paper-review-lite-codex

Name: paper-review-lite-codex
Author: scdenney

34stars0forks

How to install paper-review-lite-codex?

scdenney/open-science-skills/paper-review-lite-codex

$npx -y skills add scdenney/open-science-skills --skill paper-review-lite-codex

Installs into the current project.

Summary

Run a cross-model adversarial pre-submission audit from Codex. Use when a user explicitly requests the heavier paper-review-lite-codex workflow, independent Codex and Claude review passes, or cross-model verification of a manuscript. Apply the paper-review-lite protocol, cross-ch

Files · 1

View on GitHub

SKILL.md

1	# Paper review lite: Codex lead, cross-model audit
2
3	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.
4
5	Read [`../paper-review-lite/SKILL.md`](../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.
6
7	## Sandbox constraint — read before the first `claude -p` call
8
9	`claude -p` is a different binary than `codex exec`, so it does not hit the identical `codex exec` 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. This finding is less rigorously isolated than the nested-`codex exec` failure (it wasn't captured as a 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 in this skill hangs rather than returning, do not assume it will eventually resolve — consider requesting escalation (`sandbox_permissions: require_escalated`) for that call, or running from an unsandboxed session, before retrying.
10
11	## Preflight
12
13	1. Confirm that the user explicitly requested `$paper-review-lite-codex` or a cross-model audit. Otherwise use `$paper-review-lite`.
14	2. Locate the manuscript, supplement, bibliography, figures, preregistration, and replication archive.
15	3. Check `command -v claude` and run a harmless authentication/status check supported by the installed CLI. Do not print credentials.
16	4. Explain that `claude -p` is an external model call that may consume separate credits. Obtain confirmation before the first call unless the user already explicitly authorized Claude or cross-model execution.
17	5. If Claude is unavailable or authorization is declined, offer the fallback in “Reduced-diversity mode.”
18
19	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](https://code.claude.com/docs/en/headless).
20
21	## Orient once
22
23	Follow `$paper-review-lite` Phase 1. Read the manuscript before spawning reviewers. Record actual absolute paths, design family, source format, target journal, and missing inputs. Create:
24
25	```text
26	.review-tmp/
27	├── codex/
28	├── claude/
29	└── cross-check/
30	```
31
32	Treat this directory as disposable workflow state. Preserve it only if the user asks.
33
34	## Phase 1: independent Red Teams
35
36	### Codex cohort
37
38	Run the nine `$paper-review-lite` review dimensions in parallel batches that respect the runtime's agent limit. Each subagent receives only:
39
40	- the relevant dimension block from `$paper-review-lite`;
41	- the manuscript and related absolute paths;
42	- the common severity and quote requirements; and
43	- its assigned output path under `.review-tmp/codex/`.
44
45	Keep agents blind to one another. Require direct manuscript evidence for every critical or recommended issue.
46
47	### Claude cohort
48
49	Run three independent Claude CLI calls concurrently when safe:
50
51	1. argument, claims, and numerical consistency;
52	2. references, DOI status, writing, figures, and tables;
53	3. methods, CONSORT/preregistration where applicable, and replication readiness.
54
55	Construct each prompt from the matching dimension blocks in `$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/`.
56
57	Use this command shape, substituting absolute paths and a complete prompt file:
58
59	```bash
60	claude -p \
61	--output-format text \
62	--no-session-persistence \
63	--allowedTools "Read" \
64	"$(< /absolute/path/to/prompt.txt)" \
65	> /absolute/path/to/.review-tmp/claude/review-N.md
66	```
67
68	Do not use `--bare` unless API-key authentication is configured explicitly; bare mode skips normal OAuth and keychain discovery. Never pass secrets in the prompt or command line.
69
70	Validate every output file. A non-zero exit, empty file, or refusal is a failed reviewer, not evidence that the manuscript passed.
71
72	## Phase 2: blind cross-check
73
74	After both cohorts finish, launch four veri

People who liked this also viewed

6 picks

Type

paper-review-lite-codex

34stars0forks

View on GitHub

How to install paper-review-lite-codex?

scdenney/open-science-skills/paper-review-lite-codex

$npx -y skills add scdenney/open-science-skills --skill paper-review-lite-codex

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Paper review lite: Codex lead, cross-model audit
2
3	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.
4
5	Read [`../paper-review-lite/SKILL.md`](../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.
6
7	## Sandbox constraint — read before the first `claude -p` call
8
9	`claude -p` is a different binary than `codex exec`, so it does not hit the identical `codex exec` 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. This finding is less rigorously isolated than the nested-`codex exec` failure (it wasn't captured as a 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 in this skill hangs rather than returning, do not assume it will eventually resolve — consider requesting escalation (`sandbox_permissions: require_escalated`) for that call, or running from an unsandboxed session, before retrying.
10
11	## Preflight
12
13	1. Confirm that the user explicitly requested `$paper-review-lite-codex` or a cross-model audit. Otherwise use `$paper-review-lite`.
14	2. Locate the manuscript, supplement, bibliography, figures, preregistration, and replication archive.
15	3. Check `command -v claude` and run a harmless authentication/status check supported by the installed CLI. Do not print credentials.
16	4. Explain that `claude -p` is an external model call that may consume separate credits. Obtain confirmation before the first call unless the user already explicitly authorized Claude or cross-model execution.
17	5. If Claude is unavailable or authorization is declined, offer the fallback in “Reduced-diversity mode.”
18
19	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](https://code.claude.com/docs/en/headless).
20
21	## Orient once
22
23	Follow `$paper-review-lite` Phase 1. Read the manuscript before spawning reviewers. Record actual absolute paths, design family, source format, target journal, and missing inputs. Create:
24
25	```text
26	.review-tmp/
27	├── codex/
28	├── claude/
29	└── cross-check/
30	```
31
32	Treat this directory as disposable workflow state. Preserve it only if the user asks.
33
34	## Phase 1: independent Red Teams
35
36	### Codex cohort
37
38	Run the nine `$paper-review-lite` review dimensions in parallel batches that respect the runtime's agent limit. Each subagent receives only:
39
40	- the relevant dimension block from `$paper-review-lite`;
41	- the manuscript and related absolute paths;
42	- the common severity and quote requirements; and
43	- its assigned output path under `.review-tmp/codex/`.
44
45	Keep agents blind to one another. Require direct manuscript evidence for every critical or recommended issue.
46
47	### Claude cohort
48
49	Run three independent Claude CLI calls concurrently when safe:
50
51	1. argument, claims, and numerical consistency;
52	2. references, DOI status, writing, figures, and tables;
53	3. methods, CONSORT/preregistration where applicable, and replication readiness.
54
55	Construct each prompt from the matching dimension blocks in `$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/`.
56
57	Use this command shape, substituting absolute paths and a complete prompt file:
58
59	```bash
60	claude -p \
61	--output-format text \
62	--no-session-persistence \
63	--allowedTools "Read" \
64	"$(< /absolute/path/to/prompt.txt)" \
65	> /absolute/path/to/.review-tmp/claude/review-N.md
66	```
67
68	Do not use `--bare` unless API-key authentication is configured explicitly; bare mode skips normal OAuth and keychain discovery. Never pass secrets in the prompt or command line.
69
70	Validate every output file. A non-zero exit, empty file, or refusal is a failed reviewer, not evidence that the manuscript passed.
71
72	## Phase 2: blind cross-check
73
74	After both cohorts finish, launch four veri

People who liked this also viewed

6 picks

Type