Files
pipdock/CLAUDE.md
T
poli0981andClaude Opus 5.5 2539950cd5 fix: what an independent review of 1.6.0 found
A venv made with --symlinks is run and stored as itself, not as the base
Python its python.exe links to: validate_interpreter returned the
canonical path, so `--env .venv update --all --yes` would have changed
the system Python. Canonicalizing now only decides where a link leads
and whether it is PipDock's own; a link to another program is refused,
and a Microsoft Store alias (os error 1920) is judged by its own path.

The GUI merges and removes an added row by its path rather than by
envHash, which follows links: a re-add replaces a broken row, removal
sticks for an entry whose file is gone (a share), and neither runs
during a scan. Added entries are validated again on every listing in
both heads (envs::probe_listed), and the CLI checks its stored default
when it uses it. `env remove` makes a typed path absolute first.

SECURITY section 2 and env_probe's doc claimed an enforcement point over
IPC; they now say what is checked, and backlog X36-b records the rest.
notify.yml reports a Triage failure on a pull request.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-21 23:22:12 +07:00

39 KiB
Raw Blame History

PipDock — working notes for Claude

Windows GUI + CLI for bulk-managing Python packages. Tauri 2 + Rust core + React 19. Repo poli0981org/pipdock (moved from poli0981/pipdock on 2026-11-21) · GPL-3.0-only · Status: 1.6.0, and in maintenance mode since that release. M1, M2, Phase 3, Phase 4, the P1 wave, the Security tab and the dependency view are all complete, and 1.6.0 closed the last P0 gap, Browse… (env_add_manual). Read ROADMAP's Maintenance mode section before starting anything: dependency, security, major-upgrade, defect and wording work is taken; features are not, and triage.yml closes requests for them. 1.0.0 shipped the release pipeline, a protected main, the manual charter and legal documents that match the code; 1.1.0 added pin auto-suggest, requirements export/import, the cache manager and the Ctrl+K palette; 1.2.0 added the Security tab (PRD P1-1); 1.3.0 added the dependency view (P1-6), which is a mode of the package screen rather than a tab and adds no runtime dependency — read its ROADMAP record before touching graph/, because the shape of the screen is a consequence of measurements rather than of taste. 1.4.0 added no feature: it is the audit release — about 170 findings, every Critical and High fixed with a test that failed first, and the rest in ROADMAP's Audit backlog, with file:line and a reason each. Read that table before starting a slice; most of what is wrong with the codebase is already written down there. 1.5.0 added no feature either: it is the accessibility pass (backlog U45, U47, U48) — focus, names, keyboard patterns and high contrast, axe-core in the suite, and ACCESSIBILITY.md, which is a record of what was tested and has to be kept true like one. 1.6.0 added Add interpreter… (GUI) and pipdock env add/env remove (CLI) over a manual_envs table and envs::validate_interpreter, moved the repository to the poli0981org organization with CI that calls nothing outside it, and put the project in maintenance mode. Nothing is left of Post-1.0: the scheduled check (P1-7) was closed as not planned, for the reasons ROADMAP's survey of it gives — no toast component, no notification/autostart/tray plugin, nothing on a timer, and no surface that can report a background failure.

Read the docs before changing anything

The design is fully specified. Read in this order; do not invent behaviour that contradicts them.

  1. docs/PRD.md — goals G1–G7, personas, P0/P1/P2 feature matrix, non-goals
  2. docs/ARCHITECTURE.md — crate layout, Engine trait, probe.py, storage, Tauri IPC, version manifest
  3. docs/DATA-FLOW.md — update/install/uninstall/rollback state machines, engine command mapping, invariants §9
  4. docs/UI-SPEC.md — design tokens, screens, click budgets, keyboard map
  5. docs/CLI-SPEC.md — commands, flags, exit codes, JSON contracts
  6. docs/ERROR-CATALOG.md — the launch codes and their stderr classifiers. 35 in Rust (Code::ALL, pinned by two tests); the doc tabulates fewer rows because it folds PD-HLT-001..004 into one, so it shows 32 rows. Adding a code means the catalog, Code::ALL, both counts, the "no code without a fixture" gate in classifier_corpus.rs, both locales/*/errors.json, a regenerated codes.json (cargo run -p xtask -- ipc-fixtures), regenerated bindings, and re-blessed goldens — golden__schema-Code.snap and every schema that embeds it, which since 1.4.0 is five: StepResult, ExecutionSummary, CheckReport (its error), HealthReport and AuditReport, the last two via ToolProblem.code. Six snapshot files in all. This line has been wrong about the count twice — grep PD-ENV-001 across crates/pipdock-cli/tests/snapshots/ rather than trusting the list. Nine places; the gates will tell you which you forgot, including i18n.test.ts, which fails on a code with no copy in either language
  7. docs/CODE-HEALTH-SPEC.md, docs/SECURITY.md, docs/TESTING.md, docs/RELEASE-CI.md, docs/I18N.md, docs/ROADMAP.md
  8. legal/ — EULA, Disclaimer, Privacy, Third-Party Notices (public; the in-app legal gate links to these GitHub URLs)

Hard invariants — never violate these

  1. No mutating engine call without (a) a ResolutionReport accepted this session and (b) a successful snapshot write. Snapshot failure aborts the plan, executing nothing (PD-SNP-001). See DATA-FLOW §9.
  2. Subprocess via argv arrays only — never a shell. tokio::process::Command with explicit args. Package names validated against PEP 508 and versions against PEP 440 before they reach argv. See SECURITY §2.
  3. Never reimplement dependency resolution. The selected engine resolves in dry-run mode; PipDock parses and explains. See ARCHITECTURE §1.2.
  4. Rust emits catalog codes and structured data only. All human phrasing lives in the i18next catalogs under ui/src/locales/. Codes, package names, versions and stderr are never translated. See I18N §1.
  5. --break-system-packages is never passed implicitly. PEP 668 environments are blocked by default; the flag is only sent when the user has explicitly enabled the Settings override. See SECURITY §3.

Two more that follow from them: every user-visible failure carries exactly one catalog code, and a failed package never aborts a batch (skip-and-continue, then a "13 successful, 2 failed" summary).

What the spikes proved (read spikes/README.md before writing an adapter)

  1. Every plan must restate the full installed set as explicit requirements. install -U <pkg> ignores installed packages' constraints in both engines and will break them at exit 0.
  2. Every pip invocation needs PYTHONIOENCODING=utf-8 + PYTHONUTF8=1. Without it --dry-run --report - crashes on Windows/cp1252 (pip 25 and 26), data-dependently.
  3. uv writes its plan to stderr, so classify errors only after a plan-parse attempt.
  4. env_hash lowercases the interpreter path on Windows, or one env splits its pins and snapshots in two.
  5. PipDock enforces Requires-Python itself (src/compat.rs) — the engines disagree, and the preview must not depend on which one is selected.

Conventions

  • One session, one slice, one version. A session takes a single item — a docs/ROADMAP.md entry, or something that surfaced while doing one — takes it deep, tests it properly, and ships it as one version, plus a patch release if it needs a hotfix. Not several shallow items. The record of what a slice cost goes in ROADMAP; what a user would notice goes in changelog/.

  • The changelog lives in changelog/, one file per major line (1.x.md), with a pre-release line getting its own file (2.0-beta.md). Root CHANGELOG.md is a pointer only. Split a file by minor range at roughly 600 lines and never renumber an existing range file — changelog/README.md is the authority.

  • Crates pipdock-*; React components prefixed Pd (PdPackageRow, PdConflictRow)

  • Tauri identifier com.skullmute.pipdock; CLI binary pipdock

  • Frontend lives in ui/, but package.json sits at the repo root — first for the ops repo's reusable CI workflow, which had no working-directory input, and now because SonarCloud, Qodana, the scripts and Tauri's config all assume it; Vite is configured with root: './ui'

  • Rust ↔ TS types are generated by cargo run -p xtask -- bindings over schemars — never hand-mirror a type. (ARCHITECTURE §9 originally said specta/tauri-specta; bindings.rs records why it is not, and drift is a cargo test failure rather than a compile error)

  • No component calls invoke directly; everything goes through the typed wrappers in ui/src/ipc/

CI reality

docs/RELEASE-CI.md is authoritative. Every workflow is project-local since 2026-11-21, when the repository moved to the poli0981org organization: ci-node.yml's web job, codeql.yml and the Discord notifier were callers into the owner's personal ops repo (poli0981/.github, at @main) until then, and a project in the organization no longer runs code another account can change. Every action is pinned by commit with its version in a comment, GitHub's own included, and a write permission belongs on the job that needs it, not the workflow. Dependency bumps are Dependabot (not Renovate); notifications are Discord-only, through the vendored .github/scripts/discord_notify.py (DRY_RUN=1 prints instead of posting).

Eight checks are required on main — fmt · clippy · test, cargo audit, cargo deny, lint · typecheck · test · build, npm audit, npm licences, tauri lockstep, Analyze (javascript-typescript) — and a required name is this repository's own now, so renaming a job means changing the protection payload in the same change (RELEASE-CI §2). npm licences is the npm half of deny.toml: scripts/check-licenses.mjs with the same allowlist, entry for entry.

notify.yml lists its workflows by name, and a new workflow has to be added to it. workflows: ['*'] — the ops repo's old example — is not a wildcard to workflow_run: it matched nothing, and main was red from 2026-09-23 for three weeks with no one told. Start any CI question with gh run list --branch main, and remember that security-scan.yml exists for the case no PR will catch: an advisory against something already merged.

Dependabot gives a dependency to the most specific group, not the first. A group with update-types beats one with patterns alone, so the tauri groups caught nothing until the catch-alls got exclude-patterns (2026-11-21); the order of groups in dependabot.yml decides nothing. Its log says which group took what ("belongs to more specific group").

Qodana runs as two Qodana Cloud projects — qodana-rust.yaml (RustRover in Docker; it cannot see #[cfg(windows)] code) and qodana-js.yaml (WebStorm, native), one token each — and is not a required check. A new workflow cannot be dispatched before it exists on main, so a workflow's first full run happens in its own PR.

An attestation stays with the account that made it. After the transfer, 1.4.0 and 1.5.0 verify only with gh attestation verify <file> --owner poli0981 --signer-repo poli0981/pipdock; every --repo form returns 404 for them. 1.6.0 on verifies with --repo poli0981org/pipdock.

SonarQube Cloud is CI-based since 2026-10-25 — sonar.yml and sonar-project.properties; since 2026-11-21 the project is poli0981org_pipdock in the poli0981org organization (the old poli0981_pipdock keeps the history to 1.5.0). Automatic analysis is off in the project's settings and must stay off, or the CI analysis is refused. Automatic analysis never analysed a line of Rust and imported no coverage, which is why it went. The gate's "new code" is previous version, with the version from package.json, and a bump takes effect one analysis late: the analysis that first sent 1.4.0 still measured from 2026-09-27, and the next measured from the last analysis of the version before it (2026-10-20, a314689). So the merge carrying a release bump is gated on the whole previous period, and periods in api/qualitygates/project_status says which baseline a gate used. New code needs 80 % coverage and no new issue, so a slice whose new lines no test reaches turns main's gate red, and so does a std::fs call inside an async fn (rust:S7493) or a tracing field named message (rust:S6677). Dependabot PRs have no SONAR_TOKEN and skip the analysis.

Current state

M1 and M2 complete (IPC bridge 2026-07-30; Installed+Updates, the mutation spine and search+install all 2026-08-04). The app discovers, lists, previews, decides, executes, streams, summarises and installs over real commands. "Update everything" is 4 clicks and "install one" is 4 clicks, both counted by hand in the running app; search is 22 ms median per keystroke against a 50 ms budget. 15 of UI-SPEC §6's 16 components exist — PdEnvSwitcher alone does not, and nothing needs it while the header shows the selected interpreter. (This line said 16 of 16 for two milestones; §6 itself was corrected in P4 and is the authority.) docs/ROADMAP.md Phase 2 has a table per stage and says where to pick up; read it before starting a slice.

Phase 3 is complete — six slices, exit criteria met and recorded. P4 put Code Health on screen: PdHealthReport is UI-SPEC §6's fifteenth of sixteen, EnvRow.healthProject remembers the folder per environment, and health_save_report writes Markdown + JSON. NOT_YET was down to two after it, both M3-general — P5 landed the gated ruff --fix, PipDock's first write outside site-packages and %LOCALAPPDATA%. P4 did need new commands, contrary to the earlier claim — one shipped, and cancellation was deferred on a measurement (a warm three-tool run is 1.3 s in release).

Four P4 defects worth not repeating, all found by running: an empty state that said no issues found before anything had run; a project declaring no dependencies reported as a deptry failure rather than as deptry not being applicable; env_probe hardcoding EnvSource::Manual, so Upgrade pip relabelled a registry-discovered interpreter; and a test asserting a version Dependabot owns, which made merging a ruff bump break main after CI had gone green on a base predating the code that would have caught it. Three tool facts are pinned by tests because the specs had them wrong: all three exit non-zero on findings and vulture uses 3; deptry emits a flat list keyed by module to a file path and cannot be told which environment to compare against; ruff's docs URL is keyed by rule name.

P2 built the tools venv — health::sync_tools_venv fills %LOCALAPPDATA%\PipDock\data\tools\.venv from the shipped pins over pipdock tools sync, and there is no TOOLS_PYTHON_MAX because deptry ships an abi3 wheel (ROADMAP's premise for that const was wrong; the record says why). P1 closed P0-10: pip's version is on every Environments row and Upgrade pip is 2 clicks. P3–P5 followed — the three runners and HealthReport, the Health screen, the gated ruff --fix — and are recorded in ROADMAP. P2 touched no NOT_YET entry: it is CLI-only on purpose, so P3's health_run owns the implicit sync rather than the frontend getting two ways to reach one operation.

NOT_YET is empty since 1.6.0: env_add_manual shipped with Add interpreter…, beside a second command, env_remove_manual. NOT_YET in src-tauri/src/lib.rs is the authority and three tests keep it honest in both directions — this file has said "four", "two" and "two" in different paragraphs at once, each true when written and none updated. A future command that is declared before it is built goes back on that list, not into a sentence here.

1.4.0, the audit release, left these behind — most found by running rather than by reading, and each rule below backed by a test where a test can hold it:

  • A flow is only as safe as what can drive it. The stale-plan and drift checks (DATA-FLOW §9.3) never fired, because nothing could run a flow without a real engine, so nothing did. FlowOptions, the Engine::probe seam and tests/flows.rs's scripted engine exist so that every flow decision has a test needing no pip. A new flow decision without one is a regression.
  • A check that cannot run must not report a pass. unwrap_or(ok: true) said "verified" exactly when nothing was — a pip broken by the run it was checking. CheckReport.error says so. The empty-plan path kept its own copy of the bug until the last UI commit of the release.
  • python -I ignores PYTHONUTF8 and PYTHONIOENCODING (it implies -E), so invariant 7's variables did nothing for the probe, and every venv under a folder named in Vietnamese crashed it. -X utf8 survives isolation; probe.py runs under it.
  • A grant is a contract no frontend test can see. Decline on the legal gate did nothing in the installed app while its test passed: the test mocks close(), and core:default never granted core:window:allow-close. lib.rs pins the webview's exact permissions now — widening them is a test change, on purpose.
  • Data drift and format drift look identical in a red drift job. Seven weeks of red were two causes: PyPI data (idna's newest release, scipy's version list) and uv 0.13 rewriting its error rendering. Data is redacted in capture.py; a format change gets a frozen corpus of its own (tests/fixtures/uv-0.11/, the last miette-style uv) beside the live one (uv/) — or, when only some scenarios moved, a partial one: pip-26.2/ holds the one capture pip 26.3 re-rendered.
  • Normalise everything you hash, not most of it. The legal-docs hash normalised CRLF in the content and hashed the raw length, so a Windows checkout re-showed the gate for nothing.
  • --json is one document. Scripts parse stdout with one ConvertFrom-Json; prose, prompts and warnings go to stderr. Every command's shape is in CLI-SPEC §6, and the goldens hold it.
  • Never put a backslash in a Bash heredoc that becomes source code: the tool collapses \\ to \, which turned b"\r\n" test strings into real line breaks and a CRLF test into one that could not fail. Write such files with the Write/Edit tools, then check the failing-first run actually failed.

1.5.0, the accessibility pass, left these behind — the first found by reading, the rest by driving the installed build:

  • jsdom is wrong about focus and blind to the cascade, and both cost a shipped defect. A browser takes focus from a control that becomes disabled; jsdom does not. So nine buttons that disabled themselves when pressed dropped focus to <body> with every test green. Busy is aria-disabled plus a click guard; disabled is for unavailable. And jsdom applies no stylesheet: the forced-colors border on destructive buttons sat in @layer base, lost to Tailwind's border utility in a later layer, and had never been drawn. A rule that must beat a utility goes outside every @layer, and destructive.test.ts counts the braces.
  • The recipe that found them. Install the bundle, launch with WEBVIEW2_ADDITIONAL_BROWSER_ARGUMENTS=--remote-debugging-port=<n> and a scratch LOCALAPPDATA (a release build honours it), and speak CDP to http://127.0.0.1:<n>/json: Input.dispatchKeyEvent for real key presses (Enter needs a char event with \r), Runtime.evaluate for document.activeElement, Accessibility.getFullAXTree for names, and Emulation.setEmulatedMedia with forced-colors: active. No system setting changes. It cannot hear a screen reader — say so wherever the result is written down. A scratch profile still discovers the owner's real interpreters: select the scratch venv before starting any plan.
  • Anything that replaces what <main> shows without changing nav belongs in App.tsx's second focus effect. There were four; the legal gate was a fifth, found after the first four were fixed and tested. The effect acts only when focus was lost, so adding a dependency is safe.
  • A destructive button carries data-destructive, whatever colour it is drawn in, and a test reads the source for it. Do not add a data-action to a stylesheet list; there is no list.
  • An accessible name is assembled without spaces. Inline spans join as Updates(5); put a literal {' '}. A browser adds one before visually-hidden text, so : numpy in an sr-only span read "Remove packages : numpy" — use aria-label for a name with punctuation in it.
  • sr-only text needs a positioned parent. It is position: absolute, so with no positioned ancestor it is laid out against the viewport, outside every scroller: a list of pin chips on Pins made the whole window scroll. Put relative on whatever holds the hidden text.
  • ACCESSIBILITY.md is hand-kept and says who tested what. A change to keyboard, focus, names or contrast is a change to it, the way a behaviour change is a change to legal/. It currently says Narrator was not re-run on 1.5.0, and only a person can change that sentence.
  • eslint-plugin-jsx-a11y peers ESLint 9 at most, like react/jsx-no-literals. axe-core is called directly from ui/src/test/axe.ts; color-contrast is off there, and is tokens.test.ts's job.
  • Zoom is off on a measurement (zoomHotkeysEnabled): at 200 % in the default window the table's name column collapses to zero. Backlog A5 says what has to change first.

1.6.0, Add interpreter…, left these behind — found by an independent review after every suite and the installed-build smoke were green:

  • The path you check is not the path you run. validate_interpreter canonicalized to compare against PipDock's own venvs and returned that, so a --symlinks venv — whose python.exe links to its base — was run and stored as the base, and --env .venv update --all would have changed the system Python. Resolve to decide; run and store what was given, made absolute. envHash follows links for the same reason, so nothing that must tell a venv from its base may key on it.
  • Anything that resolves a path needs a fallback to the path. A Microsoft Store Python is an app-execution alias that canonicalize() refuses (os error 1920). canonical_interpreter always fell back; the new validator did not, and env use would have broken for every Store user.
  • A check made once is a claim about the past. An added entry is validated again on every listing (envs::probe_listed, both heads), and the CLI's stored default when it is used: a link can be re-pointed after it was accepted.
  • Remove a row by what it shows, not by an identity recomputed from it. For a file that is gone, env_hash hashes the stored text, which is not the identity it was added under whenever canonicalizing changed it (a share: UNC\server\…); removal by hash reported success, and the row came back on the next scan.

Three rules P1 left behind, all found by running rather than reading:

  • A venv is never externally managed, and sysconfig.get_path("stdlib") inside one resolves to the base. probe.py was missing the early return pip's own check_externally_managed opens with, so every venv built from a uv-managed, Debian, Homebrew or Fedora Python refused every mutation. Anything reading a path out of sysconfig inside a venv is asking about the base.
  • envs::venv_scan matches exactly .venv, venv, env, .env. A scratch venv named anything else is not discovered; since 1.6.0 it can be added — Add interpreter… in the GUI, pipdock env add in the CLI — and that is how hand-verification reaches a venv outside the working directory. A path a person supplies goes through envs::validate_interpreter (named like Python, not PipDock's own venvs) before anything runs it; the picker is not the check.
  • Every pip below 22.2 is broken on Python 3.12+ (distutils, pkgutil.ImpImporter), so the one case Upgrade pip exists for is also a pip that cannot run. upgrade_pip falls back to ensurepip, and decides by asking pip --version rather than matching a traceback.

Three rules from S2/S3 that bind everything after them:

  • "N pinned excluded" in the table is presentation, not enforcement. DATA-FLOW §9.5 is enforced by pins::filter_upgrades at the plan boundary; a preview must show the flow's excluded_pins(), not a number the UI computed, or the two drift.

  • Never render a state you have not loaded. Row state is three-valued (unknown/current/outdated) because pkg_list is local and pkg_outdated is networked. Treating "not yet known" as "up to date" dims 200 rows and un-dims a handful a second later.

  • A Tauri command's future must be Send, and Store is not Sync. Never hold the store guard across an await — it does not compile at the command boundary, and where it does compile (inside core) it serializes every other command behind a network call. Read what you need, drop the guard, then await. This is why UpdateFlow::start takes pins rather than a &Store, and why index::metadata/refresh take the app-data path. It has bitten twice; assume it will again.

  • Two slices that key on different things need two resets. The package slice is keyed to the selected environment and the snapshot timeline to the open one, and they move independently — so folding the timeline into NO_PACKAGES meant a rescan wiped a freshly-loaded timeline whenever the package slice happened to be stale. After a rollback that is exactly the case. A shared reset constant is only correct for fields with a shared key.

  • A spec can promise a field that has never existed. SECURITY §6 asked for a severity per finding and PRD P1-1 for "severity-sorted"; pip-audit publishes no severity under either vulnerability service, and never has. §6 also said pip-audit runs from the Code Health tools venv, which the code had refused since P2 with three tests enforcing the refusal. Both were found by running the pinned tool and reading its JSON, not by reading the specs against each other. When a spec sentence describes an external tool, the tool is the authority — and the correction belongs in the spec, not in a workaround that invents the missing field.

  • A watchdog is a unit, not a number. audit::run inherited Code Health's 120 s TOOL_TIMEOUT and hit it on the first real run, reporting PD-HLT-003 for a tool that was working. Code Health is a linter walking a source tree; an audit is a network fetch. Measured: 68 s cold for one package, 20.0 s warm for twelve — twelve cost 1.4 s more than one, so the cost is a one-off database fetch and package count is not the risk. The same measurement reopened cancellation, which P4 had deferred on Code Health's 1.3 s. A borrowed constant carries its original units with it.

  • Adding a field to a struct that is summed is a bug the suite will not see. Target::Audit gave cache::Usage a fourth entry and total_bytes still added three — every cache test passed and the number on screen was wrong with nothing to say so. The guard asserts on distinct powers of two so a missing term appears as its own bit rather than as a plausible total. Same shape as the entryOf ternary whose last branch was a hand-maintained negation of every case above it.

  • A dialog is not an enforcement point. DATA-FLOW §5's three options are what the user sees; what stops a removal is GuardAck inside UninstallFlow::execute, beside SnapshotProof. The same shape three times now — SnapshotProof, GuardAck, and P5's FixConsent: the report and the proof are produced in one call and consumed in another, so the only way "somebody forgot to look" cannot happen is a value the executing call demands. Any future flow with a decision between two IPC messages needs its own.

  • ruff --fix takes no snapshot, and that is the reasoned answer rather than an oversight. DATA-FLOW §9.1/§9.2 are scoped to a mutating engine call; a fix rewrites the user's source tree, which no snapshot in this application describes, so one taken here would have no consumer that could use it — invariant 2's own argument for the pip-upkeep exemption, a second time. The safety net is the user's own git, asked about twice (once to decide what the dialog renders, once inside health_fix to decide what it allows), and the exemption is stated in the confirm rather than left silent. DATA-FLOW invariant 6 is where this lives.

  • Measure in --release, and say which build a number came from. A debug build measures bounds checks: the same index load is 572 ms in debug and 140 ms in release, and the same keystroke is 176 ms against 16 ms. A design was justified with a figure four times too large before anyone noticed. crates/pipdock-core/tests/search_latency.rs refuses to run in debug for that reason.

Things worth knowing before you change any of it:

  • cargo test fails when ui/src/ipc/generated.ts is stale. Fix with cargo run -p xtask -- bindings; the failure names that command and the first differing line.

  • The L4 goldens (crates/pipdock-cli/tests/golden.rs, 70 snapshots at 1.6.0) are the CLI's output contract. A diff there is a real behaviour change — re-bless deliberately, never reflexively. They are what made the core::flow refactor provably behaviour-preserving.

  • Two tests hold the wire format: Code::ALL must serialize as as_str(), and no SCHEMA_TYPES property may contain _. Both fail at cargo test.

  • A control that is not a tab stop is a control a keyboard user does not have. Making the package row the only tab stop is right — 200 rows would otherwise be 600 — but every button inside one was tabIndex={-1}, so Pin and Remove were mouse-only, which is a WCAG 2.1.1 failure. The ARIA grid answer is a roving tabindex (←/→, Esc back out), and P6 added it. Enter on a row is the non-destructive primary — pin, never uninstall.

  • The seam between a tested component and its untested parent is where bugs live. PdSummarySheet had tests; PdPlanPanel had none; both rendered plan.done, so a successful run showed Back to packages twice and every suite stayed green. Found by a human using the installed build. When adding a component test, ask what renders it.

  • A hand-written fixture that matches the docs while the code does not is worse than none. fixtures.rs says so in its own module doc, and guard_report() is computed for that reason — but the plan fixture's blockers were hand-written to the shape plan::Blocker's doc describes while graph::blockers_for had drifted to building an English sentence. So PdPreviewDiff asserted the right sentence and proved nothing, the preview shipped reading scipy requires scipy 1.11.4 requires numpy <1.28.0,>=1.21.6 from S3 to 1.1.0, and no suite could see it. If a fixture's value can be computed by the function that produces it in production, compute it.

  • A sentence assembled in Rust is a sentence the other head assembles again. That was the mechanism above: the CLI prints Blocker.constraint alone, so building the phrase in core made the CLI look right and left the GUI to wrap it in plan.blocker a second time. Invariant 4 is not only about translation — a joined string cannot be un-joined by a catalog. Structured fields, one composition per head, as BrokenDependent/PdUninstallDialog have always done.

  • Never sed -i anything under tests/fixtures/. Git Bash's sed rewrites in text mode and strips every CR, which is exactly what .gitattributes' -text exists to prevent; one command flattened ten engine fixtures and reported 160 changed lines where one was intended. Use a binary-mode read/write and assert the CRLF count per file.

  • A parser that is correct for trusted input is a hole for untrusted input. graph::Requirement::parse splits the name at the first space, so this is not a requirement parses as package this with constraint is not a requirement — exactly right for Requires-Dist metadata, which never looks like prose, and a spec containing spaces heading for argv the moment the input is a file a user typed. Reuse the parser, then re-validate at the new boundary; SECURITY §2's obligation follows the input, not the function.

  • Before adding a size or a "clear" for something, check it exists. LOG_RETENTION_DAYS has had no reader since it was written, so P1-C's cache manager covers three artefacts rather than the four the roadmap listed. A row reading "logs — 0 B" would have invented a subsystem.

  • A test can assert a substring you did not think of. PdPins.test.tsx proves the screen offers no way to create a Hold pin with queryByText(/hold/i) — and "threshold" matches that. P1-A's Settings label is "Suggest pinning at" for that reason alone. Before adding copy to a screen, grep its test for regex assertions; a broad /word/i over a whole screen is a trap laid for a feature that does not exist yet.

  • A doc comment describing an intent nothing implements is worse than none. ReverseDeps::build said auto-suggest "may still count" extra-gated edges while dependent_count had always filtered them. Whoever built the feature would have believed the comment or the code, and had no way to know which. When you find one, settle it in the comment rather than leaving both readings alive.

  • A green badge is not a green main. The README badge is per workflow file and shows that file's last run, which may be a schedule on a different commit. main was red for a day at 3eeda05 — every cargo job failing before a crate compiled — because Cargo.lock carried a dependency edge to thiserror 2.0.19 while its package entry said 2.0.20, which --locked refuses to repair. Start with gh run list --branch main, not the badge. And when bumping a version: it lives in five places (Cargo.toml, Cargo.lock, package.json, package-lock.json, tauri.conf.json) and they move together or you have reproduced this bug. cargo metadata --locked is the one-command check.

  • A docs/ fix and a legal/ fix are two different changes. Removing the self-updater updated THIRD-PARTY-NOTICES.md and not PRIVACY-POLICY.md, so the shipped privacy policy described a feature that no longer existed — and, worse, still told users to delete %LOCALAPPDATA%\PipDock\ for "a complete reset", which is the folder the program installs into. docs/SECURITY.md had recorded the data-root move correctly the whole time. Anything under legal/ is what a user reads; grep it separately whenever behaviour changes. Every legal/*.md and LICENSE is hashed into PIPDOCK_LEGAL_DOCS_HASH, so any edit re-shows the gate once — that is the mechanism working, not a regression.

  • A spec sentence is not a test. UI-SPEC §8 asserted for two milestones that Ctrl+1..N refused while a plan owned the screen. It did not: App.tsx guarded on guardOpen, and PANEL_PHASES excludes guard on purpose, so the two sets are disjoint and guarding on one guarded neither. Invisible until the plan ended and dropped the user on a tab they never chose.

  • Appending to NAV_KEYS is free; inserting is not. Ctrl+9 for About cost one array entry. Four comments claimed "a ninth entry would renumber Ctrl+1..8" and had the reason wrong — Snapshots stays a mode of Environments because it would sit beside Environments, an insert.

  • Building and installing are their own test tier, and nothing else substitutes. The first bundle ever produced exposed three defects in one afternoon: release.yml could not have run (--locked is cargo's flag and goes after --), the installers shipped no pipdock.exe at all, and the program installed into the data directory — making SECURITY §8's "delete the folder to reset" advice uninstall the app. None was visible in a diff. See docs/ROADMAP.md's Phase 4 section for the layout rules that came out of it.

  • Run it, don't just test it. Every unplanned bug in Stage 1 and Stage 2 came from executing against a real environment, a real runner or a real browser — never from reading code or from a passing suite. Stage 1: the watchdog that never killed anything, --json that never parsed, attribution that told 3.12 users something false, a drift job that could not go green. Stage 2: a probe 10× slower than before (found by timing it), a size that was wrong rather than missing for editable installs and a package listed twice (both found by one real pip install -e), and every screen fetching twice on mount (found by logging what actually crossed the bridge). Stage 3: a cancellation test that passed while the orphan lived on. Stage 4: an index load reported four times too large because it was measured in debug, a keystroke budget that was being missed by 30 ms, and a plan started from Search that resolved into a screen which does not render it — the command ran, the flow parked, and the user saw nothing change.

  • probe.py is on the hot path. The Installed screen re-probes on every environment open, so anything added to _dists() is paid 200+ times per open. Distribution.files looks harmless and costs 10×; read the metadata files directly and measure before committing.

Things that look like bugs but are not

  • probe.py reporting fewer packages than pip list on a non-venv Python is the documented -I trade-off; hidden_user_site is non-null exactly then.
  • snapshot diff showing entries a rollback cannot restore (editable installs, direct URLs) is deliberate: they are excluded from the pin set and reported separately rather than dropped.
  • pipdock engine uv failing when uv is not on PATH is intentional — storing a preference that cannot be honoured just moves the failure to the next command.
  • The app has no self-update and no tauri-plugin-updater. Removed on purpose (SECURITY §5): a self-updater is a standing remote-code-execution path into a tool that already runs subprocesses against the user's interpreters, and it has to be defended forever — a key that must never leak, an endpoint that must never be spoofable. Tauri makes adding it back a three-line change, so this is written down. Users get new builds from GitHub Releases and verify SHA256SUMS.txt. Do not reintroduce it without an explicit owner decision.
  • Fixtures under tests/fixtures/{pip,uv}/ are not literal engine bytes: spikes/capture.py redacts absolute paths, object addresses, progress bars and uv's timings. Without that they differ on every run and every machine, so the weekly drift job could never go green — and pip's "[notice] To update, run: …" shipped the capturing user's home directory to a public repo. It is safe because nothing reads a path or a duration out of them: every CLASSIFIERS entry keys off message text, and the uv adapter only tests contains("Resolved "). Line endings are still byte-exact, which is what .gitattributes' -text protects.
  • meta.json next to capture-provenance.json is a deliberate split. meta.json is the contract the drift gate watches; provenance records engine/interpreter versions and argv, churns on every release, and is excluded from the gate.
  • snapshot::Meta carries #[serde(alias)] on its renamed fields. It is the on-disk .meta.json and an IPC type; without the aliases every snapshot written before the camelCase change becomes unparseable, and the rollback the user was relying on silently is not offered. Removable only once no pre-1.0 snapshot can still be on disk.
  • take_snapshot() and execute() are two calls on purpose, so the CLI can print the snapshot id before execution and the GUI can draw DATA-FLOW §3's distinct states. The cost is that skipping the snapshot becomes expressible, which is why flow::proof_from refuses NotTaken with PD-SNP-001. Do not "simplify" them back into one.
  • ScanProgress.label and every path, version and engine id in the UI are data and are never translated (I18N §2).