feat(ce-doc-review): add cross-model judgment pass#1092
Conversation
Run the conditional judgment trio (adversarial, product-lens, security-lens) through a different model family in a read-only process, folding each in as <reviewer-name>-<peer> so agreement with its in-process twin promotes the finding (synthesis 3.4). Mirrors ce-code-review's cross-model design, adapted for documents: content is embedded (no diff/base-ref), the peer is tiered per lens (security-lens -> flagship/medium; adversarial/product-lens -> mid/high), and the pass is non-blocking and runs in both interactive and headless modes. feasibility and the convergent lenses (coherence, scope-guardian) stay single-model so the pass only fires on documents that activate a conditional judgment lens. The bundled script derives the persona-brief filename from the allowlisted reviewer-name rather than accepting it as an argument, so no caller value reaches the brief-read path (closes a path-traversal -> third-party-egress vector). It embeds full document content into the peer prompt and emits a one-line stderr audit log of the cross-model send so the egress is auditable even in headless mode.
Add the cross-model judgment-pass section and FAQ to the skill doc page, and land the implementation-ready unified plan the feature was built from.
![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: 7072049f8e
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 cross-model peer prompt composed only the persona brief plus a hand-built <review-context>, omitting the <context-slots-rules> block the in-process reviewer receives from subagent-template.md. On unified-* artifacts the trio personas (which only define adaptation for bare requirements/plan) would skip their adaptation entirely, and the peer could not interpret non-path Origin values (product_contract_source:<v>). Extract the canonical block at runtime (single source of truth) and fold it into the peer prompt. Addresses PR review feedback (#1092)
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: c5b8452c7e
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".
| run_codex() { | ||
| local prev; case "$-" in *m*) prev=1;; *) prev=0;; esac | ||
| set -m | ||
| command codex exec - -C "$RUN_DIR" -s read-only -o "$OUT" -m "$CODEX_MODEL" \ |
There was a problem hiding this comment.
When the peer is codex on Claude/Cursor hosts, this sets Codex's workspace root to the scratch run dir. The reference tells callers to pass a /tmp/compound-engineering/... run dir, while the Codex CLI reference documents --cd/-C as setting the workspace root and --skip-git-repo-check as the flag that allows running outside a Git repository; with no skip flag, codex exec exits before reviewing in the common /tmp case (or lacks repo context if that check is disabled), so the cross-model pass silently produces no findings. Use the reviewed repo/document root for -C and keep only the output path in the run dir, or explicitly opt into non-repo execution while still granting the repo as context.
Useful? React with 👍 / 👎.
1 participant
What
Ports
ce-code-review's cross-model adversarial pass toce-doc-review. Whichever of the conditional judgment trio —adversarial-document-reviewer,product-lens-reviewer,security-lens-reviewer— activates for a document also runs on a different model family (Codex↔Claude) in a separate read-only process, folding in as<reviewer-name>-<peer>where agreement with its in-process twin promotes the finding (synthesis 3.4).Why
ce-doc-review's judgment lenses run single-model, so their blind spots are the host model's.ce-code-reviewalready closes this for its adversarial lens; running the same review through another model manually on documents reliably surfaces issues the single-model pass misses.ce-doc-reviewdiffers in one way that shapes the port — its high-value judgment is spread across several lenses, not concentrated in one — so a faithful adversarial-only mirror would leave most of the upside unclaimed.Key decisions
security-lensis knowledge-bound → flagship model (gpt-5.6-sol/opus);adversarial+product-lensare reasoning-bound → mid model at high reasoning (gpt-5.6-terra/sonnet). Concrete IDs are a documented maintenance point, not the contract.Security
The peer prompt embeds full document content and sends it to a third-party provider (a wider egress than a diff-only review). Mitigations: the peer runs strictly read-only (codex
-s read-only; claudedontAskwith mutators/MCP/subagents denied), so impact is bounded to disclosure not repo mutation; the script emits a one-line stderr audit log of each send so the egress is auditable even in headless mode.The bundled script derives the persona-brief filename from the allowlisted
reviewer-namerather than accepting it as an argument — closing a path-traversal→arbitrary-file-egress vector an independent code review surfaced. This simplified the script signature from the plan's stated 7 args to 6 (<peer> <reviewer-name> <document-path> <document-type> <origin> <run-dir>); the plan body was left as its point-in-time decision record.Validation
bun test→ 1909 pass / 0 fail;bun run release:validate→ in sync (29 skills, no count drift).../-bearing value now hitting the allowlist); KTD5 normalization confirmed viajqfixtures (reviewer rewrite to<reviewer-name>-<peer>, soft-array backfill, non-array-findings drop, contract predicate).references/cross-model-eval.md); running it is the deferred interactive step, sincebun testdoes not exercise SKILL.md/reference prose (perAGENTS.md"Validating Agent and Skill Changes").Built via the compound-engineering pipeline (brainstorm → plan → work → simplify → review → ship).