agy-proxy: failed agy run reported as successful 'stop' — stream handler never checks the exit code, non-stream masks failure when partial output exists #678

Finding

bin/agy-proxy.mjs reports a failed agy run as a successful completion in most shapes:

• Stream mode — the child.on('close', code => ...) handler (bin/agy-proxy.mjs:467) checks code only for the resume-retry case (code !== 0 && !hasOutput && allowRetry && targetConvId, :471). Every other non-zero exit falls through to the unconditional success ending at :481-510: a chunk with finish_reason: 'stop' plus data: [DONE]. The stderrBuffer collected at :463-465 is never read anywhere in the stream path, so the failure cause is discarded entirely.
• Non-stream mode — the failure branch is gated on empty output (if (code !== 0 && !responseText), :553), so a run that fails after producing a partial response returns HTTP 200 with finish_reason: 'stop' and the truncated text. (This is the sighting noted for separate judgement in muse-proxy: a run that fails after producing partial output is reported as a successful 'stop' — failure is masked whenever any content exists #672; the stream half above was not covered there or anywhere else.)

This is the agy-proxy counterpart of the defects already fixed in the sibling bin/muse-proxy.mjs (#670, #672, #674) and bin/codex-proxy.mjs (#676): a failed run presented as a clean stop.

Evidence (reproduced on main 4b85ef69)

Ran bin/agy-proxy.mjs with a stub agy at $HOME/.local/bin/agy (the proxy's own binary resolution) whose behaviour keys on the requested --output-format:

• Stub prints stub agy: auth expired to stderr and exits 3, no stdout:
  • stream: HTTP 200, one chunk with finish_reason: "stop", then data: [DONE] — no error anywhere
  • non-stream: HTTP 502 {"error":"agy process exited with non-zero code","code":3,"stderr":"stub agy: auth expired"} (correct — the control case)
• Stub emits one step_update text delta (partial answer) / {"response":"partial answer"}, prints stub agy: run failed midway to stderr, exits 3:
  • stream: HTTP 200, content chunk partial answer, then a finish_reason: "stop" chunk + [DONE]
  • non-stream: HTTP 200, content: "partial answer", finish_reason: "stop"
• Stub success case (delta + result event / {"response":"full answer"}, exit 0) correctly returns 200 with the full text and stop in both modes.

Dedup: searches for "agy-proxy" / "agy" across open and closed issues return only #434/#435/#437/#483/#661/#672 — #672 notes the non-stream gate at :553 as "judge separately" and explicitly excludes it from its acceptance criteria; none covers the stream handler or failure detection in this proxy. Cross-repo: bin/agy-proxy.mjs exists only in this repo (no copy in developing-today/opencode, developing-today/openchamber, dezren39/nix).

Why it matters

OpenCode consumes this bridge in streaming mode as a provider. A failed agy run (expired auth, CLI crash, model rejection) presented as a clean stop with empty or truncated text is worse than an error: downstream code and users treat it as the complete answer, and an auth expiry presents as "the model said nothing" instead of a diagnosable failure. The two modes also disagree with each other — and with the proxy's own correct empty-output non-stream case (502 with stderr).

Acceptance criteria

• A child that exits non-zero (or is killed by a signal — code is null then) is reported as a failure in both modes, regardless of how much content was already produced: non-stream returns an error status (the existing 502 shape, with stderr, is fine) instead of 200/stop; stream mode emits a data:-framed error event naming the failure (last stderr line when present) followed by data: [DONE], and no chunk claims finish_reason: "stop".
• The resume-retry behaviour is preserved: a failed resume (code !== 0, no output, a conversation was being resumed) still retries once as a fresh conversation before any failure is reported.
• Successful runs (exit 0) are unchanged in both modes.
• A regression test with a stub agy (fail-with-no-output and fail-after-partial-output cases, both modes — the bin/codex-proxy.test.mjs / bin/muse-proxy.test.mjs stub pattern; there is currently no bin/agy-proxy.test.mjs) passes; the existing proxy test suites still pass.