headless dispatch worker crashes at startup with thinking-block API 400 #150

New issue

Closed

opened 2026-05-28 12:48:17 +00:00 by coilysiren · 0 comments

coilysiren commented

2026-05-28 12:48:17 +00:00

Owner

A coily dispatch headless worker crashed on its first API call, before doing any work, with:

API Error: 400 messages.1.content.28: `thinking` or `redacted_thinking` blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response.

Details

Run: coilysiren/backend#25, pid 97602, 2026-05-28 05:36 PT.
Log was a single line (the error); no branch/PR created. Dry-run seeding looked correct.
Re-dispatch (pid 14568) launched as recovery; watching whether it recurs.

Why it matters

Headless dispatch is fire-and-forget; a startup crash leaves a single-line log and no PR. Silence looks identical to 'still working' unless something watches the pid + PR state. Suspect the prompt seeding / thinking-block handling on the first turn.

Asks

Determine whether this is transient (Anthropic API) or a seeding bug in dispatch.
Consider a dispatch-side guard: detect immediate child exit with no commit/PR and surface it (nonzero status, or a marker line) rather than leaving a bare error log.

A `coily dispatch headless` worker crashed on its first API call, before doing any work, with: ``` API Error: 400 messages.1.content.28: `thinking` or `redacted_thinking` blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response. ``` ## Details - Run: coilysiren/backend#25, pid 97602, 2026-05-28 05:36 PT. - Log was a single line (the error); no branch/PR created. Dry-run seeding looked correct. - Re-dispatch (pid 14568) launched as recovery; watching whether it recurs. ## Why it matters Headless dispatch is fire-and-forget; a startup crash leaves a single-line log and no PR. Silence looks identical to 'still working' unless something watches the pid + PR state. Suspect the prompt seeding / thinking-block handling on the first turn. ## Asks - Determine whether this is transient (Anthropic API) or a seeding bug in dispatch. - Consider a dispatch-side guard: detect immediate child exit with no commit/PR and surface it (nonzero status, or a marker line) rather than leaving a bare error log.

coilysiren referenced this issue from coilyco-flight-deck/cli-guard

2026-05-28 13:24:40 +00:00

dispatch: detached worker startup crash leaves a bare log, reports success #41

coilysiren referenced this issue from a commit

2026-05-28 13:28:43 +00:00

feat(dispatch): guard against immediate startup crash of detached child

coilysiren referenced this issue from a commit

2026-05-28 13:32:28 +00:00

chore(deps): bump cli-guard for detached-dispatch startup-crash guard

coilysiren closed this issue

2026-05-28 13:32:28 +00:00