fix(threat-model): bound document_endpoint child await with a liveness watchdog

[jorgeraad]

Problem

A recon run gets stuck at N-1/N apps with a permanent spinner: one document_endpoint tool call hangs forever even though the threat-model sub-agent it spawned finished.

Root cause — a tool-await / agent-lifecycle divergence:

• document_endpoint.execute() blocks on await generateThreatModelForEndpoint(...), which awaits a spawned CodeAgent's consume() (threatModelGenerator.ts) with no timeout.
• The only global guard, the 5-min stream idle timeout, is deliberately suppressed while a tool is in flight (createToolExecutionGate), so there is no wall-clock bound on the parent while the child runs.
• When the child wedges in a way its own idle timeout doesn't catch — a hung post-stream step, a stuck provider call, or a consume() that never settles even after the child's session already reached terminal — the await blocks forever. That freezes the tool-call, the parent agent session (status stuck running), and the whole recon.
• PR fix(offSecAgent): emit synthetic tool-result on stream abort/error #779's synthetic-tool-result safeguard can't fire here: the parent's stream is blocked inside execute() and never reaches the finally that emits synthetics.

Confirmed in production data: a stuck "API Endpoints" session with 23 threat-model children all completed, but 2 document_endpoint tool parts with no output.

Fix

Bound the child await on liveness, not a wall-clock deadline (so a slow-but-healthy threat model is never killed):

• A child-scoped AbortController (parent aborts still propagate) lets us cancel just this threat model.
• A watchdog resets on every child stream event (text-delta, tool-call-*, tool-result). text-delta fires per token, so a healthy run resets it constantly.
• If the child goes completely silent past THREAT_MODEL_LIVENESS_TIMEOUT_MS (8 min — comfortably above the child's own 5-min model-idle timeout + auto-resume, so a recoverable stall produces activity and resets the timer first), we abort the child and return null.
• document_endpoint already degrades to the heuristic risk score on a null return, so the tool-call completes and the recon advances instead of hanging.

Single call site (document_endpoint), no behavior change on the happy path. tsc + biome clean.

---

Note

Medium Risk
Changes failure handling for long-running subagents; happy path is unchanged, but silent wedges now drop full threat models in favor of heuristic scoring.

Overview
generateThreatModelForEndpoint no longer waits indefinitely on a wedged threat-model child. It races agent.consume() against a liveness watchdog that resets whenever the child emits stream events (text-delta, tool-call events, tool-result). After 8 minutes of silence, the child is aborted via a child-scoped AbortController (parent abort still propagates), the subagent is marked failed, and the function returns null so document_endpoint can finish with the existing heuristic risk fallback instead of freezing recon.

Watchdog timers and listeners are torn down in finally; errors also abort the child explicitly.

^{Reviewed by Cursor Bugbot for commit 05ff46a. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{Bugbot Autofix is ON. A cloud agent has been kicked off to fix the reported issues.}

^{Reviewed by Cursor Bugbot for commit 05ff46a. Configure here.}

[cursor]

Orphaned consume after race

Medium Severity

When the liveness watchdog wins Promise.race, the function returns after childAbort.abort() without handling the still-running agent.consume() promise. That promise can later reject or resolve, producing an unhandled rejection or stray completion after the threat model was already abandoned.

 

^{Reviewed by Cursor Bugbot for commit 05ff46a. Configure here.}

[cursor]

Long tools trip liveness

Medium Severity

The watchdog only resets on stream events, not while a child tool is executing. The child’s own stream idle timeout pauses during in-flight tools, but this timer keeps counting after tool-call-complete. A legitimate tool run longer than eight minutes (e.g. execute_command without timeout or a long whitebox job) can be misclassified as stuck and return null.

Additional Locations (1)

• src/core/agents/offSecAgent/tools/threatModelGenerator.ts#L387-L393

 

^{Reviewed by Cursor Bugbot for commit 05ff46a. Configure here.}

[Yuvanesh-ux]

@jorgeraad Would you mind cleaning up the comments/making them much more concise.

[jorgeraad]

@Yuvanesh-ux hold up, my b. This should still be a draft. But will do.