Vendors the upstream ARS v3.20.0 release at 3af9f03d5aadb0bca51af1440f20b5cbf97d6dba: evidence-bound review and revision contracts, the contained consent-gated Codex citation-integrity transport, human-subjects authority boundaries, and the optional process-isolated PDF content advisory.
Preserves the single-root Codex router, WORKFLOW.md entrypoint overlays, inactive Claude-only hooks, separately vendored experiment-agent v1.1.0, and the materialized Desktop plugin bundle. Re-pins overlay-sensitive content locks and keeps live Codex smoke tests and 24-call measurement outside automatic validation.
Synchronizes README_ZH-CN.md with the current English README and adds language navigation links. Verified against the current v0.1.24 main with matching heading and table structure.
Documents MiniMax-M3 and both regional OpenAI-compatible endpoints while preserving the existing fail-closed ungrounded-provider contract. Canonical and Desktop plugin trees remain synchronized.
Vendors upstream experiment-agent main (3 commits past the previous pin):
the v1.1.0 session-resume capability for study_manager_agent with the
study-state protocol, templates, and design docs. Preserves the WORKFLOW.md
entrypoint rename with its data_access_level/task_type frontmatter overlay;
.claude/ and .github/ upstream dev files stay excluded. ARS vendored tree
unchanged (pinned at 5769d7b).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qs4Uvp7YArR9R2TUBPVPTo
Vendors upstream main (44 commits past v3.19.0): the #574/#576/#608-#611
reviewer hardening track, #610 arithmetic-receipt grammar + script adapter,
held-out eval cohorts, medical venue disclosure policies, Chinese-literature
resolver, CARE/STARD/TRIPOD+AI guidance, bare /ars-* aliases, and the Pi
wrapper (traceability only). Extends the WORKFLOW.md path adaptation to the
new reviewer-contract lint family, re-pins the pipeline content locks, and
declares check_calibration_tiers.py inactive (needs non-vendored
.claude/CLAUDE.md). Local pytest manifest: 91/91 entries pass.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Qs4Uvp7YArR9R2TUBPVPTo
Add repo-local Codex Desktop plugin marketplace metadata and an Academic Research Skills plugin manifest without duplicating the vendored ARS skill tree. The plugin root links back to the existing skills directory so direct skill installs and plugin installs share one source of truth.
Suggested by GrLime909 in PR #29.
Replace positional aliases[1] extraction with normalized iteration over all manifest aliases (strip leading slash, preserve order, dedupe). Adds hook wrapper tests including a no-secret-echo guard.
Expand UNCLEAR_QUESTION_RE so phrasings like "my research question is still unclear" keep vague paper-topic requests in deep-research/socratic. Adds a regression test.
Align the Codex package with upstream ARS v3.9.4.2.
- pin the manifest to upstream commit 96b82e82142dc95f117595c207d3e150b078e411
- bump the Codex package to 0.1.8 and update README/CHANGELOG metadata
- preserve Codex-specific router, WORKFLOW, lint, and architecture overlays
- include upstream v3.9.1-v3.9.4 runtime updates and temporal verification tooling
Match ARS main-repo convention. Resume prompts and inter-session
context notes live in .context/ as local-only artifacts and should
not enter the public repo.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* align ARS Codex with ARS 3.9.0
* fix: re-apply codex-adapter overlay after vendor sync (codex review #13 P2)
The v3.9.0 vendor sync (6382c30) re-imported upstream artifacts that
reference `<dir>/SKILL.md`, but the codex adapter convention (set in
3cd63b7 + 8cd376d) is to rename those entry files to `WORKFLOW.md`.
The overlay was not replayed on the new commit, so 23 active artifacts
pointed at files that don't exist in this package.
Also reverts a shebang regression: check_audit_artifact_consistency.py
ships under `#!/usr/bin/env python3` to match its 23 sibling lint
scripts and upstream; the vendor sync silently changed it to `python`,
which breaks direct execution on hosts without a Python 2 fallback.
Fixes:
- 10 commands/ars-*.md "Skill entry" pointers → WORKFLOW.md
- 11 markdown artifacts (agents, references, shared, WORKFLOW.md
cross-references) → WORKFLOW.md
- 3 lint scripts that read SKILL.md as filesystem paths
(check_collaboration_depth_rubric.py, check_spec_consistency.py 5 sites)
- check_audit_artifact_consistency.py shebang → python3
- test_check_collaboration_depth_rubric.py fixture mirrors the rename
Verification:
- pytest scripts/: 1460 passed, 111 subtests passed (zero regression)
- check_collaboration_depth_rubric.py: OK
- ./scripts/check_audit_artifact_consistency.py exec test passes
Closes codex review P2 #1 (path) + P2 #4 (shebang).
P2 #2 (openalex_client response-read OSError escape) and P2 #3
(crossref_client symmetric) are upstream ARS bugs surfaced via this
vendor sync; tracked separately against the main academic-research-skills
repo and intentionally not patched here.
P3 #5 (check_claim_audit_consistency manifest_id setdefault TypeError on
malformed passport) deferred — non-blocking, lint reports the schema
finding cleanly through validate_passport().
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* examples: add Codex-platform full pipeline run evidence
Satisfies upstream CONTRIBUTING.md § Platform ports L55: "at least one
full academic-pipeline run on the target platform, committed under an
examples/ path in the sibling repo, so regressions are detectable."
The traversal exercises 8 stage transitions (1 → 2 → 2.5 → 3 → 4 → 4.5
→ 5 → 6) via 6 codex exec calls on Codex CLI 0.130.0, smoke-level
deliverables, both MANDATORY integrity gates engaged (Stage 2.5 caught
2 HOLD; Stage 4.5 PASSed with 2 watchpoints). Topic: HEQA Scope
Transformation Matrix in the agentic AI era.
The Stage 5 finalize artifact (abstract.md + index.html) is the real
output from a prior ars-codex research session on 2026-05-10, included
so the full-pipeline evidence terminates at a concrete deliverable.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(examples): address codex round-1 findings (4 P1 + 3 P2)
Codex review round 1 on PR #10 surfaced:
P1 fixes:
- Re-frame as "Codex-platform pipeline smoke traversal" not "full
pipeline run". Each routed stage used the lightest mode of its
workflow (outline-only / quick / revision-coach); Stage 3' and 4'
re-review/second-revision are absent; Stage 5 is imported, not
produced in this run. README "Honest scope" section enumerates the
trade-offs.
- 7-mode numbering disclosure on Stage 2.5: user prompt enumerated
modes in a non-canonical order. Codex response is verbatim per the
prompt order; canonical Lu 2026 numbering added as a mapping column
and called out in disclosure box. Stage 4.5 already used canonical
numbering.
- Add verbatim "User prompt (verbatim, reproducible)" sections to
Stage 2.5, 3, 4, 4.5, 6 (previously only Stage 1 + 2). README
re-run claim now lines up with what is committed.
- Reframe Stage 4 "How this aligns with the actual abstract.md" as
convergent observation, not causal same-run revision. README and
Stage 4 now both state abstract.md is imported (2026-05-10), not
downstream of the 2026-05-11 transcripts.
P2 fixes:
- README link paths corrected (../abstract.md → abstract.md from
README, ../ paths preserved inside stages/).
- Stage 6 watchpoint wording aligned with Stage 4.5 W1/W2 verbatim.
- Stage 1/2/3 "visible in transcript" claims reworded as "inferred
from transcript" with note that verbatim transcripts not committed
but verifiable by re-running the committed prompt.
Gemini-3.1-pro-preview citation audit run separately on this PR:
PASS for Lu et al. 2026 Nature + Wang & Zhang 2026 IJETHE; PRISMA-trAIce
upstream reference flagged as NOT-FOUND but out of scope for this PR
(belongs to upstream ARS docs, not added in this PR).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(examples): address codex round-2 findings (1 P1 + 2 P2)
P1: examples/README.md still claimed "full 10-stage" and "real Stage 5".
Reframed to "smoke traversal" with explicit Stage 5 imported provenance.
P2: examples/codex/.../README.md "router classification at every stage
transition (1 → 2 → 3 → 4 → 5 / 6)" overstated current-run traversal —
Stage 5 was imported. Reworded to enumerate only the stages actually
routed in this run.
P2: README stage map + Stage 3 continuity section used non-canonical
"mode 5 + mode 6" labels without flagging them as prompt-numbered.
Added prompt-numbered ↔ canonical mapping pointers.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add sponsor funding metadata, root security policy, public README badges/sections, and contributor acknowledgements for the Codex sibling distribution.
Fixes#2.
- Route vague topic alias requests through Socratic scoping first
- Record docs/superpowers exclusion in the Codex manifest
- Add regression coverage for the alias router contract
Imbad0202