Files
pipdock/docs/AI-DISCLOSURE.md
T
poli0981andClaude Fable 5.2 4732f05197 docs: sync the specs with the accessibility pass
UI-SPEC §7 no longer describes a forced-colors list that is "not complete": the
mark is data-destructive. §8 gains the five rules the pass left behind — focus
on every screen replacement, names that contain the visible text, composite
widgets as one tab stop, chip meanings as text, status regions that exist
before their first message.

TESTING says which of the three new tests holds what, and that none of them
knows what a screen reader says; the manual charter gains a ten-minute
accessibility pass. CLI-SPEC writes down that the CLI has no colour and nothing
that animates, which was true of the code and stated nowhere.

RELEASE-CI: the conversation-resolution rule the protection table never listed,
the labels box that had been true for weeks, and .nvmrc, which still said 24 a
month after every workflow moved to 26.

README links the new files and gains Accessibility and Contributing sections.
AI-DISCLOSURE and About name the model that produced 1.5.0.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
2026-11-10 20:01:03 +07:00

7.4 KiB

AI disclosure

Last updated 2026-10-17, for PipDock 1.4.0.

PipDock was written with substantial help from an AI coding assistant, under human review. This page says which model, what it did, what it did not do, and how you can check any of it yourself.

It exists because the fact is not obvious from the outside and you should not have to infer it from commit trailers. Nothing here changes the licence or the warranty position — see legal/EULA.md for those.

The model

Model Claude Opus 5 — first commit through 1.3.0 · Claude Opus 5.5 — 1.4.0 and the maintenance after it, through 2026-11-04 · Claude Fable 5.2 — from 2026-11-10, which is 1.5.0
Vendor Anthropic (anthropic.com)
Interface Claude Code, Anthropic's CLI/desktop coding agent
Period 2026-07-27 → 2026-08-21 (first commit to 1.3.0) · 2026-10-16 → 2026-10-17 (1.4.0) · 2026-10-20 → 2026-11-04 (CI and dependency maintenance) · 2026-11-10 (1.5.0)

The trailer on each commit records which model produced it, and git log is the authoritative history — the table above is a summary of it, not a substitute.

What it was used for

  • Source code — Rust core, CLI, Tauri bridge, React frontend. Most of the implementation.
  • Tests — unit, integration and the CLI golden snapshots.
  • Debugging — reading failing CI logs, reproducing defects, and diagnosing them.
  • Translation — the Vietnamese catalogue (ui/src/locales/vi/), reviewed by a native speaker who is also the project owner.
  • Documentation — the specifications under docs/, this page included.

What review meant in practice

Every change reached main through a pull request. Through 1.3.0 the owner read and merged each one. 1.4.0 was different, and is recorded as such: it was an audit of the whole codebase, and the owner reviewed and approved a written plan — scope, every decision the plan asked them to make, and the instruction to carry it through to a release — and the assistant then committed, opened the pull request, merged it once CI was green, tagged, and published, under that instruction. The pull request, its checks and every commit are public. Beyond reading diffs, four things were load-bearing, and all four caught real defects the review-by-reading did not:

  1. A green test suite was never sufficient. Thirty-five defects reached main or nearly did, and not one was caught by the suite that was green at the time. They were found by running the thing — against a real interpreter, a real CI runner, a real browser, a real installer. docs/ROADMAP.md records each one against the stage it was found in.
  2. Specifications were treated as claims, not facts. Several were wrong about the tools they described: all three Code Health tools exit non-zero on findings and vulture uses 3; deptry reports a module name rather than the distribution providing it; ruff's rule URL is keyed by rule name, not code. Each is now pinned by a test because a document was not enough.
  3. CI is a second reviewer that cannot be talked round. Five checks block every pull request — formatting, clippy and the test suite; cargo audit; npm audit; lint, typecheck, tests and build of the frontend; and CodeQL for JavaScript/TypeScript — and main is protected so that the owner cannot bypass them. Integration tests against real pip and uv in disposable venvs run on every pull request that touches what they cover, and cargo deny, CodeQL for Rust and Actions, and a Tauri version check run on all of them without being required.
  4. The built application was driven by hand before release. Before 1.0.0 the owner did it — on a real machine, in Windows Sandbox and in a Windows 10 VM — following docs/TESTING.md §4's charter, and several defects were found that way and only that way. For 1.4.0 the assistant drove the installed build itself, through screen control on the owner's machine; the 1.4.0 record in docs/ROADMAP.md lists what was checked and what was not.

What was not delegated

  • Product decisions. Scope, the feature matrix, the non-goals, and what 1.0 would and would not ship are the owner's. So is the decision to remove the self-updater, which is why PipDock has no auto-update path (see SECURITY.md §5).
  • The decision to release. Through 1.3.0 every merge, tag and publish was the owner's own action. For 1.4.0 the owner delegated the mechanics, in writing, with the conditions attached — CI green, and the installers matching SHA256SUMS.txt before the draft was published — and the decision that 1.4.0 should ship remained theirs. 1.5.0 was released the same way: the owner chose its scope from three options, approved a written plan that named every outward step — the two pull requests, the tag, the published release, the repository description and label — and the assistant carried it out under the same two conditions. The owner's own keyboard and Narrator testing of 1.4.0 is theirs and is recorded as theirs in ACCESSIBILITY.md; the assistant cannot run a screen reader, and that file says which checks were a person's and which were a machine's.
  • Anything requiring a legal or security judgment was raised for a decision rather than taken — for example the PEP 668 override's default, and whether a mailto: scheme belonged in the browser-launch allowlist.

Copyright in PipDock is held by Kokone (poli0981), and the work is licensed GPL-3.0-only. Using an AI assistant does not change either. Commits carry a Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> trailer through 1.3.0 and Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> from 1.4.0, and Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> from 2026-11-10, as an accurate record of how the work was produced — it is attribution, not an assignment of rights, and Anthropic claims no ownership of output produced with its models.

Third-party code

PipDock's dependencies are other people's work under their own licences, listed in legal/THIRD-PARTY-NOTICES.md. Nothing in this repository was copied from another project. Where an approach came from upstream documentation or a specification — PEP 508, PEP 440, PEP 668, PEP 691, PEP 514 — the relevant document is cited in the code or in docs/.

Checking any of this

git log --format='%h %an %s'                      # every commit and its author
git log --format=%b | grep -c 'Co-Authored-By'    # how many carry the trailer
gh pr list --state merged                         # every change, and the review it went through

At the v1.0.0 tag: 177 commits, 168 of them carrying the trailer. The nine without are Dependabot bumps and a handful of hand-written mechanical commits.

Those two figures are stated as of a tag rather than as of main on purpose. A count written into a file changes the moment the file is committed — which caught this page out twice while it was being written — so it is pinned to something that does not move, and the commands above are the thing to trust.

Questions

code@poli0981.dev, or open a discussion at github.com/poli0981/pipdock. If you think something in this page is inaccurate, that is worth an issue.