Academic Writing Toolkit 0.6.0-rc.2

v0.6.0-rc.2

Still a release candidate, for the same two reasons as rc.1. Gate A §7's
Windows repeat (#56) has not been run, and no one has yet written a real
chapter with this (#35). None of the four merges since rc.1 claims either;
#63 says so in its own description.

What changed since v0.6.0-rc.1 (9dda326 → ce7b9ba)

• #64 — the audit scripts fail closed. Checking nothing is not a pass:
  audit-citation-fidelity.mjs exits 2 with nothing_checked when zero
  citations were checked and reports the \cite commands it does not read;
  audit-prose-fingerprint.py withholds every percentile when the baseline
  contains a draft of the target, accounts for every baseline candidate, and
  names a target/baseline pipeline mismatch; audit-claim-positioning.py
  no longer counts a hyphenated compound or a refused claim as a novelty
  claim. A claim ledger binds a manuscript sentence to a verbatim snippet in
  an archived source; a number ledger binds a printed number to the artifact
  it comes from. scripts/check-fails-closed.py runs every check script
  against an empty target in CI and requires a non-zero exit, and every
  script under .claude/skills/*/scripts/ must be registered as a check or
  as not-a-check with a reason. build-venue-baseline.py assembles a corpus
  of papers a venue actually published, from arXiv records whose DOI
  resolves to the venue's registered container-title, so a manuscript can be
  measured against where it is going and not only against what it cites
  (spec: docs/specs/2026-09-20-venue-baseline.md, status draft).
• #63 (RubyYii) — native Windows setup: scripts/setup.mjs, junction-based
  skill links, a bundled word counter for map. CI now runs first-time
  setup on Windows. Installation and E0 checks only; no Gate A acceptance.
• #62 (RubyYii) — README states the review modes and what the isolation
  evidence covers.
• #59 — README names the release candidate and what makes it one.

The five release facts this project requires of itself

Git ref. ce7b9ba (full sha in the tag object).

Artifact. None. The toolkit is used from a clone. rc.1's message named a
guard tarball by sha256; no file was ever attached to its release page, so
this tag does not name one.

Evidence state, in the classes this project keeps separate:

Class	State
E0 — deterministic gates green in CI	met on ce7b9ba: test (170 regression tests, guards 162), the global-skills matrix on Linux, macOS and Windows, and the Windows-runtime job, all green; every check script is in the fails-closed registry and refuses an empty target
E1 — paired comparison on real inputs	unchanged since rc.1: run once, negative, under a model too small to draw from
E2 — the author's own chapter cycle	not run (#35)
E3 — evidence from someone who is not us	does not exist. The venue baseline has been run by the author on one manuscript; that is use, not evidence

The gate that approved it. Nothing beyond rc.1's. Gate A §7 passed on
macOS at rc.1 and has not been repeated on Windows. What #64 adds is a gate
on the tools, not on the workflow: the fails-closed registry is a required
CI check.

Remaining human decisions, and whose they are. Unchanged. The Windows
acceptance repeat is RubyYii's (#56). The E2 chapter cycle is the author's
(#35).

What this release does not claim

It does not claim the workflow improves anyone's writing. It does not claim
that a manuscript's position in a venue corpus predicts anything about its
review; the corpus is whatever arXiv records resolve to the venue's DOI
container-title from the year given, and under twenty documents the builder
refuses to produce one. It does not claim Windows works beyond what the CI
matrix covers. It is not a hosted service and stores no manuscript anywhere.

---

The README paragraph on main at ce7b9ba still names rc.1 as the latest
release. rc.1 discovered this after the fact; this time it is stated before
it. The bump follows as its own pull request, and this tag is not moved.

Academic Writing Toolkit 0.6.0-rc.1

A release candidate, and the label is deliberate. The daily loop is verified
end to end on macOS and every deterministic gate is green on three platforms,
but §7's Windows repeat has not been run and no one — including the author —
has yet written a real chapter with this. Those are the two things standing
between this and v0.1.0, and both are tracked below.

The first release of the dsh-distribution architecture. The previous tag,
v0.5.0 (2026-07-29), is a different product: it carried the ChatGPT App, the
packaged Codex plugin and the Workbench wheel, all decommissioned since. Anyone
following that tag has been on the wrong thing.

On the number: the design document calls this architecture "v0.1", which is a
generation, not a version — v0.1.0 was taken in May 2026 by the packaged
plugin. This continues the release line from v0.5.0.

What this is

A thesis workflow that runs as a DeepSeek Harness profile. Nine skills, and a
guard layer that refuses rather than advises: a chapter write citing a source
with no conforming notes is denied, a read past the page budget is denied, a
write outside the active edit contract is denied, and an export runs only when
every cited source resolves.

git clone … && cd academic-writing-toolkit
make setup
npm ci --prefix guards && npm run build --prefix guards
npm ci --prefix e2e
node scaffold/awt.mjs init ~/thesis
node scaffold/awt.mjs install-profile
node scaffold/awt.mjs verify ~/thesis        # six stages, keyless
export DEEPSEEK_API_KEY=...
node scaffold/awt.mjs run ~/thesis "task"    # or: web ~/thesis

The five release facts this project requires of itself

Git ref. 9dda326dfccc1ed4a101f608c804350f0611ec5c

Artifact. awt-guards-0.1.0.tgz, sha256
149b8c7ca461814e2404cc12d5161e735bbd0be87b92b100a684e73a8771dd0d — the guard
bundle, installable into a dsh profile. The toolkit itself is used from a
clone; nothing else is packaged.

Evidence state, stated in the classes this project keeps separate and never
merges:

Class	State
E0 — deterministic gates green in CI	met: guards 157, regression 132, live denial table, credential probe, skill-scope probe, pack smoke, on Linux, macOS and Windows
E1 — paired comparison on real inputs	run, and negative: with one small local model neither arm produced lint-conforming notes. No writing-efficacy claim follows
E2 — the author's own chapter cycle	not run (#35)
E3 — evidence from someone who is not us	does not exist

The gate that approved it. Gate A of
docs/specs/2026-09-06-gate-a-workspace-contract.md, whose §7 acceptance was
run twice on macOS: the first failed on the contract mechanism and produced the
fixes in #54, the second passed. §7 also requires a Windows repeat, which has
not been done.

Remaining human decisions, and whose they are. The Windows acceptance
repeat is RubyYii's (#56). The E2 chapter cycle is the author's (#35) and is
the only thing that can move the evidence past E0 for the workflow as a whole.

What this release does not claim

It does not claim the workflow improves anyone's writing. The enforcement is
CI-proven; the benefit is unmeasured, and the one measurement so far was
negative under a model too small to draw from. It does not claim Windows works
beyond what the CI matrix covers. It is not a hosted service and stores no
manuscript anywhere.

---

One thing this tarball does not contain. The README paragraph describing
this release landed in 3db9e48, one commit after the tag — a release cannot
describe itself from inside itself unless the description is written first, and
here it was not. The tag is not being moved: it is the ref the Windows
acceptance (#56) is being run against, and moving it would silently change what
is under test. Read the current README on main for the release state; nothing
else differs between 9dda326 and 3db9e48.

Academic Writing Toolkit 0.5.0

Academic Writing Toolkit 0.5.0 is the first stable release of the local, author-controlled review workbench introduced during the 0.5.0 release-candidate cycle.

What ships

• Local Workbench: import a selected manuscript, choose one of five writing tasks, run the authenticated Codex agent, review issue-level suggestions, and export a new copy without overwriting the source.
• Codex plugin: 20 local academic-writing skills, packaged with strict frontmatter validation.
• ChatGPT App package: five bounded, pasted-text tools. The hosted App runtime remains separately deployed and reviewed; this GitHub release does not silently change the endpoint currently under OpenAI review.

Stable-release fixes

• Align Workbench, Codex plugin, and ChatGPT App package versions at 0.5.0.
• Fix and guard against invalid YAML in packaged skill frontmatter.
• Clarify that GitHub release, hosted App deployment, OpenAI review, and publication are separate states.

Install

python3 -m venv .venv
. .venv/bin/activate
python -m pip install https://github.com/yha9806/academic-writing-toolkit/releases/download/v0.5.0/academic_writing_toolkit-0.5.0-py3-none-any.whl
awt --check
awt

The Codex plugin ZIP is attached as academic-writing-toolkit-plugin-0.5.0.zip. Verify both assets with SHA256SUMS before installation.

Evidence boundary

The release checks establish build, packaging, permission, and workflow behaviour. They do not establish general human writing-effect improvement, cross-domain generalisation, or model-independence.

Academic Writing Toolkit 0.5.0rc4

AWT 0.5.0rc4 synchronizes the release candidate across three intentionally different surfaces.

• Local Workbench: Agent-powered, author-controlled review and apply-copy workflow (0.5.0rc4)
• Codex plugin: 20 bundled academic-writing skills (0.5.0-rc.4)
• ChatGPT App: 5 deterministic pasted-text MCP tools (0.5.0-rc.4)

Release changes:

• aligns release identity across all three surfaces
• updates the plugin manifest for current OpenAI directory limits
• refreshes App production dependencies to zero known npm audit findings
• raises the MCP request boundary to a tested 1 MiB and returns bounded JSON errors
• adds plugin/App CI, version-alignment checks, and truthful surface-boundary docs
• deploys and live-verifies the Hugging Face MCP endpoint

Verification:

• main GitHub CI: all 4 jobs passed
• public regression suite: 121/121
• Workbench runtime: 11/11 plus fresh wheel install
• ChatGPT App: 8/8, 0 npm audit vulnerabilities
• live MCP: health/handshake RC4, 5/5 tool calls passed
• wheel and plugin ZIP: two deterministic builds matched

This remains a pre-release. Engineering checks validate packaging and runtime behavior; they do not establish scientific writing-effect improvement.

Academic Writing Toolkit v0.5.0rc3

AWT v0.5.0rc3 is the first public release candidate of the lean, local-first review workbench.

Included

• one Chinese workbench for contribution positioning, literature integration, accuracy checking, discussion/evidence alignment, and whole-manuscript review
• explicit accept / modify / reject / defer decisions per issue
• hash-bound local session recovery and confirmed apply-copy export
• deterministic bounded retrieval for selected supporting files
• loopback-only HTTP service with Host and Origin checks
• awt --version and no-model awt --check preflight

Install

python3 -m venv .venv
.venv/bin/pip install https://github.com/yha9806/academic-writing-toolkit/releases/download/v0.5.0rc3/academic_writing_toolkit-0.5.0rc3-py3-none-any.whl
.venv/bin/awt --check
.venv/bin/awt

Requires Python 3.9+ and an installed, authenticated, compatible Codex CLI. The original manuscript is never overwritten; only an explicitly confirmed new copy can be exported. Selected manuscript and evidence files are sent to the configured Codex provider, while workbench sessions remain in the configured local filesystem location.

This is a pre-release. Engineering checks establish packaging, interaction, local HTTP boundaries, source preservation, recovery, and apply-copy behaviour. They do not establish general writing improvement, publication outcomes, or commercial-service readiness.

See SHA256SUMS.txt for the wheel checksum.

Academic Writing Toolkit v0.4.0

Academic Writing Toolkit v0.4.0 turns the repository into one coherent author-control system, from source evidence to exact release.

Highlights

• A new explanation-first README with three editable Figma-backed SVG diagrams.
• 20 local/Codex skills, including new argument governance, peer review, clean-room self-review, and standalone revision escalation.
• Thesis-control schema v3 with strict drift-audit states, deterministic migration, failure-atomic writes, and human-approved cycle gates.
• A clean 10-minute evidence demo and an explicit three-surface boundary for local skills, the Codex plugin, and the ChatGPT App.
• The canonical GitHub branch is now main; every remote branch with unique work has been integrated.

Release artifacts

• academic-writing-toolkit-v0.4.0.zip — complete source archive at tag v0.4.0.
• academic-writing-toolkit-codex-plugin-v0.4.0.zip — plugin-only archive.
• SHA256SUMS — verification hashes for both packages.

Verification

• GitHub Actions on main: passed.
• make test: 112/112 passed.
• make chatgpt-app-check: 7/7 passed.
• make doctor, make plugin-check, public-content audit, citation audit, reference verification, evidence-review packet, release packet, thesis-control strict packet, SVG validation, and git diff --check: passed.
• Independent Gemini review of the complete branch diff: pass, no missing tests or unsafe claims.

The ChatGPT App package is version-aligned at 0.4.0, but its hosted /health endpoint and Apps Manage submission remain separate maintainer-owned deployment actions.

Editable README visual source: https://www.figma.com/design/HhaFm0uorv5oS7MsezWDN5

Academic Writing Toolkit v0.3.2

Academic Writing Toolkit v0.3.2 hardens /thesis-control for AI-assisted thesis editing control.

Included

• Adds a /thesis-control scaffold command for draft control packets: spine_cards.csv, edit_contracts.csv, drift_audits.csv, and a review packet from one manuscript unit.
• Rejects empty scaffold excerpts, unsafe identifiers, orphan edit contracts, uninspectable source paths, and packet paths that escape the thesis-control review root.
• Keeps ChatGPT App and Codex plugin package versions aligned at 0.3.2.
• Keeps the ChatGPT App as a pasted-text-only review surface; the full /thesis-control workflow remains local-agent and Codex-plugin only.

Verification

• make doctor
• make plugin-check
• make chatgpt-app-check 7/7
• python3 scripts/audit-public-content.py --base-dir . --json issue_count 0
• make test 69/69
• Real thesis-control positive packet: issue_count 0
• Real thesis-control drift negative case: blocked missing-human-review, unsafe-accept, unsupported-claim-passed, and missed-adjacent-passed
• git diff --check
• gemini-agent diff-review --smart-diff pass

Publication Notes

The repository tag and GitHub release are public. Updating the already-published ChatGPT App still requires redeploying the Hugging Face Space so /health reports 0.3.2, then submitting the existing MCP base URL through OpenAI Apps Manage after organization verification.

Academic Writing Toolkit v0.3.1

Patch release for the app/plugin distribution surface.

Changes:

• Adds the /manuscript-reframe workflow to the packaged Codex plugin.
• Repairs the packaged Agent skill entry so .agents/skills/manuscript-reframe is a symlink to the canonical skill.
• Aligns ChatGPT App and Codex plugin package metadata at version 0.3.1.
• Refreshes ChatGPT App, Codex plugin, and release-readiness handoff docs for the Hugging Face Space update path.
• Records the v0.3.1 release governance packet.

Verification:

• make plugin-check passed.
• make chatgpt-app-check passed with 7 Node tests.
• make test passed with 60 regression tests.
• python3 .claude/skills/release-governance/scripts/check_release_packet.py . passed.
• GitHub Actions test passed on master at 678533e and on the post-release docs commit ffe2036.

Hosted app:

• Hugging Face Space updated to commit f5ab7c9cb5bd554bf380698b039b3a56403ded1e.
• /health returns version 0.3.1 and status: ok.
• GET /mcp returns 405 Method not allowed, as expected for MCP POST-only traffic.
• /.well-known/openai-apps-challenge returns the configured challenge token.

Boundary:

• The repository release and hosted Hugging Face app are verified. OpenAI Apps dashboard submission still requires maintainer-controlled business verification and a manual Submit for Review action.

v0.3.0

Academic Writing Toolkit v0.3.0

Highlights:

• Adds release-governance workflows and release packet validation.
• Adds human-eval handoff repair skill to local agents and the Codex plugin package.
• Aligns ChatGPT App and Codex plugin metadata at version 0.3.0.
• Adds v0.3.0 release governance packet under release/.

Verification:

• make doctor
• make plugin-check
• make chatgpt-app-check (7 tests passed)
• make test (60 tests passed)
• root and demo release packet validators returned issue_count 0
• gemini-agent diff-review --smart-diff returned pass

Distribution notes:

• Render is configured to auto-deploy master after checks pass.
• Cloud Run deployment remains manual through the workflow_dispatch deployment workflow.
• OpenAI Apps dashboard publication requires maintainer credentials and post-push confirmation.

Academic Writing Toolkit Codex Plugin v0.1.0

Summary

Initial Codex plugin release for Academic Writing Toolkit.

Included

• Codex plugin manifest and marketplace entry
• 11 academic writing skills for reading, note-taking, mapping, integration, auditing, style checks, logic review, reference verification, progress tracking, and export
• Bundled helper scripts for citation checks, British English checks, paragraph logic checks, reference verification, and DOCX export
• Plugin publishing assets, privacy policy, terms of use, and publishing checklist
• make plugin-sync and make plugin-check maintenance targets

Validation

• GitHub Actions test passed on master
• make plugin-check passed
• make test passed with 50/50 tests

Install Marketplace Source

codex marketplace add yha9806/academic-writing-toolkit --ref v0.1.0 --sparse .agents/plugins --sparse plugins/academic-writing-toolkit