fix(delegation): salvage a partial summary when a subagent times out

[juniperbevensee]

The bug

When a child exceeds delegation.child_timeout_seconds, the parent abandons its future — so run_conversation never returns and its final_response is lost. The parent got summary: None regardless of how much work the child had completed.

Found while debugging a live agent. Three research subagents hit the 600s cap:

WARNING tools.delegate_tool: Subagent 1 timed out after 602.9s
WARNING tools.delegate_tool: Subagent 0 timed out after 602.9s
WARNING tools.delegate_tool: Subagent 2 timed out after 600.1s
  ✗ [1/3] Research whether Andreessen Horowitz (a1  (602.94s)
  ✗ [2/3] Research the full scope of the Rød-Larse  (602.93s)
  ✗ [3/3] Research the 'attorney-client privileged  (600.07s)

Each had completed dozens of tool calls. All three produced nothing.

It compounds: because no summary file is spilled on timeout, the parent's follow-up read_file fails too —

Tool read_file returned error: {"error": "File not found:
  /opt/data/cache/delegation/subagent-summary-0-20260729_120946_475..."}

So the model sees a missing-file error stacked on top of the timeout.

The fix

run_conversation publishes its live transcript as agent._session_messages at turn start. This is a one-line addition — the loop already re-published at the sole rebind site, so the reference stays valid for the whole turn (verified: exactly two binds in the function).

On timeout, _salvage_partial_summary rebuilds the child's trailing assistant prose plus a tool trace from that transcript, spills the full text to the delegation cache, and returns it as summary with partial: true. Summary-budget trimming is already status-agnostic, so the partial flows through the existing context-protection path unchanged.

On honesty of the salvaged output

The text is prefixed with an explicit banner:

[PARTIAL — subagent timed out before writing a final summary. 47 tool call(s): 41 ok, 6 errored. The text below is the child's last reasoning, not a considered wrap-up; treat conclusions as provisional and verify anything load-bearing.]

A timed-out child never wrote a wrap-up. Its last reasoning turn is mid-thought, and the parent must not read it as a conclusion — hence the tool-call tally (including error count) so the parent can judge how much to trust it.

Salvage is best-effort and never raises; a failure there must not mask the underlying timeout.

Testing

• 15 new tests — trailing-turn retention, tool-call/result pairing by tool_call_id (deliberately out of order, to prove it isn't positional), the PARTIAL banner, trace-without-prose, and 7 malformed-transcript cases including a property that raises.
• One test pins the _session_messages publish contract itself, so the salvage can't be silently broken by removing that line.
• Ran every suite touching _session_messages plus the delegate suites: 1232 passed, 3 skipped.
• test_approval.py::TestDetectDangerousRm::test_nonrecursive_verification_artifact_cleanup_is_not_dangerous fails — it reproduces on a pristine origin/main and is unrelated to this change.

Related

Complements #14726 (0-API-call timeout diagnostics). That covers children that never started; this covers children that did substantial work and got cut off.

🤖 Generated with Claude Code