Records the background jobs a session starts; after a crash or restart one key resumes what died, a usage-limit stall continues by itself at the reset, and…

Výbava is FixIt Technologies' portable engineering environment: small tools, agent skills, and workstation diagnostics, distributed as one catalog where every item installs independently.
| ID | Kind | What it does |
|---|---|---|
memorylint | applet | Validate and maintain AI memory homes — schema, indexes, wikilinks, fixtures, write hooks. → docs/memorylint.md |
claude-guards | applet | PreToolUse guard hooks for Claude Code — destructive git/docker, secret dumps, host-input automation, commit secrets, whole-disk walks, per-look Appium sessions, uncapped local test runners, and the context-budget rules (no shell-rewritten files, no whole-file dumps, no transcript reads); doctor re-wires hooks a settings.json rewrite dropped, vybava setup mac applies host settings. → docs/claude-guards.md |
lok | applet | Locale catalogs for AI sessions — configured in vybava.config.ts, queried by key, written by verb with every locale kept in sync, extracted from source, gated in CI. → docs/lok.md |
merge-assist | applet | Merge main without the mechanical conflicts — catalogs merge by key, generated files take theirs and regenerate, unmerged migrations renumber past the base; one table of what is left. → docs/merge-assist.md |
vybava config | verb | The shared per-repo vybava.config.ts every applet reads — init scaffolds it with typed helpers, check gates CI, show prints the evaluated JSON. → docs/config.md |
shrt | applet | Terminal-safe short links on luko.to — offline repo rules, team-shared dynamic rules, minted codes; also the redirector server. → docs/shrt.md |
posta | applet | Drive a shared test mailbox end to end — mint a per-run plus-address, wait for the mail a journey triggered, take its links and attachments. → docs/posta.md |
fontfreeze | applet | Freeze variable webfonts at rendered axis positions and subset per language. |
perfrig | applet | Performance drills from a testing/<project>/perf manifest — ramp to first failure, percentile report. |
framestats | applet | Android frame metrics for agents - parse gfxinfo framestats dumps (cadence, present intervals, per-gesture first-input and release-window frames) and perfetto traces (per-frame UI and RenderThread work, texture uploads, ART pauses, FrameTimeline jank). → docs/framestats.md |
perflab / expo-device-perf | applet + skill | Physical-device performance lab for Expo / React Native - one lease holder per phone (raw adb/devicectl on a leased phone is refused by claude-guards), native builds keyed by a portable Expo fingerprint, JS bundle variants without native rebuilds, measured runs through the project's vybava.config.ts perflab adapter, Perfetto and xctrace probes, and every number re-derived from kept evidence with compare and budget gates. The skill carries the loop, the instrument choice and the render-cost rules. → docs/perflab.md |
ingressgen | applet | Render and drift-check complete default-deny Docker ingress policies from a manifest. |
reconcile | applet | Pull-based GitOps for the infra boxes — converge a VPS to its infra repo's merged main from a per-box manifest: HELD hotfixes, transactional nginx hooks, commit rollback, textfile metrics, mesh-only status page + estate hub. → docs/reconcile.md |
readiness / release-readiness | applet + skill | Release prep of a whole integration branch: the skill runs authority questions, an inventory with a completeness critic, one named lane agent per journey cluster (own story, QA task, usertest board), capacity governance, a merged-branch roll-up and a readiness board. The applet validates the vybava.config.ts readiness adapter, freezes production..integration ranges, and seeds and renders the run directory. → docs/readiness.md |
stores | applet | Store release verbs for App Store Connect, Google Play and EAS - versions, builds, the live listing and its store.config.json drift check, prepare/submit an App Store version, Play completed or staged rollouts that keep the live release (5xx retried with a fresh edit), and the one-distribution-certificate iOS signing check; credentials injected by onyx or read from gitignored files, never printed. → docs/stores.md |
ui-loop | applet | The UI polish loop's deterministic layer: syncs the TypeScript/Playwright capture harness (typed screen manifest, capture + lint, app map) into a repo with a drift gate, runs passes from the vybava.config.ts uiLoop section (on a Devbox too), split-publishes them to vitrinka and scores the review backlog. The vitrinka map / review-loop workflows drive it. → docs/uiloop.md |
issues / issue-sweep | applet + skill | A repository's whole GitHub issue backlog to verified outcomes: Codex lanes triage every open issue against a frozen base with a skeptic on each not-live verdict, one human gate settles every decision, live bugs are fixed test-first (RED proven, Claude fixer, independent Codex verify) into one PR per issue or dependency group, stale issues get evidence comments and approved closes, every issue is matched to its vitrinka task and the run is reported as a vitrinka artifact. → docs/issues.md |
polish-kit | applet | The polish skill's deterministic layer: infers app/ui/api targets from the diff through the vybava.config.ts polish globs, resolves device lanes (iOS simulators and phones, Android devices and emulators, URLs) with the exact create/boot fix, keeps the pass ledger (run.json cells: lane x screen x theme x nav x text size, plus adverse-condition matrix cells), shoots simulators and Android natively, renders contact and edge-zoom sheets and the Markdown report with a delta. → docs/polish-kit.md |
prm / push-all / sync / push-back | skills | The git family — prm is the one PR verb (create → review rounds → gated merge → teardown), push-all the commit doctrine, sync the pull/merge-up flow, push-back the claim verifier. Install with vybava install ai-git. |
gitkit | applet | The deterministic layer those skills execute — PR selectors, review triage, merge gates, worktrees, path classification — as vybava gitkit <script>. → docs/gitkit.md |
vybava setup team | verb | One-click henderson-tech Mac — the henderson group's tools (Onyx + Helium with the lazy browser MCPs, vitrinka, SwitcherooBar, Pultík; devbox opt-in) through their own channels plus the git family, as a checklist or --yes --json. → docs/setup-team.md |
vybava setup agents | verb | Plan, apply and check lazy Appium and shared cmux badges; safe fleet finish/resume and codexsync audit. → docs/agent-setup.md |
codexsync | applet | Render ~/.claude skills and commands into ~/.agents/skills, the structure Codex discovers — nesting preserved, each command a source-command skill, duplicate discovery suppressed. → docs/codexsync.md |
ccx | skill | Claude Code stays orchestrator and app-source author; usertests, UI verification, computer use, app mapping and e2e writing/running go to Codex — the native codex subagent inside cc sessions, never the Codex CLI — routing table, briefs per mode, status handling, browser etiquette, commit duties, Opus fallback. |
codexusage | applet | Explain where the Codex plan limit went — per-thread spend from ~/.codex rollouts, the derived allowance, and the runway left at the measured burn rate. → docs/codexusage.md |
tokentime | applet | Where your AI tokens went — Claude Code transcripts and Codex rollouts indexed incrementally into permanent hour × project × model buckets (worktrees fold into their repo), rolled up by day, hour, project and model with an API-equivalent USD value. → docs/tokentime.md |
readeff | applet | How agents navigate code — what Claude Code and Codex reads and searches put into context against what they changed: re-reads, whole-file reads of big files, search hit rates and the most-read files, per repository, session and file. → docs/readeff.md |
find-session | applet + skill | Find the Claude Code session a pasted conversation came from — its author ranked above sessions that only quoted it — and the one line that resumes it from its launch directory under the switcheroo preset it started with (cc, cco, ccoo, …); the /find-session skill wraps it. Install with vybava install find-session-cli find-session. → docs/find-session.md |
repolicy | applet | Hold GitHub repository settings to a declared policy across whole owners — GitHub inherits no organization default, so audit reports the drift (exit 1) and apply converges it, touching only the settings the policy names. → docs/repolicy.md |
menubar-doctor | applet | Find and fix macOS menu-bar items that run but never appear — since macOS 26 Control Center files a status item under the process that launched the app, so anything started from a terminal is filed under the terminal and stays invisible while its switch is off. → docs/menubar-doctor.md |
reclaim / reclaiming-disk-space | applet + skill | Emergency disk reclaim for a dev Mac — a fixed ladder of regenerating caches deleted biggest-first with live df after every step, an early-stop target, and by-hand notes for what it refuses to touch. The skill adds the read-only audit (whole-Data-volume size ranking, orphaned Docker volumes by compose project name, session residue: shared-temp scratch, untracked ~/Exports evidence, transcripts, idle simulators) and the confirm-then-delete workflow. Install with vybava install reclaim reclaiming-disk-space. → docs/reclaim.md |
macwatch | applet | Mac load sampler with owner attribution - append load, memory and the heaviest processes to a TSV every interval, each tagged with project, directory and the claude/codex session that spawned it; report integrates the file into offenders, per-project and per-session totals and a spike ledger. → docs/macwatch.md |
plugin-gc | applet | Garbage-collect the Claude Code plugin cache — every version ever installed is kept behind a PID refcount that abandoned sessions never release, and the payload is almost all node_modules; reports by default, deletes only on --apply. → docs/plugin-gc.md |
vpn | applet | Persistent macOS WireGuard tunnels outside WireGuard.app — install pulls the Onyx profile into a root-only LaunchDaemon that survives reboots; status reports the route truth (up via wg-quick (utun11)) beside the app's Disconnected, the daemon kind and whether the tunnel DNS answers. → docs/vpn.md |
uplink | applet | Send one process's traffic out of an iPhone's USB Personal Hotspot while the Mac's default route and VPN stay put — devices finds attached phones by UDID and their hotspot interface, relay is a loopback CONNECT relay bound to it (IP_BOUND_IF: no root, no routes), measure times a bounded random upload. claude-switcheroo's daemon drives it. → docs/uplink.md |
onyx-rest | applet | One REST call with an Onyx-injected secret — Bearer or HTTP Basic, the response in a 0600 --out file; --base first and once, so an Onyx allowed_commands prefix pins the host. → docs/onyx-rest.md |
vybava vault-stdin | verb | Deliver Onyx-injected JSON through a private process pipe; no plaintext credential files or terminal output. → docs/vault-stdin.md |
handoffs | applet | Handoff ledger upkeep — handoffs reconcile judges every open handoff by whether its branches and PRs are still alive and archives the dead ones; unknown is never touched. → docs/handoffs.md |
press | applet | Deterministic state for the document family — project resolution, ~/Exports/<project>/ config and index, ARES lookups, shared doctrine. → docs/press.md |
press-pdf / press-logo / press-offer / press-email | skills | Offer, documentation and legal PDFs; brand marks; Czech commercial DOCX; Outlook-paste client emails. Issuer identity stays machine-local. → docs/press.md |
Groups (recommended, experimental, claude-hooks, ai-git, press-family, everything) are composable presets in the catalog — never code. claude-hooks is what a Claude home's hooks call; on Linux (a Devbox portal box) the same release archive installs it: vybava install claude-hooks.
Homebrew (everything is public — no authentication needed):
brew install --cask FixIt-Technologies/tap/vybava
CI images, workflows and provisioning scripts use the release installer in ci/ — never a checkout of this repository:
curl -fsSL -o /tmp/vybava-install.sh \
https://raw.githubusercontent.com/FixIt-Technologies/vybava/v0.3.3/ci/install.sh \
&& bash /tmp/vybava-install.sh --version 0.3.3 --bin-dir /usr/local/bin --install memorylint,hotfix
From source:
go build -o ./bin/vybava ./cmd/vybava
./bin/vybava catalog list
./bin/vybava install recommended
install takes item or group selectors (default: the recommended group) and supports --agent claude|codex|all, --scope user|project, --dry-run, and --json. Installed applets are links to the vybava binary, so vybava memory lint . and memorylint . are equivalent.
catalog/catalog.yaml package and group source of truth
cmd/vybava/ multicall entrypoint
internal/<id>/ one focused Go package per capability
skills/<id>/ canonical cross-agent skill payloads (SKILL.md plus any
references/ and assets/ the skill ships)
docs/ per-tool references, release flow, decisions
Dockerfile the luko.to redirector image (deployik app "luko")
One payload + one catalog entry = one package; presets are one more catalog line. Contract: docs/decisions/0001-modular-catalog.md · checklist: CONTRIBUTING.md · releases: docs/homebrew.md.
hooks/register.tsx 537 lines1import { atom, read, update } from 'claude-code'
2import type { EngineInterface, Register, Timer } from 'claude-code'
3
4import type { LazarusJob, LazarusKind, LazarusStall } from '../types'
5import type { Envelope, JobStatus, Ledger, LedgerEvent, LedgerJob } from '../types/fleet.gen'
6
7const mine = atom({ plugin: 'lazarus', key: 'mine' } as const, [] as LazarusJob[])
8const stoppedByModel = atom({ plugin: 'lazarus', key: 'stoppedByModel' } as const, [] as string[])
9const dead = atom({ plugin: 'lazarus', key: 'dead' } as const, [] as LazarusJob[])
10const diedAt = atom({ plugin: 'lazarus', key: 'diedAt' } as const, 0)
11const stall = atom({ plugin: 'lazarus', key: 'stall' } as const, null as LazarusStall | null)
12const lastTurnAt = atom({ plugin: 'lazarus', key: 'lastTurnAt' } as const, 0)
13
14const PANE = 'lazarus-jobs'
15const LIMIT_ID = 'usage-limit'
16const JITTER_MAX_MS = 60_000
17const RUN_TIMEOUT_MS = 15_000
18const FOLLOWED: readonly LazarusKind[] = ['workflow', 'shell', 'monitor', 'agent']
19const AUTH_ERRORS = new Set(['authentication_failed', 'oauth_org_not_allowed', 'account_on_hold', 'verification_required', 'billing_error'])
20const TASK_ID = /<task-id>([^<]+)<\/task-id>/
21const TASK_STATUS = /<status>([^<]+)<\/status>/
22const WORDS: Record<LazarusKind, readonly [string, string]> = {
23 workflow: ['workflow', 'workflows'],
24 shell: ['shell', 'shells'],
25 monitor: ['monitor', 'monitors'],
26 agent: ['agent', 'agents'],
27}
28
29// The one timer lazarus keeps: the wake at a usage limit's reset. A reload
30// drops it with the module; session.start re-arms it from the ledger.
31let limitTimer: Timer | null = null
32let ledgerWarned = false
33
34// lazarus records the background jobs the main loop launches in the
35// per-session ledger (`vybava fleet ledger`), so that after a crash or a
36// restart one key hands what died back to the model. Every tool call passes
37// through unchanged; the ledger is written only when a background job
38// starts or ends, never on a timer.
39export const register: Register = on => {
40 on('session.start', async ($, e, next) => {
41 await $.command.register({
42 name: 'park',
43 description: 'Is it safe to restart? Lists the background jobs a restart would kill',
44 immediate: true,
45 })
46 await scan($, await $.session.id())
47 return next(e)
48 })
49
50 // `claude --resume` and /resume land here; /resume fires no session.start.
51 on('classic.SessionStart', async ($, e, next) => {
52 const ran = await next(e)
53 if (e.source === 'resume') {
54 await scan($, e.session_id)
55 }
56 return ran
57 })
58
59 on('turn.start', async ($, e, next) => {
60 const now = await $.clock.now()
61 await update($, lastTurnAt, () => now)
62 const stalled = await read($, stall)
63 if (stalled?.kind === 'limit') {
64 // The engine's own auto-continue, the person or our wake started it.
65 await clearStall($, 'completed')
66 }
67 return next(e)
68 })
69
70 on('turn.complete', async ($, e, next) => {
71 const ran = await next(e)
72 if (e.agentId === undefined && e.reason === 'answer' && (await read($, stall))?.kind === 'auth') {
73 await update($, stall, () => null)
74 }
75 return ran
76 })
77
78 on('tool.call', { tool: 'Workflow' }, async ($, e, next) => {
79 const ran = await next(e)
80 if (e.agentId === undefined && ran.deny === undefined && ran.isError !== true && ran.result.status === 'async_launched') {
81 await launched($, {
82 kind: 'workflow',
83 id: ran.result.taskId,
84 runId: ran.result.runId,
85 scriptPath: ran.result.scriptPath,
86 description: ran.result.workflowName ?? 'workflow',
87 startedAt: await $.clock.now(),
88 })
89 }
90 return ran
91 })
92
93 on('tool.call', { tool: 'Bash' }, async ($, e, next) => {
94 const ran = await next(e)
95 if (e.agentId === undefined && e.run_in_background === true && ran.deny === undefined && ran.isError !== true) {
96 const taskId = ran.result.backgroundTaskId
97 if (taskId !== undefined) {
98 // Never the command itself: it can carry a secret, and the ledger
99 // keeps only its digest.
100 const description = e.description ?? 'background shell'
101 await launched($, { kind: 'shell', id: taskId, description, startedAt: await $.clock.now() }, e.command)
102 }
103 }
104 return ran
105 })
106
107 on('tool.call', { tool: 'Monitor' }, async ($, e, next) => {
108 const ran = await next(e)
109 if (e.agentId === undefined && ran.deny === undefined && ran.isError !== true) {
110 await launched($, { kind: 'monitor', id: ran.result.taskId, description: e.description, startedAt: await $.clock.now() }, e.command)
111 }
112 return ran
113 })
114
115 on('tool.call', { tool: 'Agent' }, async ($, e, next) => {
116 const ran = await next(e)
117 // A model's background Agent call answers `async_launched` with its id; a
118 // plugin's spawn answers an AgentCallRecord, which has no status.
119 if (e.agentId === undefined && ran.deny === undefined && ran.isError !== true && 'status' in ran.result && ran.result.status === 'async_launched') {
120 await launched($, { kind: 'agent', id: ran.result.agentId, description: e.description, startedAt: await $.clock.now() })
121 }
122 return ran
123 })
124
125 on('tool.call', { tool: 'TaskStop' }, async ($, e, next) => {
126 const id = e.task_id ?? e.shell_id
127 if (e.agentId === undefined && id !== undefined) {
128 await update($, stoppedByModel, ids => [...ids.filter(one => one !== id), id].slice(-200))
129 }
130 return next(e)
131 })
132
133 on('prompt.submit', async ($, e, next) => {
134 if (e.origin?.kind === 'task-notification') {
135 const id = TASK_ID.exec(e.text)?.[1]
136 const status = TASK_STATUS.exec(e.text)?.[1]
137 if (id !== undefined && status !== undefined) {
138 await ended($, id, status)
139 }
140 }
141 return next(e)
142 })
143
144 // The turn's end carries the engine's list of background work still in
145 // flight: a job of ours missing from it ended without a notification we
146 // saw. Absence proves that only when nothing is in flight, or when the list
147 // names a job of ours (so its ids are the ones the launches answered);
148 // otherwise the job is left to its notification, never closed on a guess.
149 // Agents are always left to their notifications; their task id need not be
150 // the agent id the launch answered.
151 on('classic.Stop', async ($, e, next) => {
152 const ran = await next(e)
153 if (e.agent_id === undefined && e.background_tasks !== undefined) {
154 const live = new Set(e.background_tasks.map(task => task.id))
155 const ours = (await read($, mine)).filter(job => job.kind !== 'agent')
156 const isProven = live.size === 0 || ours.some(job => live.has(job.id))
157 const gone = isProven ? ours.filter(job => !live.has(job.id)) : []
158 for (const job of gone) {
159 await update($, mine, list => list.filter(one => one.id !== job.id))
160 await record($, { kind: job.kind, id: job.id, status: 'stopped' })
161 }
162 }
163 return ran
164 })
165
166 on('classic.StopFailure', async ($, e, next) => {
167 const ran = await next(e)
168 if (e.agent_id !== undefined) {
169 return ran
170 }
171 if (e.error === 'rate_limit') {
172 await limitStall($)
173 } else if (AUTH_ERRORS.has(e.error)) {
174 const now = await $.clock.now()
175 await update($, stall, (): LazarusStall => ({ kind: 'auth', at: now, error: e.error }))
176 }
177 return ran
178 })
179
180 on('command.run', { command: 'park' }, async $ => {
181 return { text: await park($) }
182 })
183
184 on('ui.render', { component: 'AbovePrompt' }, async ($, e, next) => {
185 const jobs = await read($, dead)
186 const stalled = await read($, stall)
187 const theirs = await next(e)
188 if (e.props.hasSurvey || (jobs.length === 0 && stalled === null)) {
189 return theirs
190 }
191 const { Box, Text, Button } = $.ui.resolve(e)
192 if (jobs.length > 0) {
193 const at = await read($, diedAt)
194 return (
195 <Box flexDirection="column">
196 <Box flexDirection="row" columnGap={1}>
197 <Text>⟲ {summary(jobs)} died {clock(at)}</Text>
198 <Button key="lazarus-resume" label="resume all" hotkey="1" variant="primary" onPress={() => { void resumeAll($) }} />
199 <Button key="lazarus-list" label="list" hotkey="2" onPress={() => { void $.ui.open({ id: PANE, title: 'lazarus · what died', closeOnEscape: true }) }} />
200 <Button key="lazarus-dismiss" label="dismiss" hotkey="3" onPress={() => { void dismiss($) }} />
201 </Box>
202 {theirs}
203 </Box>
204 )
205 }
206 if (stalled === null) {
207 return theirs
208 }
209 const text = stalled.kind === 'auth'
210 ? `⚠ Claude auth failed (${stalled.error}) — switch the account in a terminal: switcheroo cs <account>`
211 : stalled.until === null
212 ? '⏸ usage limit hit'
213 : `⏸ usage limit — continues by itself at ${clock(stalled.until)}`
214 const go = stalled.kind === 'auth' ? 'retry' : stalled.until === null ? 'continue' : 'continue now'
215 return (
216 <Box flexDirection="column">
217 <Box flexDirection="row" columnGap={1}>
218 <Text>{text}</Text>
219 <Button key="lazarus-continue" label={go} hotkey="1" variant="primary" onPress={() => { void continueNow($) }} />
220 <Button key="lazarus-cancel" label={stalled.kind === 'limit' && stalled.until !== null ? 'cancel' : 'dismiss'} hotkey="3" onPress={() => { void clearStall($, 'stopped') }} />
221 </Box>
222 {theirs}
223 </Box>
224 )
225 })
226
227 on('ui.render', { component: 'Pane', requestId: PANE }, async ($, e) => {
228 const { Box, Text, Button } = $.ui.resolve(e)
229 const jobs = await read($, dead)
230 return (
231 <Box flexDirection="column">
232 {jobs.length === 0 && <Text dimColor>Nothing died.</Text>}
233 {jobs.map(job => <Text wrap="truncate-end">{describeJob(job)}</Text>)}
234 <Button key="lazarus-close" label="close" role="dismiss" onPress={() => { void $.ui.close({ id: PANE }) }} />
235 </Box>
236 )
237 })
238}
239
240// scan reads a session's ledger: an open job this process did not launch
241// died with the process before it; an open usage-limit wait is re-armed.
242async function scan($: EngineInterface, session: string): Promise<void> {
243 const ledger = await showLedger($, session)
244 if (ledger === null) {
245 return
246 }
247 const known = new Set((await read($, mine)).map(job => job.id))
248 const now = await $.clock.now()
249 const lost = ledger.jobs.filter(job => job.status === 'started' && isFollowed(job.kind) && !known.has(job.id)).map(fromLedger)
250 if (lost.length > 0) {
251 await update($, dead, list => [...list.filter(one => !lost.some(job => job.id === one.id)), ...lost])
252 await update($, diedAt, () => now)
253 }
254 const wait = ledger.jobs.find(job => job.kind === 'limit-wait' && job.status === 'started')
255 if (wait === undefined) {
256 return
257 }
258 const until = wait.until === undefined ? null : Date.parse(wait.until)
259 const at = Date.parse(wait.startedAt)
260 if (until !== null && until > now) {
261 await update($, stall, (): LazarusStall => ({ kind: 'limit', at, until }))
262 armLimit($, until - now)
263 } else {
264 // The reset passed while nothing ran: the person is back, the key decides.
265 await update($, stall, (): LazarusStall => ({ kind: 'limit', at, until: null }))
266 }
267}
268
269async function launched($: EngineInterface, job: LazarusJob, command?: string): Promise<void> {
270 await update($, mine, list => [...list.filter(one => one.id !== job.id), job])
271 await record($, {
272 kind: job.kind,
273 id: job.id,
274 status: 'started',
275 runId: job.runId,
276 scriptPath: job.scriptPath,
277 description: job.description,
278 command,
279 })
280}
281
282// ended closes a job of ours on its notification. A job killed or stopped
283// that the model did not stop itself died while the session ran.
284async function ended($: EngineInterface, id: string, word: string): Promise<void> {
285 const job = (await read($, mine)).find(one => one.id === id)
286 const status = toStatus(word)
287 if (job === undefined || status === null) {
288 return
289 }
290 await update($, mine, list => list.filter(one => one.id !== id))
291 await record($, { kind: job.kind, id, status })
292 if ((status === 'killed' || status === 'stopped') && !(await read($, stoppedByModel)).includes(id)) {
293 await update($, dead, list => [...list.filter(one => one.id !== id), job])
294 const now = await $.clock.now()
295 await update($, diedAt, () => now)
296 }
297}
298
299// resumeAll hands what died to the model as facts: workflows with the exact
300// call that resumes them, the rest for it to judge. The prompt is framed as
301// lazarus's; nothing is re-run behind the model's back.
302async function resumeAll($: EngineInterface): Promise<void> {
303 const jobs = await read($, dead)
304 if (jobs.length === 0) {
305 return
306 }
307 await update($, dead, () => [])
308 if (!(await submit($, resumePrompt(jobs)))) {
309 // Nothing reached the model: the jobs stay offered.
310 await update($, dead, list => [...jobs, ...list.filter(one => !jobs.some(job => job.id === one.id))])
311 return
312 }
313 for (const job of jobs) {
314 await record($, { kind: job.kind, id: job.id, status: 'stopped' })
315 }
316}
317
318// submit starts a turn framed as lazarus's; a refusal is logged and shown,
319// never lost in a press handler or a timer.
320async function submit($: EngineInterface, text: string): Promise<boolean> {
321 try {
322 await $.prompt.submit({ text })
323 return true
324 } catch (err) {
325 $.ui.log(`could not start a turn: ${String(err)}`, { to: 'debug' })
326 $.ui.toast('could not start a turn; see the debug log')
327 return false
328 }
329}
330
331async function dismiss($: EngineInterface): Promise<void> {
332 const jobs = await read($, dead)
333 await update($, dead, () => [])
334 for (const job of jobs) {
335 await record($, { kind: job.kind, id: job.id, status: 'stopped' })
336 }
337 await $.ui.close({ id: PANE })
338}
339
340async function limitStall($: EngineInterface): Promise<void> {
341 const now = await $.clock.now()
342 const usage = await $.session.usage()
343 const resets = usage.rateLimits
344 .filter(limit => limit.resetsAt !== undefined && limit.percentUsed >= 100)
345 .map(limit => Date.parse(limit.resetsAt ?? ''))
346 .filter(at => Number.isFinite(at) && at > now)
347 if (resets.length === 0) {
348 await update($, stall, (): LazarusStall => ({ kind: 'limit', at: now, until: null }))
349 await record($, { kind: 'limit-wait', id: LIMIT_ID, status: 'started' })
350 return
351 }
352 // Spread the sessions one limit stalled over a minute after its reset.
353 const until = Math.max(...resets) + jitter(await $.session.id())
354 await update($, stall, (): LazarusStall => ({ kind: 'limit', at: now, until }))
355 await record($, { kind: 'limit-wait', id: LIMIT_ID, status: 'started', until: new Date(until).toISOString() })
356 armLimit($, until - now)
357}
358
359function armLimit($: EngineInterface, delay: number): void {
360 limitTimer?.cancel()
361 limitTimer = $.clock.after(Math.max(0, delay), () => {
362 limitTimer = null
363 void limitReset($)
364 })
365}
366
367// limitReset is the wake at the reset: it continues the session only when
368// no turn started since the stall (the engine's own auto-continue, or the
369// person, would have cleared the stall already).
370async function limitReset($: EngineInterface): Promise<void> {
371 const stalled = await read($, stall)
372 if (stalled?.kind !== 'limit') {
373 return
374 }
375 const started = (await read($, lastTurnAt)) > stalled.at
376 await clearStall($, 'completed')
377 if (!started) {
378 await submit($, 'continue — the usage limit has reset')
379 }
380}
381
382async function continueNow($: EngineInterface): Promise<void> {
383 const stalled = await read($, stall)
384 if (stalled === null) {
385 return
386 }
387 await clearStall($, 'completed')
388 await submit($, stalled.kind === 'limit' ? 'continue — the usage limit has reset' : 'continue')
389}
390
391async function clearStall($: EngineInterface, status: JobStatus): Promise<void> {
392 const stalled = await read($, stall)
393 limitTimer?.cancel()
394 limitTimer = null
395 await update($, stall, () => null)
396 if (stalled?.kind === 'limit') {
397 await record($, { kind: 'limit-wait', id: LIMIT_ID, status })
398 }
399}
400
401async function park($: EngineInterface): Promise<string> {
402 const session = await $.session.id()
403 const ledger = await showLedger($, session)
404 const known = await read($, mine)
405 const running = ledger === null
406 ? known
407 : ledger.jobs.filter(job => job.status === 'started' && isFollowed(job.kind) && known.some(one => one.id === job.id)).map(fromLedger)
408 const lines = running.map(job => `- ${describeJob(job)}`)
409 const verdict = running.length === 0
410 ? 'Safe to restart: no background job of this session is running.'
411 : `Not safe to restart: ${running.length === 1 ? '1 job' : `${running.length} jobs`} would die (lazarus offers them for resume afterwards).`
412 const parts = [verdict, ...lines]
413 if (ledger === null) {
414 parts.push('(The ledger was unreadable; this lists only what this process launched.)')
415 }
416 await record($, { kind: 'park', id: `park-${await $.clock.now()}`, status: 'completed', description: verdict })
417 return parts.join('\n')
418}
419
420async function showLedger($: EngineInterface, session: string): Promise<Ledger | null> {
421 const out = await vybava($, ['fleet', 'ledger', 'show', '--session', session, '--json'])
422 if (out === null) {
423 return null
424 }
425 try {
426 const envelope = JSON.parse(out) as Envelope<Ledger>
427 if (envelope.ok && envelope.data !== undefined) {
428 return envelope.data
429 }
430 $.ui.log(`fleet ledger show answered not ok: ${envelope.diagnostics.map(d => d.code).join(', ')}`, { to: 'debug' })
431 } catch (err) {
432 $.ui.log(`fleet ledger show printed no envelope: ${String(err)}`, { to: 'debug' })
433 }
434 return null
435}
436
437async function record($: EngineInterface, event: LedgerEvent): Promise<void> {
438 const session = await $.session.id()
439 const out = await vybava($, ['fleet', 'ledger', 'record', '--session', session], JSON.stringify(event))
440 if (out === null && !ledgerWarned) {
441 ledgerWarned = true
442 $.ui.toast('the job ledger cannot be written; crash recovery is off for this session (see the debug log)')
443 }
444}
445
446// vybava runs one fleet verb; a failure is logged to the debug log and
447// answered as null, never thrown into the tool call it rides on.
448async function vybava($: EngineInterface, argv: readonly string[], stdin?: string): Promise<string | null> {
449 try {
450 const ran = await $.process.run(['vybava', ...argv], stdin === undefined ? { timeoutMs: RUN_TIMEOUT_MS } : { stdin, timeoutMs: RUN_TIMEOUT_MS })
451 if (ran.exitCode === 0) {
452 return ran.stdout
453 }
454 $.ui.log(`vybava ${argv.slice(0, 3).join(' ')} exited ${ran.exitCode}: ${firstLine(ran.stderr === '' ? ran.stdout : ran.stderr)}`, { to: 'debug' })
455 } catch (err) {
456 $.ui.log(`vybava ${argv.slice(0, 3).join(' ')} failed: ${String(err)}`, { to: 'debug' })
457 }
458 return null
459}
460
461function resumePrompt(jobs: readonly LazarusJob[]): string {
462 const lines = jobs.map(job => {
463 if (job.kind === 'workflow' && job.runId !== undefined && job.scriptPath !== undefined) {
464 return `- workflow "${job.description ?? job.id}": resume it with Workflow({ scriptPath: ${JSON.stringify(job.scriptPath)}, resumeFromRunId: ${JSON.stringify(job.runId)} })`
465 }
466 return `- ${describeJob(job)}: died; start it again only if it is still needed`
467 })
468 return [
469 `${summary(jobs)} this session started died with it (a crash, a restart or a kill). The person pressed resume.`,
470 ...lines,
471 'Resume the workflows; decide on the rest from where the work stands.',
472 ].join('\n')
473}
474
475function describeJob(job: LazarusJob): string {
476 const what = `${WORDS[job.kind][0]} "${job.description ?? job.id}"`
477 return job.runId === undefined ? `${what} (${job.id})` : `${what} (run ${job.runId})`
478}
479
480function summary(jobs: readonly LazarusJob[]): string {
481 return FOLLOWED.map(kind => {
482 const count = jobs.filter(job => job.kind === kind).length
483 return count === 0 ? null : `${count} ${WORDS[kind][count === 1 ? 0 : 1]}`
484 })
485 .filter((part): part is string => part !== null)
486 .join(' + ')
487}
488
489function fromLedger(job: LedgerJob): LazarusJob {
490 const started = Date.parse(job.startedAt)
491 return {
492 kind: job.kind as LazarusKind,
493 id: job.id,
494 runId: job.runId,
495 scriptPath: job.scriptPath,
496 description: job.description,
497 startedAt: Number.isFinite(started) ? started : 0,
498 }
499}
500
501function isFollowed(kind: string): boolean {
502 return (FOLLOWED as readonly string[]).includes(kind)
503}
504
505function toStatus(word: string): JobStatus | null {
506 switch (word.trim()) {
507 case 'completed':
508 return 'completed'
509 case 'failed':
510 return 'failed'
511 case 'killed':
512 return 'killed'
513 case 'stopped':
514 return 'stopped'
515 default:
516 return null
517 }
518}
519
520// jitter spreads the sessions one limit stalled: a stable 0–60 s per session.
521function jitter(session: string): number {
522 let hash = 0
523 for (const char of session) {
524 hash = (hash * 31 + char.charCodeAt(0)) >>> 0
525 }
526 return hash % JITTER_MAX_MS
527}
528
529function clock(ms: number): string {
530 const at = new Date(ms)
531 return `${String(at.getHours()).padStart(2, '0')}:${String(at.getMinutes()).padStart(2, '0')}`
532}
533
534function firstLine(text: string): string {
535 return (text.trim().split('\n')[0] ?? '').slice(0, 200)
536}
537types/index.d.ts 43 lines1// lazarus's state contract: the session values its hooks keep in $.state.
2
3/** The background work lazarus follows; the ledger's job kinds less its own. */
4export type LazarusKind = 'workflow' | 'shell' | 'monitor' | 'agent'
5
6/** A background job as the ledger keys it (kind + id). */
7export type LazarusJob = {
8 kind: LazarusKind
9 id: string
10 /** A workflow's run, what resumeFromRunId takes. */
11 runId?: string
12 scriptPath?: string
13 description?: string
14 /** When it started, in $.clock.now() milliseconds. */
15 startedAt: number
16}
17
18/**
19 * What stopped the session: a usage limit, which continues by itself at
20 * `until` when the reset is known, or an auth failure, which waits for a key.
21 */
22export type LazarusStall =
23 | { kind: 'limit'; at: number; until: number | null }
24 | { kind: 'auth'; at: number; error: string }
25
26declare module 'claude-code' {
27 interface PluginState {
28 lazarus: {
29 /** Jobs this process launched that have not ended. */
30 mine: LazarusJob[]
31 /** Task ids the model stopped itself (TaskStop): their end is no death. */
32 stoppedByModel: string[]
33 /** Jobs that died with a previous process, or were killed while this one ran. */
34 dead: LazarusJob[]
35 /** When lazarus found them dead, in milliseconds. */
36 diedAt: number
37 stall: LazarusStall | null
38 /** The last main-loop turn's start; a limit wake that sees a newer one stands down. */
39 lastTurnAt: number
40 }
41 }
42}
43