lsm / neokai / 27853891521

82%

Build Type

push

github

Committed by web-flow

Commit Message

feat(forge): break repeated tool-use errors with conversation_friction evidence (#2137)

* feat(forge): break repeated tool-use errors with conversation_friction evidence

Adds RepeatedToolErrorGuardrail to SDKMessageHandler. After 2 consecutive
identical tool-use errors (same tool name + normalized error substring)
within a session, the guardrail emits a synthetic assistant recovery
message prompting re-validation / alternative path / operator escalation,
and logs a conversation_friction evidence row on the owning task's
evolution scope with metadata {tool, error, count}.

The guardrail records tool_use ids -> names from assistant messages and
observes user-message tool_result blocks for is_error=true, resetting
the streak on any successful result or differing error.

* fix: scope repeated tool error guardrail to Forge tasks and cap state growth

- Only observe errors and intervene when the session is attached to a Forge task with an evolution scope.

- Cap tool_use_id -> tool_name lookup map to avoid unbounded growth.

- Add a per-(tool,error) cooldown after intervention to prevent evidence/message spam.

- Fix the 'error message changes' unit test and add coverage for no-task/no-scope/map-cap cases.

* fix: route recovery through message queue and resolve task from session context

- Rename recovery callback to routeRecoveryMessage and enqueue the recovery text via MessageQueue so the running SDK turn receives it.

- Resolve Forge tasks from session.context.taskId as a fallback to getTaskBySessionId for node-agent worker sessions.

- Deduplicate identical (tool, error) blocks within a single user message so parallel batched failures do not count as retries.

- Add unit tests for batched-error deduplication and cooldown behavior.

* fix: reset stale error streaks on non-error user messages and clean merge artifacts

- observeToolResultErrors now resets the streak for plain-text user messages and any message without error tool_result blocks, so a sin... (continued)

Coverage Stats

10812 of 15739 branches covered (68.7%)

Branch coverage included in aggregate %.

188 of 215 new or added lines in 2 files covered. (87.44%)

1 existing line in 1 file now uncovered.

92386 of 110631 relevant lines covered (83.51%)

297.28 hits per line