Its python.exe links to the Python it was made from, and env_hash followed the link, so the venv's pins, snapshots and cached graph, report and audit were filed under its base. Nothing showed it while no screen could list the two apart; Add interpreter now can. A venv's own interpreter, recognised by its pyvenv.cfg, is resolved only up to its folder. No other path's identity changes, and a linked venv's earlier pins and snapshots stay with the base they were filed under. Checked against a real --symlinks venv: `env use` stores the venv with its own prefix, and the venv and its base keep two identities. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
39 KiB
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.
docs/PRD.md— goals G1–G7, personas, P0/P1/P2 feature matrix, non-goalsdocs/ARCHITECTURE.md— crate layout,Enginetrait,probe.py, storage, Tauri IPC, version manifestdocs/DATA-FLOW.md— update/install/uninstall/rollback state machines, engine command mapping, invariants §9docs/UI-SPEC.md— design tokens, screens, click budgets, keyboard mapdocs/CLI-SPEC.md— commands, flags, exit codes, JSON contractsdocs/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 foldsPD-HLT-001..004into one, so it shows 32 rows. Adding a code means the catalog,Code::ALL, both counts, the "no code without a fixture" gate inclassifier_corpus.rs, bothlocales/*/errors.json, a regeneratedcodes.json(cargo run -p xtask -- ipc-fixtures), regenerated bindings, and re-blessed goldens —golden__schema-Code.snapand every schema that embeds it, which since 1.4.0 is five:StepResult,ExecutionSummary,CheckReport(itserror),HealthReportandAuditReport, the last two viaToolProblem.code. Six snapshot files in all. This line has been wrong about the count twice — grepPD-ENV-001acrosscrates/pipdock-cli/tests/snapshots/rather than trusting the list. Nine places; the gates will tell you which you forgot, includingi18n.test.ts, which fails on a code with no copy in either languagedocs/CODE-HEALTH-SPEC.md,docs/SECURITY.md,docs/TESTING.md,docs/RELEASE-CI.md,docs/I18N.md,docs/ROADMAP.mdlegal/— EULA, Disclaimer, Privacy, Third-Party Notices (public; the in-app legal gate links to these GitHub URLs)
Hard invariants — never violate these
- No mutating engine call without (a) a
ResolutionReportaccepted this session and (b) a successful snapshot write. Snapshot failure aborts the plan, executing nothing (PD-SNP-001). See DATA-FLOW §9. - Subprocess via argv arrays only — never a shell.
tokio::process::Commandwith explicit args. Package names validated against PEP 508 and versions against PEP 440 before they reach argv. See SECURITY §2. - Never reimplement dependency resolution. The selected engine resolves in dry-run mode; PipDock parses and explains. See ARCHITECTURE §1.2.
- 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. --break-system-packagesis 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)
- 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. - Every pip invocation needs
PYTHONIOENCODING=utf-8+PYTHONUTF8=1. Without it--dry-run --report -crashes on Windows/cp1252 (pip 25 and 26), data-dependently. - uv writes its plan to stderr, so classify errors only after a plan-parse attempt.
env_hashlowercases the interpreter path on Windows, or one env splits its pins and snapshots in two — and since 1.6.0 stops at a venv's ownpython.exe, or a--symlinksvenv files its pins and snapshots under the Python it was made from.- PipDock enforces
Requires-Pythonitself (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.mdentry, 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 inchangelog/. -
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). RootCHANGELOG.mdis a pointer only. Split a file by minor range at roughly 600 lines and never renumber an existing range file —changelog/README.mdis the authority. -
Crates
pipdock-*; React components prefixedPd(PdPackageRow,PdConflictRow) -
Tauri identifier
com.skullmute.pipdock; CLI binarypipdock -
Frontend lives in
ui/, butpackage.jsonsits at the repo root — first for the ops repo's reusable CI workflow, which had noworking-directoryinput, and now because SonarCloud, Qodana, the scripts and Tauri's config all assume it; Vite is configured withroot: './ui' -
Rust ↔ TS types are generated by
cargo run -p xtask -- bindingsoverschemars— never hand-mirror a type. (ARCHITECTURE §9 originally saidspecta/tauri-specta;bindings.rsrecords why it is not, and drift is acargo testfailure rather than a compile error) -
No component calls
invokedirectly; everything goes through the typed wrappers inui/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, theEngine::probeseam andtests/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.errorsays so. The empty-plan path kept its own copy of the bug until the last UI commit of the release. python -IignoresPYTHONUTF8andPYTHONIOENCODING(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 utf8survives isolation;probe.pyruns 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(), andcore:defaultnever grantedcore:window:allow-close.lib.rspins 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.
--jsonis one document. Scripts parse stdout with oneConvertFrom-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 turnedb"\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 isaria-disabledplus a click guard;disabledis for unavailable. And jsdom applies no stylesheet: the forced-colors border on destructive buttons sat in@layer base, lost to Tailwind'sborderutility in a later layer, and had never been drawn. A rule that must beat a utility goes outside every@layer, anddestructive.test.tscounts the braces. - The recipe that found them. Install the bundle, launch with
WEBVIEW2_ADDITIONAL_BROWSER_ARGUMENTS=--remote-debugging-port=<n>and a scratchLOCALAPPDATA(a release build honours it), and speak CDP tohttp://127.0.0.1:<n>/json:Input.dispatchKeyEventfor real key presses (Enterneeds acharevent with\r),Runtime.evaluatefordocument.activeElement,Accessibility.getFullAXTreefor names, andEmulation.setEmulatedMediawithforced-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 changingnavbelongs inApp.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 adata-actionto 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: numpyin ansr-onlyspan read "Remove packages : numpy" — usearia-labelfor a name with punctuation in it. sr-onlytext needs a positioned parent. It isposition: 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. Putrelativeon whatever holds the hidden text.ACCESSIBILITY.mdis 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 tolegal/. It currently says Narrator was not re-run on 1.5.0, and only a person can change that sentence.eslint-plugin-jsx-a11ypeers ESLint 9 at most, likereact/jsx-no-literals. axe-core is called directly fromui/src/test/axe.ts;color-contrastis off there, and istokens.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_interpretercanonicalized to compare against PipDock's own venvs and returned that, so a--symlinksvenv — whosepython.exelinks to its base — was run and stored as the base, and--env .venv update --allwould have changed the system Python. Resolve to decide; run and store what was given, made absolute. The identity followed the same link:env_hashfiled a linked venv's pins, snapshots and reports under its base, which nothing showed while no screen could list the two apart. A venv's own interpreter (recognised bypyvenv.cfg) is now resolved only up to its folder. Rows are still matched by path, never by hash: two names for one Python share one. - 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_interpreteralways fell back; the new validator did not, andenv usewould 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_hashhashes 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.pywas missing the early return pip's owncheck_externally_managedopens with, so every venv built from a uv-managed, Debian, Homebrew or Fedora Python refused every mutation. Anything reading a path out ofsysconfiginside a venv is asking about the base. envs::venv_scanmatches 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 addin the CLI — and that is how hand-verification reaches a venv outside the working directory. A path a person supplies goes throughenvs::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_pipfalls back toensurepip, and decides by askingpip --versionrather 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_upgradesat the plan boundary; a preview must show the flow'sexcluded_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) becausepkg_listis local andpkg_outdatedis 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, andStoreis notSync. 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 whyUpdateFlow::starttakes pins rather than a&Store, and whyindex::metadata/refreshtake 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_PACKAGESmeant 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::runinherited Code Health's 120 sTOOL_TIMEOUTand hit it on the first real run, reportingPD-HLT-003for 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::Auditgavecache::Usagea fourth entry andtotal_bytesstill 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 theentryOfternary 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
GuardAckinsideUninstallFlow::execute, besideSnapshotProof. The same shape three times now —SnapshotProof,GuardAck, and P5'sFixConsent: 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 --fixtakes 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 insidehealth_fixto 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.rsrefuses to run in debug for that reason.
Things worth knowing before you change any of it:
-
cargo testfails whenui/src/ipc/generated.tsis stale. Fix withcargo 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 thecore::flowrefactor provably behaviour-preserving. -
Two tests hold the wire format:
Code::ALLmust serialize asas_str(), and noSCHEMA_TYPESproperty may contain_. Both fail atcargo 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 (←/→,Escback out), and P6 added it.Enteron a row is the non-destructive primary — pin, never uninstall. -
The seam between a tested component and its untested parent is where bugs live.
PdSummarySheethad tests;PdPlanPanelhad none; both renderedplan.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.rssays so in its own module doc, andguard_report()is computed for that reason — but the plan fixture's blockers were hand-written to the shapeplan::Blocker's doc describes whilegraph::blockers_forhad drifted to building an English sentence. SoPdPreviewDiffasserted the right sentence and proved nothing, the preview shipped readingscipy requires scipy 1.11.4 requires numpy <1.28.0,>=1.21.6from 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.constraintalone, so building the phrase in core made the CLI look right and left the GUI to wrap it inplan.blockera 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, asBrokenDependent/PdUninstallDialoghave always done. -
Never
sed -ianything undertests/fixtures/. Git Bash's sed rewrites in text mode and strips every CR, which is exactly what.gitattributes'-textexists 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::parsesplits the name at the first space, sothis is not a requirementparses as packagethiswith constraintis not a requirement— exactly right forRequires-Distmetadata, 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_DAYShas 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.tsxproves the screen offers no way to create aHoldpin withqueryByText(/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/iover 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::buildsaid auto-suggest "may still count" extra-gated edges whiledependent_counthad 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.mainwas red for a day at3eeda05— every cargo job failing before a crate compiled — becauseCargo.lockcarried a dependency edge tothiserror 2.0.19while its package entry said2.0.20, which--lockedrefuses to repair. Start withgh 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 --lockedis the one-command check. -
A
docs/fix and alegal/fix are two different changes. Removing the self-updater updatedTHIRD-PARTY-NOTICES.mdand notPRIVACY-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.mdhad recorded the data-root move correctly the whole time. Anything underlegal/is what a user reads; grep it separately whenever behaviour changes. Everylegal/*.mdandLICENSEis hashed intoPIPDOCK_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..Nrefused while a plan owned the screen. It did not:App.tsxguarded onguardOpen, andPANEL_PHASESexcludesguardon 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_KEYSis free; inserting is not.Ctrl+9for About cost one array entry. Four comments claimed "a ninth entry would renumberCtrl+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.ymlcould not have run (--lockedis cargo's flag and goes after--), the installers shipped nopipdock.exeat 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. Seedocs/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,
--jsonthat 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 realpip 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.pyis 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.fileslooks harmless and costs 10×; read the metadata files directly and measure before committing.
Things that look like bugs but are not
probe.pyreporting fewer packages thanpip liston a non-venv Python is the documented-Itrade-off;hidden_user_siteis non-null exactly then.snapshot diffshowing 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 uvfailing 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 verifySHA256SUMS.txt. Do not reintroduce it without an explicit owner decision. - Fixtures under
tests/fixtures/{pip,uv}/are not literal engine bytes:spikes/capture.pyredacts 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: everyCLASSIFIERSentry keys off message text, and the uv adapter only testscontains("Resolved "). Line endings are still byte-exact, which is what.gitattributes'-textprotects. meta.jsonnext tocapture-provenance.jsonis a deliberate split.meta.jsonis the contract the drift gate watches; provenance records engine/interpreter versions and argv, churns on every release, and is excluded from the gate.snapshot::Metacarries#[serde(alias)]on its renamed fields. It is the on-disk.meta.jsonand 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()andexecute()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 whyflow::proof_fromrefusesNotTakenwithPD-SNP-001. Do not "simplify" them back into one.ScanProgress.labeland every path, version and engine id in the UI are data and are never translated (I18N §2).