Files
pipdock/docs/UI-SPEC.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

49 KiB
Raw Blame History

PipDock — UI Specification

Version 0.1 · 2026-07-17 · Direction: dark terminal-tech (owner decision 2026-07)

1. Design intent

PipDock should feel like a precision instrument docked over a terminal — the calm confidence of a good TUI with the affordances of a native app. Not a neon cyberpunk cliché, not a generic admin dashboard: restrained phosphor-green on near-black, monospace where data lives, one accent doing real work. Signature moves: a terminal-style status line pinned to the bottom, a live execution console that streams real engine output, and a Ctrl+K command palette (shipped 1.1.0).

Anti-goals: gradients-everywhere SaaS look, rounded-blob cards, decorative glow. Any glow is functional (focus rings, live-activity pulse) and subtle.

2. Design tokens (Tailwind 4 @theme)

Token Value Use
--color-bg #0A0E0C app background
--color-surface #101612 panels, rows
--color-surface-2 #161E18 hover, elevated
--color-border #1F2B23 hairlines
--color-accent #4ADE80 primary actions, focus, live pulse
--color-accent-dim #22935B secondary accents, links
--color-text #E7F0EA primary text
--color-text-dim #8AA394 secondary text, dimmed up-to-date rows
--color-warn #F5B94A held-back, warnings
--color-danger #F26A63 impossible, failures, force actions. #F0574F until 1.4.0, which read at 3.92:1 on its own badge tint over a selected row
--color-info #5BC0DE pinned, informational chips

Type: Inter (UI), JetBrains Mono (package names, versions, console, status line — both OFL-licensed, bundled). Base 14 px, mono data 13 px. Radius 6 px; borders 1 px; shadows minimal. Motion: 120–160 ms ease-out; live-activity pulse on the status-line engine badge during execution (not built: the status line says nothing of a running plan, only scanning or idle — audit backlog U52); prefers-reduced-motion respected.

A badge is checked as what it is drawn as, not as its token (1.4.0). A badge is its text on a tint of its own colour — badgeTones.ts: 20 % for warn, info and accent, 15 % for danger — and the tint lowers the contrast the plain token has. Danger read 3.92:1 on its own 20 % tint over a selected row (surface-2), under AA, while the token test passed, because it checked tokens on plain surfaces only. tokens.test.ts now composites every tone's tint over every surface and asserts AA on the result. Both changes were needed: the new #F26A63 at the old 20 % tint is still 4.29:1 there, and at 15 % it is 4.67:1.

--container-content (87.5rem) is the content column's widest. It is what stops a table stretching across a maximized 4K window, now that the window itself is the user's to size (§5).

3. Layout

┌──────────────────────────────────────────────────────────────┐
│ ⌂ PipDock      [env: .venv (Py 3.12) ▾]        [engine: uv]  │  top bar
├───────────┬──────────────────────────────────────────────────┤
│ Environments        │                                        │
│ Installed           │                                        │
│ Updates (7)         │              content area              │
│ Search              │                                        │
│ Pins                │                                        │
│ Health              │                                        │
│ Security (P1)       │                                        │
│ Settings            │                                        │
│ About               │                                        │
├───────────┴──────────────────────────────────────────────────┤
│ ▸ .venv · Py 3.12.4 · uv 0.9.x · idle            log ⌄  ⚠ 0  │  status line
└──────────────────────────────────────────────────────────────┘

Sidebar is icon+label, collapsible to icons. Status line (mono) always shows env · python · engine · state, plus a log-drawer toggle.

⚠ n is the number of error rows currently on screen — defined in S7, having been drawn in this diagram and specified nowhere since. It is deliberately not a session tally: that would read ⚠ 47 after a batch where 47 packages failed, stay there, and tell the user nothing the summary in front of them does not. Counting live rows means it returns to zero when the problems do, so the number and the screen always agree. PdSummarySheet's per-package rows opt out — one failed run is one problem, not forty-seven. This definition is an invention; nothing specified it.

Code Health's failed tools count (P4). A HealthReport carries at most three problems, one per tool, and each renders as a PdErrorRow above the tabs — so they are error rows currently on screen and the rule applies unchanged. This is not the PdSummarySheet case: three separable tools are three problems, where forty-seven failed packages are one run. They render above the tabs rather than inside them so the explanation is never reachable only from the tab it explains.

Health owns its own console drawer, and that is a deviation worth naming. The drawer above is specified against plan-progress, and the status line's toggle is disabled unless a plan is running. Health claims a separate session slot by design — a health run and an install can be in flight at once — so one toggle cannot address two drawers. PdHealth renders its own instance with its own toggle rather than teaching the status line to choose. The Security tab does the same, since 1.2.0 and for the same reason: an audit holds a slot of its own as well.

The log toggle opens the console drawer, and is disabled when no run is in progress: a toggle that opened an empty panel would be the same empty gesture the plain-text version was. The drawer was once the only log surface there was. Since 1.4.0 PipDock also writes its own log — one file per UTC day under data\logs\, kept 14 days — and Settings' Logs row (§4) says where it is and copies its recent lines. The toggle still opens the drawer, which streams the run in progress; the file is the record of what PipDock ran, which outlives the window. The console drawer slides up over the status line during execution, streaming plan-progress lines with per-package section markers; collapsible, never modal.

4. Screens

Environments. Auto-discovered list (source chip: registry / py launcher / uv / venv / manual) with Python version, package count and pip version; PEP 668 envs carry a red MANAGED chip and a line pointing at a virtual environment instead. Actions: Use, Open (the detail view below), Rescan, Add interpreter…, Remove from list on a row it added, and Upgrade pip when pip is below the planner's floor (§5).

This line used to promise three things the screen does not have. A managed environment is not disabled: it can be selected and read, and a change to it is refused with PD-ENV-002 by the flow — decided on a fresh probe since 1.4.0, not on whatever the row last cached — unless the Settings override is on. Browse… shipped in 1.6.0 as Add interpreter…, beside Rescan and as the empty state's one action (§7): a native picker filtered to .exe, then env_add_manual, which refuses anything not named like a Python interpreter or inside PipDock's own venvs before it runs it, and selects the row it adds. An added row carries the manual chip and Remove from list — kept even when its file is gone, so the PD-ENV-001 row can be removed — and the outcome is said in a role="status" line mounted before its words. Both are busy while a scan runs, and a scan waits for either: a scan that read the list before the change would undo it. An added row is matched to the list by its path, never by envHash, which follows links — a venv made with --symlinks is listed beside the base Python it links to, not merged into it. Discovery is unchanged: the registry, the py launcher, uv python list, and a .venv, venv, env or .env in the working directory. And there is no star-as-default: the selection lasts for the session, and a rescan keeps it or falls back to the first usable row. The CLI's env use stores a default, honoured since 1.4.0; the GUI does not read it.

Installed. Virtualized table: name (mono) · version · latest · size · chips. When the env's hidden_user_site is non-null (non-venv system Pythons only — see SECURITY §2), an inline --color-info note above the table says user-account packages are not listed and names the path; it is informational, never a block, and never appears for a venv. Up-to-date rows use --color-text-dim (owner's "dimmed" requirement); outdated rows show an UPDATE badge; pinned rows a 🔒 --color-info chip. Row actions: pin/unpin, uninstall, and details, as three inline buttons rather than a ⋮ menu. A menu would spend one click opening and one choosing, leaving one of §5's three for a confirm that has to name what breaks.

This paragraph used to say the opposite, and the correction is worth keeping rather than overwriting. It said "details" had no panel to open yet, "and when it does all three fold into a menu with the budget intact". P1-6 built the panel and the fold was wrong: it makes uninstall Installed → ⋮ → ✕ → Remove, four clicks where §5's table records three, hand-counted, for the second most destructive action in the application. Four is under the ceiling of 5, which is exactly why the budget alone would have waved it through — a ceiling is not a ratchet, and a table of measured counts is. The third button costs nothing instead: the row is the only tab stop (§8), so its controls are a roving sequence rather than tab stops, and a third is one more arrow step, not 200 more stops. Ordered non-destructive first — details, pin, remove.

No button is a tab stop — the row is (see §8). Multi-select via checkboxes; bulk action bar appears on selection.

The header and every row share one set of column widths (packageColumns.ts, 1.4.0). The header was a hand-written mirror of the row and had drifted: one 2rem slot stood above three action buttons, so every fixed column's heading sat about 76 px right of its cells. And the fixed widths summed to more than the table has at the 900 px minimum window, with the name column absorbing the whole deficit — so at that width it showed no names at all. A narrow table drops Size first and tightens the rest, by container query against the table's own width rather than the window's (Size returns at 48rem): the plan panel or a split view changes the one without the other. Header and scroller reserve the same scrollbar gutter, so a scrollbar appearing does not push the cells out from under their headings.

Updates. Same table filtered to outdated, badge count in sidebar. Header: Select all (excludes pinned; shows "3 pinned excluded"), Update selected. After resolve → Preview panel replaces the table: grouped sections — Will upgrade (A 1.2 → 2.0), Will downgrade, Will install, New dependencies, then Will break, Left out — needs a different Python and Withdrawn releases (1.4.0, below), and last Needs decision (held-back in --color-warn with one-line blocker sentence; impossible in --color-danger). Each needs-decision row hosts a segmented control: Keep compatible · Skip · Force latest (Force opens an inline confirm naming what breaks). Footer: Back, Apply choices while decisions are pending, and Confirm (n changes) — Confirm (n changes, breaks m), with the danger treatment, when the plan forces something. The panel's heading names the interpreter it changes, with the PEP 668 chip beside it where that applies: the panel replaces the screen it was started from, so a plan started from Search's dock bay gave no sign of which environment it was for.

Will downgrade was added in S3. ChangeKind has always had four variants and this section named three, leaving Downgrade with nowhere to go — and it is not a rare case: a compatible resolve routinely moves a package down to satisfy something else's constraint. Folding it in with the upgrades would have been the worst outcome, since a row reading 2.0 → 1.9 under a heading that says "Will upgrade" is actively misleading about the one change most likely to surprise. It carries --color-warn for the same reason.

On impossible rows, Keep compatible is disabled. UI-SPEC once said every needs-decision row hosts the full 3-way control, but default_decision(is_impossible = true, …) returns Skip, because an impossible package has no compatible version to keep — there is nothing for the control to select. The UI mirrors the core rather than offering a choice that cannot be honoured.

The round counter is visible. MAX_CONFLICT_ROUNDS is 3; when a decision round is applied the preview says how many remain, and the exhausted state keeps the preview on screen with the controls disabled rather than dead-ending. Without it a user hits the cap unwarned.

Three sections arrived in 1.4.0, each rendering a field ResolutionReport had and nothing showed.

  • Will break (n), in --color-danger: each package forced past what held it back, as name → latest, and under it every dependent named with the constraint the forced version violates. n counts broken dependents, not forced packages (the heading reads Forced (n) when no blocker names one), and the same count is on Confirm — DISCLAIMER §2 requires what forcing breaks to be said before it runs, and a section the user has not scrolled to has not said it. There was nothing to show before 1.4.0: Force latest was recorded in the request and never read, so a "forced" resolve was the held-back one again.
  • Left out — needs a different Python (n), in --color-warn: each release left out because its Requires-Python excludes this interpreter, with the range set against the interpreter's version (hard invariant 10). It is one of three things: an exact pin the user asked for, the latest of a package they chose to force, or a release the engine planned — uv ignores an upper bound — which is ruled out by name and the engine asked again. A request that is nothing but such pins is refused with PD-PKG-001; one with anything else in it plans the rest, and a plan whose only content is a left-out row is not "nothing to change", because the user asked for something and is owed the reason it is not happening.
  • Withdrawn releases (n), in --color-warn: PD-PKG-003 for each yanked release in the plan, with the maintainers' reason when the engine printed one. A warning, not a refusal — both engines install a yanked release (SP-2). Both parsers had always extracted these; the adapters dropped them.

Back from a preview is not a finished run (1.4.0). The selection is kept and nothing is refetched, where it used to be cleared as if something had changed. Leaving work the backend still holds — Back during a resolve, or out of a parked preview — stops it there (plan_cancel), and a late answer no longer re-opens the panel the user closed. A decision round that fails leaves the preview usable, because the flow stays parked; its Confirm used to answer "no plan".

After a run the panel shows the summary (PdSummarySheet): the counts, a row per package, the snapshot with Roll back to this, and what the post-run check found. Since 1.4.0 each row says what it moved from — the engines cannot know it, so every upgrade used to read as a fresh install — and a check that could not run says so ("could not check the environment afterwards", with its code) instead of rendering nothing, which read as a pass in exactly the case nothing was verified: a run that broke pip itself.

Search. Search field autofocused; results stream under 50 ms per keystroke from the local index. Result row: name · summary (from cached metadata) · latest · INSTALLED ✓/UPDATE chip when applicable · [Add]. Right detail panel: description, requires-python, license, links. Queue ("dock bay") docks along the right edge as a slim column of added packages with editable version fields and Install (n).

Install (n) keeps the queue until the resolve has answered (1.4.0): a plan already in flight, or an index that could not be reached, used to cost the user the list they had just built. Once there is a preview, the preview is the list. A search typed while the index is still warming is asked again when it is ready, rather than waiting for another keystroke. The field's focus on arrival is the shell's job — see §8's data-autofocus.

Pins. Pin list with reason field, committed on blur; unpin per row. An Exclude pin keeps its package out of every update, a named one included — DATA-FLOW §9.5, enforced by pins::filter_upgrades — not only out of Select all. Exclude pins only: a Hold pin is rendered faithfully if one exists, but creating one is not offered, here or by the CLI, whose pin add writes Exclude and nothing else. The reason this paragraph gave until 1.4.0 has gone: pins::hold_requirements had no caller, so a hold was a promise nothing kept, and since 1.4.0 every resolve restates a held package at its version. A hold is now honoured if one exists; nothing makes one. With no environment selected the screen says so rather than "no pins": pins are keyed by env_hash, so that is not an empty list.

The auto-suggest section — "urllib3 — 12 packages depend on it. Pin?" — landed post-1.0 as P1-A. It sits above the pin list, because it is a prompt to act and the list is a record of having acted, and it is absent entirely when there is nothing to suggest. Accepting writes an ordinary Exclude pin with the count as its reason (PRD P1-2's "suggest pin with reason"), which the reason field below can then replace.

It shows the top 8 and says how many more qualify. Measured against the 352-package fixture, the default threshold of 5 qualifies 94 — a quarter of the environment. Raising the default would be the wrong fix, because a deep dependency tree is exactly where a bulk update is most dangerous; the count in the "N more" line comes from the full list, so a capped view never misreports a total. Same rule as RUFF_ROWS_SHOWN in the Health report.

It costs one probe.py run, and that is why it is here rather than on a sidebar badge: only someone who opened this tab pays for it. It is also silent on failure — this screen's job is listing pins and it does that either way, so an error row about an advisory would be the loudest thing on it. Counts come from graph::dependent_count, which counts in-force edges only: the same set the uninstall guard warns about, because a user pinning on one rule and being warned on another is the failure worth designing against.

Dependencies (surfaced under Installed → row → details): the focused package between two columns — what requires it, and what it requires — each row carrying the specifier written on that edge, and each clickable to re-centre. Below them, the two transitive counts: how many packages a removal would leave with a missing dependency, and how many the package pulls in itself. A requirement that nothing installed satisfies is called out; nothing else in PipDock reports one. Platform markers are read since 1.4.0 — probe schema 3 reports the marker environment — so a requirement gated to another platform, such as ipython's pexpect; sys_platform != "win32", is no longer reported missing from every Windows environment.

Drawn at depth 1, and depth is navigated rather than drawn. Measured on the 352-package fixture before the screen was designed: a depth-1 neighbourhood is a median of 7 nodes and a p90 of 25, while depth 2 is a median of 172 and exceeds 60 nodes for 212 of 352 packages. No layout algorithm makes 172 nodes legible, so the node-link diagram PRD P1-6's wording suggests would have cost an eleventh runtime dependency, a THIRD-PARTY-NOTICES.md edit that re-shows the legal gate, and a CSP argument — tauri.conf.json sets no script-src, so anything reaching for eval is refused — to buy a picture nobody could read. Paths are not offered at all: setuptools has more than 2000 distinct paths to a root in that environment, so "why is this here" has no path-shaped answer at this scale, and the honest answer is a count.

Like Snapshots, it is a mode rather than a tab, for the reason below: NAV_KEYS is already nine, and a Dependencies entry would have to sit beside Installed to read as related — an insert.

Columns are capped at DEPS_ROWS_SHOWN with the overflow counted from the full set, the SUGGESTIONS_SHOWN rule for its reason: setuptools has 150 dependents, and a column of 150 is a second package list. Nothing is carried by colour — §8's forced-colors block collapses warn, danger and info to CanvasText, so every specifier is present as text.

Snapshots (surfaced under Environments → env detail): timeline of snapshots with trigger label, diff viewer (added/removed/changed in mono), Rollback… with its own preview per DATA-FLOW §8.

The detail view is a mode of the Environments tab, not a sidebar entry of its own. The reason is adjacency, not arity: §8's shortcuts are positional over NAV_KEYS, so appending is free and inserting is not — Phase 4 put About on the end and it became Ctrl+9 without moving a single existing binding, while a Snapshots entry would have to sit beside Environments to read as related, renumbering everything after it. Which environment is open lives in useEnvStore, because the plan panel replaces the whole content area while a rollback runs and component state would be unmounted with it.

The trigger label is load-bearing, not decoration. A restore snapshots the current state before restoring, so one rollback moves latest twice and two entries a minute apart are otherwise indistinguishable — while one is the state the user wants back and the other is the state they are leaving. Nothing in the UI ever names latest; the timeline hands ids to every command.

An environment whose interpreter is gone still reaches the detail view: snapshots are keyed by env_hash and outlive the Python that made them, so the timeline lists while the diff and the rollback are disabled with a line saying why. The diff's copy names the direction — "here but not in the snapshot — restoring removes it" — because "added" is ambiguous the moment you read it from the snapshot's side.

Health. Project-folder picker (persisted per env — on disk, through EnvRow.healthProject; on screen the folder is taken from the row only while none is set, so after a switch it stays the previous environment's for the rest of the session, audit backlog U16) → run panel with three result tabs (deptry / vulture / ruff) per CODE-HEALTH-SPEC; Fix with ruff gated behind explicit confirm listing file count.

Security. Run audit → advisories grouped by package, each with its id, the versions that fix it (or no fix, stated rather than left blank), its aliases — which is where the CVE lives, since pip-audit's primary id is a PYSEC-* — the advisory prose verbatim, and a link to its OSV entry. No severity column: pip-audit publishes none under either vulnerability service, so the order is package → fixable → id, and inventing a severity would put a number PipDock made up beside real ones (PRD P1-1 is amended to say so). A Cancel appears while a run is going, which Code Health has no equivalent of: an audit is 18-68 s against Health's 1.3 s. The body says nothing while a run is in flight — not "no audit has run", which is what it said until someone clicked Run against a slow bridge and read it.

It scrolls (1.4.0). The screen's root was a bare <div> inside a <main> that clips, so a report taller than the window could not be reached at all; it is now one scroll container with its console drawer pinned beneath, as Health's is. Cancel is honoured from the start as well: the cancellation token used to be registered after the environment's freeze, so a Cancel during it did nothing, and a run stopped that early now skips the audit venv's sync and reports itself cancelled.

Settings. In screen order: engine (pip/uv radio, each with its version for the selected environment), language (en/vi), pin auto-suggest count (P1-A; zero turns suggestions off), PEP 668 override (off by default, scary copy), Disk usage, Logs. Every control saves on change — the override, once its confirm is answered — which is what keeps Switch engine at three clicks (§5).

This line used to list two settings that do not exist. There is no snapshot retention — snapshots are kept until cleared under Disk usage — and no index-refresh setting: the index is downloaded from Search when it is missing, or by pipdock index refresh, and its seven-day staleness is reported, never acted on. Open logs folder became the Logs row below.

Each engine's version is beside its radio (1.4.0), for the selected environment — pip's comes from <python> -m pip --version, so there is no answer without one — and an engine that is not installed reads not found. engine_info had a typed wrapper and no caller, so nothing said uv was missing until a plan failed for want of it. The answer is keyed by interpreter, so switching environments never shows one's versions under another. Nothing preselects uv yet: a fresh install drives pip, and the first-run detection ARCHITECTURE §3 specified is owed (PRD P0-9). settings.rs puts it here, as a choice made where it can be seen, rather than in a read of an empty store.

The language is stored (1.4.0) as the user's choice, null meaning "follow the OS", and applied at start. save() sent locale: null every time, so choosing Vietnamese lasted until the window closed. Settings now load when the app starts, not only when this screen mounts, and before the legal gate — so the gate is in the chosen language too, and the status line's engine and the PEP 668 chip are right before anyone has opened Settings.

Turning the PEP 668 override on asks first (1.4.0, SECURITY §3): a PdDialog with Cancel focused, saying that --break-system-packages will be passed for every environment marked EXTERNALLY-MANAGED and that rolling back to a snapshot may not repair what that breaks. Turning it off saves at once — restoring protection never needs a second thought. Until 1.4.0 the override was one click on a checkbox that saved and nothing read; every mutating flow now probes first, decides on what the probe says, and carries the override to the engine call only when it is on (hard invariant 5).

While the override applies, a chip says so — PdOverrideChip, SECURITY §3's persistent warning — on Installed and Updates, Search's dock bay, the plan panel and an environment's detail. Applies means the setting is on and the environment carries the EXTERNALLY-MANAGED marker. A venv never does, and a chip there would claim the flag is being sent when it is not.

Disk usage (P1-C) shows what PipDock has written — snapshots, the Code Health tools venv and, since 1.2.0, the Security tab's audit venv, each with a Clear, and index.db reported without one because it holds settings, pins and the consent record in the same file. Since 1.4.0 a Clear is refused with PD-RES-003 while the feature that owns its target is running — Health for the tools venv, an audit for the audit venv, a plan for the snapshots — rather than deleting the venv a run is executing from; a preview that is merely parked is put back afterwards. Logs are not a row. This paragraph once said there were no log files and that a "logs — 0 B" row would invent one. Since 1.4.0 there are, and they look after themselves, so a Clear would only take away the record a bug report needs.

Logs (1.4.0) shows the log folder as a path — data, never translated (I18N §2) — and Copy recent log, which re-reads the tail at the moment of copying, so the clipboard holds the run that just went wrong rather than what the log said when the screen opened. The folder is shown rather than opened: the opener is scoped to four https hosts and nothing local, and the copy uses the clipboard permission Copy full log already had, so no capability was widened for the row.

The threshold is the app's first numeric input, and it sets the rule for the next one: reject at the boundary, never store junk. <input type="number"> reports '' for anything it cannot parse, so coercing the raw value writes 0 — which here is a meaningful setting, "off", that the user did not ask for. Its label avoids the word threshold on purpose; see PdPins.test.tsx's Hold assertion, which matches /hold/i.

Legal & About left Settings in Phase 4 and became the About tab below. Neither is a setting: they are read-only surfaces, and folding them into a screen of controls is what kept them unbuilt for four milestones while every control around them shipped.

Legal gate (first run). Modal listing the five documents with GitHub links and a summarized disclaimer sentence; single checkbox "I have read and accept" → Continue. Decline exits the app. Consent stored with docs-version hash; hash bump re-triggers the gate.

Decline has exited the app only since 1.4.0. The webview was granted core:default — some sixty core commands nothing called — and not core:window:allow-close, so in the installed app Decline did nothing, while its component test, which mocks close(), passed. capabilities/default.json now holds exactly what the frontend calls, and a test in lib.rs pins the union of every capability file. A consent that cannot be read or recorded is shown, as an error row on the gate; the failure used to be stored and never rendered, so Continue looked like a button that did nothing. And the hash ignores line endings: a build from a CRLF checkout used to carry a different hash from an LF build of the same commit, so two builds of one commit could disagree about whether the documents had changed.

It has a second, presentational mode since Phase 4: About re-opens it so the documents can be read again, and in that mode the checkbox and both buttons are replaced by a single Close. Consent is neither read nor written. This is a flag of its own (useLegalStore.review) rather than setting accepted = false, which would have been one line and would have made a review indistinguishable from a revocation — offering Accept and a Decline that closes the application to someone who only wanted to re-read their privacy policy.

About. The ninth tab (Ctrl+9): what PipDock is; this build — version from app_info, the SPDX identifier GPL-3.0-only, and the legal-documents hash consent is recorded against; a button that puts the gate back on screen; and contact — two addresses and the author's links page. The addresses are copied to the clipboard, not opened: mailto: would widen the opener capability by a whole URL scheme, and SECURITY §4 asks for a reason per widening. They are also rendered from a TypeScript constant rather than the catalogs, because an address is data (I18N §2) and a catalog invites a translator to change it. Only https://poli0981.dev/* was added to the allowlist, for the links page.

5. Click budgets (owner ceiling: 5)

Every click counts, including the one that opens the tab — except on Environments, which is where the app lands. Counted by hand in the running app, never inferred from the markup.

Flow Clicks Path
Update everything 4 Updates → Select all → Update → Confirm (app auto-scans on env open)
Update everything, 1 conflict kept compatible 4 default choice needs no extra click
Install one package 4 + typing Search (autofocus) → result [Add] → Install → Confirm
Uninstall one package 3 Installed → row ✕ → Remove (4 when the guard trips: → Remove the dependents too → Remove)
Rollback last snapshot 4 Environments → Open → timeline entry → Rollback… → Roll back (2 from a run's summary, via Roll back to this)
Switch engine 3 Settings → engine radio → (auto-saved) back
Run Code Health 3 Health → Run (folder persisted) → view (4 the first time in a project: Health → Choose folder… → the OS dialog, which is not PipDock's click → Run → view. The folder is remembered per environment from then on.)
Fix ruff findings 3 ruff tab → Fix N… → Fix (from a report already on screen; 5 from the landing screen with a folder remembered, which is the owner's ceiling). Cancel holds default focus, so Enter without Tab cancels.
Run a security audit 2 Security → Run audit. Counted in the running app against a stubbed bridge. The first run also builds the audit environment, which is a wait rather than a click.
See what depends on a package 2 Installed → row ⇄ (Installed is not the landing screen, so its tab click counts). Re-centring on a neighbour is one more click and no fetch — the graph is one value per environment.
Upgrade pip 2 Upgrade pip → Upgrade (Environments is the landing screen, so no tab click). The button appears only when pip is below the 22.2 planner floor — the case the ordinary Updates path cannot fix, because the planner it needs will not run.

Shell rules established by the first installed build (2026-08-13)

One window, always. tauri-plugin-single-instance is registered before every other plugin; a second launch unminimizes and focuses the window that exists. Five launches used to give five windows, which is five AppStates over one machine — five Sessions slots each certain it owned the mutation in flight, and five PD-RES-003 guards that cannot see each other. Verified: five launches, one process, one window.

The sidebar is disabled while PANEL_PHASES owns the content area. Navigating away left the plan parked in Rust with nothing driving it, and the next Update answered PD-RES-003 about a plan the user could no longer see. §8's Ctrl+1..9 is the same rule for the keyboard.

And it did not, until Phase 4. This paragraph said the shortcuts "already refuse" for two milestones while they did not: App.tsx returned early on the guard dialog alone, and PANEL_PHASES excludes guard on purpose, so the two sets are disjoint and guarding on one guarded neither. Nothing looked wrong — the sidebar was disabled and the panel kept rendering — until the plan finished and dropped the user on a tab they never chose. The shortcuts now also refuse while the legal gate is on screen, which matters once About can re-open it.

The window maximizes and snaps; the content column does not stretch. Until 1.4.0 the window was bounded at 1600×1100 and could not be maximized, because the layout — a fixed sidebar plus one scrolling column — stretched across a 4K panel is a column of text beside a metre of empty surface. That answered the right worry with the wrong control: the column now stops at max-w-content (87.5rem) and centres in what is left, and the window is the user's to size. Fullscreen stays off.

The window opens at a size that fits. 1180×760 is taller than the work area of a 1920×1080 panel at 150 % and of a 1366×768 laptop, so the status line and the plan panel's Confirm opened under the taskbar. The window is created hidden and src-tauri/src/window.rs shrinks it to the monitor's work area — never below minWidth/minHeight, 900×560 — centres it there, then shows it.

6. Component inventory (prefix Pd)

PdSidebar, PdStatusLine, PdEnvSwitcher, PdPackageTable (virtualized), PdPackageRow, PdBadge, PdPreviewDiff, PdConflictRow (segmented 3-way), PdDockBay (queue), PdConsoleDrawer, PdSummarySheet, PdSnapshotTimeline, PdHealthReport, PdLegalGate, PdEmptyState, PdOfflineBanner.

15 of those 16 exist as of Phase 3 · P4, PdHealthReport having arrived with the Health screen. Absent: PdEnvSwitcher alone, which nothing needs while the header shows the selected interpreter and the Environments tab switches it. (ROADMAP and CLAUDE.md both claimed 16 of 16 after S4; the count had quietly folded in components that are not on this list.) Six components not on this list have turned out to be load-bearing and are listed here so the inventory stays honest — the number said four while naming five for two milestones, which is the counting trap this section exists to avoid: PdErrorRow (ERROR-CATALOG §3's row, used by every error surface), PdDialog (the shared modal §7's destructive confirms need), PdUninstallDialog (§5's three options), PdPinChip (the 🔒 chip, shared by the table and the Pins screen), PdRollbackPreview (DATA-FLOW §8's preview, including its PD-SNP-002 list) and PdAuditReport (P1-1's advisory list). Three more shipped in 1.1.0 and are named nowhere in this section: PdCacheUsage, PdPalette and PdRequirements. PdDepsFocus (1.3.0, PRD P1-6's focus view) makes ten off-list components; it is written here in the same breath as the count so the two cannot drift apart again. 1.4.0 adds PdOverrideChip (SECURITY §3's chip, §4) and PdLogs (Settings' Logs row), which makes twelve — and 15 + 12 is the 27 Pd* components in ui/src/components/, which is the check to run rather than trusting either number. Screens (ui/src/screens/, PdPlanPanel among them) are not counted.

7. States & feedback

  • Loading: skeleton rows in surface-2, never spinners longer than 300 ms without text.
  • Empty states: one mono glyph + one sentence + one action (e.g. Updates empty: ▸ all packages up to date).
  • Offline: banner chip in top bar; search works (local index), metadata panel shows cached-at timestamp.
  • Errors: inline with catalog code, human message (localized), and Details expanding stderr tail. Toasts only for background completions. Report bug opens a prefilled GitHub issue that nothing sends until the user submits it; since 1.4.0 the link is rebuilt at the click with the selected environment, so the issue carries its Python version and the configured engine's version — one subprocess per report, rather than one per error row mounted, which a summary of many failures would multiply.
  • Destructive confirms (Force latest, uninstall with breakage, PEP 668 override, and — since P5 — Fix with ruff, the only one that writes outside an environment) use --color-danger primary buttons and require the dialog's default focus to be Cancel. The PEP 668 override was on this list while it was a bare checkbox with no confirm at all; it has been a PdDialog since 1.4.0. A destructive control carries data-destructive, and under forced-colors styles.css gives every button with it a double border, because high contrast drops every tint and a destructive primary that looks like Cancel is worse than no colour at all. Until 1.5.0 the stylesheet kept its own list of five data-action values instead, and the list was never complete: the uninstall dialog's plain Remove, the Force latest inline confirm, the override's Allow, Disk usage's Clear confirm and the plan panel's danger-treated Confirm (a rollback, or a forced plan) were not on it, so in high contrast each was told from the control beside it by its label alone (audit backlog U45). A list of names beside a list of buttons is two lists; ui/src/destructive.test.ts reads the source instead and fails on a button drawn in the danger colour without the attribute. The same pass fixed the active sidebar tab, which carries aria-current="page" under an outline rule that matched 'true' alone. Stop carries the attribute too: it interrupts a run part-way, and is drawn in the danger colour for that reason. And the border was never drawn at all until 1.5.0, for the old five as much as the missing five: the rule sat in @layer base, every button carries Tailwind's border utility, and a later layer wins whatever the selectors say, so border: 2px lost to border-width: 1px. It is outside every layer now — an unlayered rule beats every layered one, and destructive.test.ts counts the braces — and it is a 3 px double border, because 2 px beside 1 px is a difference nobody sees at a glance. Found by emulating forced-colors in the installed build and looking at the button; this paragraph described a border for two milestones on the strength of a rule nobody had seen take effect.

8. Keyboard & accessibility

Full keyboard traversal: Ctrl+1..9 sidebar tabs (by physical key), Ctrl+K opens the command palette, / focuses the Search field, Space toggles row selection, Ctrl+A select-all-visible, Enter primary action, Esc closes drawers, and ←/→ move within a focused row (P6). Since 1.5.0: ←/→/Home/End move between Code Health's tabs, and the arrow keys move between a conflict's three choices. ACCESSIBILITY.md is the user-facing statement of this section, with what has and has not been tested; About links to it.

Four rules from 1.4.0:

  • Ctrl+1..9 match the physical key — e.code, Digit1..9 or Numpad1..9 — not the character the layout types. On AZERTY, Ctrl+1 reports &, so the tab shortcuts did nothing on every French keyboard. Ctrl+K still reads e.key, and that is an open question rather than an oversight: for a letter the physical key is the wrong answer on Dvorak (audit backlog U49).
  • / focuses the Search field, from Search itself too. There nav does not change, so the focus effect did not run and / did nothing on the one screen it names. It is ignored while typing in a text field, so a version specifier can still contain a slash.
  • A screen can name its first focus stop with data-autofocus, and that wins over the heading the shell otherwise focuses on every tab change. §5's four-click install depends on Search's field having focus on arrival, and it lost that focus every time to the shell's effect, which runs after its children's.
  • A release build swallows the browser's own shortcuts (browserKeys.ts): F5 and Ctrl+R (reload), Ctrl+P (print), Ctrl+F, F3 and Ctrl+G (find), Ctrl+U (source), Ctrl+S (save), Alt+←/→, and the hardware browser keys for back, forward and refresh. WebView2 is a browser and kept them: a reload dropped a preview mid-decision along with everything the stores had not persisted, and find opened over a virtualized table whose off-screen rows are not in the DOM. The listener is in the capture phase and above the legal gate, so no screen or modal lets one through. Development keeps them, because a reload is how a developer recovers from a bad hot update. Right-click is not covered: WebView2's context menu still offers Reload, Back and Print in a release build, and removing it would take Copy and Paste out of every text field with it — a decision, not yet made.

A dialog — PdDialog, and the palette — returns focus to whatever opened it when it closes; it used to fall to <body>, and a keyboard user lost their place in a 200-row table. The package grid itself is focusable but not a tab stop: the row is.

Six rules from 1.5.0, the accessibility pass:

  • Focus moves whenever the screen is replaced, not only when the tab changes. The shell focused the new screen's heading on a nav change and on nothing else — but the plan panel and each of its phases, the dependency view and an opened environment all replace what <main> shows without changing nav, and each unmounts the control that was just pressed. Focus fell to <body>. The legal gate is a fifth: it is rendered instead of the shell, so accepting it replaces the screen with nav unchanged, and the first screen of a first run had focus on nothing. A second effect in App.tsx covers them, and acts only when focus was actually lost: a run advancing from resolving to preview must not pull a keyboard user off Back, and an open modal owns focus until it closes. The console drawer returns focus to its opener the same way, and the legal gate — rendered instead of the shell — focuses its own heading.

  • An accessible name contains what is on screen (WCAG 2.5.3). Health's tabs were named by an aria-label reading "Dependencies, 3 findings" over a tab that says deptry 3. The name is the content now: the tool, its count, then what the tab is about, visually hidden. A PdDialog is named by its heading, followed by what it is about, and described by its body — it was named by the package alone, so a reader heard "numpy, dialog" and never the question.

  • A composite widget is one tab stop. Health's tabs follow the tabs pattern (one stop, arrows, a tabpanel), and a conflict's three choices follow the radio group's — with one deliberate difference: the arrows move focus and choose nothing. Checking whatever focus lands on would open Force latest's destructive confirm from an arrow key.

  • What a chip means is text, not a tooltip. A title on a <span> that cannot take focus is reachable by a pointer and by nothing else; a held pin read as a bare version number.

  • A status region exists before its first message. Search's is mounted from the first render and carries "preparing", "searching", the result count and "nothing matches". A live region announces changes to itself, so one that arrives together with its text has nothing to differ from. Several older regions still mount with their text; whether a reader says them is in the Audit backlog, because only a reader can answer it.

  • Busy is aria-disabled; unavailable is disabled. A browser takes focus away from a control that becomes disabled, so a button that disables itself when pressed — Use, Rescan, Run, Run audit, Add, Download the index, Stop, the two requirements buttons — dropped focus to <body> on the keypress that started its work. Those carry aria-disabled while busy and ignore a second press; disabled is kept for a control that cannot be used at all, which nobody is focused on when it happens. jsdom does not move focus off a disabled control, so no test could see this; it was found by pressing Download the index in the installed build.

Zoom is not on. zoomHotkeysEnabled (WebView2's own zoom control on Windows — no permission, no code) was set for 1.5.0 and taken out again on a measurement: at 200 %, which in the default 1180 px window is a 590 px viewport, the package table's name column (min-w-0 flex-1) collapsed to zero width and the sidebar, which does not scroll, cut off its last two tabs. At a maximised 1920 px window the same zoom is the layout's own 900 px floor and holds. Shipping it means giving the name column a minimum and the sidebar an overflow — U51's container-query work — and until then the Windows display scale, which enlarges the window with its content, is the supported way.

The arrow keys are not decoration, and the gap they close was real. Making the row the only tab stop is what keeps 200 rows from becoming 600 — but every control inside one is tabIndex={-1}, and that left Pin and Remove reachable by mouse only: a WCAG 2.1.1 failure, found by tabbing to a row in the running app and discovering there was nowhere further to go. ←/→ are the ARIA grid pattern's roving tabindex, Esc returns focus to the row, and no tab stops are added.

Enter on a row is the non-destructive primary. On a package row that is pin, never uninstall: a row has one reversible annotation and one irreversible removal, and a destructive operation one keypress from a focused row is an accident waiting rather than a primary action. Everywhere else in the app Enter needs no binding at all, because every other primary action is a real <button>.

"Select-all-visible" means the current filtered set, not the rendered window. Resolved in S2: the package table is virtualized, so only ~25 of a 200-package environment are in the DOM at any moment and "visible" would otherwise mean whatever the scroll position happens to be. Ctrl+A selects every selectable row in the set the screen is showing — so on Updates that is the outdated rows, and pinned rows are excluded there as they are from the Select all button. The row, not its checkbox, is the tab stop and the target of Space; tabbing through two controls per row is not traversal at 200 rows. Contrast: all text tokens meet WCAG AA on their surfaces, and since 1.4.0 so does every badge's text on its own tint over every surface (§2; both verified in the design-token test); focus ring 2 px accent, never removed — and since 1.5.0 a package row's ring is drawn inside the row in every mode, because a ring 2 px outside a row that is absolutely positioned in a scroller and stacked against its neighbours was clipped on three sides. Screen-reader labels on all icon buttons; live region announces execution progress summaries ("13 of 15 complete"). Windows high-contrast mode: tokens fall back to system colors. Structure is held by axe-core over every screen (ui/src/a11y.test.tsx, TESTING L3) — which checks roles, names and required parts, and cannot check speech, forced-colors or focus order.