Retry /ship's second-model review on Astra, then Gemini, when OpenAI refuses it as a cyber-security risk (GST-2649) (#4350)

* Retry /ship's outside voice on Astra, then Gemini, when OpenAI refuses it under cyber_policy (GST-2649)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* Register the cyber_policy refusal fixtures as free test inputs (GST-2649)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* Retry a refused second-model review only when it left no review (GST-2649)

A voice the host marks refused under cyber_policy but that still wrote findings is now collected as that review, with its own outside record, instead of starting a redundant retry and losing its record. The cyber_policy test header now matches what the classifier does with a host error.

Review-Finding: p1/red-team#2 reader=red-team
Review-Finding: p1/specialist-batched#1 reader=testing
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* Skip the Gemini retry when the lane or native reviewer is already a Google model (GST-2649)

On a Google-family lane, Step 11 refuses a Gemini outside voice, so two OpenAI refusals used to plan a Gemini retry that stopped the pass on every resume. The retry chain now skips a step whose family the lane or native reviewer holds and records the second-family read as missing.

Review-Finding: p1/outside-voice#1 reader=outside-voice
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* Say in troubleshooting that a retry is skipped when its model family is the lane's (GST-2649)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

---------

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
This commit is contained in:
Garry Tanandcapy-ai[bot] authored and GitHub committed 2026-11-14 02:23:48 +00:00
1 parent 654ad4b6fe
commit 4bee369d4a
13 files changed
+390 -20

No files matched your search

+27
View File
@@ -339,6 +339,33 @@ network.
**Expected result.** `OUTSIDE_STATUS: completed`.
<a id="outside-review-cyber-policy"></a>
### `... outside review unavailable: OpenAI declined the review under its cyber-security policy (cyber_policy) (...)`
**Meaning.** OpenAI flagged the review brief as possible cybersecurity risk
and refused it (Capy relays it as `OpenAI key declined the request:
cyber_policy: ...`). It is not an auth, quota or model problem, and sending the
same brief to another OpenAI model gets the same answer: GPT-6.1 Sol and GPT-6
Astra both refused the same two briefs in the 2026-11-13 test.
**Fix.** Rerun the review once on GPT-6 Astra with the brief reworded as a
read-only correctness review (no attacker role-play, no exploit code or
reproductions; each defect as file:line, the condition under which the code
returns the wrong answer, and the smallest fix), or on Gemini 4 Carbon with the
brief unchanged. Capy's /ship does both by itself: the pass prints
`RETRY: reader=outside-voice/astra-read-only ...`, then
`RETRY: reader=outside-voice/gemini-4-carbon ...` if Astra refuses too. A
retry whose model family is already the lane's or the native reviewer's (an
OpenAI or Google lane, say) is skipped, because Step 11 refuses an outside
voice of either family. Other
hosts' outside voices (the Codex CLI, the api runner) only report this reason;
rerun them by hand. When every attempt is refused, `RETRY: exhausted` and the
Step 11 record's `outside_note` say the second-family read is missing: the
native review still runs and the PR body lists the gap as missing coverage,
never as a pass.
**Expected result.** `OUTSIDE_RECORD: ... "outside_status":"completed","outside_note":"second-family read by ..."`.
<a id="outside-review-truncated"></a>
### `... outside review unavailable: the reviewer hit its output token limit and the response is cut off (...)`