Single-tier gpt-5.5 review (retire cheap->strong cascade) - #3
Conversation
Testing showed the cheap tier added no value online: it fails closed on large diffs and otherwise just escalates anyway, so every PR paid latency for no decision. Collapse the two-pass cascade to a single enforced pass that injects the strong-tier facts, so the DSL routes one model (gpt-5.5). PROMOTE/HELD stay false and cheap_gate is never written, so the promotion label and cheap report steps no-op and the gate does a plain single-pass gate; exhaustive extra passes are preserved. The injected facts and DSL rule shape are kept so a future size-based front-routing policy can re-diverge without an Action change. The recipe notes the model in each branch is freely swappable. Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
Add a pull_request-triggered workflow using `uses: ./` so a PR is reviewed by the action code it changes — the only way to test action changes end-to-end before moving the release tag. Uses pull_request (not pull_request_target) so it runs the PR's code, not the base branch's. Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
There was a problem hiding this comment.
🐳 Orca-Code-Review
Found 2 issues in this PR: 🔴 1 P0 · 🟠 1 P1.
Reviewed via OrcaRouter — Route Smarter. Ship Safer. Spend Less.
Orca-Code-Review — push 4
Tier: STRONG (final pass) — blocked ❌ 1 finding blocks merge |
|
To use Codex here, create a Codex account and connect to github. |
|
@codex review |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9df4d7e830
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Fork PRs never receive ORCAROUTER_API_KEY, so gate the pull_request job on head repo == this repo — otherwise they reach the action with an empty key and fail the required check rather than skipping. Correct the header comment that wrongly claimed forks were "effectively skipped". Also note in the recipe that the router DSL lives in the OrcaRouter workspace, not the action, so an action bump does not update it — re-paste to adopt single-tier gpt-5.5. Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
There was a problem hiding this comment.
🐳 Orca-Code-Review
Found 2 issues in this PR: 🟠 1 P1 · 🟡 1 P2.
Reviewed via OrcaRouter — Route Smarter. Ship Safer. Spend Less.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d9bc177173
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
On /orcarouter-review (issue_comment) the event carries no PR ref, so the default checkout landed on the base branch and `uses: ./` resolved the action from main — running the OLD action code instead of the PR's, defeating the self-review purpose for the comment trigger. Pin checkout to refs/pull/N/merge on the comment path so both triggers test the PR's own action definition. Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
There was a problem hiding this comment.
🐳 Orca-Code-Review
Found 1 issue in this PR: 🔴 1 P0.
Reviewed via OrcaRouter — Route Smarter. Ship Safer. Spend Less.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 7e58980c22
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
The /orcarouter-review (issue_comment) path let a maintainer trigger a run on a FORK PR, which then checked out the fork's ref and executed its PR-controlled `uses: ./` action code with the real ORCAROUTER_API_KEY + write token — a secret-exfil vector (this repo runs the LOCAL action, unlike adopters who pin a published tag). It also resolved the action from the default branch on the comment path, running stale code. Remove the comment trigger entirely: the pull_request auto-trigger already covers self-review, and re-review on demand is a re-run from the Actions tab. Gate to same-repo PRs, add `reopened`, and document the accepted (push-access-bounded) self-test risk. Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
There was a problem hiding this comment.
🐳 Orca-Code-Review
Found 1 issue in this PR: 🟠 1 P1.
Reviewed via OrcaRouter — Route Smarter. Ship Safer. Spend Less.
1 participant
Summary
openai/gpt-5.5(the validated quality bar) and document that the model in eachuse:is freely swappable — re-diverging into a real cost cascade is just a matter of giving branches different models.PROMOTE/HELDstay false andcheap_gateis never written, so the promotion-label step and the cheap report step no-op and the Enforce step does a plain single-pass gate. Exhaustive extra passes are preserved.Why
Testing against real PRs showed the cheap tier added no value online: it fails closed on large diffs, and otherwise just escalates anyway — so every PR paid latency/cost for no decision. The injected facts and DSL rule shape are kept intact so a future size-based front-routing policy (small→cheap, large→strong, decided up front) can re-diverge without an Action change.
Test plan
action.yml+ recipe parse as valid YAML.block-onseverities from the single pass.🤖 Generated with Claude Code