Paper review lite codex

Paper review lite: Codex lead, cross-model audit

Run Codex as the lead and use Claude Code's non-interactive CLI as the independent model family. This is the Codex-side equivalent of the Claude skill with the same name; the direction of delegation is reversed.

Read ../paper-review-lite/SKILL.md completely before starting. It defines the review dimensions, severity rubric, evidence requirements, and final report format. If the sibling skill is unavailable, stop and tell the user to install it; do not reconstruct a partial protocol from memory.

Sandbox constraint — read before the first claude -p call

claude -p is a different binary than codex exec, so it does not hit the in-process IPC failure that breaks nested codex exec calls under sandbox. But under workspace-write sandbox, an outbound claude -p network call was observed (July 2026) to hang rather than complete or fail cleanly — Codex's sandbox restricts network access, and claude -p needs it to reach Anthropic's API. That observation is less rigorously isolated than the nested-codex exec failure (no distinct error message, just an unresponsive process that had to be killed), so treat it as a strong warning rather than a certainty. If a claude -p call hangs rather than returning, do not assume it will eventually resolve — request escalation (sandbox_permissions: require_escalated) for that call, or run from an unsandboxed session, before retrying.

Preflight

1. Confirm that the user explicitly requested $paper-review-lite-codex or a cross-model audit. Otherwise use $paper-review-lite.
2. Locate the manuscript, supplement, bibliography, figures, preregistration, and replication archive.
3. Check command -v claude and run a harmless authentication/status check supported by the installed CLI. Do not print credentials.
4. Before the first call, confirm the user accepts an external model call that may consume separate credits — unless they already authorized Claude or cross-model execution.
5. If Claude is unavailable or authorization is declined, offer the fallback in “Reduced-diversity mode.”

Claude Code documents claude -p as its non-interactive interface. Use --output-format text, --no-session-persistence, and read-only tools for review calls. See Run Claude Code programmatically.

Orient once

Follow $paper-review-lite Phase 1. Read the manuscript before spawning reviewers. Record actual absolute paths, design family, source format, target journal, and missing inputs. Create:

.review-tmp/
├── codex/
├── claude/
└── cross-check/

Phase 1: independent Red Teams

Codex cohort

Run the nine $paper-review-lite review dimensions in parallel batches that respect the runtime's agent limit. Each subagent receives only:

• the relevant dimension block from $paper-review-lite;
• the manuscript and related absolute paths;
• the common severity and quote requirements; and
• its assigned output path under .review-tmp/codex/.

Keep agents blind to one another. Require direct manuscript evidence for every critical or recommended issue.

Claude cohort

Run three independent Claude CLI calls concurrently when safe:

1. argument, claims, and numerical consistency;
2. references, DOI status, writing, figures, and tables;
3. methods, CONSORT/preregistration where applicable, and replication readiness.

Construct each prompt from the matching dimension blocks in $paper-review-lite; do not paraphrase away requirements. Tell Claude to read the named files, print structured Markdown only, and never edit them. Redirect stdout to the assigned file under .review-tmp/claude/.

Use this command shape, substituting absolute paths and a complete prompt file:

claude -p \
  --output-format text \
  --no-session-persistence \
  --allowedTools "Read" \
  "$(< /absolute/path/to/prompt.txt)" \
  > /absolute/path/to/.review-tmp/claude/review-N.md

Do not use --bare unless API-key authentication is configured explicitly; bare mode skips normal OAuth and keychain discovery.

Validate every output file. A non-zero exit, empty file, or refusal is a failed reviewer, not evidence that the manuscript passed.

Phase 2: blind cross-check

After both cohorts finish, launch four verification passes:

• two Codex subagents verify Claude's content/methods and technical findings;
• two Claude CLI calls verify Codex's corresponding findings.

For each critical or recommended issue, require this sequence:

1. Find the quoted span at the cited location. If absent, mark QUOTE FAILED and drop it.
2. Check whether the manuscript already answers the concern elsewhere. If yes, mark REFUTED or DOWNGRADE with the contradicting location.
3. Check the issue on its merits. Mark CONFIRMED, REFUTED, or DOWNGRADE.
4. Do not add new issues during cross-checking.

Cross-checkers read the manuscript plus the other cohort's files, never their own cohort's findings.

Phase 3: adjudicate

Apply these rules in order:

• Mutual and confirmed: retain at the stronger justified severity; label Mutual (Codex + Claude).
• Single-cohort and cross-confirmed: retain; label the originating cohort and confirmation.
• Single-cohort and cross-refuted: drop unless the lead rereads the source and records specific contrary evidence. If retained, demote one tier.
• Quote failed: drop.
• Reviewer failed: record reduced coverage; never convert missing output into a pass.

Deduplicate by underlying issue, not wording. Then follow $paper-review-lite Phase 4 to write the recommendation, editor's note, severity-ranked issues, strengths, readiness checklist, and “What Still Needs Your Input.” Add a Confidence column to critical and recommended findings.

Reduced-diversity mode

If Claude cannot run, ask whether to continue with two blind Codex cohorts. If approved:

• give each cohort fresh context and identical specifications;
• keep cohorts blind until cross-checking;
• label the report same-model independent review, never cross-model; and
• state that agreement is weaker evidence because model-family blind spots are shared.

If the user declines, run ordinary $paper-review-lite or stop as requested.

Completion checks

• Every retained critical or recommended issue has a verified quote and location.
• Both directions of cross-check completed, or reduced coverage is stated explicitly alongside any failed or empty reviewer output.
• No agent edited the manuscript.
• The final report is self-contained and does not expose scratch prompts or agent chatter.
• .review-tmp/ is removed after delivery unless the user asked to keep it.