The stage record said four items were open; they are not. `engine_info`, the `COMMANDS` drift gate, focus-follows-tab and Search's empty state all landed, so ROADMAP records what shipped rather than what was owed an hour ago. Two things are still deliberately out, and named: the `Enter` primary-action map and folding six `aria-live` regions into one. Both are cross-screen behaviour whose correctness is a screen reader and the VI sweep, not a passing test — landing them blind would be the kind of green suite this repo has learned to distrust. M2 closes here. Every P0 feature except Code Health is in both heads, all seven UI-SPEC click budgets are met by hand count, and `NOT_YET` names the five commands that remain against the slice that owes each. Where to pick up now points at Phase 3, with the two decisions worth not re-deriving and the ten-minute check that retires M3's biggest unknown before any of it is written. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
39 KiB
PipDock — Roadmap
Version 0.1 · 2026-07-17
Phase 0 — Spike week (before any feature code) — RUN 2026-07-27
The entire architecture stands on assumptions that must be verified on a real machine first (the FrameLedger/QuoteAtlas discipline). Each spike is a small script + captured fixtures committed to crates/pipdock-core/tests/fixtures/.
Outcomes — full write-up in spikes/README.md.
| Spike | Verdict | The one thing that changed |
|---|---|---|
| SP-1 | GO — ship both engines | Neither engine holds packages back on install -U <pkg>; both silently break installed dependents. Every plan must restate the full installed set as explicit requirements. uv's conflict attribution is better than pip's; uv writes its plan to stderr, not stdout. uv minimum pinned at 0.10.0. |
| SP-2 | Corpus captured; blocker found and fixed | pip install --dry-run --report - crashes with UnicodeEncodeError on Windows/cp1252 (pip 25 and 26). Every pip invocation must set PYTHONIOENCODING=utf-8 + PYTHONUTF8=1. Also: yanked releases exit 0 (a preview warning, not a failure), and pip/uv disagree on Requires-Python enforcement. |
| SP-3 | PASS | 858 k projects; cold refresh 2.95 s (budget 60 s); worst keystroke 42.1 ms (budget 50 ms) — but only in memory: scanning SQLite per keystroke costs 218 ms. Margin is thin on fast hardware, and raw nucleo ranking puts requests-ntlm above requests. |
| SP-4 | Answered | --path cannot be combined with --no-deps; freeze-file mode (-r <freeze> --no-deps -f json) is the only option. SECURITY §6 confirmed. Findings arrive with duplicate advisory ids. |
| SP-5 | Answered by SP-1 | No engine reports held-back items at all. Resolved version, latest version and blocker must all be derived by PipDock — for pip as much as for uv, contrary to ARCHITECTURE §3's implication. |
| SP-6 | Answered | uv python list returns downloadable entries and duplicate shims; env_hash must lowercase the interpreter path on Windows or pins and snapshots split in two. probe.py -I hides 24 user-site packages on a system Python — open decision. Microsoft Store Python still unverified. |
Both adapter-behaviour questions were decided 2026-07-27 and are implemented: PipDock enforces Requires-Python itself so the preview is engine-independent (crates/pipdock-core/src/compat.rs, ARCHITECTURE §3), and probe.py keeps -I while reporting hidden_user_site so the Installed screen can disclose the gap accurately (SECURITY §2). Two non-blocking questions remain at the end of spikes/README.md: search ranking scope, and how to obtain the five error fixtures this machine cannot produce.
| # | Spike | Question to answer | Exit criterion |
|---|---|---|---|
| SP-1 | uv dry-run shape | Exact stdout of uv pip install -U --dry-run (current + previous uv) across: clean upgrade, held-back, impossible. Is held-back attribution derivable, or does uv only show the resolved set? |
Fixture set committed; adapter parsing strategy written; uv minimum version pinned. Go/no-go for launching with both engines — if uv's plan output is too lossy, v1.0 ships pip-primary with uv behind a "beta engine" flag. |
| SP-2 | pip report & stderr corpus | Capture --dry-run --report JSON + real stderr for every ERROR-CATALOG build/network/permission case (broken sdist, MSVC-missing, PEP 668, locked file, yanked, hash mismatch). |
≥ 1 fixture per launch catalog code. |
| SP-3 | PEP 691 index economics | Full name-index size, download time on VN broadband, SQLite ingest time, fuzzy-search latency on ~600 k names. | < 60 s cold refresh, < 50 ms search on the i7-14700KF and on a low-end reference VM; else design delta/compression fallback. |
| SP-4 | pip-audit foreign-env mode | Correct invocation for auditing an env the tool isn't installed in (freeze-file --no-deps vs newer options). |
Command line + JSON fixture pinned (P1 feature, but flags decided now). |
| SP-5 | Held-back attribution accuracy | On a real numpy/scipy/pandas pin tangle: does graph-based blocker attribution match reality? | Preview sentences verified by hand against the tangle; ambiguity fallback path exercised. |
| SP-6 | Windows env discovery sweep | PEP 514 registry + py -0p + Microsoft Store Python + uv-managed on a real machine; Store-Python aliasing quirks. |
Discovery module design note updated; Store Python either supported or explicitly detected-and-explained. |
Phase 1 — M1 "Core + CLI" — substantially complete 2026-07-27
Engine trait + both adapters (per SP-1 outcome), probe.py, discovery, index cache + search, plan/resolve/execute two-phase, snapshots, pins, guard graph, error catalog, CLI over all of it. Exit: TESTING L1+L2+L4 green in CI; pipdock update --all survives the SP-5 tangle env with correct summary; dogfood on Kokone's own bot venvs begins.
What works
Every P0 CLI command except health: env list|use, list [--outdated], search, info, install, update, uninstall, pin add|remove|list, snapshot list|create|diff|rollback, doctor, pip-upgrade, engine [pip|uv], index refresh, schema, self report-bug. 173 tests; cargo fmt, clippy -D warnings and npm audit clean.
Verified against real environments, not only unit tests:
- the held-back conflict from SP-1 is reported and attributed —
httpcore 0.15.0 (latest 1.0.9)/httpx 0.23.0 requires httpcore <0.16.0,>=0.15.0— which no engine produces; update --allresolves the same tangle to latest, applies it, and leavespip checkclean;snapshot rollbackrestores an environment exactly (PRD §6's "env diff vs snapshot = ∅");- the uninstall guard refuses to break
requestsand exits non-zero, per CLI-SPEC §7; - exit codes match CLI-SPEC §5.
Findings that changed the design
plan_requirements |
Every resolve restates the installed set, or both engines break installed dependents at exit 0 (SP-1). |
-U in the dry-run argv |
Without it pip plans nothing and every package looks held back for no reason. The unit test had encoded the omission instead of the document. |
| Search ranking | Tiered, not score-ordered, and ranked by selection rather than a full sort: 90.5 ms → 16.5 ms against a 50 ms budget. |
SnapshotProof |
--no-snapshot is a named waiver, not an Option, so forgetting a snapshot cannot look like deliberately waiving one. |
Not yet done
health— belongs to M3 with the tools venv (CODE-HEALTH-SPEC).TESTING L2 in CI— done 2026-07-29. It had run nightly on 07-28 and 07-29 and failed both times; five causes were found and fixed (see L2's first runs below). Run30460248664onmainis green on all three jobs, including the fixture-drift re-capture that had never passed.- TESTING L4 —
assert_cmdis a dependency and the clap surface is covered, but the golden-output tests per command are not written. The SP-5 tangle env— done 2026-07-29. A 38-package numpy/scipy/pandas/matplotlib/scikit-learn/statsmodels environment on Python 3.12, with scipy pinned so numpy had to be held back.update --allreported 15 successful, 0 failed, 0 skipped, exited 0, and leftpip checkclean; numpy stayed at 1.26.4, the scipy pin held (invariant 5), and the snapshot was written before any mutation (invariant 2). See What the dogfood found below — the answer to SP-5's question was "yes, but the explanation was wrong".- Settings beyond engine choice — locale, thresholds and the PEP 668 override land with the GUI in M2.
L2's first runs — what actually failed (2026-07-29)
The job ran and failed three times. No cause was a product bug — all five were in the test harness, and all are fixed. Worth reading before touching capture.py or the workflow, because most of them look like the product misbehaving.
1. The rollback assertion targeted the wrong snapshot. The step did snapshot create → update --all → snapshot rollback latest → snapshot diff latest, asserting the prose matches snapshot. But latest moves twice during that flow: update --all writes a Trigger::Plan snapshot before mutating, and snapshot rollback writes a Trigger::Rollback one before restoring (DATA-FLOW §8 — rollback is itself reversible). So the final diff latest compared the restored environment against the pre-rollback state and correctly reported every package as changed. The rollback itself had worked. Fixed by pinning the id from snapshot list --json immediately after create, and asserting structurally on diff --json rather than on prose.
2. fixture-drift could never have gone green. spikes/capture.py built each scenario in a mkdtemp directory whose name carries a random suffix, and redact() replaced only the temp root — so the suffix survived into every committed sidecar, and every re-capture differed. Four further sources of churn were found underneath it: pip's [notice] To update, run: <venv>\Scripts\python.exe (which also leaked the capturing user's home directory into a public repo), CPython object addresses in urllib3 retry warnings, pip's download progress bar with its transfer rate, and uv's per-phase timings (Resolved 3 packages in 775ms). Fixed by redacting all of them in capture.py, splitting the sidecar into meta.json (contract, gated) and capture-provenance.json (versions and argv, excluded from the gate), and capturing cache-free so a warm dev machine and a cold runner agree.
3. The guard step failed on the outcome it was asserting. Fixing the rollback let the job reach the uninstall-guard step, which requires pipdock uninstall urllib3 to exit non-zero (CLI-SPEC §7). The guard did exactly that — but the runner ends a pwsh step with exit $LASTEXITCODE, and nothing had reset it. The other seven end-to-end steps were audited; each ends on a command that legitimately exits 0.
4. The redaction regex ate prose. With the drift diff finally readable, it showed Some things HTTP Core does d<PATH> — pip's --report embeds package descriptions, and does do:\n\n* Sending… reads as drive o: followed by \n path segments. A lookbehind requiring a non-alphanumeric before the drive letter is what distinguishes a path from prose.
5. report-encoding-crash cannot be re-captured on CI. The SP-2 UnicodeEncodeError needs a cp1252 console and a runner's is already UTF-8, so re-capturing there recorded an ordinary successful report and destroyed the evidence. Scenarios now carry a reproducible flag; host-dependent ones are skipped in bulk runs and stay capturable with --only. pip's report also echoes the PEP 508 environment markers, which carry the host OS — normalized, while the interpreter version markers are deliberately left alone, since those can change a resolution.
Verified two ways: locally, both engines' captures run twice back to back differ in 0 of 96 files; on CI, run 30460248664 re-captured against the newest pip and uv and matched the committed fixtures byte for byte.
Two things worth keeping in mind, both recorded in .gitattributes and CLAUDE.md: -text is what preserves the CRLF bytes the fixtures need, and -diff — which was also set — only suppressed the textual diff, so the drift job's own diagnostics could report nothing but "Binary files differ". Faults 3–5 were all found because that was fixed. And the weekly drift job's if: matched the nightly cron, so the PyPI-heavy job the comment says would be "rude" to run daily was running daily; it now has its own '17 4 * * 1' entry.
What the dogfood found (2026-07-29)
SP-5 asked whether graph-based blocker attribution matches reality on a real tangle. It does — and at scale it was also saying things that were not true. The held-back numpy came with eight constraints, four of which do not apply to Python 3.12:
numpy 1.26.4 (latest 2.5.1)
pandas 2.1.4 requires numpy <2,>=1.22.4 <- python_version < "3.11"
pandas 2.1.4 requires numpy <2,>=1.23.2 <- python_version == "3.11"
pandas 2.1.4 requires numpy <2,>=1.26.0 <- the only pandas branch in force
statsmodels 0.14.1 requires numpy <2,>=1.22.3 <- python_version == "3.10" and Windows
Requirement has always carried its marker, but only extra == was honoured, so every marker-gated branch of a dependency was reported as though it were in force. That is not just noise — telling someone on 3.12 that "pandas requires numpy >=1.22.4" is false, and ERROR-CATALOG's premise is that what PipDock says about a failure can be trusted. The same bug sat under the uninstall guard, which would refuse a removal on the strength of a dependent that only needs the package on another Python.
Fixed in graph/markers.rs: python_version and python_full_version are evaluated, with and/or/parentheses. Platform markers are left alone — they are constants for a Windows-only v1 and reading them would mean plumbing more out of probe.py for no present gain. An unrecognised marker keeps the requirement, because over-reporting a constraint is noise while dropping one hides the reason a package is stuck. The same environment now reports five constraints, one per package, each genuinely in force.
One thing deliberately not changed: blockers are still derived from the installed set, so a package being upgraded by the very same plan can still be named. Here that meant matplotlib, pandas, scikit-learn and statsmodels appeared even though the plan moves all four to numpy-2-compatible versions, leaving pinned scipy as the only real blocker. Narrowing that needs post-plan metadata the preview does not have, and it makes the sentence over-broad rather than false. Worth revisiting when PdConflictRow is built.
Where to pick up
Pushed to main and green: CI / Rust, CI / Node and CodeQL all pass on the published commit. Three configuration bugs were fixed getting there, each recorded in its own commit message: caller workflows must grant at least what the callee declares (a startup_failure with no logs), cargo audit was pinned to a toolchain too old to build itself, and its informational advisories needed scoping rather than silencing.
Done 2026-07-29, closing out M1's CI debt:
Dependabot triage— six PRs resolved. #3 (TypeScript 7) and #4 (rusqlite 0.40) closed per ARCHITECTURE §10; #1, #2, #5, #6 merged. No PRs are open.L2 green— all three jobs, both matrix legs, onmain.TESTING L4—tests/golden.rs, 46 snapshots. See What L4 covers and what it does not below.
Immediate, in rough order:
- Repo settings — branch protection, the Discord webhook secrets, and CodeQL/Discussions, per RELEASE-CI §5. None of these are committable and all are owner-only. The updater keypair that used to be on this list is gone: PipDock no longer updates itself (SECURITY §5).
M1 is otherwise closed. Everything else on this list is done, and Stage 1 of M2 — the IPC bridge — is the next piece of work.
What L4 covers and what it does not
TESTING §2 asks for goldens "against a mocked core". There is no seam to mock at — run.rs builds its engine from the global options and drives a real interpreter, and print_preview / print_summary are private functions that println! rather than return. So tests/golden.rs covers what is deterministic without an environment: the clap surface, the exit-code table, the --json error envelope, and schema <T> for all fifteen SCHEMA_TYPES. Preview and summary rendering stays with L2.
Those schema snapshots are also the review surface for M2's wire-format change. They pin two contract bugs as they stand today:
ExecutionSummaryserializesplan_id/stderr_tail, where DATA-FLOW §6 documentsplanId/stderrTailCodeserializes as its Rust variant name ("EnvInterpreterMissing"), where ERROR-CATALOG documentsPD-ENV-001
The error envelope, which main.rs hand-builds rather than deriving, already has the documented shape — so the binary currently disagrees with itself across its two --json paths. Fixing that is Stage 1 work, and the goldens make it a reviewable diff.
When the flow refactor moves rendering out of the CLI ("the flow never prints"), the preview and summary renderers become pure functions of core types and should gain their own snapshots there.
Then M2. The core carries everything the GUI needs; ui/src/ipc fixes the command names, the design tokens and EN/VI catalogs are scaffolded, and the shell renders.
Phase 2 — M2 "GUI shell + core flows" (~7–8 weeks)
Tauri app, design tokens, Environments/Installed/Updates/Search/Pins screens, preview + 3-way conflict UX, console drawer, summary sheet, snapshots UI, legal gate, EN/VI catalogs, settings. Exit: all UI-SPEC click budgets met by manual count; L3 green; VI sweep clean.
The estimate was 3–4 weeks until 2026-07-29, when surveying the code for M2 showed the bridge was scaffolding rather than substance: 26 command names with zero wrappers, one registered Tauri command, 2 of 16 Pd* components, 1 of 5 stores. Scope is unchanged; the estimate moved to match it.
Stage 1 — the IPC bridge — done 2026-07-30
Ten commits, 0ccee04..a553bf7. Everything M2's screens stand on now exists, and there is a running app.
| What landed | |
|---|---|
| Wire format | #[serde(rename)] on all 30 Code variants so they serialize as PD-*, and rename_all = "camelCase" on every IPC-crossing struct. --json had never matched DATA-FLOW §6 or ERROR-CATALOG §3 — and the two --json paths in the binary disagreed with each other. Held by a Code::ALL round-trip test and a test that walks every SCHEMA_TYPES schema rejecting any property containing _. |
| Codegen | ui/src/ipc/generated.ts from schemars via cargo run -p xtask -- bindings. Deviates from ARCHITECTURE §9's specta; reasoning and the revisit condition are in crates/pipdock-core/src/bindings.rs. Drift is a test failure, so it fires locally as well as in CI. |
core::flow |
UpdateFlow, UninstallFlow, RollbackFlow — resumable state machines, because the GUI's two decision points are IPC round trips with a render in between. plan::execute_rollback is new: core had rollback_plan() and nothing to run it, so the CLI hand-assembled an ExecutionSummary. The flow never prints. |
| Cancellation | CancellationToken through exec.rs, plus kill_on_drop(true) — the 600 s watchdog had never killed anything. A cancelled Phase A does not fall through to Phase B, and a step killed by us is Skipped, not Failed. |
| Progress | ProgressSink carries step/total; scan-progress finally has a producer and a payload (ScanProgress). |
| Vertical slice | Legal gate, Environments, Settings. core::settings, docsHash computed in build.rs, capabilities split and scoped. |
Four bugs turned up that were not in the plan, all found by running things rather than reading them: the watchdog leak; --json unparseable because snapshot … written went to stdout; blocker attribution naming constraints from other Pythons (the SP-5 dogfood); and updater:default granted to a plugin whose pubkey was a placeholder.
Stage 2 — Installed + Updates (read-only) — done 2026-08-04
Three PRs, d7da4d7... Both screens render from a real environment, and the slice's exit criteria
are met.
| What landed | |
|---|---|
| Docs | ARCHITECTURE §7's command list rewritten as a table of 32 — it was prose with a|b|c groupings, ambiguous enough that Stage 1 added app_info and plan_decide without amending it. Closes open owner decision #1. DATA-FLOW §7's "List installed" row corrected: it claimed an engine command neither head has ever called. |
| Core | engine::for_id — the CLI held the only copy of engine selection plus a duplicate KEY_ENGINE and a byte-for-byte copy of store::default_app_data. Dist.sizeBytes, summed from RECORD. PkgName validates on deserialize. |
| Bridge | pkg_list, pkg_outdated, pin_list|add|remove, registered and wrapped. pkg_list reads the probe, not the engine; both take a PyEnv so the outdated path never probes. |
| UI | PdBadge, PdEmptyState, PdPackageRow, PdPackageTable (virtualized) — 8 of 16 components now exist. One screen serving both tabs, useEnvStore grown, EN+VI catalogs, Space/Ctrl+A. |
| L3 | The repo's first rendered-component tests, fed from fixtures generated out of the real Rust types by cargo run -p xtask -- ipc-fixtures and held current by a staleness test. |
Four bugs turned up that were not in the plan, three of them found by running:
Distribution.filesis a 10× trap. The obvious way to sum RECORD took the probe from 551 ms to 5,492 ms on a 352-package environment — on the path the Installed screen runs on every environment open. Parsing RECORD text directly costs 79 ms.- The naive size is wrong for editables, not merely missing. A PEP 660 install has a valid
RECORD listing only its import shim: 240 bytes for a package whose sources are 8.1 MiB.
Detected via
direct_url.jsonand reported as absent, because a wrong number is worse than none. - A distribution could be listed twice. An editable puts its own
.egg-infoonsys.path, soimportlib.metadatafinds it there and at the venv's.dist-info.pip listshows it once. For the table it is an ambiguous join key and two rows a user cannot tell apart. - Every screen fetched twice on mount. React runs effects twice in development and the
loadedForguard is set only after the await, so both runs proceed.env_scanhad done this since Stage 1, invisibly, because both scans return the same rows.
Two decisions worth knowing before S3:
- Row state is three-valued, not two.
pkg_listis local,pkg_outdatedis networked, so there is a window where outdatedness is unknown. Treating it as "up to date" dims all 200 rows and un-dims a handful a second later — visibly wrong, and untrue while it lasts. - "N pinned excluded" is presentation, not enforcement. DATA-FLOW §9.5 is enforced by
pins::filter_upgradesat the plan boundary. S3's preview must report the flow'sexcluded_pins(), not the number the table computed.
Also resolved: UI-SPEC §8's "select-all-visible", which has no meaning under virtualization. It is the current filtered set, and §8 now says so.
Stage 3 — the mutation spine — done 2026-08-04
Two PRs. The app now previews, decides, executes, streams and summarises — PRD G1's promise that every mutating operation is previewed before it runs is real for the first time in the GUI.
| What landed | |
|---|---|
| Cancellation | A Windows Job Object, so a cancel reaches the build backends python -m pip spawns and not only pip. win32job is a safe wrapper, so unsafe_code = "forbid" stays absolute. |
plan-progress |
A tagged lifecycle (stepStarted / line / stepFinished) instead of a bare line, plus stream. ARCHITECTURE §7 amended. |
| Wire types | FlowStep, Intent, Decision and ProgressEvent cross IPC; roundsRemaining makes MAX_CONFLICT_ROUNDS visible for the first time. |
| Plan session | PlanSlot in AppState holds the resumable flow between the four calls that drive it; PD-RES-003 refuses a second plan. |
| UI | usePlanStore, PdPreviewDiff, PdConflictRow, PdConsoleDrawer, PdSummarySheet — 13 of UI-SPEC §6's 16 components now exist. |
| L3 | The last three of TESTING §2's five obligations, so the list is complete and §3's PR gate means something. |
Exit criteria, counted and measured in a running app: "update everything" is exactly 4
clicks — Updates → Select all → Update → Confirm — and it is still 4 with a conflict left at its
default, which is §5's second row. Cancelling mid-execution yields a coherent summary with
Skipped rows and the cancelled banner. The live region reads "2 of 4 complete" and the drawer
groups output by package, both of which were unimplementable before the lifecycle enum existed.
Three specification gaps closed, each recorded in UI-SPEC §4:
Will downgradehas a section.ChangeKindhas four variants and §4 named three groups. A compatible resolve routinely moves a package down, and2.0 → 1.9under a heading reading "Will upgrade" is misleading about the change most likely to surprise.Keep compatibleis disabled on impossible rows. §4 said every needs-decision row hosts the full 3-way control, butdefault_decision(is_impossible = true, …)returnsSkip— there is no compatible version to keep. The UI mirrors the core rather than offering a choice it would refuse.- The round counter is visible.
MAX_CONFLICT_ROUNDShad been in core since M1 with nothing surfacing it, so a user could hit the cap unwarned.
And the third Stage 1 deferral, ExecutionSummary.cancelled copy, which had no specified
wording: the summary now says the run stopped part-way, that a package may have been left
half-written, and points at the snapshot.
One design change forced by the compiler, worth knowing. UpdateFlow::start took a &Store;
Store wraps a rusqlite::Connection and is Send but not Sync, so a future holding one
across an await is not Send and a Tauri command cannot return it. It takes the pins instead —
one synchronous read at each call site, and the flow no longer touches a database at all.
Stage 4 — search, the dock bay and install — done 2026-08-04
Five commits. Both exit criteria met and measured in the running app. (The "16 of 16 components" claim recorded here was wrong: the count had folded in components that are not on UI-SPEC §6's list. It was 13 of 16 then and after S5 — see §6, which now says so.)
| What landed | |
|---|---|
| Index | IndexSlot in AppState holds the 864k-name index; warmed on first Search open, never blocking a search. index_search, index_refresh, pkg_metadata. |
| UI | useIndexStore, PdSearch, PdDockBay, PdOfflineBanner, the metadata panel, / to focus. |
| Install | The dock bay resolves through Intent::Install, into the same preview and summary S3 built. |
The keystroke budget, measured rather than assumed — and it was being missed.
| Median | p95 | Over 50 ms | |
|---|---|---|---|
| First measurement | 57.4 ms | 79.8 ms | 18 of 19 |
| After two fixes | 22 ms | 24.5 ms | 0 of 19 |
The two fixes were a leading-edge debounce — it was trailing-only, spending 16 ms of a 50 ms
budget doing nothing on every keystroke, while the comment above it had claimed leading-edge
behaviour since the day it was written — and SEARCH_LIMIT 50 → 20, because React commits
every row on every keystroke and the row count is the largest lever on the render half.
A number I reported was wrong, and it mattered. The 613 ms index load was a debug build; in
release it is 140 ms, and the worst keystroke is 16 ms rather than 176 ms. The design — hold the
index, warm on demand — is unchanged and still right at 140 ms, but it had been justified with a
figure four times too large. crates/pipdock-core/tests/search_latency.rs now refuses to run in
debug rather than printing a number that means nothing.
Two bugs found by counting clicks, neither reachable by any test written for them:
- A plan started from Search resolved into nowhere.
PdPlanPanelwas owned byPdPackages, so the command ran, the flow parked, and the user saw unchanged search results. A plan belongs to the app — there is one at a time, and which tab is showing is not part of that. - Every result offered [Add], including installed packages, because only
PdPackagesever loaded the installed set. That is exactly the mistake DATA-FLOW §4's chips exist to prevent.
Stage 5 — uninstall, the guard dialog and the Pins screen — done 2026-08-09
Twelve commits. Every mutating command in ARCHITECTURE §7 now exists, and removal is the first flow whose whole value is refusing to do what was asked.
| What landed | |
|---|---|
| Core | GuardReport.breaks carries BrokenDependent { pkg, version, constraint } — the specifier, not just the name. GuardAck makes the guard an enforcement point rather than a dialog. PEP 668 refused on removal. New code PD-RES-004. |
| Execution | execute_uninstall emits the stepStarted/stepFinished lifecycle it never had, checks the cancellation token in its loop, takes a base step index shared with execute_rollback, and honours --no-snapshot. |
| Bridge | One Session enum for all four flows, with typed claims; uninstall_guard, uninstall_execute; plan_cancel async, and now discarding a parked session. |
| UI | PdDialog (the repo's first modal), PdUninstallDialog, PdPinChip, the Pins screen, the row's ✕, and the failed phase. |
| L3 | The first mocked @/ipc, the first store test, the first screen test. 6 files → 9; 59 tests → 79. |
Exit criteria, counted and verified in a running app (driven in the Browser pane against a
stubbed bridge, per the recipe that found Stage 2's double fetch): uninstall is exactly 3 clicks
— Installed → row ✕ → Remove — and 4 when the guard trips, by design. The dialog reads
pandas 2.1.4 requires numpy<2,>=1.26.0, naming the dependent and its constraint. Remove
dependents too issues a second uninstall_guard over the widened set with no
uninstall_execute in between. At the CLI, against a real venv, the guard refuses and exits 2 with
every package still installed.
Five live bugs came out of it, none of them S5 features:
- A claim that found nothing wedged the slot.
plan_decide/plan_executewroteBusybefore discovering there was no plan, and never released — so one out-of-order call from the UI made every later command answerPD-RES-003for a plan that had never existed. execute_uninstallemitted no lifecycle markers, and every line carriedstep: 0. The CLI never noticed because it printsevent.line()and nothing else; the console drawer groups onstepStartedand the live region countsstepFinished, so the GUI would have shown an empty drawer against a green suite.- A cancelled removal removed everything, then reported
cancelled: true. The loop had no token check at all. --no-snapshotwas parsed and ignored on the uninstall path — the same waiver meant "waive" for an update and nothing for a removal, the one operation with no way back.- The pin button was a tab stop, so Space on it toggled the row's selection instead of pinning. The rule was enforced on the checkbox and forgotten one element later.
And one specification gap closed: DATA-FLOW §9.1 required "a ResolutionReport accepted in this
session", which the uninstall path cannot produce — there is nothing to resolve. It now names the
proof per plan shape, so the one flow with no preview is no longer the one flow the invariant did not
describe.
Stage 6 — snapshots, the env detail and rollback — done 2026-08-09
Eight commits. Every snapshot the app has been writing since S1 is now reachable from it, and P0-8 is complete in both heads.
| What landed | |
|---|---|
| Core | RollbackPreview and RollbackPlan cross IPC; snapshot::Meta gains a schema rename so embedding it does not mint a second interface. parse_freeze and unrestorable_lines rebuilt on one classifier. |
| Bridge | snapshot_list|create|diff|rollback_preview|rollback. ARCHITECTURE §7's single rollback row becomes two, 32 → 33. |
| UI | PdSnapshotTimeline, PdRollbackPreview, the env detail as a mode of the Environments tab, the diff viewer, and the rollback the summary's copy has promised since S3. |
| L3 | Two more generated fixtures; 9 → 11 test files, 79 → 91 tests. Goldens 52 → 54. |
Exit criteria, verified by running. At the CLI against a real venv: create → mutate → snapshot rollback <id> → snapshot diff <id> prints environment matches snapshot — PRD §6's metric, and
the timeline afterwards shows the rollback of … entry above the one it restored, which is
latest moving twice, live. In the app: rollback = 4 clicks (Open → entry → Rollback… →
Roll back), 2 from a run's summary; the preview lists PD-SNP-002 with the verbatim freeze line
before the confirm.
Four bugs, three of which no test would have found:
- A cancelled rollback killed the package in flight and installed the rest. The restore loop had no token check — S5 gave one to the removal half and not to the install half. Found by writing the cancellation test the plan asked for.
parse_freezeandunrestorable_linesdid not partition their input. Two independent filters that disagreed: a line likeFoo Bar==1.0was dropped by both and vanished. In a rollback preview, whose whole job is saying what it cannot put back, that silence is the worst possible failure.snapshot_listfired twice on open — Stage 2's double fetch, reintroduced exactly, because the guard is written after the await. Invisible in the UI, since both calls return the same rows.- A rollback left the timeline empty. The refetch and the rescan race, and the rescan wiped the
result: the snapshot fields were inside
NO_PACKAGES, whose reset is keyed to the package slice's environment. The two slices key on different things and now reset separately.
Deliberately not done: DATA-FLOW §8's dry-run resolve of the restore set, which is implemented
in neither head. It is a network round trip on an interaction budgeted as instant, and it would
resolve against the pre-removal environment — reporting conflicts the two-phase execution never
hits. §8 now says so, and the preview's promise is precise instead: it lists what cannot be
restored, not everything that might fail. Snapshot retention (snapshot::delete + keep-N) also
stays out; a destructive command against the user's only way back needs its own confirm.
Stage 7 — the shell — done 2026-08-09
Eight commits. What the shell was missing was rarely a feature — it was a promise the docs had made and nothing had kept.
| What landed | |
|---|---|
| Fonts | Inter and JetBrains Mono vendored. --font-sans/--font-mono had named them since the tokens were written and both fell through to a system fallback, because no @font-face was ever registered. |
| i18n | 26 of 32 catalog codes had no copy in either locale. All 32 now do, in locales/*/errors.json, asserted against a generated codes.json — both directions, plus I18N §4's ≤ 90-char cap. |
| Bug report | pipdock_core::report shared by both heads; LogRing in AppState fed by three writers; report_bug_url; the error row's Copy full log and Report bug. |
| a11y | Every --color-* rebound under forced-colors, with borders on the destructive controls whose tint the UA erases. The pseudo-locale, dev-only and verified absent from a production build. |
| Status line | The log toggle does something, and ⚠ n means something. |
| Bridge | engine_info, and three tests that make COMMANDS mean something. It listed 32 names while 19 were registered — a wrapper for any of the other 13 typechecked and would have failed at runtime, on a command that looked implemented. |
| Keyboard | Focus follows the tab: Ctrl+3 changed the screen and left focus behind, so the next Tab resumed inside the screen the user had just left. |
Deliberately left for later, and named: the Enter primary-action map and folding the six
aria-live regions into one — both are cross-screen behaviour that wants the VI sweep and a real
screen reader to verify, not a passing test. S7b (Settings' remaining controls) was always a
follow-up PR. PdEmptyState now covers every surface that renders a list.
One thing worth keeping: i18next's nsSeparator is :, so I18N §1's seven namespaces would
have meant t('errors:PD-NET-001') at every call site — churn across every screen, bought with lazy
loading a desktop app that bundles both catalogs cannot use. The app ships one namespace and two
files per locale; §1 is amended to say so rather than left describing something that was never built.
M2 is complete. Every P0 feature except Code Health (P0-11) is in both heads, all seven UI-SPEC
click budgets are met by hand count, and NOT_YET names the five commands that remain — each
against the slice that owes it.
Where to pick up
Phase 3. Its first slice, P1 (pip upkeep), is the smallest: pip_upgrade is the only command it
needs, Engine::upgrade_pip and the CLI already exist, and engine_info — which the Environments
row needs to show a version — landed here. The plan's decomposition (P1 pip upkeep · P2 tools venv ·
P3 runners + HealthReport · P4 the Health screen · P5 the gated ruff --fix) is in
~/.claude/plans/b-n-l-n-plan-cho-eventual-wadler.md, along with the two decisions worth not
re-deriving: the tools venv is populated with pip unconditionally rather than the configured
engine, and the health report type is HealthReport, because CheckReport is already taken by
engine.check() and is a published pipdock schema contract.
Ten minutes before any of it: install tools-requirements.txt's pin set into a scratch venv on
3.14 and on 3.13. py -0p reports 3.14 first on this machine and deptry ships compiled wheels; no
cp314 wheel means an sdist build and PD-BLD-001 with no user action. That answer sets
TOOLS_PYTHON_MAX and is the top unknown in all of M3.
All three Stage 1 deferrals are closed, each in S3 as planned — the plan-progress lifecycle enum, the Windows Job Object, and the ExecutionSummary.cancelled copy. Deferring them to the slice that could verify them worked: each was finished against a running UI or a real process tree rather than against a guess.
Phase 3 — M3 "Health + polish" (~2 weeks)
Tools venv, deptry/vulture/ruff runners + report UI + gated fix, pip upkeep, bug-report deep link, offline states, keyboard map, icon/branding pass. Exit: CODE-HEALTH flows pass on two real projects (one pyproject, one requirements-only).
Phase 4 — RC → v1.0 (~1–2 weeks)
Release pipeline live (bundling, checksums), manual charter executed, docs/README screenshots, legal files public, a clean install of the RC verified on a machine that has never run PipDock. Exit: RELEASE-CI §5 checklist fully checked; v1.0.0 published + notify fan-out.
Post-1.0 (P1 wave)
Security tab (pip-audit), pin auto-suggest, requirements/constraints export-import, cache manager, command palette, dependency graph view, scheduled check. Then P2 candidates by demand: macOS/Linux, per-env engine, elevation broker, winget, JP locale.
Standing risks
| Risk | Mitigation |
|---|---|
| uv output churn between releases | weekly latest-engine parser CI job (TESTING L1) + PD-ENG-003 graceful degrade + "switch to pip" escape hatch |
| pip UX flags deprecations (e.g. report format changes) | same job covers pip; adapter version-gates per engine version |
| Index size growth | SP-3 fallback design kept in the drawer |
| Scope creep from Health into a linter IDE | CODE-HEALTH §7 non-goals are contractual |