Investigate Devin writer exit zero without a final exported response

[arjitj2]

Observed failure

A configured devin:swe-2@high isolated writer ran from 2026-09-26 23:57:44.243 UTC to 2026-09-27 00:35:34.255 UTC (2,270,012 ms). The direct child exited 0 with no signal. Pstack returned exit 65, status: malformed-output, failurePhase: postprocess, processStarted: true, and terminalSuccess: false, with devin export did not end with a final agent response.

The worker left substantial incomplete edits. They were preserved in the draft work for #57 (PR #59, checkpoint a9a3f9ca51dfabce483166fa7964f02bc4ff3f1b). That task owns those edits and its separately authorized continuation; this investigation must not change its checkout or replay its writer.

The receipt records no explicit deadline (timeoutMs: null) and apiSpend: deny. Its prompt assigned Git mutations to the parent. The owning task reports that inherited refusal fallback was explicitly unset. Neither proves a cause for this exit.

The named private ATIF export, its directory, and accepted-output file are absent on inspection; the failed-writer snapshot remains. Runner cleanup deletes its private export directory in finally. Only the bounded narration prefix in the old receipt remains, which does not expose the final ATIF step or establish the producer's termination reason. Do not infer an export race, permission denial, quota exhaustion, context limit, or provider timeout from these facts.

Relationship to existing work

• Investigate Devin SWE-2 nonfinal read-only exits and writer transport failures #30 / PR Fix Devin read-only tool selection #31 fixed read-only selection of a denied shell tool. This recurrence is an isolated writer; the same cause has not been established. Track it separately instead of reopening the resolved read-only fix.
• Investigate apparent Devin writer stall during a long generation #48 concerns a still-running worker that eventually resumed. Here the direct child has exited.
• Disable inherited Devin refusal fallback in assigned worker lanes #53 prevents inherited refusal fallback; Require Codex turn completion before accepting worker output #51 fixes Codex completion. Neither diagnoses this Devin exit.
• Exclude intermediate provider transcripts from failure receipts #54 owns privacy-safe failure evidence. Coordinate any structural diagnostic need there; do not retain raw exports by default or restore transcript prefixes.
• Expose live worker lifecycle progress without changing completion or retry semantics #55 adds lifecycle visibility, not missing final answers or automatic recovery.
• Enforce parent-owned worker Git and support bounded operational continuation #57 owns worker contracts and explicitly authorized operational continuation. A generic malformed export is not verified permission interruption and must not gain retry eligibility through this investigation.

Investigation and completion criteria

1. Preserve the original receipt/event/decision and reviewed partial-work checkpoint outside the repository. Record exact runner revision and Devin CLI version for any new observation; do not infer the historical version from today's executable.
2. Inspect the current print/export contract and existing tests for final-step acceptance, process exit, pipe draining, and cleanup. Determine which structural evidence is missing before attempting a reproduction.
3. If a fresh isolated reproduction is warranted and allowed by the saved model/spending policy, use a small disposable task, fresh paths, parent-owned Git, and no modifications to the original failed writer. Capture only allowlisted terminal/export structure and source-backed termination codes, without prompts, reasoning, file contents, or raw tool output. Preserve evidence before cleanup without weakening Exclude intermediate provider transcripts from failure receipts #54's privacy contract.
4. Distinguish a provider stopping without a final response from an adapter rejecting a valid final export or reading it prematurely. Fix only a demonstrated adapter defect. If the provider is responsible, prepare a sanitized reproducible upstream report; do not guess from the historical truncated receipt.
5. Keep nonfinal exports rejected. Child exit zero alone cannot publish a complete task. No implicit timeout, automatic retry, unsaved fallback, broadened permissions, ACP migration, or default raw transcript retention.

There is currently no demonstrated implementation fix. The historical cause is unverified because the original export is unavailable. Any eventual product change needs the repository's exact-candidate tests, versioning, and installed-parent verification gates before merge.

[arjitj2]

Related but distinct: #72 now owns the directly reproduced denied-MCP termination, upstream CognitionAI/devin-cli#15 . Do not mark #60 fixed or duplicate solely from that reproduction: this writer's original export is unavailable, so its cause remains unknown. Its reviewed partial-work checkpoint and historical evidence remain preserved.

[arjitj2]

Read-only assessment completed on September 28. The historical cause remains unknown, and the available artifacts do not justify a fix, another writer run, or an upstream report.

The preserved receipt confirms a started devin:swe-2@high isolated writer, child exit 0, no signal, no explicit deadline, apiSpend: deny, and postprocess malformed-output with terminalSuccess: false. The saved event is terminal-failure; the saved inspection is unsafe; the decision is stop/not-eligible. Those records do not authorize replay or permission-denial recovery. The prompt assigns Git mutations to the parent and directs sandboxed exec for edits, but instructions do not prove which tools the worker actually used.

The original ATIF export, its private directory and accepted output remain absent. The receipt's old narration prefix is not terminal evidence and is not reproduced here. The prompt names baseline 84662d0; checkpoint a9a3f9c preserves partial work. Neither attests the exact launched runner bytes or historical Devin version. The original owner confirms no additional trustworthy historical version or terminal evidence is available. Today's executable version must not be assigned retroactively.

In the current parser at 12a8f97, the recorded error covers several final-step failures: no final agent step, nonempty or malformed tool calls, or a missing, non-string or empty message. The error cannot distinguish them. The current runner waits for child exit and stdout/stderr draining before reading the export, then removes the private export directory in finally. Source inspection establishes that ordering, not the historical provider's export-completion behavior. It provides no basis to diagnose a race, quota, context limit, permission refusal or provider timeout.

PR #59's new native-write-denial proof is separate. Its inspected classifier at 829c189 needs a specific terminal native write call and its matching provider observation. #60 retained neither. Later support for narrated call steps does not recover the missing historical facts. No duplicate implementation is proposed, and #72/upstream Devin #15 remain distinct.

The smallest useful next step is capture at the existing export-assessment boundary during a future concrete, authorized disposable reproduction, before cleanup. Record the exact runner/package and CLI version/hash, exit/signal/deadline, export-read success, schema recognition, a bounded final-source category, message type/nonempty boolean, tool-call shape/count, observation count and whether references match. A known provider-owned denial may become a fixed allowlisted diagnostic code; unmatched observations must remain unknown. Do not retain prompt text, message content, reasoning, arguments, paths, raw observations or raw exports. Preserve reviewed partial work separately. If a read/export-ordering claim becomes testable, it needs producer-completion evidence, not an added sleep or retry.

No historical writer was resumed, no provider inference ran, and no other implementation checkout or session database was inspected. Three existing offline process tests passed with eight assertions on unchanged 12a8f97, covering omitted deadlines, an explicitly bounded descendant-held pipe, and post-exit cancellation. The earlier 43-test Devin suite on that same source also passed; neither result diagnoses this historical failure. Original artifact hashes and a sanitized structural assessment are retained outside the checkout at /Users/arjitjaiswal/open-pstack-verification-issue60-20260928.

Keep #60 open as an unresolved historical failure with insufficient terminal evidence. Stop this investigation here until a concrete reproduction or trustworthy new artifact makes a causal test possible. No PR, merge request, raw-export retention change or upstream report is warranted by this assessment.

[arjitj2]

Reproduced while preparing #104. A devin:swe-2@high writer finished its CLI process with exit zero after approximately 726 seconds, but the ATIF export did not end with a final agent response. The runner correctly recorded malformed-output; no provider-owned tool denial was present and no saved fallback applied. Partial scoped edits were retained and independently inspected, without retrying that assignment. Private export/system context is not published. This occurrence supports keeping final-response validation strict while investigating the provider export behavior separately.