Subagent reports success with repetitive output and no task artifacts

[acoliver]

During #2616 work on September 22, 2026, functions.task with subagent_name: coder returned terminate_reason: GOAL for two test-preservation assignments without the requested handoff file or verifiable implementation. A subsequent MCP cutover task (coder-t9v77l) returned GOAL with an extremely long repetitive response containing phrases such as “The verification is complete” and “Sixteen,” but did not update tmp/verify2616-consolidated/mcp-credential-owner-cutover/handoff.md or provide commands, logs or changed-file evidence. The last handoff still explicitly marks migration and verification as pending. Earlier coder-nypx48 failed after six retries with zai-org error: 3021 (inference request/minute rate limit). This followed a provider/model change intended to restore cheap subagent work.

Expected: GOAL should mean the requested deliverable exists, or the worker should report a concrete failure/blocker. Repetitive generation should stop rather than falsely claiming tests passed. No work was accepted as verified from these outputs.

The tasks were run in branch-2 llxprt-code, SAMEPR3741, using the coder profile. Relevant checkpoints: tmp/verify2616-consolidated/mcp-credential-owner-cutover/handoff.md; requested but absent: tmp/verify2616-consolidated/legacy-task-pending-launch-preservation/handoff.md.

[github-actions]

LLxprt Issue Planner — infrastructure failure

The autonomous planner failed with exit code 1. Please inspect the workflow logs (Run planner agent section) and re-run once resolved.

[impartshadow]

This is a false-completion failure: the task returned terminate_reason: GOAL while the requested deliverable never appeared on disk.

My hypothesis is that GOAL is set when the worker stops generating, not by any check on the output. The issue describes an "extremely long repetitive response" containing "The verification is complete" and "Sixteen". That reads like a generation collapse after the provider/model swap, and the text sounds finished, so it passes as done. The issue's own evidence contradicts the status: "The last handoff still explicitly marks migration and verification as pending", and the two requested handoffs are missing. The zai-org error: 3021 failure on coder-nypx48 at least surfaced as a failure. The GOAL cases are the dangerous ones.

The deterministic check belongs at the task tool's return boundary, before GOAL is shown to the parent:

for p in assignment.deliverable_paths:
    assert exists(p) and mtime(p) > task_start and hash(p) != hash_at_launch
assert not claims_tests_passed(final_text) or transcript.has_executed_test_cmd()
# any failure -> terminate_reason = GOAL_UNVERIFIED, not GOAL
# stream guard: abort if n-gram repetition over the last ~2k tokens exceeds a threshold

That would have turned all three runs into explicit failures without anyone reading the transcript.

I haven't seen the raw transcripts, so I'm unsure whether the cause is the model or the status mapping. If the coder-t9v77l log shows real file writes elsewhere, or a max-token or stop-sequence exit being mapped to GOAL, the cause is the mapping and not model collapse.

Advisory SFA-2026-D9CB7D229A · Full redacted write-up: https://impartshadow.github.io/echo-site/failure-cases/subagent-reports-success-with-repetitive-output-and-d9cb7d229a.html
If this diagnosis is useful, react 👍 to this comment; if it is wrong, react 👎 — I use the verdict to choose what I investigate next.