fix(coding-agent): report a subagent's output-token-limit stop as failed, not completed

[jasonkneen]

Problem

packages/coding-agent/src/features/step-subagent.ts isFailed (~292) only treated exitCode !== 0 and stopReason "error"/"aborted" as failure. A child whose final assistant message stopped with stopReason: "length" (the provider's output token limit — the answer is truncated mid-generation, and agent-loop.ts refuses to execute any tool calls the truncated message may contain, see agent-loop.ts:240-246) has exitCode: 0 and a non-error/aborted stopReason, so isFailed returned false. statusForResult in packages/coding-agent/src/features/subagent/execute.ts (~56-59) then reported the record as "completed", and the parent (blocking subagent tool result, resultText/summaries in execute.ts ~400-410, and background-lane notifications via lane-events.ts laneOutputText/notifyLaneFinal) treated the truncated partial answer as a finished result — violating the invariant that hitting a limit must be a distinct terminal status, never "completed".

packages/coding-agent/src/features/workflow/agent-runner.ts also calls the shared isFailed, so a workflow step with a truncated child had the same bug.

Stop-reason propagation itself was already correct: parseJsonEvent (step-subagent.ts) copies message.stopReason from each message_end/tool_result_end event into the running StepSubagentRunResult, and rpc-adapter.ts's settleTurn only overwrites stopReason for the aborted case, so a "length" stop from the last assistant message already reached the parent's StepSubagentRunResult.stopReason. The only propagation gap was errorMessage: providers don't set one for a plain length stop, so the record had no message describing the truncation.

Fix

• isFailed now also treats stopReason === "length" as a failure. This is a minimal, single-source-of-truth change: statusForResult (subagent/execute.ts) already derives its status from isFailed, so the success/failure classification, the success count (records.filter(status === "completed")), telemetry (reportFinished), the TUI status icon, and background-lane routing (background_done vs background_failed in lane-events.ts) all pick up the new outcome for free — no separate status enum was introduced, since the existing 4-value union (running/completed/failed/aborted) already has a distinct non-"completed" terminal value and is switched on in several small consumers.
• parseJsonEvent now fills in a default errorMessage ("Subagent response hit the output token limit; result is truncated.") when the final assistant message's stopReason is "length" and the provider didn't supply its own errorMessage. A provider-supplied errorMessage is preserved.
• resultText now special-cases a "length" stop: it returns the truncation message followed by the partial assistant output (via finalOutput), so the parent-visible text (single-task result, parallel-task summaries, and background-lane <agent-notification> bodies, which all route through resultText) clearly says the answer was truncated and still surfaces the partial content instead of hiding it behind a bare error string.
• Exported statusForResult from subagent/execute.ts (was module-private) so it has direct test coverage.

Test

New file packages/coding-agent/test/subagent-length-status.test.ts:

• isFailed treats a length stop as failure / a normal stop as success.
• statusForResult reports a length stop as "failed" (not "completed") and a normal stop as "completed".
• resultText for a length stop contains "truncat" and the partial output text; unaffected for a normal stop.
• parseJsonEvent sets the default truncation errorMessage when the child provides none, and preserves a provider-supplied errorMessage.
• An end-to-end test driving createStepSubagentExtension's subagent tool (parallel mode, one task) with a mock runner returning a length-stopped result: asserts the tool's returned text contains "0/1 succeeded", mentions truncation, includes the partial output, and details.results[0].status === "failed".

Before the fix, expect(isFailed({ exitCode: 0, stopReason: "length" })).toBe(true) failed (expected false to be true), expect(statusForResult(...)).not.toBe("completed") failed (expected 'completed' not to be 'completed'), and the resultText assertion failed (expected 'partial answer cut off mid-sen' to contain 'truncat'). All 9 new tests pass after the fix (npx vitest --run test/subagent-length-status.test.ts).

Risk / behaviour change

• workflow/agent-runner.ts's WorkflowAgentRunResult.status is now "failed" (not "completed") for a length-truncated workflow step, since it shares isFailed. This is the correct behavior per the invariant, but it is a visible change: a workflow step that previously silently passed truncated output downstream as "completed" will now fail that step.
• Chain-mode subagent calls (if (record.status !== "completed") break; in execute.ts) now stop the chain on a truncated step instead of feeding the truncated text into the next task's {previous} placeholder.
• Telemetry's reportFinished maps the new "failed" outcome to status: "error" (existing mapping, unchanged code) rather than "completed".
• No new status value was added to the "running" | "completed" | "failed" | "aborted" union, so all existing switches/filters over StepSubagentResultRecord["status"] / BackgroundAgentLane["status"] needed no changes.

https://claude.ai/code/session_01GUdnnHEaDThHUATSwXBpV9

[Copilot]

Copilot review overview

🔵 Needs a closer look

Always include the canonical truncation message, even when a provider supplies an error message.

Review effort: Lite
Findings: None

What changed in this PR

Updates subagent handling so output-token-limit stops are reported as failures while preserving truncated output.

Changes:

• Treats stopReason: "length" as a failure.
• Adds truncation messaging and partial-output propagation.
• Adds regression and end-to-end test coverage.

File	Summary
packages/​coding-agent/​test/​subagent-length-status.test.ts	Adds coverage for classification, parsing, output, and tool results.
packages/​coding-agent/​src/​features/​subagent/​execute.ts	Exports status classification for direct testing.
packages/​coding-agent/​src/​features/​step-subagent.ts	Handles length-stop failures and truncated output. Moderate issue (1 vote): provider-supplied error messages can omit the canonical truncation message.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.