[no-mistakes#407] Agent exit-1 with empty stderr fails the whole run — retry the agen...

[github-actions]

Decision needed - no-mistakes#407

Issue triage by Blakeolson21 · issue-triage

Agent exit-1 with empty stderr fails the whole run — retry the agent and surface the real error (usage limit / structured-output retry cap)

Situation

• Compliance: n/a
• Tests: n/a
• Notes: open issue, no linked PR

Triage

• Summary: The reporter asks that a bare exit-1 with empty stderr from the claude CLI agent (seen both when a usage-limit banner is only on stdout and when the CLI's structured-output retry cap is exhausted after all real work is done) be retried and have its real error surfaced instead of silently failing the whole run.
• Product implications: Worth owner discussion: it's a well-documented, well-evidenced report (multiple run IDs, binary strings inspected) proposing a behavior change to the retry/error-classification path in internal/agent/retry.go, which already has a narrow transient-error classifier (Pipeline fails the whole run on a transient Anthropic overloaded_error from the doc-check stage instead of retrying no-mistakes#154 precedent) that could plausibly be extended, but doing so risks masking real failures or retrying non-idempotent side effects if not scoped carefully.
• Recommended next step: discuss - solid repro evidence and a concrete extension point (retry.go's classifyTransient/claudeRetryClassifier), but the fix touches error-classification semantics and overlaps three other open issues (test step: claude exited: exit status 1 when the agent structured result contains an unescaped < (e.g. </dev/null) no-mistakes#391, Agents fallback list no-mistakes#398, v1.31.2 field report: post-receive hook sends --gate . (+ session-limit handling, error classification, trust preflight, default_skips) no-mistakes#400), so scope should be agreed before implementation.

Recommended action

Triage - open issue with no linked PR yet.

Your decision

Tick one box for a quick call, or reply with a slash-command (/close, /decline <reason>, /hold, /comment <text>):

• Close / decline
• Investigate - deep code-grounded review (leaves this card open)
• Hold - I'll handle this manually

_{Only the repository owner can drive this decision - everyone else's edits and comments are ignored.}