Repository navigation
PATCH as the mock's canvas, every function kept (Plan-008 C2a) - #117
Merged
Merged
Conversation
… ⚡ ▾
Plan-008 C2a, first step. PATCH's two toolbar rows (subject block,
callout, ⚡, teach pair, camera group, locks) become the specimen's head:
PATCH · FAMILY (a preset's category, only as made), the name at the
display size, and a subtitle the rack counts ("4 modules · 1 modulator,
in signal order", with the layout's own words when it is not the chain,
and the states the caption used to carry). On the right: what to do next
above the acts; ⚡ EVOLVE FROM THIS with a ▾ holding lock knobs, lock
wiring, clear locks and, while a stoppable ⚡ runs, stop; ADD MODULE (/);
NEW PATCH (CLEAR, BACK TO ‹name› in a new patch); HOW TO READ THIS, a
legend over the well's foot that says what each mark is, measured or not.
The edit bar ("N changes · KEEP AS NEW · pick the edit ☐ · undo to as
opened") shows once the bench is edited, in a row the head always keeps,
so nothing moves. N is the undo steps since the open (the stack starts
empty at every open); "undo to as opened" is one edit_set_tree to the
bottom of the stack, settled as that many undos (settleRestore with
`all`), so ⇧⌘Z brings them back in order. Past the stack's 60 steps it
is withheld and says why. The commit pair stays.
The camera group moves to the well's bottom-left corner unchanged (its
own commit restyles it); the budget joins the subtitle (near a ceiling,
or under Show measurements, today's rule); the belief line stays above
the well until C2b gives it a home under ⌥.
Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW
…plays Plan-008 C2a. The rack's frame becomes the specimen's well: a rounded, recessed canvas, the viewport inside it without an edge of its own. The plates take the specimen's look over the same SVG (a flat plate with a hairline edge and its name in silk caps; no bevel, screws or pocket; modulators dashed), and a knob is its travel and value arcs and a pointer. Every hook the specs and the motion system read stays on its node. Past the amp the build draws OUT (a lead, the jack, its name) in the amp's own group, so it travels with it, and leaves a zone for the face at OUT (OUT_ZONE, counted in the rack's size and in every fit). The face is #rack-play itself, an HTML button the camera places over that zone (placeOutFace, from applyView): the bench's measured face, the same target as the menu bar's chip (an edit's once its render lands, never an estimate), and a press plays the bench's phrase, waits for an edit with ▶'s dotted ring and says why when it can't, as the rack's ▶ did. The head's small face beside the name goes; faces.spec reads the one at OUT. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW
…dges count what is past them Plan-008 C2a. A modulation cable into a module now carries the specimen's label, "depth 25% · 0.51 Hz": the host's mod depth and the modulator's rate (up a chain to its source), each printed as its knob prints it, so the words are the knobs as set and never a measurement. It is repainted with the knobs on a value-only redraw, and leaves under the silkscreen floor with the readouts. When the patch is wider than the view, the well's edges say how many modules' middles lie past each side (‹ 2 · 3 ›), and a press glides the nearest one to the middle of the view (--d-move, instant under reduced motion). Refreshed from applyView, so a pan, a zoom or a fit updates them. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW
…spec dock retires Plan-008 C2a. One module is selected at a time: the plate the keyboard stands on, or the one a press landed on (not on a knob, jack or button); a press on the well's bare floor that does not become a pan clears it, and a rebuild keeps it by key. The selected plate shows its ⋯ (the structure menu) and its lock, with its edge and name lit; other plates show them only under the pointer (on a coarse pointer, always). A locked plate's edge is solid amber, no longer dashed. The readout (`#pt-read`) sits at the well's foot, beside the camera's corner: the module under the pointer or the selected one, by name, with its setting and what it does in one sentence. Pointed at in the catalogue, a module's longer description (ports, settings, what is heard, how your taste leans) opens over the well's bottom edge, so the well never resizes on hover; at rest it says nothing. The spec dock under the rack and its resizer (and `auracle-spec-h`) go: the armed line lives in the readout until it moves to the well's top line. The toast lane steps over the foot. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW
…nd on its top line
Plan-008 C2a. The module rail beside the rack becomes the specimen's
catalog (`pt-cat`): a panel over the well's left, closed at every start,
opened by ADD MODULE and /, closed by ✕ and Esc (after a module in hand,
before a new patch). It keeps everything the rail had: the search
(#nb-q), what is in this patch, the groups, the port pips, the legend,
the status line, the keyboard grid, drag a module to a socket, click to
take it in hand, the walkthrough, the guess's mark on its kind. θ is the
model view's now: the bars show only under ⌥ (the walkthrough's step
says so). A new patch opens it, as the specimen's np-desk does, and
leaving the new patch closes it. While it is open, fits keep its width
clear, and opening it over plates makes room with a jump. The rail's
collapsed tab and its width handle go with the rail.
The thing in hand moves from the strip under the rack to the well's top
line (the pick chip, no longer pinned to the plate): the module, its
price ("prices what, not where"), ▶ to hear it at the socket with that
socket's scope, how many sockets are lit, what happens at the socket
under the pointer, Esc or ✕ to put it down. The guess's line steps aside
for it. Fits keep the top line and the foot clear as well.
Specs that place a module open the catalog first (openCatalogue in
shell.js); taste_marks reads θ under the model view.
Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW
… the layout's ▾ Plan-008 C2a, round 2's camera decision. The toolbar's camera group becomes quiet corner controls in the well's bottom-left: fit · − · + · map · layout ▾. The ▾ is a menu (arrows, Esc, a press outside): chain, compact, by hand; snap to the grid and reset positions, which act on a layout by hand only; the level of detail; leans (which C2b moves under ⌥). The map stands over the corner, its bookmarks with it (⇧-click, ⇧1–9); the fit hint becomes a note over the corner. Every gesture on the canvas is unchanged, and the subtitle names the layout. The compact level of detail is now the specimen's plate: each module shows its first three dials to read (their arcs and pointers as set, repainted with the knobs), not to grab, where it used to show none. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW
…, Enter in, Esc out Plan-008 C2a, round 2's keys decision. On a module: ←/→ walk the modules in signal order (left to right as drawn, the model's guess among them), ↑/↓ go to the nearest module below or above (into a module's modulators and back out), Home/End the first module and the last (the amp, at OUT), Enter into its controls, F2 or the context-menu key its structure menu (right-click still opens it), Delete removes it through the survivor choice, l locks it, / puts a module after it, and Esc leaves it. Inside a module the arrows walk its own controls and stop at its ends; ↑/↓ turn a knob (⇧ finely, a run one undo step), Enter or Space cycles a setting, Esc goes back to the module. Home and End reach the modules from anywhere in PATCH, and fit-all moves to ⇧Home; `.`, ⇧1–9, ⌘0/⌘−/⌘= and ⌘Z/⇧⌘Z are unchanged. Esc is one chain: a knob, then the module, then the catalog, then a new patch. A pointer's turn still leaves no focus, and PERFORM's pad keys stay PERFORM's. The ? card and the rack's accessible name say the new keys; audio_in_takes enters a module's controls with Enter before walking them. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW
…ps; the sheet's face Plan-008 C2a. What was unplugged, deleted or bypassed is no longer a strip under the rack: SET ASIDE n sits at the well's foot beside the camera (only while something is set aside) and opens the shelf to drag from; the catalog lists it first too. Every item keeps its jack to drag back onto a ○ and its ✕, kept across a reload as before, with a finger's reach on a coarse pointer. The TEACH strip folds into a chip at the foot's right end (shown while a pair is on the table) that unfolds it over the well, with ✕ and Esc to fold it again; "Teach it your taste" opens it. The bench's one-line tour becomes PATCH's first steps in the guide pill, one at a time, each ticked when it happens: turn a knob, lock what you love, ⚡ (guide.js now holds steps per level and says each level's closing line; a tour already dismissed counts them done). The pill shows on PATCH under its well, PERFORM's steps on PERFORM. "Play it first" is PERFORM's first step, so the head's next step says nothing until a key is played; what is not onboarding (refit, breed a generation, what it bred) stays pressable there. The touch sheet shows the bench's face beside its settings, the one at OUT: "as made", "measured" once edited, "measuring…" while an edit is on its way. AUDIO IN modules that read one input draw one stream fanning out to each, as the specimen does, read off their input knobs. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW
…eet, a finger's reach; specs for the canvas Plan-008 C2a. The edit bar's compare, which the inventory placed on paintKnobKept, could not live there: that files PERFORM's performed value behind a knob, not the value before an edit. So each continuous knob now keeps its value as the sound was opened (by node identity, so a structural edit leaves it on its knob) and, once turned, shows a pale tick there; HOW TO READ THIS says what it is. On a touch screen the module sheet carries AUDIO IN's input line, MONITOR, NEW CLIP and ALLOW INPUT and CAPTURE's RECORD as 44 px buttons that press the lane's own on the rack (one of each, its state read back after the press), since the lane's buttons are 20 rack units tall. The head's acts, ⚡ and its ▾, and a plate's ⋯ and ▢ pads reach 40 px under a coarse pointer. A menu item that disables itself on its own press (clear locks) still folds its menu, and the face at OUT keeps the focus ring with the waiting cue on its edge, as ▶ did. Specs, one per area: patch_canvas (the head's counts, the face at OUT is the bench's and plays, the edit bar and KEEP AS NEW, undo to as opened in one restore, ⋯ on the selected module reaching every verb, L and ⚡'s ▾), patch_catalogue (ADD MODULE and /, search, the top line's price and ▶, a drag placed, θ only under ⌥, set aside first), patch_keys (signal order, into modulators, Enter and Esc, ⇧Home, F2 and Delete, the Esc chain, armed placement, keys yield to the search and notes still play), patch_camera (the corner, the layout menu, map and bookmarks, ctrl-wheel zooms the camera not the level, the edges) and patch_touch (sizes, ⋯ by tap, the sheet's face and lane buttons). patch_audible reads the face's look in place of ▶'s glyph. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW
…ntime notes and the plan Plan-008 C2a changed what PATCH shows, so every page that describes it says what is there now (ADR-004): the PATCH page is rewritten for the head, the edit bar, the well (modules in signal order, OUT and the face, the cables' depth and rate, the edges, the guess and the thing in hand on the top line, the foot's camera, SET ASIDE, readout and TEACH), the catalog, first steps and the sheet; rack.md for getting around (the corner, the keyboard's walk), the layout's ▾, locks in ⚡'s ▾ and the selected module's ⋯; wiring.md for the catalog in place of the module rail; keyboard.md for PATCH's keys and the Esc chain; and the pages that named the rail, the strip under the rack or the quick-pick strip (accessibility, faces, glossary, playing through, reading the model, teaching, the bank, first session, your data, the guess's reference, the READMEs). web-runtime.md gains how the canvas is built; the plan records C2a as built (where it differs from the mock and round 2, and the film selectors it leaves stale); the CHANGELOG says what changed. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW
…he catalog The fan from one input to several AUDIO IN modules is drawn left of the first plate, past the rack's origin, and contentBox floored the fit at the origin, so its name was shaved by the well's edge. The fit now includes the fan's box (rackExtras) and its floor follows it. Beside the open catalog the well's top line kept the well's whole width as its limit and started past the catalog, so the price, the preview's play button and "N sockets lit" ran off the well's right edge at 1000 px. The line now takes what is left and wraps. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW
Opening a preset fits it with a tween, and three of the camera's claims read the viewBox right after: the bookmark's zoom, ctrl-wheel's "smaller than before" and the edge's pan. Read mid-tween, the baseline was a frame of the fit rather than the fit, and one in fifteen runs failed. The spec now waits for two reads a quarter second apart to agree (20 of 20 passed). Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW
…arrow At 1000 px with the catalog open, the readout beside it is about 330 px wide, and the longer description over the well's foot kept its three columns: each a word wide, rising past the top of the well. The readout is now a size container and below 560 px the columns stack, with the card scrolling at 300 px. At 1440 nothing changes. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW
…do, keys, focus, the lean and first steps Review of C2a against the mock and the code. The mock's plates read at the fit a patch opens at; ours were drawn for a 2x zoom and fitted at 0.6-0.7x at 1440, so the silkscreen floor took every value and label and the settings were blank chips. The plates are re-sized (PLATE_W 84-240, KNOB_R 11, a 30-unit head, 56-unit rows, LAYER_GAP 56, FIT_MAX 1.25): every stock preset now opens at 0.84-1.19x at 1440 (Loom, the widest, 0.69x) with each name, value and label at its token's size. The setting that names a module's kind is the control in its head; each knob prints its value then its label; a setting prints its value (no chip); the amp's head draws its envelope from its four knobs. Audio cables are level S-curves, a two-input module's a/b sit outside it, plain in/out are not printed. Compact is where readouts stop reading (under 0.68x, with its band above the line) and draws names and settings larger, keeping AUDIO IN's and CAPTURE's lanes for the touch sheet's buttons. ⋯ and the lock moved to a pocket on the top edge, finger-wide on touch. The scope is off until asked for (it sat over the toasts), the belief line moved onto the well's top line and is empty without a guess, and the picks to the first refit have one home, the TEACH chip, which counts them. The armed line lets the pointer through to the sockets under it. Undo to as opened and ⌘Z waited only for the lane, so during an open they restored the leaving sound's tree under the arriving id. Both wait for benchSettled now; a refused ↺ says why (aria-disabled, so the press arrives), its toast is said when it lands, and its UNDO retires on the next step or restore. Global Home/End ran before the steered-control check and under modals: End on VOL walked to the amp, Home in the comparison focused a plate behind it. They are the canvas's now. A pointer's ✕ on the catalog or TEACH no longer hands the focus to a chip (Space played nothing); the lean shows under the model view only, as the copy said; × on the pill is per level, and PATCH's first step asks a player who has not played to play. Lows: TEACH resets with its pair, closing the catalog cancels a ⋯ handoff, the tour's lean step points at MODEL, the keyboard card stands beside the catalog, the selection follows the module's uid, one Esc closes one thing, lock-clear's reason is the new one, dead pick-chip code is gone, mod-cable words step aside while a plate is dragged, patch_touch waits on state, and catalogue is catalog (patch_catalog.spec.js). Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW
…w left them The guide's PATCH page, rack.md, keyboard.md, accessibility.md, wiring.md and the first session say what the canvas now shows and does: the kind's setting in the head, values then labels, the envelope drawn from the knobs (not the sound), curved cables with a/b by them, compact where labels stop reading, the scope put away until asked for, the belief line on the well's top line, TEACH counting the picks, ↺ saying why it can't go, Home/End belonging to the canvas, the lean under ⌥ only, × per level and PATCH's play-it step. The runtime notes, the plan's as-built and the CHANGELOG follow. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW
drawLane painted the lane's state but not its meter, and the meter loop runs only while an input is open. Lanes are now drawn in compact too, so a level-of-detail switch after an input was unplugged left a fresh lane with the meter as built (height 0) rather than as painted (0.0), and the same for its face. drawLane paints that lane's meter and face as they stand. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW
…p aside under a dragged plate Two of the review's low fixes had no spec. Closing the catalog now cancels a socket ⋯ handed it (the line goes, and the next module clicked is only in hand, nothing posted), and the modulation cable's words hide while a plate is in the hand (they are placed at the build, the cables move every frame) and come back when it is put down. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW
The catalog replaces the module rail; first steps are per level; the edit bar, the well, the camera corner, detail, TEACH and SET ASIDE get rows. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW
…he Browser job gets 25 minutes CI went red on two specs that turn the knob the bench opens with (keys_are_not_notes 102, evolve_truth 181) and on patch_editing's Loom spec. The last round moved automatic detail's compact line to 0.68x, where the readouts stop reading. The pool sound a first visit opens (ten modules) fits at 0.51x at 1440, so it opened compact, with no knob to turn from the mouse or the keyboard; Loom, at 0.69x locally, crossed the line on the runner. Compact is the specimen's plate to read, not to edit, so it has to stay where a knob really is too small for a hand: the line is 0.45x (a 10 px ring in a 16 px target, about the old knob at its floor), with the hysteresis band below it again. Between the two, the silkscreen floor leaves the values and labels off as it always did, and the names and head settings now print larger there, as compact draws them, so each module still says what it is. The new spec zooms Reese out to 0.48x and checks the knobs are drawn, the values are off, the names show, and the keyboard still turns one. The Browser job's timeout goes from 20 to 25 minutes. Shard 1 already took 19.6 of its 20 minutes before C2a's twenty-odd specs, each a fresh boot of 7-17 s with only state waits. Shard 2 was cancelled at 20 minutes, after spending 75 s on the two failures above. Sharing a boot across tests would make them depend on each other's order, so the job gets the time instead. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW
alexnodeland
added a commit
that referenced
this pull request
Oct 5, 2026
…e one source; the ship skill (#149) * docs: how work flows, written down once (RFC-009, ADR-019, docs/process.md) The flow from an idea to main was spread over CONTRIBUTING, testing.md, area AGENTS.md files, agent briefs and a lead session's notes; outstanding work lived in plan prose, and several rules (no retries, quarantine with an issue, voice rows approved first, agents commit while the operator ships, review before every PR) existed only in conversation. The maintainer asked for the process to be made canonical. docs/process.md is the living page: twelve stages from an issue to a closed issue, the issue labels and milestones, building in worktrees, the review checklist, PRs, CI and merging, flakes, dependencies, releases, status updates and the machine's rules. RFC-009 argues it, ADR-019 records the rules. The docs index gains them and the missing ADR-018; plans say their tasks become issues; CONTRIBUTING links the page and no longer says make check is all CI runs. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW * docs(agents): the front door points at how work flows Root AGENTS.md gains rule 9 (issues, a worktree per branch, agents commit and the operator ships, review before every PR, merge on green, flakes fixed or quarantined, new words approved) linking docs/process.md, and its commands table says what make check runs and how a PR is waited on and merged. CLAUDE.md listed a `site` skill that does not exist; it now lists `ship`, and says agents hand back while the operator pushes, opens PRs and merges. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW * docs(skills): a ship skill, and the skills say how CI and flakes work now `ship` takes one task from its issue to a merged PR as docs/process.md says: the worktree, the brief's essentials, review before the PR, the PR body, waiting on CI by state, the checked squash merge and the clean-up. auracle-strategy points at process.md and ADR-019 and lists the skills that exist (`site` never did); check says CI is the gate and adds the spec and workflow rows; browser-test covers make browser-changed, the eight runners, spec-only PRs, the merged HTML report, the shared fixture when present, and quarantine instead of retries. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW * docs(skills): changelog says which changes get no entry CI, specs, refactors and the contributor docs leave the release notes alone (as #118, #119 and #121 did); a fix a player could see gets an entry even when a test found it. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW * docs(agents): builders commit and hand back; the reviewer has the checklist web-engineer, engine-engineer, docs-writer and film-producer say where their work goes: their own worktree and branch, commits only (never push, PR or merge), Refs #N, no hand-written attribution, the fast gates and their own specs rather than the full suite, voice rows drafted not committed, and nothing moved without a home by mouse, keyboard and touch; each report starts with the head SHA. The reviewer gains the drop-nothing, ADR-004/012, spec-robustness and process checks it has been asked for in practice, a verdict line, and how to review only the fixes. docs-writer no longer points at a `site` skill that does not exist. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW * chore(github): the PR template and a task and a flake issue template follow process.md The PR template still asked for `make check` with release tests as the gate; it now asks for what docs/process.md requires: Closes #, the review before the PR, the fast gates and the specs the change reaches (CI runs the rest), a test named for the behaviour, descriptions true, nothing dropped by mouse, keyboard or touch, voice rows approved, wasm and cache-busters, no retried test. New issue templates: Task (plan, decisions, brief, done when, depends on) and Flaky test (the test, the run, why, the fix, quarantined?), plus a config link to the process page. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW * docs: the plans' open tasks link their issues; the cache-buster rule as practised Plan-008's progress table gains an Issue column (C2a merged as #117; C2b in review; #130–#132, #136, #139, #133/#134 to do). Plan-005 lists its open tasks (8 touch #134, 11 the own-sound card #133). process.md says which of two branches bumping ?b= takes the next value. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW * docs: every CLAUDE.md is the one-line import; the Claude Code setup moves to AGENTS.md and .claude/README.md The root CLAUDE.md carried the skills, agents, hooks and plugins tables after its @AGENTS.md import, so the rules lived in two kinds of file. Now every CLAUDE.md, the root one too, is `@AGENTS.md` (Claude Code loads CLAUDE.md, so the shims stay); the root AGENTS.md gains a short Tooling section, and .claude/README.md holds the detail. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW * make: a branch's AURACLE_TEST_PORT reaches browser-changed, -fast and -slow The browser targets ran on BROWSER_PORT (8690) whatever port the environment had set, so a worktree's own port never reached them. Now the port is AURACLE_TEST_PORT when set, else BROWSER_PORT, else 8690. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW * docs(process): the merge rule as GitHub enforces it, and who pushes what main's ruleset now requires the CI check with no bypass, and merged branches delete themselves; process.md says so. Before a merge, a PR that touches the app, tests, crates or CI is rebased if main moved, force-pushed with lease and re-checked; a docs-only PR may merge behind main, where main's run verifies it in full (reuse needs the same tree). While building or in review, a rebase only resolves a conflict. Agents' branches are pushed by the operator; a contributor pushes their own. Most PRs close an issue; a Dependabot bump or a small fix seen in passing may stand alone. Small fixes skip stages 2-4; stage 3 is an ADR per decision and a plan when the work spans PRs. Voice approvals are for words voice.md's table governs. The full-ci label, the backlog (no milestone), gh pr create --title, and where a branch's port comes from. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW * docs: ADR-019, RFC-009, CONTRIBUTING and AGENTS agree on pushing, issues, rebasing and voice Agents commit and the operator pushes; a human contributor pushes their own branch or fork. Most PRs close an issue, a Dependabot bump or a small fix seen in passing may stand alone. Rebase only for conflicts while building; before the merge, a PR touching app, tests, crates or CI catches up with main and is re-checked, a docs-only PR may merge behind. main's ruleset requires CI with no bypass. Voice approvals are for new rows in voice.md's word table, as voice.md § How this is kept says. RFC-009's context no longer says issues held "only" flakes (#62, #129), and it notes its one ADR and no plan. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW * docs(skills): ship works from any directory, opens issues non-interactively, and catches up before merging Every git and gh command names the repository ($REPO and the worktree's path for git, -R for gh). gh issue create --template fails without a terminal and drops the template's labels, so step 1 passes the body (frontmatter stripped), the labels and the milestone. CI is looked up by the head SHA throughout. Step 7 is the pre-merge rule: rebase, push with lease and re-check when main moved and the PR touches app, tests, crates or CI; a docs-only PR may merge behind. The remote branch deletes itself on merge, so clean-up removes the worktree and the local branch. The port goes in the brief as AURACLE_TEST_PORT. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW * docs(agents): hooks described as the scripts are, agents on Opus, the reviewer changes nothing in the repo AGENTS.md's Tooling and .claude/README.md now say exactly what each hook does: the session-start report (missing or stale pkg, the browser queue), all five generated paths guard-generated refuses, the post-edit checks by file type, and guard-bash's two refusals (cargo test without --release, --profile or --doc on any crate; playwright test outside one_browser.sh). Every agent definition gains `model: opus` (the maintainer's choice), and the README says so. The reviewer has one Truth bullet, and its probes live only in the session's scratchpad with their own Playwright config. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW * docs(agents): the reviewer has one Truth bullet and keeps its probes out of the repo It had two Truth bullets; they are one now, ADR-004 and ADR-012 together. It changes nothing in the repository, and a probe lives only in the session's scratchpad with its own Playwright config, through the browser queue on its own port. New voice rows are those voice.md's table governs. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW * docs(agents): builders draft only the voice rows voice.md's table governs; check_docs reads .claude/README.md web-engineer, engine-engineer, docs-writer and film-producer, and the auracle-strategy invariant, say the approval is for a new term, label or phrase the word table governs, as voice.md § How this is kept has it, not every sentence. check_docs.py now checks .claude/README.md's links too; the check skill says actionlint runs if installed; a long line wrapped. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW * docs(plans): Plan-004's and Plan-007's open items marked for their issues; the live face is done Plan-004 task 1's two "still to do" items (the landing, brand, 404 and film pages onto the size scale; the NOT_YET colour files) and Plan-007's "a clip per input" carry a TODO for the issue the operator opens and links. Plan-007 task 4's heading said "except the live face", but #104 built it (createLiveMeter, the audio_in spec "the square draws the input's live face"); the heading now says so. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW * docs: a branch's AURACLE_TEST_PORT reaches the make browser targets (browser-test skill, tests/web/AGENTS.md) Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW * docs: Plan-004's and Plan-007's open items link their issues (#145–#148) Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW * docs: the ship skill's commands work in zsh and across Bash calls; the session hook watches Cargo.lock From the re-review: `GH="gh -R …"; $GH …` does not split in zsh, so every gh line in the ship skill failed and its CI wait never ended; each command is now written whole, and the skill says each Bash call starts a fresh shell. The catch-up rebase checks main moved (merge-base --is-ancestor) and leases the pushed head explicitly. session-start.sh's find dropped Cargo.lock (its -name filter) and the root Cargo.toml; both are watched now, as the README says. The reviewer's scratchpad probe needs NODE_PATH; the Makefile comment's port precedence was backwards; reuse is when main did not move; the backlog is open issues with no milestone; agent branches are what's reviewed before a PR. Claude-Session: https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Plan-008 PR C2a: PATCH is rebuilt as the mock's canvas, and no PATCH function is dropped. C2b (the ⌥ model view in PATCH and the four engine facts) follows.
What changes
/), NEW PATCH, HOW TO READ THIS.edit_set_tree, and waits until the bench has settled./): search, set aside, in-this-patch, groups, port pips, spec card, legend, drag or click to arm.Tests
patch_canvas,patch_catalog,patch_keys,patch_camera,patch_touchspecs, plus guide_pill and guide.test additions.closedis per level.#nextstepis attached-only while a pair is dealt.make web-check(116 pass),make dev-checkclean,wasm-stamp. All targeted specs pass locally; CI is the gate.Not matched to the mock, on purpose
Cache-busters bumped. The voice.md rows follow in this PR once approved.
https://claude.ai/code/session_01GMg3KFS9ggSrYdvGkbBshW