eq!(output.file.content, "one\ntwo\nthree");
+        assert_eq!(output.file.num_lines, 3);
+        assert!(!output.file.truncated);
+        assert_eq!(output.file.next_offset, None);
+    }
+
     #[test]
     fn edits_file_contents() {
         let path = temp_path("edit.txt");

rust/crates/runtime/src/prompt.rs

@@ -42,7 +42,25 @@ pub const SYSTEM_PROMPT_DYNAMIC_BOUNDARY: &str = "__SYSTEM_PROMPT_DYNAMIC_BOUNDA
 pub const FRONTIER_MODEL_NAME: &str = "Claude Opus 4.6";
 const MAX_INSTRUCTION_FILE_CHARS: usize = 4_000;
 const MAX_TOTAL_INSTRUCTION_CHARS: usize = 12_000;
-const MAX_GIT_DIFF_CHARS: usize = 50_000;
+/// Character budget for the git diff snapshot embedded in the system prompt.
+///
+/// This lives in the *system prompt*, so its cost is paid on every session whose
+/// working tree has changed — and a diff changes on essentially every edit. The
+/// previous 50_000 budget was four times the entire instruction-file budget and
+/// could add ~15_000 tokens to the prompt from a single modified file, which is
+/// affordable against a hosted frontier model and ruinous against a local one.
+///
+/// Override with `CLAW_MAX_GIT_DIFF_CHARS`. `0` drops the diff snapshot entirely;
+/// git status and the recent-commit list are unaffected.
+const DEFAULT_MAX_GIT_DIFF_CHARS: usize = 4_000;
+const MAX_GIT_DIFF_CHARS_ENV_VAR: &str = "CLAW_MAX_GIT_DIFF_CHARS";
+
+fn max_git_diff_chars() -> usize {
+    std::env::var(MAX_GIT_DIFF_CHARS_ENV_VAR)
+        .ok()
+        .and_then(|value| value.trim().parse().ok())
+        .unwrap_or(DEFAULT_MAX_GIT_DIFF_CHARS)
+}
 
 /// Neutral identity for the model family line in generated prompts.
 #[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
@@ -498,8 +516,15 @@ fn read_git_diff(cwd: &Path) -> Option<String> {
 }
 
 fn truncate_diff(mut diff: String) -> String {
-    if diff.len() > MAX_GIT_DIFF_CHARS {
-        let mut end = MAX_GIT_DIFF_CHARS;
+    let budget = max_git_diff_chars();
+    // A zero budget means "omit the diff", not "truncate to nothing" -- returning
+    // an empty string here would still push an empty "Git diff snapshot:" heading
+    // into the prompt.
+    if budget == 0 {
+        return String::new();
+    }
+    if diff.len() > budget {
+        let mut end = budget;
         while !diff.is_char_boundary(end) {
             end -= 1;
         }
@@ -758,6 +783,7 @@ fn get_simple_doing_tasks_section() -> String {
         "Do not add speculative abstractions, compatibility shims, or unrelated cleanup.".to_string(),
         "Do not create files unless they are required to complete the task.".to_string(),
         "If an approach fails, diagnose the failure before switching tactics.".to_string(),
+        "Act in the turn you plan in: when the next step is a tool call, make the call. Never end a turn describing a step you have not taken, and do not stop to ask for a go-ahead on work already requested.".to_string(),
         "Be careful not to introduce security vulnerabilities such as command injection, XSS, or SQL injection.".to_string(),
         "Report outcomes faithfully: if verification fails or was not run, say so explicitly.".to_string(),
     ]);
@@ -782,7 +808,8 @@ mod tests {
         collapse_blank_lines, display_context_path, normalize_instruction_content,
         render_instruction_content, render_instruction_files, truncate_diff,
         truncate_instruction_content, ContextFile, ModelFamilyIdentity, ProjectContext,
-        SystemPromptBuilder, MAX_GIT_DIFF_CHARS, SYSTEM_PROMPT_DYNAMIC_BOUNDARY,
+        SystemPromptBuilder, DEFAULT_MAX_GIT_DIFF_CHARS, MAX_GIT_DIFF_CHARS_ENV_VAR,
+        SYSTEM_PROMPT_DYNAMIC_BOUNDARY,
     };
     use crate::config::ConfigLoader;
     use std::fs;
@@ -1444,33 +1471,69 @@ mod tests {
         assert!(rendered.contains("Project rules"));
     }
 
+    /// Serializes the tests that read or mutate `CLAW_MAX_GIT_DIFF_CHARS`; env vars
+    /// are process-global and the default-budget tests fail if another test leaves
+    /// an override set.
+    fn diff_env_lock() -> std::sync::MutexGuard<'static, ()> {
+        static LOCK: std::sync::Mutex<()> = std::sync::Mutex::new(());
+        LOCK.lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner)
+    }
+
     #[test]
     fn truncate_diff_preserves_short_content() {
+        let _guard = diff_env_lock();
+        std::env::remove_var(MAX_GIT_DIFF_CHARS_ENV_VAR);
         let short = "a".repeat(1_000);
         let result = truncate_diff(short.clone());
         assert_eq!(result, short);
         assert!(!result.contains("[diff truncated"));
     }
 
+    #[test]
+    fn truncate_diff_honors_env_override() {
+        let _guard = diff_env_lock();
+        std::env::set_var(MAX_GIT_DIFF_CHARS_ENV_VAR, "100");
+        let result = truncate_diff("z".repeat(1_000));
+        std::env::remove_var(MAX_GIT_DIFF_CHARS_ENV_VAR);
+        let marker = "\n\n... [diff truncated — too large for system prompt]";
+        assert!(result.ends_with(marker));
+        assert_eq!(result.len() - marker.len(), 100);
+    }
+
+    #[test]
+    fn truncate_diff_zero_budget_omits_the_snapshot_entirely() {
+        // Zero must yield nothing at all -- a truncated-to-empty body would still
+        // drag a "Git diff snapshot:" heading and the truncation marker into the
+        // prompt, which is the opposite of what the operator asked for.
+        let _guard = diff_env_lock();
+        std::env::set_var(MAX_GIT_DIFF_CHARS_ENV_VAR, "0");
+        let result = truncate_diff("z".repeat(1_000));
+        std::env::remove_var(MAX_GIT_DIFF_CHARS_ENV_VAR);
+        assert!(result.is_empty());
+    }
+
     #[test]
     fn truncate_diff_caps_oversized_content() {
-        let large = "x".repeat(MAX_GIT_DIFF_CHARS + 5_000);
+        let _guard = diff_env_lock();
+        std::env::remove_var(MAX_GIT_DIFF_CHARS_ENV_VAR);
+        let large = "x".repeat(DEFAULT_MAX_GIT_DIFF_CHARS + 5_000);
         let result = truncate_diff(large);
         assert!(result.contains("... [diff truncated — too large for system prompt]"));
-        // The body before the marker must be at most MAX_GIT_DIFF_CHARS bytes
+        // The body before the marker must be at most DEFAULT_MAX_GIT_DIFF_CHARS bytes
         let marker = "\n\n... [diff truncated — too large for system prompt]";
         let body_len = result.len() - marker.len();
-        assert!(body_len <= MAX_GIT_DIFF_CHARS);
+        assert!(body_len <= DEFAULT_MAX_GIT_DIFF_CHARS);
     }
 
     #[test]
     fn truncate_diff_respects_utf8_char_boundaries() {
-        // Build a string where MAX_GIT_DIFF_CHARS falls in the middle of a
+        // Build a string where DEFAULT_MAX_GIT_DIFF_CHARS falls in the middle of a
         // multi-byte character (U+1F600 = 4 bytes in UTF-8).
-        let prefix_len = MAX_GIT_DIFF_CHARS - 2;
+        let prefix_len = DEFAULT_MAX_GIT_DIFF_CHARS - 2;
         let mut input = "a".repeat(prefix_len);
         // Append a 4-byte emoji so bytes [prefix_len..prefix_len+4] are the
-        // emoji.  MAX_GIT_DIFF_CHARS lands at prefix_len+2, inside the emoji.
+        // emoji.  DEFAULT_MAX_GIT_DIFF_CHARS lands at prefix_len+2, inside the emoji.
         input.push('\u{1F600}');
         input.push_str(&"b".repeat(10_000));
 
@@ -1482,7 +1545,7 @@ mod tests {
         // inside it would be invalid UTF-8.
         let marker = "\n\n... [diff truncated — too large for system prompt]";
         let body = &result[..result.len() - marker.len()];
-        assert!(body.len() <= MAX_GIT_DIFF_CHARS);
+        assert!(body.len() <= DEFAULT_MAX_GIT_DIFF_CHARS);
         assert!(body.is_char_boundary(body.len()));
     }
 }

rust/crates/runtime/src/session.rs

@@ -718,6 +718,20 @@ impl ConversationMessage {
         }
     }
 
+    /// Text the harness injected on its own initiative, not something the user typed.
+    ///
+    /// Reaches the model as an ordinary turn — the wire protocol has nowhere else to put it — but
+    /// stays labelled in the transcript, so a replay, an export, or a compaction summary does not
+    /// present the harness's own prompting back as the user's words.
+    #[must_use]
+    pub fn system_text(text: impl Into<String>) -> Self {
+        Self {
+            role: MessageRole::System,
+            blocks: vec![ContentBlock::Text { text: text.into() }],
+            usage: None,
+        }
+    }
+
     #[must_use]
     pub fn assistant(blocks: Vec<ContentBlock>) -> Self {
         Self {

rust/crates/runtime/src/trident.rs

@@ -29,7 +29,7 @@ impl Default for TridentConfig {
 }
 
 /// Statistics from a Trident compaction run.
-#[derive(Debug, Clone, PartialEq, Eq)]
+#[derive(Debug, Clone, PartialEq, Eq, Default)]
 pub struct TridentStats {
     pub superseded_count: usize,
     pub collapsed_chains: usize,
@@ -41,21 +41,6 @@ pub struct TridentStats {
     pub final_message_count: usize,
 }
 
-impl Default for TridentStats {
-    fn default() -> Self {
-        Self {
-            superseded_count: 0,
-            collapsed_chains: 0,
-            messages_collapsed: 0,
-            clusters_found: 0,
-            messages_clustered: 0,
-            tokens_saved_estimate: 0,
-            original_message_count: 0,
-            final_message_count: 0,
-        }
-    }
-}
-
 impl TridentStats {
     pub fn format_report(&self) -> String {
         let compression = if self.final_message_count > 0 {
@@ -192,7 +177,7 @@ fn stage1_supersede(messages: &[ConversationMessage]) -> (Vec<ConversationMessag
 
     let mut obsolete_indices: BTreeSet<usize> = BTreeSet::new();
 
-    for (_path, ops) in &file_ops {
+    for ops in file_ops.values() {
         if ops.len() < 2 {
             continue;
         }
@@ -205,10 +190,8 @@ fn stage1_supersede(messages: &[ConversationMessage]) -> (Vec<ConversationMessag
 
         if let Some(last_write) = last_write_idx {
             for op in ops {
-                if op.op_type == FileOp::Read && op.index < last_write {
-                    obsolete_indices.insert(op.index);
-                } else if (op.op_type == FileOp::Write || op.op_type == FileOp::Edit)
-                    && op.index < last_write
+                if op.index < last_write
+                    && matches!(op.op_type, FileOp::Read | FileOp::Write | FileOp::Edit)
                 {
                     obsolete_indices.insert(op.index);
                 }
@@ -325,7 +308,7 @@ fn stage2_collapse(
                     usage: None,
                 });
             } else {
-                result.extend(buffer.drain(..));
+                result.append(&mut buffer);
             }
             buffer.clear();
             result.push(msg.clone());
@@ -476,7 +459,7 @@ fn stage3_cluster(
     }
 
     let total_clustered: usize = cluster_assignments.len();
-    let clusters_found = cluster_id as usize;
+    let clusters_found = cluster_id;
 
     let mut result: Vec<ConversationMessage> = Vec::new();
     let mut cluster_buffers: BTreeMap<usize, Vec<usize>> = BTreeMap::new();
@@ -677,7 +660,7 @@ fn truncate_text(text: &str, max_chars: usize) -> String {
 mod tests {
     use super::*;
     use crate::compact::CompactionConfig;
-    use crate::session::{ContentBlock, ConversationMessage, MessageRole, Session};
+    use crate::session::{ContentBlock, ConversationMessage, Session};
 
     #[test]
     fn stage1_removes_obsolete_file_reads() {
@@ -736,7 +719,7 @@ mod tests {
     fn stage2_collapses_chatty_messages() {
         let mut messages = vec![];
         for i in 0..6 {
-            messages.push(ConversationMessage::user_text(&format!("ok {i}")));
+            messages.push(ConversationMessage::user_text(format!("ok {i}")));
             messages.push(ConversationMessage::assistant(vec![ContentBlock::Text {
                 text: format!("got {i}"),
             }]));
@@ -767,9 +750,9 @@ mod tests {
                 },
             ]));
             messages.push(ConversationMessage::tool_result(
-                &format!("read_{i}"),
+                format!("read_{i}"),
                 "read_file",
-                &format!(r#"{{"path":"src/{i}.rs","content":"data {i}"}}"#),
+                format!(r#"{{"path":"src/{i}.rs","content":"data {i}"}}"#),
                 false,
             ));
         }

rust/crates/runtime/src/usage.rs

@@ -89,6 +89,19 @@ impl TokenUsage {
             + self.cache_read_input_tokens
     }
 
+    /// Total tokens the provider had to read to serve this call, i.e. the live size of the
+    /// context window.
+    ///
+    /// `input_tokens` alone is NOT that number. With prompt caching a warm call reports
+    /// `input_tokens: 2` and puts the real bulk in `cache_read_input_tokens`, so anything
+    /// budgeting against the context window must sum all three input counters.
+    #[must_use]
+    pub fn total_input_tokens(self) -> u32 {
+        self.input_tokens
+            .saturating_add(self.cache_creation_input_tokens)
+            .saturating_add(self.cache_read_input_tokens)
+    }
+
     #[must_use]
     pub fn estimate_cost_usd(self) -> UsageCostEstimate {
         self.estimate_cost_usd_with_pricing(ModelPricing::default_sonnet_tier())
@@ -243,6 +256,41 @@ mod tests {
         assert_eq!(tracker.cumulative_usage().total_tokens(), 48);
     }
 
+    #[test]
+    fn total_input_tokens_includes_cached_context() {
+        // The shape a warm prompt-cached call actually reports: input_tokens is ~nothing and
+        // the real context size is in the cache counters.
+        let cached = TokenUsage {
+            input_tokens: 2,
+            output_tokens: 1_133,
+            cache_creation_input_tokens: 1_453,
+            cache_read_input_tokens: 111_186,
+        };
+        assert_eq!(cached.total_input_tokens(), 112_641);
+        // Output tokens are not part of the input budget.
+        assert_ne!(cached.total_input_tokens(), cached.total_tokens());
+
+        // A backend without prompt caching reports the whole prompt as input_tokens.
+        let uncached = TokenUsage {
+            input_tokens: 29_056,
+            output_tokens: 73,
+            cache_creation_input_tokens: 0,
+            cache_read_input_tokens: 0,
+        };
+        assert_eq!(uncached.total_input_tokens(), 29_056);
+    }
+
+    #[test]
+    fn total_input_tokens_saturates_instead_of_overflowing() {
+        let usage = TokenUsage {
+            input_tokens: u32::MAX,
+            output_tokens: 0,
+            cache_creation_input_tokens: u32::MAX,
+            cache_read_input_tokens: u32::MAX,
+        };
+        assert_eq!(usage.total_input_tokens(), u32::MAX);
+    }
+
     #[test]
     fn computes_cost_summary_lines() {
         let usage = TokenUsage {

rust/crates/rusty-claude-cli/src/setup_wizard.rs

@@ -2,8 +2,6 @@ use std::io::{self, IsTerminal, Write};
 
 use runtime::{save_user_provider_settings, ConfigLoader, RuntimeProviderConfig};
 
-use serde_json;
-
 const PROVIDERS: &[(&str, &str, &str)] = &[
     ("1", "Anthropic", "anthropic"),
     ("2", "xAI / Grok", "xai"),

rust/crates/rusty-claude-cli/tests/compact_output.rs

@@ -7,7 +7,9 @@ use std::process::{Command, Output, Stdio};
 use std::sync::atomic::{AtomicU64, Ordering};
 use std::time::{Duration, Instant, SystemTime, UNIX_EPOCH};
 
-use mock_anthropic_service::{MockAnthropicService, SCENARIO_PREFIX};
+use mock_anthropic_service::{
+    MockAnthropicService, CONTEXT_WINDOW_RETRY_TEXT, SCENARIO_PREFIX, STREAM_THEN_ERROR_TEXT,
+};
 use serde_json::Value;
 
 static TEMP_COUNTER: AtomicU64 = AtomicU64::new(0);
@@ -189,6 +191,199 @@ fn text_prompt_mode_streams_assistant_text_once_before_spinner_completion() {
     fs::remove_dir_all(&workspace).expect("workspace cleanup should succeed");
 }
 
+#[test]
+fn text_prompt_mode_emits_the_response_body_exactly_once() {
+    // given a workspace pointed at the mock Anthropic service running the
+    // streaming_text scenario, which 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("text-prompt-once");
+    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 in normal text prompt mode
+    let prompt = format!("{SCENARIO_PREFIX}streaming_text");
+    let output = run_claw(
+        &workspace,
+        &config_home,
+        &home,
+        &base_url,
+        &[
+            "--model",
+            "sonnet",
+            "--permission-mode",
+            "read-only",
+            &prompt,
+        ],
+    );
+
+    // then the response body reaches stdout exactly once: the emitter owns the
+    // terminal while the turn runs, so the caller must not print the assistant
+    // text a second time
+    assert!(
+        output.status.success(),
+        "text prompt 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 plain_stdout = strip_ansi_codes(&stdout);
+    assert_eq!(
+        plain_stdout
+            .matches("Mock streaming says hello from the parity harness.")
+            .count(),
+        1,
+        "text prompt stdout should carry the response body exactly once ({stdout:?})"
+    );
+
+    fs::remove_dir_all(&workspace).expect("workspace cleanup should succeed");
+}
+
+#[test]
+fn text_prompt_mode_emits_retried_body_once_after_auto_compact() {
+    // given a mock service whose first answer to this scenario is a context
+    // window rejection and whose answer to the retry is a streamed response
+    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("text-prompt-auto-compact-retry");
+    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 claw is rejected for context window overflow and recovers by
+    // compacting the session and retrying the same prompt
+    let prompt = format!("{SCENARIO_PREFIX}context_window_retry");
+    let output = run_claw(
+        &workspace,
+        &config_home,
+        &home,
+        &base_url,
+        &[
+            "--model",
+            "sonnet",
+            "--permission-mode",
+            "read-only",
+            &prompt,
+        ],
+    );
+
+    // then the retry succeeds and its response body is on stdout exactly once.
+    // The recovery path builds a fresh runtime with its own terminal tail, so
+    // this is where a tail that only tracked the first attempt would leave the
+    // streamed line unterminated and let `spinner.finish` clear it.
+    assert!(
+        output.status.success(),
+        "auto-compact retry 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 plain_stdout = strip_ansi_codes(&stdout);
+    assert_eq!(
+        plain_stdout.matches(CONTEXT_WINDOW_RETRY_TEXT).count(),
+        1,
+        "auto-compact retry stdout should carry the response body exactly once ({stdout:?})"
+    );
+    assert!(
+        plain_stdout
+            .lines()
+            .any(|line| line == CONTEXT_WINDOW_RETRY_TEXT),
+        "the retried response body should land on its own line ({stdout:?})"
+    );
+    assert!(
+        plain_stdout.contains("✨ Done (after auto-compact)"),
+        "the retry should report the auto-compact completion ({stdout:?})"
+    );
+
+    // and the scenario really was requested twice, so the retry path ran
+    let captured = runtime.block_on(server.captured_requests());
+    let messages: Vec<_> = captured
+        .iter()
+        .filter(|request| request.path == "/v1/messages")
+        .collect();
+    assert!(
+        messages.len() >= 2,
+        "the scenario should be requested twice so the retry is exercised ({captured:?})"
+    );
+
+    fs::remove_dir_all(&workspace).expect("workspace cleanup should succeed");
+}
+
+#[test]
+fn text_prompt_mode_keeps_streamed_text_when_the_turn_fails() {
+    // given a mock service that streams part of an answer and then fails the
+    // response, so the turn ends with text already rendered on the terminal
+    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("text-prompt-stream-then-error");
+    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 the turn fails after the emitter already streamed a partial line
+    let prompt = format!("{SCENARIO_PREFIX}stream_then_error");
+    let output = run_claw(
+        &workspace,
+        &config_home,
+        &home,
+        &base_url,
+        &[
+            "--model",
+            "sonnet",
+            "--permission-mode",
+            "read-only",
+            &prompt,
+        ],
+    );
+
+    // then the failed turn is reported...
+    assert!(
+        !output.status.success(),
+        "a mid-stream failure should fail the run\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 plain_stdout = strip_ansi_codes(&stdout);
+
+    // ...and the text the emitter already streamed survives the failure banner.
+    // `spinner.fail` moves to column 0 and clears the current line, so a caller
+    // that does not know a partial line is pending erases that text instead of
+    // printing the banner underneath it. This is the failure arm of the same
+    // ownership rule the success arm follows.
+    assert!(
+        plain_stdout
+            .lines()
+            .any(|line| line == STREAM_THEN_ERROR_TEXT),
+        "the streamed text should keep its own line despite the failure ({stdout:?})"
+    );
+    assert!(
+        plain_stdout.contains("Request failed"),
+        "the failure banner should still be reported ({stdout:?})"
+    );
+
+    fs::remove_dir_all(&workspace).expect("workspace cleanup should succeed");
+}
+
 #[test]
 fn compact_flag_with_json_output_emits_structured_json() {
     let runtime = tokio::runtime::Runtime::new().expect("tokio runtime should build");

rust/crates/rusty-claude-cli/tests/output_format_contract.rs

@@ -451,11 +451,16 @@ fn direct_resume_safe_slash_commands_route_to_local_json_actions_831() {
         .output()
         .expect("git init should launch");
 
-    for (command, expected_kind, expected_status) in [
-        ("/version", "version", "ok"),
-        ("/sandbox", "sandbox", "warn"),
-        ("/diff", "diff", "ok"),
-        ("/status", "status", "ok"),
+    // `/sandbox`'s status reflects the host's real sandbox capability, so it is the one
+    // entry with more than one legitimate value: "ok" where namespace isolation works,
+    // "warn" where it is unsupported and only the filesystem sandbox is active (#731).
+    // This test is about #831 routing — that these commands produce a local JSON action
+    // instead of `interactive_only` — not about the kernel the suite happens to run on.
+    for (command, expected_kind, expected_statuses) in [
+        ("/version", "version", &["ok"][..]),
+        ("/sandbox", "sandbox", &["ok", "warn"][..]),
+        ("/diff", "diff", &["ok"][..]),
+        ("/status", "status", &["ok"][..]),
     ] {
         let output = run_claw(&root, &["--output-format", "json", command], &[]);
         assert!(
@@ -470,9 +475,9 @@ fn direct_resume_safe_slash_commands_route_to_local_json_actions_831() {
             .unwrap_or_else(|_| panic!("{command} must emit JSON (#831), got: {stdout:?}"));
 
         assert_eq!(parsed["kind"], expected_kind, "{command} kind: {parsed}");
-        assert_eq!(
-            parsed["status"], expected_status,
-            "{command} status: {parsed}"
+        assert!(
+            expected_statuses.contains(&parsed["status"].as_str().unwrap_or_default()),
+            "{command} status should be one of {expected_statuses:?}: {parsed}"
         );
         assert_ne!(
             parsed["error_kind"], "interactive_only",
@@ -1007,13 +1012,13 @@ fn inventory_commands_emit_structured_json_when_requested() {
     assert!(
         !plugins
             .as_object()
-            .map_or(false, |o| o.contains_key("reload_runtime")),
+            .is_some_and(|o| o.contains_key("reload_runtime")),
         "plugins list should not include reload_runtime"
     );
     assert!(
         !plugins
             .as_object()
-            .map_or(false, |o| o.contains_key("target")),
+            .is_some_and(|o| o.contains_key("target")),
         "plugins list should not include target"
     );
     // #703: structured summary replaces prose message
@@ -1707,13 +1712,13 @@ fn resumed_inventory_commands_emit_structured_json_when_requested() {
     assert!(
         !plugins
             .as_object()
-            .map_or(false, |o| o.contains_key("reload_runtime")),
+            .is_some_and(|o| o.contains_key("reload_runtime")),
         "plugins list should not include reload_runtime"
     );
     assert!(
         !plugins
             .as_object()
-            .map_or(false, |o| o.contains_key("target")),
+            .is_some_and(|o| o.contains_key("target")),
         "plugins list should not include target"
     );
     assert!(
@@ -2946,7 +2951,7 @@ fn prompt_empty_arg_json_stdout_missing_prompt_823() {
         "claw prompt empty arg must retain abort action (#823); got: {parsed}"
     );
     assert!(
-        parsed["hint"].as_str().map_or(false, |h| !h.is_empty()),
+        parsed["hint"].as_str().is_some_and(|h| !h.is_empty()),
         "claw prompt empty arg missing_prompt hint must be non-empty (#823)"
     );
 }
@@ -2984,9 +2989,9 @@ fn flag_value_errors_have_error_kind_and_hint_756() {
         "invalid --reasoning-effort must be invalid_flag_value (#756): {parsed}"
     );
     assert!(
-        parsed["hint"].as_str().map_or(false, |h| h.contains("low")
-            || h.contains("medium")
-            || h.contains("high")),
+        parsed["hint"]
+            .as_str()
+            .is_some_and(|h| h.contains("low") || h.contains("medium") || h.contains("high")),
         "hint must mention valid values (#756): {parsed}"
     );
 
@@ -3012,7 +3017,7 @@ fn flag_value_errors_have_error_kind_and_hint_756() {
         "missing --model value must be missing_flag_value (#756): {parsed2}"
     );
     assert!(
-        parsed2["hint"].as_str().map_or(false, |h| !h.is_empty()),
+        parsed2["hint"].as_str().is_some_and(|h| !h.is_empty()),
         "missing --model hint must be non-empty (#756): {parsed2}"
     );
 }
@@ -3256,7 +3261,7 @@ fn short_p_flag_swallows_no_flags_755() {
         "flag-like token after -p must be rejected as missing_prompt (#755): {parsed2}"
     );
     assert!(
-        parsed2["hint"].as_str().map_or(false, |h| !h.is_empty()),
+        parsed2["hint"].as_str().is_some_and(|h| !h.is_empty()),
         "missing_prompt hint must be non-empty (#755)"
     );
 }
@@ -3398,7 +3403,7 @@ fn config_unsupported_section_json_hint_741() {
         assert!(
             parsed["supported_sections"]
                 .as_array()
-                .map_or(false, |a| !a.is_empty()),
+                .is_some_and(|a| !a.is_empty()),
             "config {section} JSON must include supported_sections (#741)"
         );
     }
@@ -3455,7 +3460,7 @@ fn export_json_has_kind_702() {
     // On success stdout has kind:export; on failure stderr has type:error.
     // Either way, both envelopes must be valid JSON.
     let stdout = String::from_utf8_lossy(&output.stdout);
-    let stderr = String::from_utf8_lossy(&output.stderr)
+    let _stderr = String::from_utf8_lossy(&output.stderr)
         .lines()
         .filter(|l| l.starts_with('{'))
         .collect::<Vec<_>>()
@@ -3547,8 +3552,8 @@ fn config_parse_error_has_typed_error_kind_and_hint_764() {
         !output.status.success(),
         "malformed settings.json should cause non-zero exit"
     );
-    let stdout = String::from_utf8_lossy(&output.stdout);
-    let stderr = String::from_utf8_lossy(&output.stderr);
+    let _stdout = String::from_utf8_lossy(&output.stdout);
+    let _stderr = String::from_utf8_lossy(&output.stderr);
     let stdout = String::from_utf8_lossy(&output.stdout);
     let json_line = stdout
         .lines()
@@ -3582,8 +3587,8 @@ fn login_logout_removed_subcommands_have_error_kind_and_hint_765() {
             !output.status.success(),
             "claw {subcmd} should exit non-zero"
         );
-        let stdout = String::from_utf8_lossy(&output.stdout);
-        let stderr = String::from_utf8_lossy(&output.stderr);
+        let _stdout = String::from_utf8_lossy(&output.stdout);
+        let _stderr = String::from_utf8_lossy(&output.stderr);
         let stdout = String::from_utf8_lossy(&output.stdout);
         let json_line = stdout
             .lines()
@@ -3722,8 +3727,8 @@ fn resume_non_slash_trailing_arg_has_typed_error_kind_and_hint_768() {
         !output.status.success(),
         "claw --resume latest compact should exit non-zero"
     );
-    let stdout = String::from_utf8_lossy(&output.stdout);
-    let stderr = String::from_utf8_lossy(&output.stderr);
+    let _stdout = String::from_utf8_lossy(&output.stdout);
+    let _stderr = String::from_utf8_lossy(&output.stderr);
     let stdout = String::from_utf8_lossy(&output.stdout);
     let json_line = stdout
         .lines()
@@ -3762,8 +3767,8 @@ fn session_with_unknown_subcommand_returns_interactive_only_not_credentials_767(
             !output.status.success(),
             "claw session {sub} should exit non-zero"
         );
-        let stdout = String::from_utf8_lossy(&output.stdout);
-        let stderr = String::from_utf8_lossy(&output.stderr);
+        let _stdout = String::from_utf8_lossy(&output.stdout);
+        let _stderr = String::from_utf8_lossy(&output.stderr);
         let stdout = String::from_utf8_lossy(&output.stdout);
         let json_line = stdout
             .lines()
@@ -3816,7 +3821,7 @@ fn slash_only_verbs_with_args_return_interactive_only_not_credentials_770() {
             "claw {} should exit non-zero",
             args.join(" ")
         );
-        let stdout = String::from_utf8_lossy(&output.stdout);
+        let _stdout = String::from_utf8_lossy(&output.stdout);
         let stderr = String::from_utf8_lossy(&output.stderr);
         let stdout = String::from_utf8_lossy(&output.stdout);
         let json_line = stdout
@@ -3858,8 +3863,8 @@ fn agents_plugins_mcp_unknown_subcommand_have_hint_774() {
     {
         let output = run_claw(&root, &["--output-format", "json", "agents", "bogus"], &[]);
         assert!(!output.status.success(), "agents bogus should fail");
-        let stdout = String::from_utf8_lossy(&output.stdout);
-        let stderr = String::from_utf8_lossy(&output.stderr);
+        let _stdout = String::from_utf8_lossy(&output.stdout);
+        let _stderr = String::from_utf8_lossy(&output.stderr);
         let stdout = String::from_utf8_lossy(&output.stdout);
         let json_line = stdout
             .lines()
@@ -3882,8 +3887,8 @@ fn agents_plugins_mcp_unknown_subcommand_have_hint_774() {
     {
         let output = run_claw(&root, &["--output-format", "json", "plugins", "bogus"], &[]);
         assert!(!output.status.success(), "plugins bogus should fail");
-        let stdout = String::from_utf8_lossy(&output.stdout);
-        let stderr = String::from_utf8_lossy(&output.stderr);
+        let _stdout = String::from_utf8_lossy(&output.stdout);
+        let _stderr = String::from_utf8_lossy(&output.stderr);
         let stdout = String::from_utf8_lossy(&output.stdout);
         let json_line = stdout
             .lines()
@@ -3902,7 +3907,7 @@ fn agents_plugins_mcp_unknown_subcommand_have_hint_774() {
     {
         let output = run_claw(&root, &["--output-format", "json", "mcp", "bogus"], &[]);
         assert!(!output.status.success(), "mcp bogus should fail");
-        let stdout = String::from_utf8_lossy(&output.stdout);
+        let _stdout = String::from_utf8_lossy(&output.stdout);
         let stderr = String::from_utf8_lossy(&output.stderr);
         let stdout = String::from_utf8_lossy(&output.stdout);
         let json_str = if stdout.trim().starts_with('{') {
@@ -4017,7 +4022,7 @@ fn interactive_only_guard_batch_769_to_771() {
             "claw {} should exit non-zero",
             args.join(" ")
         );
-        let stdout = String::from_utf8_lossy(&output.stdout);
+        let _stdout = String::from_utf8_lossy(&output.stdout);
         let stderr = String::from_utf8_lossy(&output.stderr);
         let stdout = String::from_utf8_lossy(&output.stdout);
         let json_line = stdout
@@ -4081,7 +4086,7 @@ fn resume_plugin_mutations_are_typed_interactive_only_777() {
             !output.status.success(),
             "/plugins {mutation} in resume mode should exit non-zero"
         );
-        let stdout = String::from_utf8_lossy(&output.stdout);
+        let _stdout = String::from_utf8_lossy(&output.stdout);
         let stderr = String::from_utf8_lossy(&output.stderr);
         let stdout = String::from_utf8_lossy(&output.stdout);
         let json_line = stdout
@@ -4138,7 +4143,7 @@ fn resume_skills_invocation_is_typed_interactive_only_779() {
         !output.status.success(),
         "/skills <skill> in resume mode should exit non-zero"
     );
-    let stdout = String::from_utf8_lossy(&output.stdout);
+    let _stdout = String::from_utf8_lossy(&output.stdout);
     let stderr = String::from_utf8_lossy(&output.stderr);
     let stdout = String::from_utf8_lossy(&output.stdout);
     let json_line = stdout
@@ -4179,8 +4184,8 @@ fn acp_unsupported_invocation_has_hint_782() {
 
     let output = run_claw(&root, &["--output-format", "json", "acp", "start"], &[]);
     assert!(!output.status.success(), "acp start should fail");
-    let stdout = String::from_utf8_lossy(&output.stdout);
-    let stderr = String::from_utf8_lossy(&output.stderr);
+    let _stdout = String::from_utf8_lossy(&output.stdout);
+    let _stderr = String::from_utf8_lossy(&output.stderr);
     let stdout = String::from_utf8_lossy(&output.stdout);
     let json_line = stdout
         .lines()
@@ -4217,7 +4222,7 @@ fn init_json_envelope_has_hint_and_already_initialized_783() {
     // Fresh init — already_initialized should be false, hint should mention CLAUDE.md
     let output = run_claw(&root, &["--output-format", "json", "init"], &[]);
     assert!(output.status.success(), "init should succeed");
-    let stdout = String::from_utf8_lossy(&output.stdout);
+    let _stdout = String::from_utf8_lossy(&output.stdout);
     let stderr = String::from_utf8_lossy(&output.stderr);
     let stdout = String::from_utf8_lossy(&output.stdout);
     let raw = if stdout.trim_start().starts_with('{') {
@@ -4286,7 +4291,7 @@ fn init_json_envelope_has_hint_and_already_initialized_783() {
     // Idempotent re-init — already_initialized should be true
     let output2 = run_claw(&root, &["--output-format", "json", "init"], &[]);
     assert!(output2.status.success(), "re-init should succeed");
-    let stdout2 = String::from_utf8_lossy(&output2.stdout);
+    let _stdout2 = String::from_utf8_lossy(&output2.stdout);
     let stderr2 = String::from_utf8_lossy(&output2.stderr);
     let stdout2 = String::from_utf8_lossy(&output2.stdout);
     let raw2 = if stdout2.trim_start().starts_with('{') {
@@ -4367,7 +4372,7 @@ fn export_arg_errors_have_typed_kind_and_hint_784() {
         &[],
     );
     assert!(!out1.status.success(), "--output with no value should fail");
-    let stderr1 = String::from_utf8_lossy(&out1.stderr);
+    let _stderr1 = String::from_utf8_lossy(&out1.stderr);
     let stdout1 = String::from_utf8_lossy(&out1.stdout);
     let j1: serde_json::Value = stdout1
         .lines()
@@ -4394,7 +4399,7 @@ fn export_arg_errors_have_typed_kind_and_hint_784() {
         &[],
     );
     assert!(!out2.status.success(), "extra positional should fail");
-    let stderr2 = String::from_utf8_lossy(&out2.stderr);
+    let _stderr2 = String::from_utf8_lossy(&out2.stderr);
     let stdout2 = String::from_utf8_lossy(&out2.stdout);
     let j2: serde_json::Value = stdout2
         .lines()
@@ -4431,7 +4436,7 @@ fn unknown_subcommand_returns_typed_kind_785() {
     // "dump" is close enough to "dump-manifests" to trigger the typo suggestion path
     let output = run_claw(&root, &["--output-format", "json", "dump"], &[]);
     assert!(!output.status.success(), "unknown subcommand should fail");
-    let stderr = String::from_utf8_lossy(&output.stderr);
+    let _stderr = String::from_utf8_lossy(&output.stderr);
     let stdout = String::from_utf8_lossy(&output.stdout);
     let j: serde_json::Value = stdout
         .lines()
@@ -4479,7 +4484,7 @@ fn dump_manifests_missing_dir_has_typed_kind_and_hint_786() {
         &[],
     );
     assert!(!out1.status.success());
-    let stderr1 = String::from_utf8_lossy(&out1.stderr);
+    let _stderr1 = String::from_utf8_lossy(&out1.stderr);
     let stdout1 = String::from_utf8_lossy(&out1.stdout);
     let j1: serde_json::Value = stdout1
         .lines()
@@ -4511,7 +4516,7 @@ fn dump_manifests_missing_dir_has_typed_kind_and_hint_786() {
         &[],
     );
     assert!(!out2.status.success());
-    let stderr2 = String::from_utf8_lossy(&out2.stderr);
+    let _stderr2 = String::from_utf8_lossy(&out2.stderr);
     let stdout2 = String::from_utf8_lossy(&out2.stdout);
     let j2: serde_json::Value = stdout2
         .lines()
@@ -4561,7 +4566,7 @@ fn resume_directory_path_returns_typed_kind_and_hint_787() {
         !output.status.success(),
         "resume with directory should fail"
     );
-    let stderr = String::from_utf8_lossy(&output.stderr);
+    let _stderr = String::from_utf8_lossy(&output.stderr);
     let stdout = String::from_utf8_lossy(&output.stdout);
     let j: serde_json::Value = stdout
         .lines()
@@ -4611,7 +4616,7 @@ fn skills_show_not_found_emits_single_json_object_788() {
     assert!(!output.status.success(), "skills show unknown should fail");
     // Skills handler emits JSON to stdout; the duplicate was on stderr from the main error path.
     // After fix: stdout has 1 JSON object, stderr has none (no duplicate).
-    let stdout = String::from_utf8_lossy(&output.stdout);
+    let _stdout = String::from_utf8_lossy(&output.stdout);
     let stderr = String::from_utf8_lossy(&output.stderr);
     let stdout = String::from_utf8_lossy(&output.stdout);
 
@@ -4773,7 +4778,7 @@ fn system_prompt_unknown_option_returns_typed_kind_790() {
         &[],
     );
     assert!(!out1.status.success());
-    let stderr1 = String::from_utf8_lossy(&out1.stderr);
+    let _stderr1 = String::from_utf8_lossy(&out1.stderr);
     let stdout1 = String::from_utf8_lossy(&out1.stdout);
     let j1: serde_json::Value = stdout1
         .lines()
@@ -4800,7 +4805,7 @@ fn system_prompt_unknown_option_returns_typed_kind_790() {
         &[],
     );
     assert!(!out2.status.success());
-    let stderr2 = String::from_utf8_lossy(&out2.stderr);
+    let _stderr2 = String::from_utf8_lossy(&out2.stderr);
     let stdout2 = String::from_utf8_lossy(&out2.stdout);
     let j2: serde_json::Value = stdout2
         .lines()
@@ -4838,7 +4843,7 @@ fn config_extra_args_have_non_null_hint_791() {
         &[],
     );
     assert!(!out1.status.success());
-    let stderr1 = String::from_utf8_lossy(&out1.stderr);
+    let _stderr1 = String::from_utf8_lossy(&out1.stderr);
     let stdout1 = String::from_utf8_lossy(&out1.stdout);
     let j1: serde_json::Value = stdout1
         .lines()
@@ -4872,7 +4877,7 @@ fn config_extra_args_have_non_null_hint_791() {
         &[],
     );
     assert!(!out2.status.success());
-    let stderr2 = String::from_utf8_lossy(&out2.stderr);
+    let _stderr2 = String::from_utf8_lossy(&out2.stderr);
     let stdout2 = String::from_utf8_lossy(&out2.stdout);
     let j2: serde_json::Value = stdout2
         .lines()
@@ -5058,7 +5063,7 @@ fn plugins_uninstall_not_found_has_hint_793() {
         "plugins uninstall not-found must exit non-zero (#793)"
     );
     // Error envelope goes to stderr (propagated via ? to main error handler)
-    let stderr = String::from_utf8_lossy(&output.stderr);
+    let _stderr = String::from_utf8_lossy(&output.stderr);
     let stdout = String::from_utf8_lossy(&output.stdout);
     let j: serde_json::Value = stdout
         .lines()
@@ -5103,7 +5108,7 @@ fn plugins_install_not_found_path_returns_typed_kind_794() {
         !output.status.success(),
         "plugins install not-found-path must exit non-zero (#794)"
     );
-    let stderr = String::from_utf8_lossy(&output.stderr);
+    let _stderr = String::from_utf8_lossy(&output.stderr);
     let stdout = String::from_utf8_lossy(&output.stdout);
     let j: serde_json::Value = stdout
         .lines()
@@ -5367,7 +5372,7 @@ fn agents_create_scaffolds_toml_and_lists_locally_431() {
         .iter()
         .any(|agent| {
             agent["name"] == "my-agent"
-                && PathBuf::from(agent["path"].as_str().expect("listed agent path"))
+                && Path::new(agent["path"].as_str().expect("listed agent path"))
                     == fs::canonicalize(&agent_path).expect("canonical listed agent path")
         }));
 }
@@ -5492,7 +5497,7 @@ fn plugins_extra_args_have_non_null_hint_797() {
         !output.status.success(),
         "plugins show with extra arg must exit non-zero (#797)"
     );
-    let stderr = String::from_utf8_lossy(&output.stderr);
+    let _stderr = String::from_utf8_lossy(&output.stderr);
     let stdout = String::from_utf8_lossy(&output.stdout);
     let j: serde_json::Value = stdout
         .lines()
@@ -5558,7 +5563,7 @@ fn plugins_list_trailing_dash_text_error_stays_on_stderr_817() {
         String::from_utf8_lossy(&output.stdout)
     );
     let stderr = String::from_utf8_lossy(&output.stderr);
-    let stdout = String::from_utf8_lossy(&output.stdout);
+    let _stdout = String::from_utf8_lossy(&output.stdout);
     assert!(stderr.contains("[error-kind: cli_parse]"), "{stderr}");
     assert!(
         stderr.contains("unknown option for `claw plugins list`: --"),
@@ -5583,7 +5588,7 @@ fn empty_prompt_has_non_null_hint_798() {
         !output.status.success(),
         "empty prompt must exit non-zero (#798)"
     );
-    let stderr = String::from_utf8_lossy(&output.stderr);
+    let _stderr = String::from_utf8_lossy(&output.stderr);
     let stdout = String::from_utf8_lossy(&output.stdout);
     let j: serde_json::Value = stdout
         .lines()

rust/crates/tools/src/lib.rs

@@ -506,13 +506,17 @@ pub fn mvp_tool_specs() -> Vec<ToolSpec> {
         },
         ToolSpec {
             name: "read_file",
-            description: "Read a text file from the workspace.",
+            description: "Read a text file from the workspace. Large files come back one page at \
+                          a time: if the response has \"truncated\": true you have NOT seen the \
+                          whole file, and \"nextOffset\" is the line to pass as `offset` to \
+                          continue. Prefer grep_search to locate what you need over paging \
+                          through a big file, and never re-read a page you already have.",
             input_schema: json!({
                 "type": "object",
                 "properties": {
                     "path": { "type": "string" },
-                    "offset": { "type": "integer", "minimum": 0 },
-                    "limit": { "type": "integer", "minimum": 1 }
+                    "offset": { "type": "integer", "minimum": 0, "description": "First line to read (0-based). Pass the previous response's nextOffset to continue." },
+                    "limit": { "type": "integer", "minimum": 1, "description": "Maximum lines to return. Capped by a byte ceiling regardless." }
                 },
                 "required": ["path"],
                 "additionalProperties": false
@@ -10401,6 +10405,12 @@ mod tests {
 
     #[test]
     fn repl_executes_python_code() {
+        // Resolving the python runtime reads PATH, and PATH is process-global: the PowerShell
+        // tests below blank it while they run. Without this guard that races and this test
+        // intermittently fails with "python runtime not found" on a machine that has python.
+        let _guard = env_lock()
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner);
         let result = execute_tool(
             "REPL",
             &json!({"language": "python", "code": "print(1 + 1)", "timeout_ms": 500}),