Repository navigation
Releases
1.8.2
1.8.0 and 1.8.1 were never published: their work is here, with what followed them.
-
no-find(base, on in every project,updateincluded):find,bfs,fd,locateandmdfindare refused as commands, anywhere in a line, past wrappers andxargs, unless the user's prompt that turn asks for one by name ("use find to …"; a question about one asks nothing). The refusal says where to look: the order's sources with their paths, the layout's folders, Glob and Grep on a named folder. A BANDWARP agent'sfind / -name QwertyPlayer.hsearched a Mac's whole disk for a minute and a half for a file the order lists in~/projects/synthcore(KWB-154). Its selftest, 18 cases, each law seen red. -
The status line and
statusshow an agent from its first minute: a worktree on the order's own branch (work/<order>-<name>) is that order's agent, "not started yet" while its branch has nothing of its own and nothing runs in it; a merged one untouched for two hours is left out unless work runs in it. Where no heartbeat hook runs yet, an agent's last call comes from Claude Code's own transcripts. The lead's own work (its builds, its gates) shows on its order's row beside its agents. -
Logged for the next release: KWB-151 (the same warning flags are not the same law on every compiler), KWB-152 (the kit is not proved the same on Linux, macOS and Windows), KWB-153 (a test the kit runs has no memory or time limit), KWB-155 (a file in one worktree read as a copy of itself in another).
-
The runs as a tree, in
statusand the status line: a project, its orders under it when it has more than one, each order's agents under the order (├,└,│); a lead working alone is said on its order's row ("lead on W9 …"), never as a row of its own; a run in a worktree is named by its project (VSYNTHGT (ui-look)), never the worktree's folder. Once other machines are listed every run says where it runs, this machine's too (PC:,MAC:; remotes.json's"name", elsemacon macOS andpcelsewhere). The process check measures the lead's checkout as well, so a lead's build shows its processes, cores and memory.test/run-remotes.test.mjs. -
KWB-149: a build yields with
niceon every system; on macOS it ran undertaskpolicy -b, background work, which macOS keeps on the efficiency cores: a Mac's build used about 2.5 cores of 18. KWB-150: a job takes the core budget divided by the jobs running as it starts, its own counted, never less than its slot's share: a build alone on a quiet machine had compiled at-j 3.test/core-budget.test.mjs. -
KWB-148:
toJsonwrites an object that appears twice side by side in full both times; only an object inside itself is a circle.test/json-helper.test.mjs. -
The runs of other machines:
~/.config/claude-packs/remotes.jsonlists them (name,sshas~/.ssh/confignames the host, anddir, a project there whose kit answers).statusand the status line ask each for itsstatus --jsonover ssh, with no password prompt, a 5-second connect limit and 20 seconds in all, from the background refresh, so nothing waits on a machine that is asleep; its runs show asMAC: SHOPwith their agents, measured on that machine. A machine that cannot be reached, or whose kit has nostatus, is a red row that says so.status --dirasks no machine.test/run-remotes.test.mjs. -
The kit's status line (
base): every project shows what is working and where it is. The top line starts with the kit version that project runs (from its lock), then the model and effort, the context used, and the 5h and 7d quotas on a subscription or the session's cost on API billing; the model shows red when a fallback swapped it from the one pinned in the user's settings. The second line has the branch, its uncommitted files, ahead and behind, the last tool with its target and the last shell command, read from the session's own transcript. Below them, every unattended run on the machine:PROJECT [bar] NN% > agent > step [bar] NN%, an agent quiet for 30 minutes in red. The runs are found in the background at most every 30 seconds and drawn from a cache, so a render never waits (about 70 ms); any feed that is missing or broken draws nothing.initandupdatewire it in.claude/settings.jsonwithrefreshInterval: 1where the project has no status line of its own; a project's own is kept and the install says so;layout.json"disable": ["tools/statusline.mjs"]takes the kit's out.test/statusline.test.mjs. -
Every agent's sign of life, kept by the hooks, never by the agent:
heartbeat(base, async on every tool call before and after, on by default withupdatetoo, since it refuses nothing) writes what each agent did last into the main checkout's.claude/logs/heartbeats/<worktree>.json. Each refresh also checks what really runs in each agent's worktree (processes, cores and memory from/proc, or onelsofcall on macOS) and whether its build waits for a slot of the build queue (the slot layout is one file now,hooks/lib/build-slots.mjs, read bybuildqand the check). An agent's row:♥ 12s ⚙ 13 proc · 1.3 cpu · 322M ram READ <target>,⟳while a call runs, columns aligned and as wide as their widest value; red when it is stuck by the evidence, no call for 5 minutes or a running call with nothing working in its worktree. A project's percent counts each agent's own progress, so it moves while the lead waits on its agents.heartbeat --selftest, its worktree case,test/run-status.test.mjs. -
One JSON helper (KWB-146):
packs/base/hooks/lib/json.mjs,toJson,parseJson,readJsonandwriteJson, none of which throws: a circular reference, a BigInt, an Error and atoJSONthat throws are each written, bad or missing text gives the fallback, and a file is written whole through a part file. Everything new in this release calls it, andlib/global.mjs's strict reader is built on it.test/json-ratchet.test.mjscounts the rawJSON.parseandJSON.stringifycalls left in the kit (560) and is red if the count rises, until the migration takes it to zero.test/json-helper.test.mjs. -
The run readers (
run-status.mjs,run-discovery.mjs) live in the hooks lib, where the installed status line reaches them;lib/run-status.mjsre-exports them.status --json [--out <file>]prints the report as data. -
claude-packs status: where every unattended run stands. It finds the runs alive on the machine from theirrun.shprocesses (or one project's orders with--dir), and prints one line per project,NAME [bar] NN% X of Y items left, N agents, then one per agent,> name > step [bar] NN% what it is doing, or the lead's current step when no agent runs.--watch <minutes>prints it again on every tick. The unit is the item, a step's numbered items plus one for its Done when; the run's ownstatus.jsonand each agent'sagents/<name>.jsonin its worktree win, and an order without them is counted from itsprogress.mdlines, a step marked done counted whole. A status file that is not JSON is ignored, never a crash. The work-orderPROMPT.mdnow tells every run to keep both files.test/run-status.test.mjs.
1.7.2
- Smoothing is always on (KWB-145):
realtime-guardjudges every project withaudio-core, with no setting that switches it off. Withoutguards.smoothingthe smoothing module is any folder namedsmooth(and the copy trees) and the scope the project's audio code; the key only names them more exactly. Existing projects take it withupdate: only lines an edit adds are judged. It also refuses a bypass switched in one sample (if (bypassed)), which clicks, andsmooth-continuous-parametersnow says every switch that changes what is heard crossfades through the smoothing module, and the click gate covers stepped controls and bypass.
1.7.1
- Every work order ends its build with the README (
base's new skillwrite-the-readme, and a README step in the plugin, rebuild and generic templates, right before the last review): written last, when the UI is finished, the product's own screenshot on top (rendered by the product through its real code path, never a mock-up), what it does, install, first use, build, requirements, what is not done yet, the licence; every link and command checked, thehumanizepass recorded.test/workorder.test.mjs. - KWB-144:
updateno longer writes"settingsLayer": []into a project that never had the key (1.7.0 did). - KWB-143:
npm run releasereaches GitHub over HTTPS with$GH_TOKEN(through a credential helper, never on a command line), never through an SSH origin, and refuses to start without the token.
1.7.0
- A menu is one route table (
instrument-shell): the rulemenu-is-one-route-tableand the toolroute-table.mjs.checkjudges the table a project writes from its own: one outcome for each page, input and condition, every page reached from the start, a way back to the start from every page, every input routed or declared inert with its reason, every pulldown and window closing back to the page beneath, and no route to a page or input that does not exist.mermaiddraws one diagram per branch from the table, for a person to read; the table stays the source. Each law is seen to catch its planted fault (--selftest,test/route-table.test.mjs). - Every plugin has a settings layer with the standard settings (
audio-core): the rulesettings-layer-owns-the-devices(the standalone built as the project's own app roundStandalonePluginHolder, so no JUCE Options button; oneAudioDeviceManager; tabs along the top, each tab scrolling; the audio and MIDI tab's standard rows: devices, channels, sample rate, buffer, mute input to avoid feedback, the MIDI inputs, Enable QWERTY) and the hooksettings-layer, which refuses a write to the layer (layout.json"settingsLayer") that leaves out a standard setting, and a CMake write that adds a JUCE Standalone in JUCE's own window. The standard settings are one list (hooks/lib/settings-standard.mjs); a group a plugin has no use for is left out with its reason.doctorruns its--status: a JUCE plugin with no settings layer is red.ui-mockuprenders the settings layer as a screen of every plugin. - What a layer hides does no work (
instrument-shell): the rulehidden-under-a-layer-is-idle(under an opaque layer nothing is drawn, ticked or animated, what shows through a translucent one still is, the sound and the parameters go on, and the page beneath is drawn once at close at its current values) and the skillprove-hidden-is-idle(paints and ticks counted under every layer for a second, 0, with a planted meter left running seen red). menu-mapreads the routes as well as the pages: which press opens which page, from the firmware's own page changes, with the manual as the second source.- KWB-141:
test-changedsends a change underplanner/to the planner test. - KWB-142:
workorderandnewrefuse a name that starts with-, soworkorder --helpno longer writesdocs/workorders/--help/.
1.6.1
DISCLAIMER.md(shipped in the package) and a Disclaimer section in the README: no warranty, the user answers for what they build and for having the right to every source a project reads (firmware, manuals, recordings), reverse engineering is for learning and interoperability and the law and licences that apply must be checked before using what it finds, tools the kit points to (Ghidra among them) are their makers' and their output is a claim to prove, product names belong to their owners, and none of it is legal advice.- The README's first line links GitHub and its releases beside npm, so npm's page leads to the repo.
npm run release(scripts/release.mjs, the kit's own, not shipped): one release the same way every time: npm, the tag, and a GitHub release whose notes are the version's CHANGELOG section. It refuses offmain, on a dirty tree, or whenmainis not GitHub's, and skips each step already done.test/release.test.mjs.- Image prompts quote every word to be drawn:
ui-mockup(a new law), the walkthrough's 5.1 andstart-from-a-workorderput each title, label, scale mark and readout in double quotes, with the rule that the only text drawn is the quoted text. An unquoted word is read as description, and one mock-up came back with labels no order named. A misspelled word is never a reason to render again: the build draws every word from the order, in code.test/planner.test.mjs.
1.6.0
The kit's claude.ai side: the planner a person uses to get from "I want to build a synth" to a finished work order, and the kit commands that make its hand-over true. Work order planner.
planner/(new, shipped in the package): three skills for claude.ai.workorder-writeris the interview, nine stages from the idea toworkorder.md, every question one of two (what are you making, what makes it done), withquestions.mdandstages.mdbeside it; it writes the order's install block and a ten-linesummary.md.start-a-buildis the hand-over (the machine, WSL and LSP, the commands, where the order and the pictures go, the start sentence, or the whole set-up over a GitHub connection), withhandover.md.ui-mockupmakes the design pictures: 16:9, Nano Banana Pro through OpenArt, iterated by image, approved only on six checks (counts, depth, alignment, readability, live parts, clutter), withanalyse-a-mockup.md; never an HTML prototype.planner-skills --zip <dir>: one zip per claude.ai skill, five in all: the planner's three, anddesign-a-presetandsynth-costfrom their packs as they are.claude-ai-skillnow writes the planner'sworkorder-writer.zip(its 1.5.0 prompt file is gone: one interview, not two).workorder <name> --template generic|plugin|rebuild: the templates live inprompts/workorder/templates/;genericis the default, and an unknown name is refused with the list.pluginis new: a synth or effect with the house's decisions written in (the stack, gates and pluginval from commit one, the chart and the parameter table before any DSP, one step per sound part with its[if ...]markers, the panel from the approved render, presets, the realtime audit, a named host list). Every template carries an install block, and the synth ones the "never prototype the UI in HTML" line.workorder check <name>: the pre-flight. It refuses, naming each, a placeholder left outside a code fence, a step whose Done when holds no number and no path to a test, a skill named that the project has not installed (with the pack to add), and a reddoctor --strict; an order with no install block is warned.workorder <name>prints the same at its end;run.shruns it before the first session and starts nothing on a refusal (exit 5), andPROMPT.mdasks for it first.init --from <workorder.md>andnew <name> --from <workorder.md>: install from the order's install block (one fence with the info stringclaude-packs): its configuration and packs, then its layout keys (enable,looks,guards,commands, ... the keys a project owns), then an update. A key the kit does not read is refused by name and nothing is installed. The block is parsed withyaml2.9.1, the kit's one runtime dependency (docs/versions.md).new <name> --template <t>also writes the project's work order from that template.synth-cost(audio-core, skill): the cost ledger, units per voice against the budget, the cuts in order of least pain, effects global by default;cpu-worst-case-per-blockreplaces the ledger's estimate with the W0 measurement and holds the gate to it.design-a-preset(sound-design) is now the patch document: the method, the gate, the vocabulary with the house engines generalised, and its references loaded by kind (building-blocks.md,tables.md,what-a-synth-needs.md, fifteencategories/<kind>.md, generic recipes only).presets-are-named-and-neutralgrows the bank laws: levels measured, never typed (-6 dBFS peak on a four-note chord at v100, the bank within 1 LU, mono under 120 Hz, nothing under 40 Hz, limiter off), a pad's release at least its attack, velocity audible and never only volume on a pad, the wheel does something, the delay off the arp's subdivision, glide off by default.prove-a-bank(sound-design, tool): the bank's measurable gate on renders frombounce-a-pluginand the patch values: chord peak, bank loudness, mono lows, energy under 40 Hz, tail, first sound, velocity in level and colour, wheel, sync at 90, 120 and 140 BPM, delay against arp, pad and bell release, organ velocity.audio-core'swav.mjsreturns every channel and writes a float WAV;spectrum.mjsexports its FFT.- Two
sound-designchecks, measured on the house banks:preset-name-no-brand(a machine or brand name in a preset name; 1 hit in 2,086 presets) andpreset-not-default-envelope(a text patch that never sets its amp envelope). run-a-workorder(base, skill): the run's half ofwrite-a-workorder: the check before the first session, how to start and leave a run, what to watch, how to change one, agents in the foreground, landing once.write-a-workordernames only installed skills, carries the install block, and starts from the three templates.- The work-order prompt lands one branch per order and ends with the tests the branch's changes reach, not a branch per step and the full suite (KWB-128).
- An update no longer writes
"chart": []into a layout that set none, which madechart-firstlook for no chart at all and refuse every source write in asynth-creatorproject (KWB-136; unreleased since the public repo's first commit).