Skip to content

t1k-config-core.json

registryVersion: 1 · priority: 10

Core layer — always available. No engine-specific context required.

FlagValue
_agentsMdMigrationFrameNoteON by default (decision-record D2, Plane T1K-46). Gates agents-md-migration-nudge.cjs, the SessionStart frame that INSTRUCTS a consumer session (project or global-only install) to start the t1k-agents-md-migrateskill without asking (AGENTS.md real, CLAUDE.md =@AGENTS.md); trueis that standing approval. The hook never writes the tree; the skill does the work through a reviewed PR. Setfalsein a PROJECT config to decline for that project (the skill writes this on a decline), or in the global config to decline everywhere. Env kill switch:T1K_SKIP_AGENTS_MD_MIGRATION=1.
_autonomousPlanningNoteON by default. /t1k:planand thet1k-planneragent do not stop at design forks: each becomes a Conservative-Pick recorded in the plan README's## Deferred decisionssection, surfaced once at the end. The rule itself is SSOT'd at skills/t1k-cook/references/yolo-mode.md; the planning doctrine is skills/t1k-plan/references/autonomous-mode.md. Per-run opt-out is—interactiveon the invocation; setting this keyfalse restores ask-at-each-fork for every run in this install. In autonomous mode this flag is ALSO the standing confirmation Plane's Stage 1 needs to auto-CREATE a work item when no existing one matches — a work item only, never a project, a module, or an estimate (rules/plane-workitem-workflow.md). Irreversible/destructive decisions (deletions, repo/package creation, publishing, force-push, secrets, consumer permission/config changes) are still asked and batched at the end, with the affected phase marked BLOCKED.
_autoPruneMergedWorktreesNoteON by default (owner instruction 2026-10-01). hooks/worktree-auto-prune.cjs (SessionStart + Stop, throttled 30 min per repo, detached) removes every linked worktree AND its local branch whose PR is MERGED/CLOSED-with-merged-twin and whose work is provably on base, via scripts/lib/pr-worktree-cleanup.cjs sweep() (refuses: dirty/untracked, tip not within the merged PR head or not byte-equal on base, cwd or live session/process inside, locked, open PR, main worktree, home branch). Also removes a DETACHED worktree only when clean, idle 30 min, unoccupied, unlocked and HEAD is an ancestor of origin/<base>. A sweep skipped on a held sweep-lock does not start the 30 min throttle. Also deletes loose local branches whose upstream is gone AND whose merged PR head SHA equals the local tip. Reports ONE line [t1k:worktree-prune] removed N, skipped M (CODE n, ...); per-worktree reasons persist in the state file lastResult.keeps. This flag IS the standing cleanup instruction for a merged PR (rules/worktree-session-pinning.md); set false to opt out. Env kill switch T1K_SKIP_WORKTREE_PRUNE=1 always wins.
_autoPushKitSyncNoteDecided by Tu, 2026-10-08 — flipped false->true (reverses core#1555 for kit-sync files). Kit sync now commits the update's own files DIRECTLY on the current branch, protected or not, and pushes that commit to the branch's existing upstream. There is no t1k/kit-sync-* landing branch and no PR (the core#3006 channel was deleted; kit-sync-cleanup.cjsremoves the ones already created). The push is never forced, never pulls or rebases, never creates a remote branch, and only runs when the sync commit is the ONLY unpushed commit on the branch; a rejected or impossible push (branch protection that needs a PR, non-fast-forward, offline) keeps the local commit and stays silent, so it rides out with the user's next normal push. Set false in a user-owned fragment to keep the commit local.autoCommitKitSync remains the master kill switch.
_autoRoutingGapRemediationNoteON by default (2026-09-27, revised in PR #3130 review). Does NOT change deny/enforceAgentRouting behaviour — a generic spawn stays allowed (or denied, per that separate flag) either way; a DENIED spawn is remediated too, since a deny still means the model's first choice was generic. Every generic spawn generic-agent-detector.cjs already detects (subagent_type fork/general-purpose/claude/omitted, at the SAME fingerprint-dedup + 3/session rate-limit gate the pre-existing skill-bug queue uses, plus a second call site inside the fork-routable/lane-worker-routable shape nudges — see docs/marker-namespaces.md [t1k:routing-gap-remediate]) emits an ADVISORY frame (hookSpecificOutput.additionalContext, or permissionDecisionReason on the deny path — plain exit-0 stderr is never visible to the model) instructing the session to spawn t1k-routing-gap-remediator in the background. That agent runs .claude/scripts/t1k-routing-gap-facts.cjs, classifies the gap (missing-install / existing-agent-not-matched / missing-agent-or-skill / legit-transcript-fork / unknown-on-read-failure), and lands the fix. The skill-bug queue row is ALWAYS written when this fires (never suppressed) — only the 7-day dedup BANK is deferred until the remediator's own writeback confirms the fix landed, so an unremediated gap stays eligible for the pre-existing auto-issue path instead of being silently lost. Set false to go back to the pre-existing skill-bug-only behaviour.
_autoUpdateNoteFORCED as of 2026-09-17. features.autoUpdate(andautoUpdateMajor, and midSessionUpdateHours) are listed in FORCED_FEATURE_FLAGS / FORCED_FEATURE_NUMBERS in hooks/telemetry-utils.cjs, which readFeatureFlag()/readFeatureNumber() consult BEFORE any config scan. These keys are deliberately NOT HONOURED — setting falseor0anywhere (project or global scope) changes nothing. The keys are kept rather than deleted so existing installs do not trip stale-key warnings. Reason: a machine that silently stops updating runs a hook set the kit already fixed, and the resulting failure presents as a kit bug against code that no longer exists while the machine reports healthy — the cost is paid by whoever debugs it, not by whoever disabled it. Removing an entry from the forced table restores consumer control; do not add or remove keys there without maintainer sign-off. Since the fleet config policy (docs/fleet-policy.md) shipped, this compiled table is the NO-POLICY fallback: anenforce entry in fleet-flags.json configPolicy outranks it, and it applies whenever no policy has been fetched.
_claudeCodeAutoUpdatesNoteFORCED as of 2026-09-29, same policy as _autoUpdateNote but for Claude Code itself: check-cli-updates.cjs (SessionStart) sets top-level autoUpdatesin ~/.claude.json (or $CLAUDE_CONFIG_DIR/.claude.json) back to true whenever it readsfalse, and prints one advisory line. There is NO kit feature flag for this: a file value can be flipped by any tool or UI with no record of intent, which is exactly the silent stop FORCED_FEATURE_FLAGS exists for, so it is forced rather than offered as a doctor fix. The ONE honoured opt-out is Claude Code's own env switch DISABLE_AUTOUPDATER (a deliberate operator statement in a shell profile or managed settings): the hook reports it once a day and leaves the file alone. Edit contract, atomic write and race guard: hooks/lib/claude-code-autoupdates.cjs.
_enforceAgentRoutingNoteON by default. Originally (#949) gated ONLY a strong-confidence suggestAgent() match. Owner decision 2026-09-28 (generic-spawn-audit-11.md — an advisory frame reached all 11 real generic spawns audited and only 2 sessions acted on it) ADDED two more deny shapes behind this SAME flag: a write-shaped GENERIC brief (commit/push/PR/worktree/.claude edit, 2+ signals) denies toward t1k-lane-worker; a brief carrying a lane <value>`` directive plus the word "review" denies toward t1k-review-<lane> (security/correctness/tests/contracts — tests+contracts ties to t1k-review-tests) or t1k-code-reviewer for an unnamed lane. Both new shapes still require the resolved specialist to be an installed agent file and honour [t1k:route-override] (see docs/marker-namespaces.md [t1k:agent-routing-block]). Because this flag already defaults true here, the new deny shapes are live for every consumer on their next t1k modules updatewith no additional opt-in — set false to fall back to the pre-existing advisory-only nudges for all three shapes. A FIFTH shape (quota-reduction P02): everysubagent_type: forkspawn denies toward a fresh specialist (the suggestAgent() pick, else t1k-investigator / t1k-lane-worker) unless ONE line of the brief carries[t1k:route-override]ANDneeds-transcript: — a bare override does not lift it, and it is NOT deny-once (an identical re-spawn denies again); this same flag turns it off along with the other shapes.
_fleetPolicyNoteAllow-listed keys (features.autoUpdate, features.autoUpdateMajor, midSessionUpdateHours, modelRouter.enabled) can be set fleet-wide by the maintainer through configPolicy in fleet-flags.json on theonekit-core main, each as enforce (policy wins, user is told) or default (local value wins). Values set here are overridden by an enforce entry. T1K_FLEET_POLICY_OVERRIDE=1 skips fleet-policy enforcement locally and is reported in the config snapshot, but the kit's compiled defaults still apply to features.autoUpdate, features.autoUpdateMajor and midSessionUpdateHours (the override cannot change those three); it does change modelRouter.enabled. See docs/fleet-policy.md.
_forkBriefFrameNote`ON by default (P04, plans/260926-rules-to-skills/phase-04-core-orchestration-frames.md, Q8). Gates check-ambiguous-fork-invocation.cjs’s PreToolUse:Task
_herdrLazygitPaneNoteON by default. A herdr worktree session (t1k-worktree session) splits its agent pane and starts lazygit in the new pane, as the kitty path does. Set false to opt out; env kill switch T1K_SKIP_HERDR_LAZYGIT=1 always wins. Read by scripts/herdr-lazygit.cjs.
_mergeUpTreeNoteON by default from this release (plans/260927-tree-default-on-routing-quality P03), rolled forward via the CLI's three-way config merge. Gates the merge-up recursive work-tree flow (skills/t1k-team/references/merge-up-tree.md; the planner writes each tree's root node as tree/0-root.md, P19): root closes design, children/grandchildren refine-never-reopen, leaves merge locally into their parent, one PR at the end. Trigger, thresholds and parent-owned generated-file exemptions live in the top-level mergeUpTreeblock below, not here — this key is only the on/off switch. Per-project opt-out: setfalsein that project's own.claude/t1k-config-core.json(works on legacy-metadata projects too, since P18). Code default and absence both stayfalse(D2) — a consumer who has not updated to a fragment carrying this key, or whose config is unreadable, gets the same behaviour asfalse.
_midSessionUpdateHoursNoteFORCED to 6 as of 2026-09-17 — see _autoUpdateNote. This throttle is forced to 6, NOT to 0: the consumer decision was to ignore the throttle, not to remove it. At 0 the Stop (mid-session) arm would re-run the full update check on every turn end. The 6h window below is therefore the in-force value regardless of what any fragment declares.
_openPrAdvisoryNoteON by default (the user's open-PR follow-through gap, 2026-10-16). hooks/open-pr-advisory.cjs (SessionStart, detached, throttled per repo) lists the open PRs the signed-in gh account(s) authored in THIS repo, read through the t1k-gh.cjs wrapper (never raw gh), as ONE non-blocking [t1k:open-pr-advisory] block; it exists because remaining-work-stop.cjs sees only PRs opened in the current session's transcript. FLAGGED = older than openPrAdvisoryStaleHours (default 24), or not a draft and auto-merge KNOWN to be off (the broker projection of a pulls row omits auto_merge and draft, so they come from gh-status-client getPrState; when that cannot answer the line says auto-merge=unknown and is never flagged for it); the block is silent when nothing is flagged. openPrAdvisoryMaxLines (default 10) caps the PR lines, the rest print as '+K more'. openPrAdvisoryThrottleMinutes (default 30) is the per-repo gap between detached GitHub reads. A SessionStart hook never awaits the network, so each session shows the snapshot the PREVIOUS detached run wrote (labelled with its age); the first session in a repo shows nothing. Fails open on any gh error (the reason is kept in the state file lastError). Set openPrAdvisory false to opt out; env kill switch T1K_SKIP_OPEN_PR_ADVISORY=1 always wins. A 0 or non-numeric number falls back to its default.
_skillAutoloadBytesPerSkillNotePer-skill byte allowance for doctor check #84 (hooks/doctor-check-84-skill-autoload-bytes.cjs): the advisory skill-list budget is max(25000 B floor, this x model-invocable skills installed), so a multi-kit install is not held to a one-kit figure (core#5555). Read via readFeatureNumber; a non-positive or non-numeric value falls back to 120 with a note in the finding.
_spawnEtaNoteON by default (spawn-ETA phase 02, plans/260927-spawn-eta/). Gates hooks/lib/spawn-eta.cjs :: resolveSpawnEtaConfig()(read viaT1KUtils.readFeatureFlag(claudeDir, ‘spawnEta’, true)). Consumed by hooks/task-description-model-badge.cjs` (PreToolUse Task
_timingNoteON by default (comprehensive-timing phase 02, plans/260929-comprehensive-timing/, decisions D-A/U4). One flag per timing kind: skillTiming(K1 skill runs),turnTiming(K2 turns),toolTiming(K3 per-call local JSONL),agentPhaseTiming(K4 agent phase split),toolTimingUpload(K3 daily aggregates to D1). Local recording needs only the kind's flag; every D1 send additionally needsfeatures.telemetry. The library hooks/lib/timing.cjsreads them viaresolveTimingConfig(); nothing is wired until phases P03-P06 land.
advisorySupervisiontrue
agentFanOuttrue
agentPhaseTimingtrue
agentsMdMigrationFrametrue
agentsMdPointertrue
agentTaskGaptrue
arcAutoJointrue
autoCommitKitSynctrue
autoContributetrue
autoGlobalInstalltrue
autoIssueSubmissiontrue
autoLessonSynctrue
autoLibrarySynctrue
autonomousPlanningtrue
autoPruneMergedWorktreestrue
autoPushKitSynctrue
autoRoutingGapRemediationtrue
autoUpdatetrue
autoUpdateMajortrue
cfOnlytrue
configTelemetrytrue
contextBloatGuardtrue
delegationDrifttrue
delegationFloortrue
doctorSweepGuardtrue
enforceAgentRoutingtrue
executionTracetrue
fleetInventorytrue
foregroundWaitGuardtrue
forkBriefFrametrue
forkDepthGuardtrue
gitPrePushGuardtrue
herdrLazygitPanetrue
hookLoggingtrue
imageAnalysisRoutingtrue
kitScopeSessionGuardtrue
knowledgeNudgetrue
laneResumeGuardtrue
laneStashGuardtrue
laneWaitGuardtrue
laneWipGuardtrue
leanBriefTokenThreshold2000
mainSessionReadStreakNudgetrue
mediaRoleRoutingtrue
mergeUpTreetrue
midSessionUpdateHours6
modelUsageTelemetrytrue
openPrAdvisorytrue
openPrAdvisoryMaxLines10
openPrAdvisoryStaleHours24
openPrAdvisoryThrottleMinutes30
orphanReapertrue
processSignalGuardtrue
projectLinksChecktrue
pythonStdinStubGuardtrue
registryAuthtrue
remainingWorkPendingSweeptrue
remainingWorkStaleInProgresstrue
sessionArchivetrue
skillAutoloadBytesPerSkill120
skillTimingtrue
spawnEtatrue
staleTaskNudgeMinutes120
submoduleHeadGuardtrue
telemetrytrue
toolTimingtrue
toolTimingUploadtrue
turnTimingtrue
updateRunnerTimeoutSec1800
vaultMcptrue
worktreeBehindNudgetrue

Raw JSON (ground truth)
{
"_retiredModelsNote": "Quota-reduction P01d: model ids retired for cost. hooks/retired-model-guard.cjs blocks a prompt ONCE per session when the last assistant reply ran on an EXACT match (claude-opus-5-5 never matches claude-opus-5). Opt-out: features.retiredModelGuard false / T1K_SKIP_RETIRED_MODEL_GUARD=1.",
"_rsiSignalsNote": "Line-anchored regex signals that name a reusable symbol worth contributing to the studio library — DATA declared by engine kits (theonekit-unity, theonekit-cocos), EXECUTED by core via hooks/lib/rsi-signals.cjs. Core ships an empty array on purpose: it owns the mechanism, not the engine-specific patterns. Merge is ADDITIVE BY id across every t1k-config-*.json fragment (highest priority wins per id), NOT the override-by-priority scalars use. Each entry: {id, engine, pattern, class:'library-candidate'}; an optional named `symbol` capture group feeds the candidate row. Validate a fragment with `bun .claude/hooks/lib/rsi-signals.cjs --self-test --fragment <path>`.",
"advisorySupervision": {
"enabled": true,
"killSwitchEnv": "T1K_SKIP_ADVICE",
"level": "forks",
"maxSpawnsPerSession": 5,
"planReview": true,
"stallDetection": true,
"threshold": 3
},
"autoIssueSubmission": {
"dedupeTTLDays": 7,
"dryRunEnv": "T1K_AUTO_ISSUE_DRY_RUN",
"maxPerSession": 5
},
"autoLessonSync": {
"dedupeTTLDays": 7,
"dryRunEnv": "T1K_LESSON_SYNC_DRY_RUN",
"maxPerSession": 5
},
"cfOnly": {
"_note": "cf-only guard data (hooks/cf-only-guard.cjs, rules/cf-only.md). Everything the guard knows about shell shapes is DATA here, never a verb in the hook. retiredBins: binaries that must not be invoked. wrappers: commands that run their argument (valueFlags names the flags that take a separate value, so `timeout -s KILL 10 x` is peeled correctly). runners/packageManagers: launchers that run a package binary (pmRunVerbs: exec dlx x run; pmBinDirect: `bun <bin>` runs the bin; pmInstallVerbs: installing it is also retired; packageFlags: --package/-p name a package to run). interpreters: `bun <file>` whose file is a retired bin. shells + evalCommands: a command STRING is re-parsed (and a shell fed by heredoc/here-string/pipe re-parses its stdin). keywords: shell words stripped before the command word. commandStringFlags: flags whose value is itself a command line. lookupFlags: `command -v <bin>` is a presence probe, not an invocation. A MISSING or EMPTY block is not tolerated: the hook denies a command that mentions a retired bin and tells the model the guard is inactive, and doctor check #97 FAILs. Merge rule: override by priority, whole block; a consumer fragment that declares this key must carry every list.",
"commandStringFlags": {
"env": [
"-S",
"--split-string"
],
"npm": [
"-c",
"--call"
],
"npx": [
"-c",
"--call"
],
"pnpm": [
"-c",
"--shell-mode"
]
},
"escapeEnv": "T1K_ALLOW_WRANGLER",
"evalCommands": [
"eval"
],
"interpreters": [
"bun",
"node",
"nodejs",
"deno",
"tsx"
],
"keywords": [
"if",
"then",
"else",
"elif",
"while",
"until",
"do",
"!",
"{"
],
"lookupFlags": {
"command": [
"-v",
"-V"
]
},
"packageFlags": [
"-p",
"--package"
],
"packageManagers": [
"npm",
"pnpm",
"yarn",
"bun"
],
"pmBinDirect": [
"yarn",
"pnpm",
"bun"
],
"pmInstallVerbs": [
"i",
"install",
"add"
],
"pmRunVerbs": [
"exec",
"dlx",
"x",
"run",
"run-script"
],
"retiredBins": [
"wrangler"
],
"runners": [
"npx",
"bunx",
"pnpx"
],
"shells": [
"bash",
"sh",
"zsh",
"dash",
"ksh",
"fish"
],
"valueFlags": {
"bun": [
"--cwd",
"-c",
"--config",
"--prefix",
"-w",
"--workspace",
"--filter",
"-F",
"--dir",
"-r",
"--require",
"--preload",
"--import",
"--loader",
"--conditions",
"-e",
"--eval",
"-p",
"--print"
],
"deno": [
"--config",
"-c",
"--import-map"
],
"doas": [
"-u",
"-C"
],
"env": [
"-u",
"-C",
"--unset",
"--chdir"
],
"exec": [
"-a"
],
"ionice": [
"-c",
"-n",
"-p",
"-P",
"-u",
"--class",
"--classdata",
"--pid",
"--pgid",
"--uid"
],
"nice": [
"-n",
"--adjustment"
],
"node": [
"-r",
"--require",
"--import",
"--loader",
"--experimental-loader",
"--conditions",
"-C",
"-e",
"--eval",
"-p",
"--print"
],
"nodejs": [
"-r",
"--require",
"--import",
"--loader",
"--experimental-loader",
"--conditions",
"-C",
"-e",
"--eval",
"-p",
"--print"
],
"npm": [
"--prefix",
"-w",
"--workspace",
"--userconfig",
"--cache",
"--registry"
],
"npx": [
"--registry",
"--cache",
"--userconfig"
],
"pnpm": [
"-C",
"--dir",
"--filter",
"-F",
"--filter-prod",
"--config"
],
"stdbuf": [
"-i",
"-o",
"-e",
"--input",
"--output",
"--error"
],
"sudo": [
"-u",
"-g",
"-h",
"-p",
"-C",
"-D",
"-R",
"-T",
"-U",
"-r",
"-t",
"--user",
"--group",
"--host",
"--prompt",
"--close-from",
"--chdir",
"--chroot",
"--command-timeout",
"--other-user",
"--role",
"--type"
],
"time": [
"-f",
"-o",
"--format",
"--output"
],
"timeout": [
"-s",
"-k",
"--signal",
"--kill-after"
],
"xargs": [
"-n",
"-I",
"-L",
"-P",
"-s",
"-d",
"-E",
"-a",
"--max-args",
"--max-procs",
"--max-lines",
"--max-chars",
"--delimiter",
"--arg-file",
"--eof"
],
"yarn": [
"--cwd"
]
},
"wrappers": [
"sudo",
"doas",
"env",
"time",
"command",
"exec",
"nohup",
"nice",
"ionice",
"xargs",
"timeout",
"setsid",
"stdbuf",
"corepack",
"builtin"
]
},
"ciWait": {
"enabled": true,
"failTtlSec": 600,
"hangFloorSec": 1800,
"historyRuns": 20,
"jobSampleRuns": 6,
"liveProbeSec": 300,
"maxPages": 3,
"maxWatchMin": 360,
"minSamples": 5,
"queueFloorSec": 900,
"stuckFactor": 1.5,
"ttlSec": 21600,
"validationMinWaits": 10
},
"cli": {
"npmPackage": "@the1studio/theonekit-cli",
"repo": "The1Studio/theonekit-cli"
},
"cloudReleaseCostGate": {
"cfSubcommandsVerifiedNote": "`cf` (Cloudflare's new CLI, studio default for new Cloudflare work since 2026-09-29) subcommands above were confirmed against `cf <command> --help` at the version in the `cf-pin` marker of the t1k-cloudflare skill's references/beta-caveats-and-fallback.md, unauthenticated and read-only; see the t1k-cloudflare skill's command-mapping reference. r2/kv writes (secrets-store, queues, dns) were not added here: their exact write-subcommand syntax was not verified against --help.",
"cli": [
{
"bin": "wrangler",
"subcommands": [
"deploy",
"d1 execute",
"d1 migrations apply",
"r2 object put",
"r2 multi-object delete",
"kv key put",
"kv bulk put",
"secret put",
"queues create",
"pages deploy"
]
},
{
"bin": "cf",
"subcommands": [
"deploy",
"d1 query",
"d1 migrations apply",
"d1 import",
"d1 raw",
"r2 objects put",
"kv keys put",
"kv bulk put",
"r2 objects delete",
"r2 buckets create",
"kv keys delete",
"d1 create",
"workers secrets update",
"pages deploy",
"queues create"
]
},
{
"bin": "terraform",
"subcommands": [
"apply"
]
},
{
"bin": "gcloud",
"subcommands": [
"deploy",
"functions deploy",
"run deploy",
"app deploy",
"dns record-sets",
"firestore",
"datastore"
]
},
{
"bin": "aws",
"subcommands": [
"s3 cp",
"s3 sync",
"s3 rb",
"dynamodb put-item",
"dynamodb batch-write-item",
"dynamodb update-item",
"lambda create-function",
"lambda update-function-code",
"rds create-db-instance"
]
},
{
"bin": "flyctl",
"subcommands": [
"deploy",
"volumes create"
]
},
{
"bin": "fly",
"subcommands": [
"deploy",
"volumes create"
]
},
{
"bin": "vercel",
"subcommands": [
"deploy",
"env add"
]
},
{
"bin": "firebase",
"subcommands": [
"deploy"
]
},
{
"bin": "netlify",
"subcommands": [
"deploy"
]
},
{
"bin": "serverless",
"subcommands": [
"deploy"
]
},
{
"bin": "sls",
"subcommands": [
"deploy"
]
}
],
"enabled": true,
"enforce": true,
"planCostSection": "## Cost Estimate",
"planMarker": "t1k:cost-accounted",
"planSignalFiles": [
"plan.md",
"phase-*.md",
"*.md"
],
"pricing": {
"d1RowsReadPerM": 0.001,
"d1RowsWrittenPerM": 1,
"note": "Pricing display is advisory; the hook never computes a bill. Rates retrieved 2026-08-29 from the t1k-datastore-cost skill.",
"r2ClassAPerM": 4.5,
"r2ClassBPerM": 0.36
},
"triStateMode": true
},
"context": {
"description": "Core commands run in any context",
"requiredFeatures": [],
"requiredPaths": []
},
"delegationDrift": {
"threshold": 5
},
"delegationFloor": {
"sweepThreshold": 12
},
"delegationFloorGate": {
"enabled": true,
"exemptPrefixes": [
"plans/",
".claude/",
"docs/",
"wiki/"
],
"hardThreshold": 8,
"smallEditChars": 1200
},
"description": "Core layer — always available. No engine-specific context required.",
"discordInbound": {
"_comment": "Discord inbound router (plan 261110; contract docs/discord-inbound.md). Read ONLY through hooks/lib/discord-inbound.cjs readConfig(): defaults are mirrored there and pinned equal by discord-inbound.test.cjs. Kill switches: env T1K_SKIP_DISCORD_INBOUND=1, enabled=false, or the per-user `disabled` marker. host is read from this install file only (never a project or user fragment), https only, on studioAuth.allowedHosts. wake stays none until the phase 0 spike record (docs/discord-inbound-spikes.md) says otherwise; replies.source stays none while that record's replies-route line says not-deployed (flip it to service when the F3 route ships); readTools is the exact MCP read-tool names the triage guard lets through. stampFreshSec: a prompt stamp older than this is not evidence of the human.",
"contentTtlHours": 48,
"enabled": true,
"host": "https://discord-mentions.the1studio.org",
"idleExitMin": 10,
"idlePollSec": 300,
"linkNudgeDays": 7,
"maxBackoffSec": 900,
"maxContentChars": 1500,
"maxPages": 10,
"overlapSec": 120,
"pendingTtlDays": 7,
"pollSec": 30,
"readTools": [],
"replies": {
"ignoreBots": true,
"maxRepliesPerThreadPerPoll": 5,
"maxWatchedThreads": 30,
"source": "none",
"threadIdleDays": 7
},
"seenTtlDays": 7,
"staleOwnerSec": 600,
"stampFreshSec": 3600,
"toast": true,
"transport": "poll",
"wake": "none"
},
"eSignSendGate": {
"_note": "core#3616. Send-class e-sign tools hooks/pandadoc-send-gate.cjs asks the human about before every call. ADDITIVE across every t1k-config-*.json fragment (project + home): a kit contributes its own entries with no core code change. Entry: tool (exact name) OR toolPattern (regex), optional inputPattern (regex over the JSON tool input), label. Registration matcher covers mcp__pandadoc__documents_send, documents_status_change and mcp__docusign__.*; an unlisted tool flows through. Kit-declared tools outside those namespaces also need a registration matcher.",
"tools": [
{
"label": "PandaDoc",
"tool": "mcp__pandadoc__documents_send"
},
{
"inputPattern": "\\bsent\\b|document\\.sent",
"label": "PandaDoc",
"tool": "mcp__pandadoc__documents_status_change"
},
{
"label": "DocuSign",
"toolPattern": "^mcp__docusign__.*(send|void|resend|correct)"
}
]
},
"evidenceExit": {
"_comment": "Kill switch and audit rate for evidence exits (plan 261022-model-adaptive-strictness, contract.md section 7). auditPerMille is an INTEGER per-mille (100 = 10%), never a float. lib/audit-pick.cjs disableEvidenceExits() sets enabled:false; a person restores it after review.",
"auditPerMille": 100,
"enabled": true
},
"fanOut": {
"_comment": "Routed-hop fan-out controller (plans/261005-routed-parallel-lanes P05a). Read by scripts/lib/t1k-slots.cjs; every number is a PROPOSAL pending README U2 (perHopMB is the brief's 335 MB figure, the P05 smoke re-measures; perHeavyMB 2048 is a proposal for one full test suite at 4 workers). enabled:false restores unlimited fan-out (rollback). See docs/fanout-controller.md.",
"backoffCapSec": 120,
"backoffInitialSec": 8,
"breakerCooldownSec": 300,
"enabled": true,
"machineMax": 12,
"maxAgeSec": 1800,
"maxHoldSec": 14400,
"perHeavyMB": 2048,
"perHopMB": 335,
"perSessionMax": 8,
"pressureWindowSec": 120,
"psiSomeAvg10Max": 20,
"reserveMB": 4096,
"swapUsedPctMax": 70,
"unknownSwapMachineMax": 4,
"warmupSec": 30
},
"features": {
"_agentsMdMigrationFrameNote": "ON by default (decision-record D2, Plane T1K-46). Gates `agents-md-migration-nudge.cjs`, the SessionStart frame that INSTRUCTS a consumer session (project or global-only install) to start the `t1k-agents-md-migrate` skill without asking (AGENTS.md real, CLAUDE.md = `@AGENTS.md`); `true` is that standing approval. The hook never writes the tree; the skill does the work through a reviewed PR. Set `false` in a PROJECT config to decline for that project (the skill writes this on a decline), or in the global config to decline everywhere. Env kill switch: `T1K_SKIP_AGENTS_MD_MIGRATION=1`.",
"_autoPruneMergedWorktreesNote": "ON by default (owner instruction 2026-10-01). hooks/worktree-auto-prune.cjs (SessionStart + Stop, throttled 30 min per repo, detached) removes every linked worktree AND its local branch whose PR is MERGED/CLOSED-with-merged-twin and whose work is provably on base, via scripts/lib/pr-worktree-cleanup.cjs sweep() (refuses: dirty/untracked, tip not within the merged PR head or not byte-equal on base, cwd or live session/process inside, locked, open PR, main worktree, home branch). Also removes a DETACHED worktree only when clean, idle 30 min, unoccupied, unlocked and HEAD is an ancestor of origin/<base>. A sweep skipped on a held sweep-lock does not start the 30 min throttle. Also deletes loose local branches whose upstream is gone AND whose merged PR head SHA equals the local tip. Reports ONE line [t1k:worktree-prune] removed N, skipped M (CODE n, ...); per-worktree reasons persist in the state file lastResult.keeps. This flag IS the standing cleanup instruction for a merged PR (rules/worktree-session-pinning.md); set false to opt out. Env kill switch T1K_SKIP_WORKTREE_PRUNE=1 always wins.",
"_autoPushKitSyncNote": "Decided by Tu, 2026-10-08 — flipped false->true (reverses core#1555 for kit-sync files). Kit sync now commits the update's own files DIRECTLY on the current branch, protected or not, and pushes that commit to the branch's existing upstream. There is no t1k/kit-sync-* landing branch and no PR (the core#3006 channel was deleted; `kit-sync-cleanup.cjs` removes the ones already created). The push is never forced, never pulls or rebases, never creates a remote branch, and only runs when the sync commit is the ONLY unpushed commit on the branch; a rejected or impossible push (branch protection that needs a PR, non-fast-forward, offline) keeps the local commit and stays silent, so it rides out with the user's next normal push. Set false in a user-owned fragment to keep the commit local. `autoCommitKitSync` remains the master kill switch.",
"_autoRoutingGapRemediationNote": "ON by default (2026-09-27, revised in PR #3130 review). Does NOT change deny/enforceAgentRouting behaviour — a generic spawn stays allowed (or denied, per that separate flag) either way; a DENIED spawn is remediated too, since a deny still means the model's first choice was generic. Every generic spawn generic-agent-detector.cjs already detects (subagent_type fork/general-purpose/claude/omitted, at the SAME fingerprint-dedup + 3/session rate-limit gate the pre-existing skill-bug queue uses, plus a second call site inside the fork-routable/lane-worker-routable shape nudges — see docs/marker-namespaces.md [t1k:routing-gap-remediate]) emits an ADVISORY frame (hookSpecificOutput.additionalContext, or permissionDecisionReason on the deny path — plain exit-0 stderr is never visible to the model) instructing the session to spawn t1k-routing-gap-remediator in the background. That agent runs .claude/scripts/t1k-routing-gap-facts.cjs, classifies the gap (missing-install / existing-agent-not-matched / missing-agent-or-skill / legit-transcript-fork / unknown-on-read-failure), and lands the fix. The skill-bug queue row is ALWAYS written when this fires (never suppressed) — only the 7-day dedup BANK is deferred until the remediator's own writeback confirms the fix landed, so an unremediated gap stays eligible for the pre-existing auto-issue path instead of being silently lost. Set false to go back to the pre-existing skill-bug-only behaviour.",
"_autoUpdateNote": "FORCED as of 2026-09-17. `features.autoUpdate` (and `autoUpdateMajor`, and `midSessionUpdateHours`) are listed in FORCED_FEATURE_FLAGS / FORCED_FEATURE_NUMBERS in hooks/telemetry-utils.cjs, which readFeatureFlag()/readFeatureNumber() consult BEFORE any config scan. These keys are deliberately NOT HONOURED — setting `false` or `0` anywhere (project or global scope) changes nothing. The keys are kept rather than deleted so existing installs do not trip stale-key warnings. Reason: a machine that silently stops updating runs a hook set the kit already fixed, and the resulting failure presents as a kit bug against code that no longer exists while the machine reports healthy — the cost is paid by whoever debugs it, not by whoever disabled it. Removing an entry from the forced table restores consumer control; do not add or remove keys there without maintainer sign-off. Since the fleet config policy (docs/fleet-policy.md) shipped, this compiled table is the NO-POLICY fallback: an `enforce` entry in fleet-flags.json configPolicy outranks it, and it applies whenever no policy has been fetched.",
"_autonomousPlanningNote": "ON by default. `/t1k:plan` and the `t1k-planner` agent do not stop at design forks: each becomes a Conservative-Pick recorded in the plan README's `## Deferred decisions` section, surfaced once at the end. The rule itself is SSOT'd at skills/t1k-cook/references/yolo-mode.md; the planning doctrine is skills/t1k-plan/references/autonomous-mode.md. Per-run opt-out is `--interactive` on the invocation; setting this key `false` restores ask-at-each-fork for every run in this install. In autonomous mode this flag is ALSO the standing confirmation Plane's Stage 1 needs to auto-CREATE a work item when no existing one matches — a work item only, never a project, a module, or an estimate (rules/plane-workitem-workflow.md). Irreversible/destructive decisions (deletions, repo/package creation, publishing, force-push, secrets, consumer permission/config changes) are still asked and batched at the end, with the affected phase marked BLOCKED.",
"_claudeCodeAutoUpdatesNote": "FORCED as of 2026-09-29, same policy as _autoUpdateNote but for Claude Code itself: check-cli-updates.cjs (SessionStart) sets top-level `autoUpdates` in ~/.claude.json (or $CLAUDE_CONFIG_DIR/.claude.json) back to true whenever it reads `false`, and prints one advisory line. There is NO kit feature flag for this: a file value can be flipped by any tool or UI with no record of intent, which is exactly the silent stop FORCED_FEATURE_FLAGS exists for, so it is forced rather than offered as a doctor fix. The ONE honoured opt-out is Claude Code's own env switch DISABLE_AUTOUPDATER (a deliberate operator statement in a shell profile or managed settings): the hook reports it once a day and leaves the file alone. Edit contract, atomic write and race guard: hooks/lib/claude-code-autoupdates.cjs.",
"_enforceAgentRoutingNote": "ON by default. Originally (#949) gated ONLY a strong-confidence suggestAgent() match. Owner decision 2026-09-28 (generic-spawn-audit-11.md — an advisory frame reached all 11 real generic spawns audited and only 2 sessions acted on it) ADDED two more deny shapes behind this SAME flag: a write-shaped GENERIC brief (commit/push/PR/worktree/.claude edit, 2+ signals) denies toward t1k-lane-worker; a brief carrying a `lane `<value>`` directive plus the word \"review\" denies toward t1k-review-<lane> (security/correctness/tests/contracts — tests+contracts ties to t1k-review-tests) or t1k-code-reviewer for an unnamed lane. Both new shapes still require the resolved specialist to be an installed agent file and honour [t1k:route-override] (see docs/marker-namespaces.md [t1k:agent-routing-block]). Because this flag already defaults true here, the new deny shapes are live for every consumer on their next `t1k modules update` with no additional opt-in — set false to fall back to the pre-existing advisory-only nudges for all three shapes. A FIFTH shape (quota-reduction P02): every `subagent_type: fork` spawn denies toward a fresh specialist (the suggestAgent() pick, else t1k-investigator / t1k-lane-worker) unless ONE line of the brief carries `[t1k:route-override]` AND `needs-transcript: <reason>` — a bare override does not lift it, and it is NOT deny-once (an identical re-spawn denies again); this same flag turns it off along with the other shapes.",
"_fleetPolicyNote": "Allow-listed keys (features.autoUpdate, features.autoUpdateMajor, midSessionUpdateHours, modelRouter.enabled) can be set fleet-wide by the maintainer through configPolicy in fleet-flags.json on theonekit-core main, each as enforce (policy wins, user is told) or default (local value wins). Values set here are overridden by an enforce entry. T1K_FLEET_POLICY_OVERRIDE=1 skips fleet-policy enforcement locally and is reported in the config snapshot, but the kit's compiled defaults still apply to features.autoUpdate, features.autoUpdateMajor and midSessionUpdateHours (the override cannot change those three); it does change modelRouter.enabled. See docs/fleet-policy.md.",
"_forkBriefFrameNote": "ON by default (P04, plans/260926-rules-to-skills/phase-04-core-orchestration-frames.md, Q8). Gates check-ambiguous-fork-invocation.cjs's PreToolUse:Task|Agent branch, which replaces the deleted always-loaded rules fork-context-brief.md and lean-brief-pointer-not-payload.md with a warn-only spawn-time frame ([t1k:fork-brief-reminder] / [t1k:lean-brief]) — never a block. `leanBriefTokenThreshold` is the token count (estimateTokens, CHARS_PER_TOKEN=2.8) past which a spawn prompt is flagged as carrying payload instead of a pointer. Kill switch `T1K_SKIP_FORK_BRIEF_FRAME=1` always wins regardless of this flag.",
"_herdrLazygitPaneNote": "ON by default. A herdr worktree session (t1k-worktree `session`) splits its agent pane and starts lazygit in the new pane, as the kitty path does. Set false to opt out; env kill switch T1K_SKIP_HERDR_LAZYGIT=1 always wins. Read by scripts/herdr-lazygit.cjs.",
"_mergeUpTreeNote": "ON by default from this release (plans/260927-tree-default-on-routing-quality P03), rolled forward via the CLI's three-way config merge. Gates the merge-up recursive work-tree flow (`skills/t1k-team/references/merge-up-tree.md`; the planner writes each tree's root node as `tree/0-root.md`, P19): root closes design, children/grandchildren refine-never-reopen, leaves merge locally into their parent, one PR at the end. Trigger, thresholds and parent-owned generated-file exemptions live in the top-level `mergeUpTree` block below, not here — this key is only the on/off switch. Per-project opt-out: set `false` in that project's own `.claude/t1k-config-core.json` (works on legacy-metadata projects too, since P18). Code default and absence both stay `false` (D2) — a consumer who has not updated to a fragment carrying this key, or whose config is unreadable, gets the same behaviour as `false`.",
"_midSessionUpdateHoursNote": "FORCED to 6 as of 2026-09-17 — see _autoUpdateNote. This throttle is forced to 6, NOT to 0: the consumer decision was to ignore the throttle, not to remove it. At 0 the Stop (mid-session) arm would re-run the full update check on every turn end. The 6h window below is therefore the in-force value regardless of what any fragment declares.",
"_openPrAdvisoryNote": "ON by default (the user's open-PR follow-through gap, 2026-10-16). hooks/open-pr-advisory.cjs (SessionStart, detached, throttled per repo) lists the open PRs the signed-in gh account(s) authored in THIS repo, read through the t1k-gh.cjs wrapper (never raw gh), as ONE non-blocking [t1k:open-pr-advisory] block; it exists because remaining-work-stop.cjs sees only PRs opened in the current session's transcript. FLAGGED = older than openPrAdvisoryStaleHours (default 24), or not a draft and auto-merge KNOWN to be off (the broker projection of a pulls row omits auto_merge and draft, so they come from gh-status-client getPrState; when that cannot answer the line says auto-merge=unknown and is never flagged for it); the block is silent when nothing is flagged. openPrAdvisoryMaxLines (default 10) caps the PR lines, the rest print as '+K more'. openPrAdvisoryThrottleMinutes (default 30) is the per-repo gap between detached GitHub reads. A SessionStart hook never awaits the network, so each session shows the snapshot the PREVIOUS detached run wrote (labelled with its age); the first session in a repo shows nothing. Fails open on any gh error (the reason is kept in the state file lastError). Set openPrAdvisory false to opt out; env kill switch T1K_SKIP_OPEN_PR_ADVISORY=1 always wins. A 0 or non-numeric number falls back to its default.",
"_skillAutoloadBytesPerSkillNote": "Per-skill byte allowance for doctor check #84 (hooks/doctor-check-84-skill-autoload-bytes.cjs): the advisory skill-list budget is max(25000 B floor, this x model-invocable skills installed), so a multi-kit install is not held to a one-kit figure (core#5555). Read via readFeatureNumber; a non-positive or non-numeric value falls back to 120 with a note in the finding.",
"_spawnEtaNote": "ON by default (spawn-ETA phase 02, `plans/260927-spawn-eta/`). Gates `hooks/lib/spawn-eta.cjs :: resolveSpawnEtaConfig()` (read via `T1KUtils.readFeatureFlag(claudeDir, 'spawnEta', true)`). Consumed by `hooks/task-description-model-badge.cjs` (PreToolUse Task|Agent: `~p50 (<=p80)` description suffix, `[t1k:spawn-eta]` frame, one `estimate` history row per spawn) and the CLI (`scripts/t1k-spawn-eta.cjs`). Setting it false removes the suffix, frame and row; the model badge is unaffected. The `spawnEta` block below (historyPath/minN/quantiles/sizeBuckets/maxHistoryRows/taskShapeTags) is read ONLY through that same resolver -- never hand-roll a second fragment scan.",
"_timingNote": "ON by default (comprehensive-timing phase 02, `plans/260929-comprehensive-timing/`, decisions D-A/U4). One flag per timing kind: `skillTiming` (K1 skill runs), `turnTiming` (K2 turns), `toolTiming` (K3 per-call local JSONL), `agentPhaseTiming` (K4 agent phase split), `toolTimingUpload` (K3 daily aggregates to D1). Local recording needs only the kind's flag; every D1 send additionally needs `features.telemetry`. The library `hooks/lib/timing.cjs` reads them via `resolveTimingConfig()`; nothing is wired until phases P03-P06 land.",
"advisorySupervision": true,
"agentFanOut": true,
"agentPhaseTiming": true,
"agentTaskGap": true,
"agentsMdMigrationFrame": true,
"agentsMdPointer": true,
"arcAutoJoin": true,
"autoCommitKitSync": true,
"autoContribute": true,
"autoGlobalInstall": true,
"autoIssueSubmission": true,
"autoLessonSync": true,
"autoLibrarySync": true,
"autoPruneMergedWorktrees": true,
"autoPushKitSync": true,
"autoRoutingGapRemediation": true,
"autoUpdate": true,
"autoUpdateMajor": true,
"autonomousPlanning": true,
"cfOnly": true,
"configTelemetry": true,
"contextBloatGuard": true,
"delegationDrift": true,
"delegationFloor": true,
"doctorSweepGuard": true,
"enforceAgentRouting": true,
"executionTrace": true,
"fleetInventory": true,
"foregroundWaitGuard": true,
"forkBriefFrame": true,
"forkDepthGuard": true,
"gitPrePushGuard": true,
"herdrLazygitPane": true,
"hookLogging": true,
"imageAnalysisRouting": true,
"kitScopeSessionGuard": true,
"knowledgeNudge": true,
"laneResumeGuard": true,
"laneStashGuard": true,
"laneWaitGuard": true,
"laneWipGuard": true,
"leanBriefTokenThreshold": 2000,
"mainSessionReadStreakNudge": true,
"mediaRoleRouting": true,
"mergeUpTree": true,
"midSessionUpdateHours": 6,
"modelUsageTelemetry": true,
"openPrAdvisory": true,
"openPrAdvisoryMaxLines": 10,
"openPrAdvisoryStaleHours": 24,
"openPrAdvisoryThrottleMinutes": 30,
"orphanReaper": true,
"processSignalGuard": true,
"projectLinksCheck": true,
"pythonStdinStubGuard": true,
"registryAuth": true,
"remainingWorkPendingSweep": true,
"remainingWorkStaleInProgress": true,
"sessionArchive": true,
"skillAutoloadBytesPerSkill": 120,
"skillTiming": true,
"spawnEta": true,
"staleTaskNudgeMinutes": 120,
"submoduleHeadGuard": true,
"telemetry": true,
"toolTiming": true,
"toolTimingUpload": true,
"turnTiming": true,
"updateRunnerTimeoutSec": 1800,
"vaultMcp": true,
"worktreeBehindNudge": true
},
"foregroundWaitGuard": {
"_comment": "PreToolUse:Bash HARD BLOCK for EVERY caller (main session, sub-agents, routed hops): a FOREGROUND call whose sleep/poll shape can hold the turn open for maxBlockSec seconds or more (`sleep 600`, `for i in $(seq 29); do ..; sleep 20; done`, an unbounded `while|until .. sleep`, `while kill -0 $PID`, `tail --pid`). Evidence: 2026-10 Unity postmortems, WarShip 94.2 h of polling-shaped Bash, PirateEndlessPuncher 317 sleeps = 6.1 h, EpicIdleJurney 294 poll-waits = 1071 min, SteelBornWarShipIdle 163 loops = 5.4 h. The deny names the fix: start the work with run_in_background and wait with the Monitor tool (an until-loop on a producer-written sentinel or log line), or call the project wait helper. NOT denied: a wait under maxBlockSec, run_in_background calls, `timeout N cmd`, sleeps that are quoted data or heredoc bodies, `while read` loops, the lane-worker agents (laneWaitGuard owns them at its own limit). maxBlockSec is the threshold in seconds (default 60; the call is denied AT or ABOVE it; project overrides global; the Bash tool timeout caps an unbounded loop). Kill switch: features.foregroundWaitGuard=false. Audited waiver for a call that truly must block: `T1K_FG_WAIT_OVERRIDE=1 <command>` (leading assignment) or the same variable in the session env; one stderr line plus a telemetry row, never the command text. See hooks/foreground-wait-guard.cjs, hooks/lib/foreground-wait.cjs, hooks/lib/lane-wait-scan.cjs.",
"_env": "T1K_FG_WAIT_OVERRIDE",
"_hooks": [
"foreground-wait-guard.cjs"
],
"maxBlockSec": 60
},
"forkDepth": {
"budget": 3
},
"gateDecision": {
"_comment": "Data for hooks/lib/gate-decision.cjs decideGate() (plan 261022-model-adaptive-strictness, contract.md section 3; docs/gate-decision-contract.md). `gates` is the ONLY id list: an id absent here never relaxes. Per gate: carrier (line|marker), flag = the existing switch the hook already honours (decideGate never reads it), lexiconVeto = a destructiveTerms hit in the final paragraph vetoes. rejectPatterns are case-insensitive regex sources; an uncompilable one fails closed.",
"destructiveTerms": [
"delete",
"drop",
"force-push",
"force push",
"reset --hard",
"rm -rf",
"token",
"password",
"secret",
"credential",
"publish",
"release",
"deploy",
"merge",
"permission",
"--admin",
"truncate",
"wipe",
"overwrite",
"revoke",
"irreversible"
],
"gates": {
"ask-before-stop": {
"carrier": "marker",
"env": "T1K_SKIP_ASK_BEFORE_STOP",
"flag": "features.askBeforeStopGuard",
"lexiconVeto": true
},
"brainstorm": {
"carrier": "line",
"env": null,
"flag": null,
"lexiconVeto": false
},
"evidence": {
"carrier": "line",
"env": null,
"flag": null,
"lexiconVeto": false
},
"exact-requirements": {
"carrier": "line",
"env": null,
"flag": null,
"lexiconVeto": false
},
"forecast": {
"carrier": "line",
"env": null,
"flag": null,
"lexiconVeto": false
},
"git-route": {
"carrier": "line",
"env": null,
"flag": null,
"lexiconVeto": false
},
"plan": {
"carrier": "line",
"env": null,
"flag": null,
"lexiconVeto": false
},
"remaining-work-plan": {
"carrier": "marker",
"env": "T1K_SKIP_REMAINING_WORK",
"flag": "features.remainingWorkReport",
"lexiconVeto": false
},
"remaining-work-prose": {
"carrier": "marker",
"env": "T1K_SKIP_REMAINING_WORK",
"flag": "features.remainingWorkReport",
"lexiconVeto": false
},
"scout": {
"carrier": "line",
"env": null,
"flag": null,
"lexiconVeto": false
},
"specialist-fit": {
"carrier": "line",
"env": "T1K_DISABLE_GENERIC_AGENT_DETECTOR",
"flag": "features.detectGenericAgents",
"lexiconVeto": false
},
"tree-decision": {
"carrier": "line",
"env": "T1K_SKIP_TREE_DECISION",
"flag": "features.mergeUpTree",
"lexiconVeto": false
}
},
"minChars": 12,
"minWords": 3,
"rejectPatterns": [
"<[^>]*>",
"\\.\\.\\.",
"your reason",
"no parentheses",
"<gate>"
],
"rejectReasons": [
"n/a",
"na",
"none",
"nothing",
"not applicable",
"no reason",
"skip",
"skipped",
"skip it",
"ok",
"okay",
"fine",
"tbd",
"todo",
"because",
"obvious",
"trivial",
"simple",
"unneeded",
"not needed",
"no need"
]
},
"genericAgentDetection": {
"_broadSpecialistCodeOnlySignalsNote": "core#5787. Whole-phrase brief signals (description + prompt, case-insensitive) that a lane spawned on one of `agents` needs no engine Editor. The first row with a matching signal whose `suggested` agent is installed names the specialist in one [t1k:broad-agent-advisory] frame; a stronger keyword or review pick wins, and the spawn always proceeds. Every spawn also records a codeOnly flag for doctor #112. Highest-priority fragment wins whole, so an engine kit owns its table by restating this key.",
"_note": "core#3348. Which spawned agent types generic-agent-detector.cjs and t1k-routing-gap-facts.cjs treat as generic, resolved by hooks/lib/generic-agent-types.cjs (highest priority wins per key; no in-code default, a malformed or missing value fails loud). genericTypes: denied when a specialist fits (under enforceAgentRouting) AND routing-gap-remediated. remediateOnlyTypes: routing-gap-remediated only, never denied (owner decision 2026-09-29: remediate, never deny). A type may not appear in both. broadSpecialistTypes: named agents that are broad in practice; they never take the generic path (no deny, queue row or remediation frame) and get one [t1k:broad-agent-advisory] when a narrower installed specialist strongly matches the brief. broadSpecialistExemptMarkers: per-agent brief phrases (case-insensitive) that make the broad spawn deliberate, e.g. the hard-coder's 'already failed once'. Arrays do not merge across fragments: the highest-priority fragment declaring a key wins whole, so a kit or project adds a name by restating the full list at a higher priority.",
"broadSpecialistCodeOnlySignals": {
"agents": [
"t1k-unity-base-developer",
"t1k-hard-coder"
],
"rows": [
{
"signals": [
"csv"
],
"suggested": "t1k-unity-base-data-author"
},
{
"signals": [
"test-only",
"test only"
],
"suggested": "t1k-unity-base-test-writer"
},
{
"signals": [
"blind",
"never launch unity",
"no editor",
"rename",
"reword",
"rebase",
"regenerate",
"relock"
],
"suggested": "t1k-unity-base-feature-implementer"
}
]
},
"broadSpecialistExemptMarkers": {
"t1k-hard-coder": [
"already failed",
"failed once",
"failed twice",
"second attempt",
"third attempt",
"previous attempt failed",
"previous fix failed",
"prior fix failed",
"fix did not hold"
]
},
"broadSpecialistSpawnThreshold": {
"t1k-hard-coder": 3
},
"broadSpecialistTypes": [
"t1k-unity-base-developer",
"t1k-hard-coder",
"t1k-fullstack-developer"
],
"genericTypes": [
"general-purpose",
"claude",
"fork"
],
"remediateOnlyTypes": [
"t1k-lane-worker"
]
},
"ghBroker": {
"_comment": "GitHub event broker client (plan 261009-gh-event-broker, phase 9). `url` is the broker Worker base URL. Phase 34 (group flips g3 + g2, then g1 + g4) sets it; without a url, scripts/lib/gh-status-client.cjs skips the Worker (`reason: broker-unconfigured`) and every PR/check read behaves exactly as before (GraphQL + the shared local cache). The Bearer token is sent only to the PINNED host https://gh-broker.the1studio.org (http only to the literal loopback addresses 127.0.0.1 and [::1] for tests, never localhost); no env var or config can widen that, any other url is rejected (reason broker-url-invalid) and no token is read or sent. Redirects are refused. T1K_GH_BROKER_URL overrides it for one process; T1K_GH_BROKER_DISABLE=1 turns the client off. Phase 23 (covered-read proxy client, scripts/lib/gh-read-client.cjs): `proxyGroups` lists the groups (g1 releases/tags, g2 issues/PR lists/search, g3 runs/Actions, g4 contents/branches/protection) whose REST reads are tried against the Worker `/v1/gh` proxy first; phase 34 adds groups one by one after a recorded preview smoke and the user's confirmation: g3, g2, g1 and g4 are on (g4 is the coarse client switch for the broker scopes g4-enrol, g4-content and g4admin; they flip together). `localTtlSec` is the local single-flight cache TTL for a covered read, `proxyMaxPages` bounds a paginated read.",
"cacheMaxBytes": 268435456,
"cacheTtlHours": 24,
"localTtlSec": 300,
"proxyGroups": [
"g3",
"g2",
"g1",
"g4"
],
"proxyMaxPages": 20,
"url": "https://gh-broker.the1studio.org"
},
"gitCleanup": {
"_clonesNote": "Absolute paths of OTHER clones worktree-cleanup-facts.cjs sweeps alongside the repo it runs in (same as passing --clone per path). Needed because sync-back and cross-kit work leave worktrees in sibling kit clones that a cwd-scoped sweep never sees. UNIONED across every t1k-config-*.json in the project .claude/ and the user config dir (a list of machine paths is additive, not override-by-priority); set it in a machine-local fragment, never in a shipped one. A relative or non-string entry is reported as a warning and skipped. Each clone is fetched (--prune) first unless --no-fetch; a fetch failure is a warning, never fatal.",
"_comment": "Standing authorisation + kill switch for `t1k:git cleanup` ACTING on the facts worktree-cleanup-facts.cjs emits. ON by default: the per-item safety conditions (merged into the primary branch · no REAL uncommitted work · no process cwd inside · idle past minIdleMin · lock absent, dead-pid-held, or a current-session lock on an agent the orchestrator named finished) are enforced by the script and are NOT tunable here — autoApply only removes the per-run ask, never a safety condition. `false` is the opt-out: the sweep is then preview-only and the caller decides per item. When ON, the SessionStart hook worktree-awareness-guard.cjs (and the Stop hook stranded-work-dispatch.cjs, for items classed landed) emits a [t1k:git-cleanup-dispatch] frame (retireable worktrees >= 1 OR any local branch with a gone upstream — a cheap `git for-each-ref` trigger, never the 15 s facts script) ordering the session to spawn t1k-git-manager in the background without asking; the flag IS that standing approval. Read via readConfigFragments (lib/scope-enforcement-config.cjs), override-by-priority. minIdleMin is the one threshold that IS tunable: raise it on a machine where agents sit idle mid-round, and the caller passes it through as --min-idle-min. Harness artifacts (telemetry logs, .skip-commit-gate) are NOT listed here — the script's own DEFAULT_IGNORE is their single home, because one list serving both the dirt split and the idle walk is the only way those two can never disagree.",
"autoApply": true,
"buildServers": {
"_comment": "Lingering BUILD SERVERS the worktree pruner does not count as occupancy (lib/build-server-occupancy.cjs). A Unity batch run leaves a Roslyn compile server (VBCSCompiler, parent pid 1) alive 10-15+ min after it ends; it caches under <worktree>/.../Library/Bee/ and is not a session. A process with its cwd inside a worktree is ignored ONLY when ALL hold: its command line matches a cmdlinePatterns regex, its cwd is inside a GIT-IGNORED directory whose name is in cacheDirs, and EVERY file it holds open under the worktree is inside one too. Unity itself and AssetImportWorker are deliberately NOT patterns: they mean a live Editor and still block. Unreadable /proc, a non-Linux host or no git check-ignore answer all BLOCK. Override-by-priority per key; an invalid value is a warning and the list stays empty (nothing ignored). An engine kit adds its own build servers here.",
"cacheDirs": [
"Library",
"Temp",
"obj",
"Logs"
],
"cmdlinePatterns": [
"VBCSCompiler",
"MSBuild\\.dll.*-nodemode",
"rzc\\.dll.*\\bserver\\b",
"\\bdotnet\\b.*\\bbuild-server\\b"
]
},
"clones": [],
"minIdleMin": 30
},
"issueClaim": {
"adminMerge": {
"bypassReview": true,
"enabled": false,
"requireCiGreen": true
},
"bodyMarkerPrefix": "t1k-claim:",
"enforcementMode": "hard-block",
"inScopeRepos": [
"The1Studio/theonekit-*",
"The1Studio/t1k-*",
"The1Studio/unity-mcp",
"The1Studio/cocos-mcp-server"
],
"labelName": "t1k:claim",
"stalenessDays": 14
},
"jevRouting": {
"actFloor": {
"minLowerBound": 0.9,
"minN": 15,
"wilsonZ": 1.2816
},
"actThreshold": 0.9,
"breakerConsecutiveFailures": 2,
"breakerCooldownMs": 600000,
"hardTimeoutMs": 3500,
"instructions": "Which specialist agent is the best fit to carry out this sub-agent brief? Pick by the role each agent owns.",
"maxCriterionChars": 400,
"model": "typesafe/jev",
"note": "Jev agent routing (plans/260929-jev-agent-routing), read by hooks/lib/specialist-suggester.cjs; the highest-priority fragment declaring the block wins, a missing or malformed block skips Jev and routing falls back to legacy config-driven behavior (B3 = exactly today's detector, including the keyword deny under enforceAgentRouting). There is deliberately NO endpoint key and NO enabled key: the endpoint and host allowlist are code constants in hooks/lib/jev-client.cjs (a config fragment must never steer the gh token), and the kill switch is the T1K_SKIP_JEV_ROUTING=1 env var. instructions is the benchmarked text; change it only with a re-benchmark. actThreshold plus actFloor gate B1 (deny-capable): prob >= actThreshold AND the agent's Wilson lower bound (z=wilsonZ) of k/n in .claude/data/jev-agent-precision.json >= minLowerBound with n >= minN and a matching descriptionSha1. The breaker skips Jev for the session on 401/403/no-token, and for breakerCooldownMs after breakerConsecutiveFailures consecutive timeouts/429/529/5xx/transport errors. Act floor caveat: the shipped actFloor values (wilsonZ 1.2816, minLowerBound 0.9, minN 15) were NOT validated against a per-agent (k, n) distribution. At these values a small sample must be perfect or nearly so (15/15 passes, 14/15 and 29/30 fail, 39/40 passes, 47/50 fails, 48/50 passes), so many agents may sit below the floor or the floor may select on sample noise; the calibration follow-up prints the (k, n) histogram beside the chosen floor and picks minN and minLowerBound from it. Scope caveat: a project fragment that declares its own jevRouting block, together with a project .claude/data/jev-agent-precision.json, can make B1 live for that project's agents (a deny only redirects to an agent the project ships; the endpoint stays pinned in code). Until a precision file ships, B1 never fires and the legacy keyword path decides (D-interim).",
"timeoutMs": 2500
},
"kitName": "theonekit-core",
"knowledgeNudge": {
"_readerNote": "Read ONLY through hooks/knowledge-nudge-stop.cjs :: resolveConfig(), highest-priority fragment wins per key. Gated by features.knowledgeNudge (default true); kill switch T1K_SKIP_KNOWLEDGE_NUDGE=1. The Stop frame fires at most once per session, only when the studio-knowledge MCP is registered and the session made at least minEdits mutating tool uses (hooks/lib/subagent-transcript.cjs :: scanTranscript, MUTATING_TOOLS + git commit/push) OR minToolCalls tool uses in total. Contract: docs/marker-namespaces.md [t1k:knowledge-nudge].",
"minEdits": 3,
"minToolCalls": 25
},
"kongmingGate": {
"duplicateJaccard": 0.6,
"duplicateWindowMinutes": 10,
"failedAttemptPatterns": [
"\\b([3-9]|\\d{2,}|three|four|five|six)\\+?\\s+(failed|unsuccessful)\\b",
"\\b(failed|tried)\\s+([3-9]|three|four|five)\\+?\\s+times\\b",
"\\b([3-9]|three|four|five)\\s+(failed\\s+)?(attempts?|tries)\\b"
],
"forkPatterns": [
"\\bfork\\b",
"\\b(vs\\.?|versus)\\b",
"\\btrade-?offs?\\b",
"\\b(option|approach|alternative|design)s?\\s*\\(?[A-C1-3]\\)?(\\s*[:=)]|\\s+(vs|or|and)\\b)",
"\\b(two|three|2|3)\\s+(options|approaches|alternatives|designs)\\b"
],
"userCounselIntentPatterns": [
"\\bd[ei]c[ia]d\\w*|\\bdecisions?\\b",
"\\bcounsel\\w*|\\badvis\\w*|\\bsecond opinion\\b"
],
"userCounselModelPatterns": [
"\\b(kongming|fable|opus)\\b",
"\\b(strongest|smartest|best|top)\\s+model\\b"
]
},
"laneResumeGuard": {
"_comment": "P2 of plans/archive/261003-lane-worker-boot-trim (2026-10-05) - PreToolUse:SendMessage DENY in the MAIN session (hooks/lane-resume-guard.cjs + hooks/lib/lane-resume.cjs): SendMessage to a listed lane that already delivered its report, has been idle > maxIdleSec (its 5-minute prompt cache is cold) and carries > maxContextTokens is denied; the message names the fix (a fresh fix lane via t1k-lane-directive.cjs, brief = PR + failing check + prior head, read `gh pr diff`). Rebuild ~ 3.75 x C vs fresh ~ 3.75 x 40K, so above ~80K the rebuild dominates; 270 s leaves headroom under the 300 s TTL. A warm lane, a small lane, a running lane and a lane stopped mid-task are never denied. Kill switch: features.laneResumeGuard:false. Audited overrides: T1K_LANE_RESUME_OVERRIDE=1 or a `[t1k:lane-resume-override reason: <why>]` message prefix. Independent of features.teammateContextGuard.",
"agents": [
"t1k-lane-worker"
],
"maxContextTokens": 80000,
"maxIdleSec": 270
},
"laneRouting": {
"_readerNote": "`health` is read by scripts/t1k-lane-health.cjs and `limits` by scripts/lib/lane-size.cjs (the t1k-contract.cjs lane-size / lane-shape gate), both through the shared fragment reader (hooks/lib/scope-enforcement-config.cjs :: readConfigFragments); scopes read: home ~/.claude and this kit's .claude only (project-scope fragments are NOT consulted); per-key merge, higher priority wins, equal priority: the kit dir wins. An absent `limits` block means lane-shape does not judge size. Specs: plans/261005-routed-parallel-lanes/phase-07-measurement-and-stop-rule.md (health thresholds proposed, pending U3) and phase-04-lane-decomposition-rules.md (limits). Health contract and re-enable procedure: docs/lane-routing-measurement.md.",
"health": {
"acceptancePassRateMin": 0.8,
"mainSessionRisePctMax": 25,
"minSampleN": 10,
"rerunRateMax": 0.2,
"window": 30
},
"limits": {
"_note": "Pre-spawn size gate for a ROUTED lane, the SSOT for every number: docs cite this block, never restate it. maxAllowedFiles is a PILOT value (plan U1: the router's stale curve puts 3-5 files at 38% failure vs 1-2 at 14%); ratchet it on measured data. maxContractChars / maxBriefChars: briefs barely separate success from failure below the router's own ceiling, so these are ceilings, not targets.",
"forbidClaudeDir": true,
"maxAllowedFiles": 2,
"maxBriefChars": 3000,
"maxContractChars": 2000,
"requireDisjointAllowedFiles": true,
"requireFileLineMap": true
}
},
"laneStashGuard": {
"_comment": "core#3211/#3214/#3187 (2026-09-28) — PreToolUse:Bash HARD BLOCK for the listed agents: `git stash` in any form (pathspec included) except read-only `list`/`show` — the shared stash stack can be stolen or clobbered by a concurrent session/worktree. See hooks/worktree-session-pin-guard.cjs `LANE-STASH BAN`. Kill switch: features.laneStashGuard / T1K_LANE_STASH_OVERRIDE=1 (audited, per-command).",
"agents": [
"t1k-lane-worker"
]
},
"laneWaitGuard": {
"_comment": "plan 261003 P1b (A/B-1 closed the 1h-TTL route; the wait guard is the 5-minute-cache fallback) — PreToolUse:Bash HARD BLOCK for the listed agents: a call whose text can block longer than `maxBlockSec` seconds (`sleep N`, `timeout N ...`, `for .. seq .. sleep`, an unbounded `until|while .. sleep`, `gh run watch`, `gh pr checks --watch`, `t1k-pr-watch`) because the 5-minute prompt cache expires and the next call re-bills the whole context. `maxBlockSec` stays under the 5-minute cache TTL: never above 240. The deny message names run_in_background and Monitor, never a Bash poll loop (core#5557). See hooks/worktree-session-pin-guard.cjs + hooks/lib/lane-wait-scan.cjs. Kill switch: features.laneWaitGuard / T1K_LANE_WAIT_OVERRIDE=1 (audited, per-command). A backgrounded call (run_in_background, or a lone `&`) is never denied.",
"agents": [
"t1k-lane-worker"
],
"maxBlockSec": 240
},
"laneWipGuard": {
"_comment": "#3072 — PostToolUse advisory + PreToolUse block for `agents` (t1k-lane-worker, t1k-hard-coder): after `threshold` tool calls since the last commit in the lane's tree, with a dirty tree, emit [t1k:lane-wip-due]; re-fire every `repeatEvery` calls while dirty; at `threshold + 2*repeatEvery` calls Edit|Write|MultiEdit is DENIED ([t1k:lane-wip-block]) until a commit lands (Bash never blocked). Blocking since 2026-10-01 (fail-fast; see hooks/lane-wip-guard.cjs header). Kill switch: features.laneWipGuard / T1K_SKIP_LANE_WIP_GUARD=1.",
"agents": [
"t1k-lane-worker",
"t1k-hard-coder"
],
"repeatEvery": 20,
"threshold": 40
},
"mainSessionReadStreak": {
"bash": {
"agents": {
"analysis": "t1k-data-verifier",
"probe": "t1k-ci-watcher",
"read": "Explore"
},
"enabled": true,
"threshold": 8
},
"threshold": 8
},
"mcp": {
"optional": [
{
"daemon": {
"args": [
"-y",
"@modelcontextprotocol/server-memory"
],
"command": "bunx",
"env": {
"MEMORY_FILE_PATH": "${T1K_MCP_DIR}/memory.json"
}
},
"installCmd": "claude mcp add memory -- npx -y @modelcontextprotocol/server-memory",
"name": "memory",
"note": "Demoted from required in #1584 — a repo-wide caller audit (agents/skills/rules/hooks, not just tools: grants) found zero kit surfaces that invoke mcp__memory__* today. Kept installable for manual/user-initiated workflows; do not re-promote to required without naming a caller. `shared: true` (2026-09-27) fixes two of the three hazards this note used to list: the daemon-hosted instance is ONE process per machine, so the no-lock concurrent-write race and the per-session cold-cache reset both go away when the shared daemon is reachable; MEMORY_FILE_PATH stays per-machine (still no shared graph across a fleet) either way.",
"purpose": "Cross-session knowledge persistence via knowledge graph",
"shared": true,
"sharedRequiresExistingRegistration": true,
"verifyTool": "mcp__memory__",
"zeroKitCallers": {
"issue": "#1584",
"since": "2026-09-15"
}
},
{
"daemon": {
"args": [
"-y",
"@playwright/mcp@0.0.83",
"--headless",
"--browser",
"chromium",
"--isolated",
"--port",
"8931",
"--host",
"127.0.0.1",
"--allowed-hosts",
"127.0.0.1:8931"
],
"command": "bunx",
"host": "127.0.0.1",
"listenerMatch": "playwright-mcp",
"path": "/mcp",
"port": 8931,
"processMatch": "@playwright/mcp",
"stdioArgs": [
"@playwright/mcp@0.0.83",
"--headless",
"--browser",
"chromium"
],
"stdioCommands": [
"npx"
],
"stdioVariants": [
[
"@playwright/mcp@latest",
"--headless",
"--browser",
"chromium"
],
[
"@playwright/mcp@latest",
"--headless"
],
[
"@playwright/mcp@0.0.83",
"--headless"
]
],
"transport": "http"
},
"installCmd": "claude mcp add playwright -- npx @playwright/mcp@0.0.83 --headless --browser chromium",
"name": "playwright",
"note": "Native-http daemon class (T1K-65 item b). The daemon runs `--isolated`: every client gets its OWN browser context (own cookies, localStorage and tabs) inside ONE Chromium, so sessions cannot read each other's pages, and it also ends the 'Browser is already in use for ~/.cache/ms-playwright-mcp/mcp-chrome-<hash>' failure two sessions with one cwd used to hit. Consequences, stated plainly: (1) `--isolated` keeps the profile in memory, so saved logins and a persistent profile are gone (use a --storage-state workflow or keep your own per-project `playwright` registration with extra args: a registration with any other args is a different server and is never touched); (2) one Chromium serves every session, so a crash or OOM drops all sessions' browsers until the next tool call relaunches it; (3) the endpoint is loopback-only but UNAUTHENTICATED: any local process or local user can open a session and drive a browser from this machine; (4) the server enforces a Host allow-list, so `--allowed-hosts 127.0.0.1:<port>` is passed (`--host 127.0.0.1` alone makes the server accept only `localhost:<port>` while the kit registers the 127.0.0.1 URL; the port appears in both flags and a test pins it); (5) flags are daemon-global (headless, browser, viewport): no per-session headed mode. `daemon.port` 8931 is Playwright's own documented HTTP example port. The package version is pinned (a daemon `@latest` would re-resolve on every restart); installCmd, stdioArgs and daemon.args carry the SAME pin, asserted by a test. `stdioVariants` are the per-session shapes found in existing registrations (`@latest --headless --browser chromium`, `@latest --headless` from bunx, and the pinned `--headless` from the installCmd's npx); `stdioCommands` adds npx as a launcher next to `command` (bunx), so the installCmd's own registration is retired and the daemon is hosted for it. All of them are retired, a copy with any extra arg or env (`--user-data-dir`, `--storage-state`, `--extension`, `--cdp-endpoint`) is a different server and is kept.",
"purpose": "Browser automation for E2E testing and visual regression. One shared HTTP daemon per machine instead of one stdio server (and one Chromium) per Claude session; hosted only when a playwright registration already exists.",
"shared": true,
"sharedRequiresExistingRegistration": true,
"verifyTool": "mcp__playwright__"
},
{
"installCmd": "claude mcp add chrome-devtools -- npx -y chrome-devtools-mcp@latest --no-usage-statistics --isolated",
"name": "chrome-devtools",
"note": "Opt-in; not nudged: `optIn: true` means check-mcp-health never prints an install line for it (about 40 MB x 3 processes per session); add it yourself with the installCmd. Deliberately NOT shared (T1K-65 item b): it is stdio-only and has no per-client isolation, so any shared form makes sessions see, navigate and close each other's tabs and share cookies. `--no-usage-statistics` drops the telemetry watchdog process every session started (about 70 MB); `--isolated` gives each launch a temporary profile so two concurrent sessions stop colliding on the one default profile (`~/.cache/chrome-devtools-mcp/chrome-profile`, one browser at a time). It stays HEADED (no `--headless`): that is the behaviour installs had, and it is not changed silently. `--isolated` drops the saved profile (logins). Existing registrations are never rewritten for you: check-mcp-health prints an info line with the exact remove/add commands.",
"optIn": true,
"purpose": "Browser DevTools for console, network, and performance analysis. Opt-in; not nudged: add it yourself with the installCmd",
"verifyTool": "mcp__chrome-devtools__"
},
{
"installCmd": "claude mcp add -s local sequential-thinking -- npx -y @modelcontextprotocol/server-sequential-thinking",
"name": "sequential-thinking",
"note": "Opt-in; not nudged: `optIn: true` means check-mcp-health never prints an install line for it; add it yourself with the installCmd. Demoted from required (2026-10-07): measured 14-15 calls in 60 days across 29,594 transcripts, all from sub-agents, almost all from the planner. Not shared: a stateless server whose idle gateway process cost about 150 MB. The agent `tools:` grants stay, and the prompts say to use it if connected, else reason inline. If you still see it declared in both `required` and `optional` after updating (a locally edited entry or no baseline keeps the old copy, and `required` wins), run `claude mcp remove sequential-thinking` or delete it from your local `mcp.required`.",
"optIn": true,
"purpose": "Structured step-by-step analysis for problem-solving when stuck",
"servedTools": [
"sequentialthinking"
],
"verifyTool": "mcp__sequential-thinking__"
},
{
"_versionPinComment": "firebase-tools@15.28.2; verified 2026-08-28.",
"installCmd": "~/.claude/skills/t1k-mcp-management/scripts/install-firebase-mcp.sh",
"name": "firebase",
"purpose": "Firebase CLI/console tooling from Claude — projects, Remote Config, Firestore rules, App Distribution. Developer tooling only; NOT the Unity client SDK (see the t1k-unity-mobile-3rd-integrate skill for that). Auth uses the operator's ambient `firebase login` CLI session; no credential is stored by this entry.",
"scope": "local",
"servedTools": [
"firebase_login",
"firebase_logout",
"firebase_get_project",
"firebase_list_apps",
"firebase_list_projects",
"firebase_get_sdk_config",
"firebase_create_project",
"firebase_create_app",
"firebase_create_android_sha",
"firebase_get_environment",
"firebase_update_environment",
"firebase_init",
"firebase_get_security_rules",
"firebase_read_resources",
"firebase_deploy",
"firebase_deploy_status",
"developerknowledge_search_documents",
"developerknowledge_answer_query",
"developerknowledge_get_documents"
],
"verifyTool": "mcp__firebase__"
},
{
"_sharedNote": "gcloud MCP is stateless (auth is the operator's ambient gcloud session), so ONE gateway-hosted copy replaces one stdio copy per live Claude session (T1K-65). Opt-in only (sharedRequiresExistingRegistration): the gateway hosts gcloud, and migration adds a machine-wide user-scope gcloud entry, ONLY for operators who already ran the installer. The daemon.args policy path uses ${HOME}, expanded per user by mcp-shared-daemon.cjs (daemon.args and daemon.env both expand it). Server name stays gcloud, so mcp__gcloud__* tool names and the t1k-gcloud-engineer agent are unchanged.",
"_versionPinComment": "@google-cloud/gcloud-mcp@0.5.3; verified 2026-09-16.",
"appliesWhen": {
"moduleInstalled": "t1k-gcloud"
},
"daemon": {
"args": [
"-y",
"@google-cloud/gcloud-mcp@0.5.3",
"-c",
"${HOME}/.claude/skills/t1k-gcloud/config/gcloud-mcp-policy.json"
],
"command": "bunx"
},
"installCmd": "~/.claude/skills/t1k-gcloud/scripts/install-gcloud-mcp.sh",
"name": "gcloud",
"purpose": "Google Cloud operations from Claude — inventory, describe, logs, IAM review, and Cloud Run / GKE / Cloud SQL work through Google's local gcloud MCP server. Preview software ('may see breaking changes'); auth is the operator's ambient `gcloud auth login` session, so no credential is stored by this entry. T1K layers its own deny policy on top of Google's always-enforced default one.",
"registeredCmdContains": "@google-cloud/gcloud-mcp",
"scope": "local",
"servedTools": [
"run_gcloud_command"
],
"shared": true,
"sharedRequiresExistingRegistration": true,
"verifyTool": "mcp__gcloud__"
},
{
"appliesWhen": {
"moduleInstalled": "t1k-jenkins"
},
"installCmd": "~/.claude/skills/t1k-mcp-management/scripts/install-jenkins.sh",
"name": "jenkins",
"purpose": "Studio Jenkins CI — build jobs, pipeline status, run history",
"requiredEnv": [
"JENKINS_USERNAME",
"JENKINS_API_TOKEN"
],
"scope": "user",
"verifyTool": "mcp__jenkins__"
}
],
"recommended": [
{
"appliesWhen": {
"moduleInstalled": "t1k-ual"
},
"authMode": "studio",
"autoInstall": true,
"installCmd": "~/.claude/skills/t1k-mcp-management/scripts/install-hosted-mcp.sh --name ual --url https://assets.the1studio.org/mcp --scope user",
"name": "ual",
"purpose": "Studio asset catalogue (search, get, usage, stats) over the UniversalAssetLibs corpus. The server ships FOUR tools today — `asset_search`, `asset_get`, `asset_usage`, `asset_stats` (`registerTools` in `packages/ual-web/src/tools.ts`); a fifth, `asset_resolve`, is planned in UAL plan 07 Phase 05 and is NOT built, so it is not listed in `servedTools` until it ships — a doctor check that reconciled a stale entry against the live server would report a false missing tool. DEPLOYED since 2026-10-01 at https://assets.the1studio.org/mcp (served directly by the `ual-web` Worker, which owns `assets.the1studio.org` since 2026-10-04 and forwards non-MCP paths to the AssetHub Worker `assethub-web`). Cloudflare Access is REMOVED from the host (2026-10-19): the Worker verifies the studio assertion itself (audience `universal-asset-libs`, STUDIO_AUTH_ENABLED), so this entry is `authMode: studio` and every user with a `gh auth login` GitHub account connects with no manual /mcp browser login. The anonymous 401 `Bearer realm=\"OAuth\"` is the Worker's own constant, not Access. With `autoInstall: true` the SessionStart worker registers it by itself on a machine that has the t1k-ual module and a `gh auth login` (one background `claude mcp add-json`, once), so no manual installer run is needed. A machine with no GitHub token registers nothing (`studio-no-token`). Rollback per machine: `T1K_MCP_AUTH_MODE_UAL=access`. The old host https://assets.playablelabs.ai/mcp is dead (Cloudflare 404). **It is module-scoped: this entry applies only when the `t1k-ual` module is installed** (`appliesWhen.moduleInstalled`), so a core install that does not want the asset catalogue is never reminded of it — `t1k-ual` is `required: false` and reaches an install only via the engine kits' planned cross-kit dependency or an explicit `t1k modules add t1k-ual`. The studio-bearer `initialize` and `tools/list` (4 tools) were verified 2026-10-19; the entry stays `recommended` until UAL plan-06 Phase 05 makes the catalogue self-updating (today it is a 2026-09-27 Cocos-only snapshot). The query path that needs no MCP at all also remains: `searchAssets`/`usageOf`/`statsOf` imported from `packages/ual-web/src/tools.ts` in the UAL repo, per the t1k-ual skill. Query-only by design: there is no asset_install tool, because a pull writes into a project and stays an explicit CLI verb with a dry run. `installCmd` is the generic hosted installer, which carries the studio `headersHelper` into the registration.",
"requiredEnv": [],
"scope": "user",
"servedTools": [
"asset_search",
"asset_get",
"asset_usage",
"asset_stats"
],
"verifyTool": "asset_stats"
},
{
"authMode": "static",
"installCmd": "~/.claude/skills/t1k-mcp-management/scripts/install-knowledge-retrieval.sh",
"name": "knowledge-retrieval",
"purpose": "Studio knowledge-retrieval MCP — doc_search / doc_list_assemblies / doc_get_member over the Unity/.NET and Cocos corpora. Remote HTTP, bearer-gated at the app layer by MCP_SHARED_SECRET (Cloudflare Access left the host; no OAuth discovery, so the client sends a static bearer via headersHelper kr-mcp-headers.cjs; the secret comes from env MCP_SHARED_SECRET or t1k vault get MCP_SHARED_SECRET)",
"scope": "user",
"servedTools": [
"doc_search",
"doc_list_assemblies",
"doc_get_member",
"doc_list_indexed_repos"
],
"sessionAutoInstall": true,
"verifyTool": "mcp__knowledge-retrieval__doc_search"
},
{
"authMode": "studio",
"autoInstall": true,
"installCmd": "~/.claude/skills/t1k-mcp-management/scripts/install-studio-knowledge.sh",
"name": "studio-knowledge",
"purpose": "Studio shared knowledge service - search_knowledge / get_knowledge / lookup_project / propose_knowledge over studio process, product, client and inventory facts. Studio-bearer only (headersHelper); registered at user scope by the installer, which then runs the doctor. Not autoInstall: the audience is minted by studio-auth, so a machine without it would 401.",
"scope": "user",
"verifyTool": "mcp__studio-knowledge__search_knowledge"
}
],
"required": [
{
"authMode": "gh",
"autoInstall": true,
"installCmd": "~/.claude/skills/t1k-mcp-management/scripts/install-hosted-mcp.sh --name github --url https://api.githubcopilot.com/mcp/ --scope user",
"name": "github",
"purpose": "GitHub issue/PR management for triage, sync-back, issue reporting. Authenticates from your gh auth login session (headersHelper gh-mcp-headers.cjs); no GITHUB_TOKEN env var.",
"scope": "user",
"verifyTool": "mcp__github__"
},
{
"autoInstall": true,
"installCmd": "~/.claude/skills/t1k-mcp-management/scripts/install-hosted-mcp.sh --name context7 --url https://mcp.context7.com/mcp --scope user",
"name": "context7",
"purpose": "Library/framework documentation lookup for planning and implementation",
"scope": "user",
"servedTools": [
"resolve-library-id",
"get-library-docs"
],
"verifyTool": "mcp__context7__"
},
{
"autoInstall": true,
"installCmd": "~/.claude/skills/t1k-mcp-management/scripts/install-discord-mcp.sh",
"name": "discord-mcp",
"purpose": "Discord bridge — post updates, manage threads/channels/messages. User-scope HTTP; Discord OAuth2 broker, no headersHelper, stale studio helpers auto-removed.",
"retiredAuthModes": [
"studio"
],
"scope": "user",
"verifyTool": "mcp__discord-mcp__"
},
{
"appliesWhen": {
"moduleInstalled": "t1k-vault-mcp"
},
"autoInstall": true,
"installCmd": "~/.claude/skills/t1k-mcp-management/scripts/install-vault-mcp.sh --name studio-vault",
"name": "studio-vault",
"purpose": "Studio credential vault over MCP (ADR-0020): typed vault_* tools (list, status, kinds, get with reveal none by default, put from a value_file only, request, request-access, audit) acting as the caller's own GitHub identity. Local stdio server launched from the pinned package in vaultMcp. Scope is USER, on purpose (owner ruling 2026-10-30): the server must reach every project and global-only sessions, so the one process per session it costs is accepted and named by mcpUserScopeAllowed, which silences the user-scope WARN for it alone. Required for every t1k user from the next kit release; the t1k vault CLI stays the fallback and the kill switches are features.vaultMcp, VAULT_MCP_DISABLE=1 and the Worker flags.",
"scope": "user",
"servedTools": [
"vault_list",
"vault_kinds",
"vault_status",
"vault_get",
"vault_put",
"vault_request",
"vault_request_status",
"vault_request_cancel",
"vault_request_access",
"vault_request_access_status",
"vault_request_access_cancel",
"vault_audit"
],
"verifyTool": "mcp__studio-vault__vault_kinds"
}
]
},
"mcpDemand": {
"_note": "Data for .claude/skills/t1k-mcp-management/scripts/mcp-demand.cjs (T1K-65; changes no Claude Code state). disableCandidates (core fragment only; other fragments cannot add names) is the ONLY set of server names the scan may ever report as a per-project disable candidate; neverCandidates (kit-owned shared/gateway servers) always wins and any fragment may extend it; gatewayCommandMarkers flags a stdio registration that merely proxies a shared server; a registration whose transport is http/sse/ws is never one, whatever it is named. minSessions: OBSERVED sessions (transcript head shows an MCP surface was loaded, or it holds an MCP call; sessionMarkers, minus nonInteractiveEntrypoints) a project needs in the window before its zero-call servers count. estimatedMB is the measured per-instance PSS+swap (2026-10-05); servers without an entry use estimatedMBDefault (unverified).",
"disableCandidates": [
"playwright",
"chrome-devtools",
"whatsapp",
"squarespace",
"discord-mcp",
"figma",
"grafana",
"tripo",
"mintegral",
"blender"
],
"estimatedMB": {
"chrome-devtools": 192,
"playwright": 142,
"whatsapp": 47
},
"estimatedMBDefault": 100,
"gatewayCommandMarkers": [
"mcp-remote",
"mcp-proxy"
],
"minSessions": 3,
"neverCandidates": [
"plane",
"context7",
"github",
"knowledge-retrieval",
"jenkins",
"sequential-thinking",
"memory",
"UnityMCP",
"gcloud",
"google-workspace",
"code-graph-unity",
"code-graph-cocos",
"unity-asset-store",
"cocos-creator"
],
"nonInteractiveEntrypoints": [],
"sessionMarkers": [
{
"match": "\"type\":\"mcp_instructions_delta\""
},
{
"match": "\"type\":\"deferred_tools_delta\"",
"requiresMcpName": true
},
{
"match": "\"type\":\"deferred_tools_record\"",
"requiresMcpName": true
}
],
"windowDays": 14
},
"mcpUserScopeAllowed": [
"studio-vault"
],
"mergeUpTree": {
"_comment": "Merge-up tree thresholds (plans/260924-merge-up-tree P01). Trigger: features.mergeUpTree is true AND (planned leaves >= minLeaves OR touched files >= minFiles) — otherwise work takes today's path, zero overhead. `maxDepth` is NOT a separate key here — it IS `forkDepth.budget` (default 3) above; one SSOT, never a second cap to drift out of sync. `parentOwnedFiles` are generated/registry files no node's `Owns` may claim; only the ROOT regenerates them once, after its last child merge, via `t1k-tree.cjs root-regen` — a non-leaf's incidental rewrite of one is exempt from the leaf-merge file-set check rather than failing it. `decisionWords` (P02) is `tree-check`'s `decisionWords` check's word list over a child's goal/requirement/done_when — `\"?\"` is deliberately absent, `t1k-contract.cjs`'s `hasOpenMarker` already covers it. `postMergeOnlyValidators` (P06 smoke item 5) are validators no tree gate at ANY level can ever satisfy for a genuinely new file, because the thing they check only happens AFTER merge to main (validate-origin-injection-coverage's origin-marker comment is injected by CI post-merge) — skipped at every merge-child AND root-regen call, reported as `skipped`/`gate.skipped`/`diffGate.skipped`, never judged locally. This block is shared and sequenced across phases: P04 adds `rootOnlyValidators`; no other phase edits it. Lowered to minLeaves 2 / minFiles 6 by plans/260928-tree-adoption so ordinary multi-part work reaches the tree path.",
"_rootHeavyGateNote": "merge-first-gate-once (plans/archive/261110-merge-first-gate-once). Root tier: the engine's full compile+test (verificationTiers.heavyRun) runs ONCE per tree in root-regen, after the regen gate and before the commit/PR, in the foreground of the root-regen process (no detached process). Lanes never run it. auto (default) = off for a kit-source repo (its gate is the validators + CI); for any other repo, required when an engine declares verificationTiers.heavyRun, else off with the reason recorded (never a silent skip, never a stop for a consumer without an engine). required = a missing heavyRun makes root-regen exit 2 naming the key. off = explicit opt-out. Rule: rules/merge-first-gate-once.md (t1k-parallel).",
"decisionWords": [
"TBD",
"TODO",
"decide",
"choose",
"either",
"or maybe",
"option A",
"option B"
],
"minFiles": 6,
"minLeaves": 2,
"parentOwnedFiles": [
".claude/t1k-modules.json",
".claude/modules/*/module.json",
"docs/marker-namespaces.md"
],
"postMergeOnlyValidators": [
"validate-origin-injection-coverage"
],
"rootHeavyGate": "auto",
"rootOnlyValidators": [
"validate-all-shippable-claimed",
"validate-shipped-artifact-coverage",
"validate-module-file-claims",
"validate-module-rules-claim",
"validate-modules-registry-sync",
"validate-registry-schema",
"validate-modules",
"validate-marker-namespaces"
]
},
"modelStrength": {
"_comment": "Threshold data for hooks/lib/model-strength.cjs (plan 261022-model-adaptive-strictness). relaxFromTier is a tier NAME of TIER_RANK (haiku|sonnet|opus|fable); NOT an on/off switch (per-hook switches stay on features.* and T1K_SKIP_*). A name outside TIER_RANK, or this block missing, means strong:false. firstPartyPrefixes: a raw model id must start with one (or be a bare tier alias), so a gateway id such as acme-opus-clone never reads as strong.",
"firstPartyPrefixes": [
"claude-",
"anthropic.",
"us.anthropic.",
"global.anthropic."
],
"relaxFromTier": "opus"
},
"offscopeFindings": {
"_note": "Off-scope findings offload policy — read by rules/off-scope-finding-offload.md's enforcement surfaces (hooks/offscope-*.cjs), never restated in the rule body. SHIPS LIVE: the CLASS LIST IS THE SWITCH — there is no `enabled` key (removed 2026-09-19: a declared class was always the real gate, and a second boolean could only ever disagree with it). The always-loaded rule stays live and reads no key anywhere; its degraded path carries the finding down the existing queue instead. A MISSING block, or one declaring no `classes`, MUST read as an EMPTY class list and must never throw — that is the whole version-skew contract: nothing is offloadable, so nothing parks and no finding is lost. `dryRun` is the drain's only gate: true describes a drain instead of performing it, and an unresolvable block defaults to it (describe, never perform). classes and neverOffload are DATA (rules/code-conventions.md § 'Data-Driven Over Hardcoded') — the class list and the authorization boundary are enforced from these arrays and from code, so deleting a hardcoded copy breaks nothing and a rule that restates either list is the drift this key exists to prevent. EVERY DECLARED CLASS MUST HAVE A SUPPRESSION SITE — a class declared offloadable that nothing offloads is a promise the system does not keep, and .claude/hooks/__tests__/offscope-class-coverage.test.cjs asserts it by identity. 2026-09-19 resolved the four classes that had none: checkpoint-failure, doc-drift and kit-update-repair were REMOVED (nothing offloads them, and each has a reason a background drain is the wrong actor — docs/off-scope-finding-offload.md § The classes), and rsi GAINED its site (the [t1k:rsi-review] / [t1k:rsi-escalation] frames in lesson-queue-processor.cjs now park instead of printing). Merge rule is override-by-priority, as for every sibling block: core is priority 10, so an engine kit at 90 may override the whole block. Made UNCONDITIONAL 2026-09-19: `enabled` was removed and the shipped `dryRun` flipped to false, so a declared class parks and a drain performs. `dryRun: true` is not an option for a shipped block — with `enabled` gone a dry run re-parks everything it described, so it would produce a permanent once-per-turn receipt naming findings nothing ever drains: a nag, not a pipeline. The LEVERS are unchanged and none of them depend on this file: per session `T1K_SKIP_OFSCOPE=1` empties the class list at the source and restores the pre-pipeline frame byte-for-byte, frame included (core#2241); per install, declare an empty `offscopeFindings.classes` in any fragment, or override the whole block at a higher priority.",
"classes": [
"skill-bug",
"lesson",
"doctor-finding",
"mcp-gap",
"rsi"
],
"dryRun": false,
"maxDrainsPerTurn": 1,
"neverOffload": [
"deliverable-change",
"destructive",
"credentials",
"config-permission",
"consent-required"
]
},
"orphanReaper": {
"dryRun": false,
"minAgeMinutes": 30
},
"planValidators": {
"_comment": "Plan validators (plans/261005-plan-skill-stronger-models, contract D2/D4/D5/D12/D18). The runner skills/t1k-plan/scripts/validate-plan.cjs invokes ONLY the members listed here, never a glob: fixed = .claude/scripts/, skill = skills/t1k-plan/scripts/. notMembers are never invoked by the runner (validate-plan is the runner itself; validate-plan-ci is the env/gh-driven CI step and takes no plan dir). naCodes is the closed list of NA(<code>) a validator may print; the runner recomputes the fact and turns a mismatch into ERROR. selfImprovement.minWords is the word floor for a Loop/Fix/Guard answer and for a 'None, because' reason. evidence.externalForms are accepted as unverified-external path evidence. accept.noopCommands are commands that prove nothing as an Accept line. accept.runners are known test-runner command heads: an Accept whose head is one of them needs no path token of its own when the phase declares a test path in Owns:/Creates: (every other Accept ban still applies). maxBaselineFailShare is the user-approved go/no-go: the CI step is enabled only if the measured false-FAIL share on the baseline plans is <= it. Read through skills/t1k-plan/scripts/lib/plan-config.cjs (global base, project on top); a block missing in both scopes yields the frozen default plus the marker DEFAULT(planValidators missing). No key here is a lane/phase/width ceiling (D11).",
"accept": {
"noopCommands": [
"true",
":",
"exit 0",
"echo"
],
"runners": [
"bun test",
"bun run test",
"bun test",
"pytest",
"go test",
"cargo test",
"bunx jest",
"bunx vitest",
"bunx vitest run",
"bunx vitest --run",
"dotnet test"
]
},
"ax": {
"requiredFrom": "261015"
},
"deps": {
"requiredFrom": "261008"
},
"evidence": {
"externalForms": [
"corpus:",
"repo:"
]
},
"humanSteps": {
"requiredFrom": "261012"
},
"maxBaselineFailShare": 0.25,
"members": {
"fixed": [
"validate-reuse-search",
"validate-plan-self-improvement"
],
"skill": [
"validate-plan-waves",
"validate-plan-lane-sizing",
"validate-plan-contract",
"validate-plan-accept",
"validate-plan-deps",
"validate-plan-ax"
]
},
"naCodes": {
"validate-plan-ax": [
"legacy"
],
"validate-plan-contract": [
"serial"
],
"validate-plan-deps": [
"legacy"
],
"validate-plan-lane-sizing": [
"serial"
],
"validate-reuse-search": [
"no-phases"
]
},
"notMembers": [
"validate-plan",
"validate-plan-ci"
],
"selfImprovement": {
"minWords": 3
}
},
"plannerHardness": {
"_comment": "Planner hardness signal (plans/261005-plan-skill-stronger-models, contract D14/D15). hardAt thresholds ship null = unset: the signal is recorded first and no planner tier changes until a measured threshold is set here (instrument first, user decision Q2). The planner is pinned model: opus (owner decision 2026-11-20: architecture, security-policy and data-layer design judgment). There is deliberately NO effort value here: a subagent spawn carries no per-invocation effort (only frontmatter effort), and hardness is measured from the plan the planner is about to write, so nothing could ever apply it (skills/t1k-plan/references/plan-hardness.md). Read through skills/t1k-plan/scripts/lib/plan-config.cjs.",
"hardAt": {
"lanes": null,
"modules": null,
"newLedgerRows": null,
"phases": null
}
},
"prAutoWatch": {
"claimTtlSec": 120,
"clockSkewSec": 30,
"enabled": true,
"maxConcurrent": 2,
"mergeAuthorization": "armed-only",
"releaseTimeoutMin": 25,
"timeoutMin": 45
},
"priority": 10,
"processSignalGuard": {
"_comment": "PreToolUse:Bash HARD DENY for EVERY caller (main session, sub-agents, routed hops): a kill/pkill/killall whose PID comes from the parent chain (ps -o ppid=, $PPID, PPid), a kill of PID 1, -1 or 0, a kill of an ancestor of this Claude process or of a systemd/herdr/kitty/kwin_wayland/plasmashell/sddm process, a pkill/killall pattern naming a whole session-wide family (systemd, herdr, kitty, plasmashell, kwin, claude), and systemctl exit / loginctl terminate-user|kill-user. Incident 2026-11-03 11:12 +07: an agent killed the systemd --user manager reached through ps ppid, and exit.target stopped the whole desktop session. Kill switch: features.processSignalGuard=false (human-owned config; deliberately NO in-band waiver, since an agent can type any token it can type). See hooks/process-signal-guard.cjs and hooks/lib/process-signal.cjs.",
"_hooks": [
"process-signal-guard.cjs"
]
},
"protectedBranches": {
"globs": [
"main",
"master",
"trunk",
"develop",
"development"
]
},
"pythonStdinStubGuard": {
"_comment": "PreToolUse:Bash HARD BLOCK (Windows only, core#5089): a command that feeds python a script on stdin (`python3 - <<EOF ... || bun ...`) while `python3`/`python` resolves FIRST to the Microsoft Store App Execution Alias stub. The stub blocks on stdin forever, so the `|| bun` fallback never runs and chained edits stay armed on an orphaned shell. The deny names `bun -e` / `bun -` as the fix. No-op on every other platform; fails open on its own error. Audited waiver: a leading T1K_PY_STUB_OVERRIDE=1 assignment (or the same variable in the session env).",
"_env": "T1K_PY_STUB_OVERRIDE",
"_hooks": [
"python-stdin-stub-guard.cjs"
]
},
"registryAuth": {
"_comment": "Registry tokens from the user's GitHub identity (scripts/registry-auth-sync.cjs, SessionStart hook registry-auth-session.cjs, doctor check #99). Data-driven: a registry is touched only if listed here AND enabled. ROLLOUT STATE (owner decision 2026-10-23, docs/registry-rollout-runbook.md): utils-prod and cpm-prod are enabled:true (stages 2 and 3); upm-prod stays enabled:false until stage 4 (Unity kit pointers unreleased); the three *-staging entries stay enabled:false (opt-in through the user fragment). stage test = *-staging hosts, stage prod = prod hosts; tenant = upm|cpm|utils. The default-on CAPABILITY is built (quiet-by-default messaging, hand-made tokens kept, doctor WARN for no-login/no-grant, --probe); flipping an entry is the staged rollout step (plan 261022-registry-module phase 11), pinned by registry-auth-rollout.test.cjs against the runbook Current stage line. Opt in or out per machine with enabled in ~/.claude/t1k-config-user-overrides.json. kind: upmconfig = ~/.upmconfig.toml [npmAuth.\"<host>/\"], npmrc = ~/.npmrc //<host>/:_authToken. Read from the kit install only, never from a project file, because the studio assertion is sent to these hosts; allowedHosts is the kit-owned allow-list every registry host must match (https only, EXACT hostnames: no *.suffix wildcard, so a sibling subdomain is never a registry host; adding a registry host means adding it here too, pinned by registry-auth-fragment-trust.test.cjs and the facts test). Only fragments the install's metadata.json lists as kit-shipped (plus this one) are read; any other t1k-config-*.json beside the engine is ignored. Gate: features.registryAuth; kill switch T1K_SKIP_REGISTRY_AUTH=1. See docs/registry-auto-auth.md.",
"_studioTokenRouteNote": "package-registry#98 (2026-10-28) deleted the studio token mint route on purpose and the registry keeps no machine mint route. retired (the shipped value; any value other than \"present\" fails closed the same way): the kit calls no mint route and reports the state with the token page and the file to store a token by hand. present: the pre-#98 mint path, kept for a registry that restores the route; flip only with the registry owners.",
"allowedHosts": [
"upm.the1studio.org",
"upm-staging.the1studio.org",
"utils.the1studio.org",
"utils-staging.the1studio.org",
"cpm.playablelabs.ai",
"cpm-staging.playablelabs.ai"
],
"registries": [
{
"cutover": "pending",
"enabled": false,
"host": "https://upm-staging.the1studio.org",
"id": "upm",
"kind": "upmconfig",
"stage": "test",
"tenant": "upm"
},
{
"cutover": "pending",
"enabled": false,
"host": "https://cpm-staging.playablelabs.ai",
"id": "cpm",
"kind": "npmrc",
"stage": "test",
"tenant": "cpm"
},
{
"cutover": "pending",
"enabled": false,
"host": "https://utils-staging.the1studio.org",
"id": "utils",
"kind": "npmrc",
"stage": "test",
"tenant": "utils"
},
{
"cutover": "done",
"enabled": true,
"host": "https://upm.the1studio.org",
"id": "upm-prod",
"kind": "upmconfig",
"stage": "prod",
"tenant": "upm"
},
{
"cutover": "done",
"enabled": true,
"host": "https://cpm.playablelabs.ai",
"id": "cpm-prod",
"kind": "npmrc",
"stage": "prod",
"tenant": "cpm"
},
{
"cutover": "done",
"enabled": true,
"host": "https://utils.the1studio.org",
"id": "utils-prod",
"kind": "npmrc",
"stage": "prod",
"tenant": "utils"
}
],
"studioTokenRoute": "retired"
},
"registryVersion": 1,
"repoCreation": {
"allowPersonalOwner": false,
"defaultOwner": "The1Studio"
},
"repos": {
"primary": "The1Studio/theonekit-core"
},
"retiredModels": [
"claude-opus-5"
],
"rsi": {
"_libCandidateNote": "Auto-contribute intake (studio-lib-rsi phase 01). `enabled` gates DETECTION only — whether hooks/lesson-queue-processor.cjs computes lib-candidate signals and emits the [t1k:lib-candidate-review] frame. It is safe ON by default: core ships zero rsiSignals, so nothing matches until an engine kit declares its own. Whether that frame is ACTED ON (auto-spawn t1k-lib-contribute → --auto PR) is a SEPARATE gate, features.autoContribute (default true since 2026-10-01), read by skills/t1k-lib-contribute/references/auto-contribute.md. frameCooldownHours rate-limits the frame per project, same latch as rsi.frameCooldownHours. Read through the same submission-config reader; override-by-priority.",
"_readerNote": "Read by skills/t1k-rsi (review/apply/audit) via the resolved t1k-config-*.json stack; merge rule is override-by-priority, so an engine kit at 90 may override this whole block. autoFix is 'propose'|'apply' — 'propose' emits a diff and writes NOTHING, and is the shipped default on purpose: RSI edits creator artifacts, which are always-loaded prose every consumer pays for. maxApplyPerRun caps creator edits in ONE run (a ceiling, not a target). frameCooldownHours rate-limits the [t1k:rsi-review] frame per project. escalationThreshold is the same-class-same-target count within 30 days that means prose has failed and a mechanism is owed. The frame itself is emitted by hooks/lesson-queue-processor.cjs (P1), which reads this block through the same reader — do not re-spell these defaults at a call site.",
"autoFix": "apply",
"enabled": true,
"enabledNote": "Features-style enable/disable lives here rather than in features.* because RSI is one skill's whole policy surface, not a codepath several hooks branch on. `false` is a supported state: t1k-rsi reports 'rsi disabled' and stops.",
"escalationThreshold": 3,
"frameCooldownHours": 24,
"libCandidate": {
"enabled": true,
"frameCooldownHours": 24
},
"maxApplyPerRun": 3
},
"rsiSignals": [],
"scopeEnforcement": {
"_readerNote": "Read ONLY through hooks/lib/scope-enforcement-config.cjs — readScopeEnforcement(claudeDir). Never hand-roll a second fragment scan, and never re-spell a default at a call site. Merge rule is override-by-priority: core is priority 10, so an engine kit at 90 CAN override this whole block — that is the intended escape hatch. There is deliberately NO bloatUnionTokens key: a union-token threshold fires on 78% of installs with zero duplication, so it was removed rather than re-tuned. features.contextBloatGuard is a features.* flag and is NOT part of this block.",
"allowKits": [],
"autoRemoveKits": [
"core",
"model-router"
],
"bloatDoubleLoadedTokens": 8000,
"enabled": true,
"lockStaleMinutes": 15,
"maxRemovalsPerRun": 2,
"removalCooldownDays": 7,
"unusedLookbackDays": 7
},
"scratch": {
"_comment": "Scratch-dir hygiene (plans/261016-scratch-worktree-hygiene). DOCUMENTS the defaults; the code reads them from DEFAULTS in hooks/lib/scratch-root.cjs, so a key left out here still has its default. root: null = the per-user disk-backed platform default (Linux $XDG_CACHE_HOME or ~/.cache/t1k/scratch, macOS ~/Library/Caches/t1k/scratch, Windows %LOCALAPPDATA%\\t1k\\scratch); env T1K_SCRATCH_ROOT wins over this key; TMPDIR/TEMP/TMP are never read.",
"graceMs": 600000,
"guard": {
"mode": "deny"
},
"harness": {
"retentionDays": 7
},
"pressure": {
"critUsedPct": 95,
"warnFreeBytes": 2147483648,
"warnUsedPct": 85,
"writeRefuseBytes": 1048576
},
"reaper": {
"enabled": true,
"throttleMinutes": 30
},
"relocateHarness": true,
"root": null,
"unownedAgeMs": 86400000
},
"sessionArchive": {
"blobEndpoint": "https://t1k-telemetry.the1studio.org/ingest/session-archive/blob",
"defaultOnNote": "features.sessionArchive ships ON by default as of 2026-08-14, by explicit maintainer decision, ahead of the authz gate this key previously guarded. As of 2026-09-10 the flag is MANDATORY and vestigial: archiving no longer consults it at any scope — a false fragment here or in any consumer's own t1k-config-*.json has zero effect (opt-out withdrawn, explicit user decision: 'we want the archive to be mandatory, no opt-out'). The key is left in place rather than removed so its value is not silently misleading if ever re-read, but the resolver is resolveArchiveEnablement() in lib/session-archive-upload.cjs, not this flag. The hook's code-level fallback stays fail-closed on purpose — a machine whose config is absent or unreadable does NOT upload, and that state is now distinguishable on the coverage heartbeat as config_state: 'config-unreadable'.",
"denyProjectPaths": [
"/contracts",
"/legal",
"/client-work",
"/clients",
".env",
"-contracts",
"-legal",
"-clients"
],
"endpoint": "https://t1k-telemetry.the1studio.org/ingest/session-archive",
"knownOpenRisks": [
"R2 deletion is UNPROVEN (t1k-telemetry-worker#66): `wrangler r2 object delete` reports success unconditionally — it prints 'Delete complete.' for a bucket that does not exist — and one object survived three deletes with a subsequent read returning its full body. The OAuth session carries no r2 scope, so a permission failure would be invisible. Until a verifying delete path lands, treat the retention policy as a stated intention, not a demonstrated capability: we can write, and we have not yet shown we can erase.",
"theonekit-service has no auth middleware and per-user scoping is structurally impossible — 0 of 162 users carry both an email and a GitHub login, so email<->GitHub linkage cannot be resolved. Session archives carry full transcript bodies, so anything that serves them is currently unscoped.",
"The scrubber (hooks/lib/session-scrubber.cjs, shipped in t1k-base@2.47.0) is the ONLY control between a pasted credential and R2. Its /g+lastIndex regression suite is a CI gate and does NOT ship to consumers — __tests__/ is excluded from the release payload. Changes to the scrubber must be gated in CI, because nothing re-verifies it on a consumer machine."
],
"maxBodyBytes": 12582912,
"maxUploadsPerRun": 200,
"r2KeyNote": "sessions/{user}/{session_id}.jsonl.gz — bound server-side from the verified GitHub token; the client-sent value is ignored.",
"statusEndpoint": "https://t1k-telemetry.the1studio.org/ingest/session-archive/status"
},
"spawnEta": {
"_modelAliasesNote": "Compare-time model aliasing, read ONLY through spawn-eta.cjs :: normalizeModel(). alias -> regex source strings (case-insensitive, tested against the model name after the provider prefix is stripped; first alias in key order whose pattern matches wins, no match keeps the stripped name). It makes the runtime name an actual row carries (claude-sonnet-5-5, claude-opus-5-5, claude-haiku-4-5-20251001, claude-fable-5-1, as backfilled from transcripts) compare equal to the short ladder key (sonnet, opus, haiku, fable), and cc-deepseek-v4.1-flash equal to deepseek-v4.1-flash (one model, two write-path spellings). Stored rows are never rewritten. A fragment's aliases merge per alias over the other scope's; an invalid pattern is ignored and reported in warnings.",
"_readerNote": "Read ONLY through hooks/lib/spawn-eta.cjs :: resolveSpawnEtaConfig(). Gated by features.spawnEta (default true) above, not by a key in this block. Cross-scope merge is PER-FIELD (global is the base, project overrides field-by-field) — a project fragment that declares only `minN` still keeps whatever `taskShapeTags`/`sizeBuckets`/etc. global declared; it does NOT replace the whole resolved config, only the fields it names. `minN` is the fallback-ladder threshold (tag -> brief-size -> agent+model -> agent -> global; a level is used only at n >= minN, per plans/260927-spawn-eta/DECISIONS.md). `quantiles` are fractions (0.5/0.8), converted to the 0-100 scale `hooks/lib/stats.cjs :: percentile()` expects. `sizeBuckets` gives the S/M char-length ceilings for briefSizeBucket(); anything above `M` is bucket L — and MERGES PER-KEY (a fragment declaring only `S` keeps the OTHER scope's `M`). `maxHistoryRows` bounds readHistory()'s tail-read by EVIDENCE COUNT (kind!==\"estimate\" rows); `maxScanBytes` (P06 follow-up) additionally bounds the SAME tail-read by bytes read, so a history that is mostly \"estimate\" rows can never force a full-file scan on every call before enough \"actual\" rows accumulate — whichever of maxRows/maxScanBytes/BOF is hit first stops the scan, and readHistory()'s `truncatedByBytes` flag reports when the byte cap (not BOF) was the reason; a caller that sets `maxRows` but OMITS `maxScanBytes` entirely now gets THIS field's value by default (P06 review Must-fix #1) rather than an unbounded scan. `taskShapeTags` is core's OWN shipped keyword->tag map (unlike the empty `rsiSignals`, task shape is engine-neutral) — UNLIKE `sizeBuckets`, this field REPLACES WHOLESALE, never a per-key merge inside the map (a project fragment that wants to drop a core tag must repeat every tag it still wants — a documented asymmetry, not an oversight); override via a higher-priority t1k-config-*.json fragment (override-by-priority, same rule every config block in this kit uses), never edit the map in code. T1K-69 feedback loop: `recency` (halfLifeDays 3, floorWeight 0.1; halfLifeDays 0 disables decay) and `sourceWeights` (source -> multiplier, unknown source weighs 1) weight each evidence row in estimate()'s quantile (rows with no finite ts weigh 1 before the source weight); `calibration` is the per-key bias correction learned from estV>=2 estimate/actual pairs (`minN` pairs per level, `clamp` [lo,hi] ascending positive, `levels` walked in order from agent+model|agent, `p80Mode` own-quantile|same-factor, `enabled:false` switches calibration off alone); `gate` is read by `report --gate` only (biasLow/biasHigh bound the median actual/estimate ratio, hitMin the p80 hit rate, minN the smallest evaluated group, windowDays the actual-ts window). An invalid value in any of these reverts to its default and is reported in warnings, never trusted. `maxHistoryRows` 10000 and `maxScanBytes` 5242880 (5 MiB) cover most of the 7-day window and are the smallest caps within 0.02 of the 12000 rows / 6 MiB accuracy row (contract section 7, results-accuracy.md), sized so the badge hook stays inside its +20 ms p50 budget; findEstimate keeps its own smaller bound. Plan-level ETA (plans/261008-plan-exact-finish-eta, contract.md section 7): `planEta` and `planGate` take FLAT fields only (the resolver warns and ignores a nested or unknown field, so a nested key would be dropped silently). `planEta.humanGateMinutes` (default 10) is the allowance per human gate counted inside the headline finish, `maxParallel` (integer >= 1) the parallel-schedule width, `thinN` the evidence floor, `headlineMethod` p80-chain|rss, `reworkFactor` (>= 1) the learned review/fix multiplier, `overheadStaleDays`/`autoOverheadTimeoutSec` the delivery-overhead cache and the auto-path gh timeout, `forecastPath` the plan-eta --record file, `idleGapCapMinutes` the cap on idle gaps in plan actuals, `autoShow:false` stops only the unrequested print. `approvalAssumedHours` is no longer shipped (plan 261008 phase 7 removed the value and its PLAN-11 pin); it stays a VALID field, and a consumer fragment that still names it without `humanGateMinutes` is read as hours * 60 for one release, from the raw fragments by `spawn-eta-plan.cjs :: legacyGateMinutes` (the resolved config always carries the default `humanGateMinutes`, so the resolved value cannot tell the two apart). `planGate` is read by `report --gate --plans` only (minN is the ONE sample threshold for the live gate, the validated= tail and the ship bar; hitMin/hitMax bound the plan hit rate, biasLow/biasHigh the median ratio, windowDays the actual-ts window).",
"calibration": {
"clamp": [
0.25,
4
],
"enabled": true,
"levels": [
"agent+model",
"agent"
],
"minN": 15,
"p80Mode": "own-quantile"
},
"gate": {
"biasHigh": 1.5,
"biasLow": 0.67,
"hitMin": 0.7,
"minN": 30,
"windowDays": 7
},
"historyPath": "~/.claude/telemetry/spawn-timings.jsonl",
"maxHistoryRows": 10000,
"maxScanBytes": 5242880,
"minN": 5,
"modelAliases": {
"deepseek-v4.1-flash": [
"^cc-deepseek-v4\\.1-flash$"
],
"fable": [
"^claude-fable"
],
"haiku": [
"^claude-haiku"
],
"opus": [
"^claude-opus"
],
"sonnet": [
"^claude-sonnet"
]
},
"planEta": {
"_note": "backtest* are the measured bias of this headline on past plans (plans/261008-plan-exact-finish-eta/results.md: n=10, p80-chain hit 2/10, median actual/headline ratio 1.78; re-run after the phase-id fix, section 8). The headline prints them while it is still unvalidated (live recorded plans < planGate.minN); live evidence replaces them. backtestN 0 hides the note.",
"autoOverheadTimeoutSec": 5,
"autoShow": true,
"backtestHit": 0.2,
"backtestMedianRatio": 1.78,
"backtestN": 10,
"defaultAgent": "t1k-lane-worker",
"forecastPath": "~/.claude/telemetry/plan-eta-forecasts.jsonl",
"headlineMethod": "p80-chain",
"humanGateMinutes": 10,
"idleGapCapMinutes": 120,
"maxParallel": 4,
"overheadCacheTtlHours": 6,
"overheadGhTimeoutSec": 20,
"overheadPrLimit": 300,
"overheadStaleDays": 7,
"reworkFactor": 1,
"thinN": 10
},
"planGate": {
"biasHigh": 1,
"biasLow": 0.55,
"hitMax": 0.95,
"hitMin": 0.75,
"minN": 10,
"windowDays": 60
},
"quantiles": [
0.5,
0.8
],
"recency": {
"floorWeight": 0.1,
"halfLifeDays": 3
},
"sizeBuckets": {
"M": 2000,
"S": 500
},
"sourceWeights": {
"seed-ledger": 0.25,
"seed-router": 0.25
},
"taskShapeTags": {
"deploy": [
"deploy",
"release",
"ship",
"publish"
],
"docs": [
"docs",
"documentation",
"readme",
"wiki"
],
"fix": [
"fix",
"bug",
"debug",
"error",
"crash"
],
"plan": [
"plan",
"design",
"architecture",
"phase"
],
"refactor": [
"refactor",
"cleanup",
"simplify",
"restructure"
],
"research": [
"research",
"investigate",
"explore",
"scout"
],
"review": [
"review",
"code review",
"audit",
"critique"
],
"test": [
"test",
"unit test",
"coverage",
"spec"
]
}
},
"statusChangeCommentGuard": {
"_note": "BLOCKS by default (fail-fast, 2026-10-01) — `[t1k:status-change-needs-comment]` on every bare gh pr/issue close|reopen|ready|lock|unlock|transfer, gh pr merge --disable-auto, gh pr/issue edit --base or a state-carrying label, and a git push branch-delete that auto-closes an open PR. Exit 2; no per-call override — the fix is the comment itself. Optional per-install deferral: a user-owned fragment `blockingFrom` / env T1K_STATUS_CHANGE_COMMENT_GUARD_BLOCKING_FROM=<future yyyy-mm-dd> downgrades to warn until that date, T1K_SKIP_STATUS_CHANGE_COMMENT_GUARD=1 (kill switch).",
"enabled": true
},
"studioAuth": {
"allowedHosts": [
"*.the1studio.org"
],
"authBase": "https://auth-api.the1studio.org"
},
"submoduleHeadGuard": {
"_comment": "Submodule detached-HEAD protection (owner request 2026-10-18: a submodule sitting on a detached HEAD is not proper git practice; prevent the ways it loses work). A submodule detached at a commit that IS on a remote branch is git's NORMAL `git submodule update` result, so hooks/submodule-head-guard.cjs stays SILENT for it; it speaks only for the three shapes that lose work: (a) commits made while detached, (b) uncommitted edits in a detached submodule, (c) a superproject commit staging a submodule pointer to a SHA on no remote-tracking branch. SessionStart emits ONE compact frame naming each at-risk submodule and the exact commands (git -C <sub> switch -c <name>, push the branch to the submodule's own repo, then bump the pointer); PreToolUse:Bash DENIES a `git commit`/`git push` whose target repo is a detached submodule (that is the only way work is lost) and WARNS on the unpublished-pointer commit. It never switches, commits or stages anything itself. Facts come from hooks/lib/submodule-head-facts.cjs (no network). Kill switch: features.submoduleHeadGuard / T1K_SKIP_SUBMODULE_HEAD_GUARD=1.",
"_env": "T1K_SKIP_SUBMODULE_HEAD_GUARD",
"_hooks": [
"submodule-head-guard.cjs"
]
},
"telemetry": {
"cloud": {
"auth": "github-token",
"enabled": true,
"endpoint": "https://t1k-telemetry.the1studio.org/ingest",
"org": "The1Studio"
}
},
"tierDecision": {
"actFloor": {
"minLowerBound": 0.6,
"minN": 5,
"wilsonZ": 1.2816
},
"actThreshold": 0.9,
"cacheTtlMs": 86400000,
"efforts": {
"high": "Multi-step reasoning, cross-file design or debugging.",
"low": "Mechanical, well-specified change or lookup; little reasoning needed.",
"max": "Maximum depth: only for the hardest, highest-stakes reasoning.",
"medium": "Ordinary implementation or analysis with some judgment.",
"xhigh": "Hard problem: ambiguous requirements, deep investigation or architecture."
},
"fanOutSignals": [
"\\b(spawn|spawns|spawning)\\b",
"\\bsub-?agents?\\b",
"\\blanes?\\b",
"\\bfan[- ]?out\\b",
"\\bin parallel\\b",
"\\bdelegate[sd]?\\b",
"\\bsplit (this|it|the)\\b",
"\\b(Agent|Task)\\s*\\("
],
"instructions": {
"effort": "Which reasoning effort does this sub-agent brief need to complete correctly?",
"tier": "Which model tier and route is the smallest one that will still complete this sub-agent brief correctly? Prefer the pinned tier unless the brief clearly needs less or more."
},
"maxCriterionChars": 400,
"note": "Per-spawn model tier / effort / route decision (plans/261001-lane-worker-ai-tier P02), read by hooks/lib/tier-decider.cjs via tier-options.cjs; the highest-priority fragment declaring the block wins, a missing or malformed block disables the decider (reason 'disabled': every spawn keeps its frontmatter pin). INERT until the model-router (P03) calls the lib. Transport, timeouts, the breaker file and the T1K_JEV_ENDPOINT pin are the jevRouting ones (there is deliberately NO endpoint key here); the kill switch is T1K_SKIP_TIER_DECISION=1. The agent's own pin is always offered as option 'pin' and a native option equal to the pin is dropped. Jev's answer is acted on only when strong: prob >= actThreshold AND a passing per-option Wilson floor (actFloor, same math as jevRouting) from .claude/data/tier-precision.json (shipped by P04; absent => every spawn keeps its pin). A proposal ABOVE the pin additionally needs prob >= upgradeThreshold. sampleRate is the fraction of observe-mode spawns that call Jev (the quota is one shared fleet key); steer it fleet-wide through the fleet config policy. routableExcluded lists excludeAgents members the router may route on a strong verdict with a routed-* route; emptied 2026-10-02 (plan 261002 Q4: t1k-lane-worker is native only, and the router no longer rewrites spawns at all), so the routable:false drop below applies to every agent. The decider is an observe-only second opinion since 261002. Routed options are dropped for: excluded or routable:false agents, fan-out briefs on agents holding Agent (fanOutSignals, case-insensitive regexes), unknown or .claude/** owns, gitEgress agents without 'delivery: local-commit', and every foreground spawn. The actFloor (minN 5, minLowerBound 0.6, same z as jevRouting) is the user-chosen starting floor (replan D12, 2026-10-01), deliberately looser than jevRouting.actFloor (minN 15, minLowerBound 0.9), which is unchanged; real-usage telemetry tightens it.",
"options": [
{
"description": "Native Haiku: trivial, fully specified edits, lookups and formatting; no judgment.",
"id": "haiku",
"model": "haiku",
"rank": 1,
"route": "native"
},
{
"description": "Native Sonnet: standard implementation, tests and review with a clear brief.",
"id": "sonnet",
"model": "sonnet",
"rank": 2,
"route": "native"
},
{
"description": "Native Opus: hard debugging, architecture and ambiguous multi-file work.",
"id": "opus",
"model": "opus",
"rank": 3,
"route": "native"
},
{
"description": "Native Fable: advisory-only premium reasoning (never for agents that write files).",
"id": "fable",
"model": "fable",
"rank": 4,
"route": "native"
},
{
"description": "Routed cheap provider via the model-router: bounded, background, well-scoped work that needs no .claude/** edits.",
"id": "routed-cheap",
"model": "routed",
"rank": 1,
"route": "routed"
}
],
"routableExcluded": [],
"sampleRate": 0.1,
"upgradeThreshold": 0.95
},
"timing": {
"_readerNote": "Read ONLY through hooks/lib/timing.cjs :: resolveTimingConfig() (global scope is the base, the project overrides field-by-field; arrays replace wholesale, toolPhase/bashClassPhase merge per key). A malformed bashClasses entry is dropped with one stderr warning, never fatal. bashClasses is ORDERED, the first match wins, default class is bash:other. Hook code carries no maps: extend these from a higher-priority t1k-config-*.json fragment. Decisions: plans/260929-comprehensive-timing/DECISIONS.md (D-C bash classes incl. wait, D-F retention, D-B grace window).",
"bashClassPhase": {
"gh": "ci-pr",
"test": "test",
"wait": "ci-pr"
},
"bashClasses": [
{
"class": "test",
"pattern": "(^|[;&|]\\s*|\\s)(npm|pnpm|yarn|bun)\\s+(run\\s+)?test\\b|\\bjest\\b|\\bvitest\\b|\\bpytest\\b|test-runner\\.cjs|node\\s+--test"
},
{
"class": "wait",
"pattern": "\\bsleep\\s+\\d|gh\\s+run\\s+watch|gh\\s+pr\\s+checks\\b.*--watch|\\buntil\\b.*;\\s*do"
},
{
"class": "gh",
"pattern": "(^|[;&|(]\\s*|\\s)gh\\s"
},
{
"class": "wrangler",
"pattern": "\\bwrangler\\b"
},
{
"class": "git",
"pattern": "(^|[;&|(]\\s*|\\s)git\\s"
},
{
"class": "node",
"pattern": "(^|[;&|(]\\s*|\\s)node\\s"
}
],
"excludeTools": [
"Agent",
"Task",
"Skill"
],
"hooksRetentionDays": 7,
"mcpServerAllowList": [],
"orphanMergeHours": 24,
"phasePriority": [
"ci-pr",
"test",
"edit",
"read",
"other"
],
"retentionDays": 14,
"rollupGraceHours": 26,
"toolPhase": {
"Edit": "edit",
"Glob": "read",
"Grep": "read",
"MultiEdit": "edit",
"NotebookEdit": "edit",
"Read": "read",
"Write": "edit"
}
},
"vaultMcp": {
"_note": "THE single pin for the studio-vault MCP (ADR-0020, plans/261029-vault-mcp phase 9). The installer (install-vault-mcp.sh) and doctor check #110 both read this block; no other file may hard-code the version. version 0.1.2 is the release that adds vault_request_read and the four-label check for new secrets. 0.1.1 (file-only reveal) and 0.1.0 (release run 41028719864, GitHub Packages, 2026-10-30) were published before it. vault-mcp-pin.test.cjs asks the registry (npm view) and goes RED for any version that is not published, so a pin can never name a missing package.",
"minVersion": "0.1.0",
"package": "@the1studio/vault-mcp",
"registryScope": "@the1studio",
"version": "0.1.2"
},
"verificationTiers": {
"_comment": "core#3771 ask C, read by scripts/t1k-verify-tier.cjs (policy: skills/t1k-test/references/verification-tiers.md). heavyRunSkipGlobs: a diff whose changed paths ALL match needs no HEAVY run (the engine batch run / full suite that goes through `t1k gate run`); one path outside them makes it `run`. Anchored globs, `*` stays inside a directory, `**/` is zero or more directories. An engine project narrows or widens this in its own fragment (the higher `priority` wins per key). heavyGateClass is the `t1k gate run --class` for that run. generated: [{ \"glob\": \"**/Replica.asset\", \"regen\": \"<command>\" }] names generated files that are regenerated AFTER the merge and never hand-merged; empty here because core ships none. heavyRun (OPTIONAL, absent here: core ships NO value, its own command is whatever the agent passes after `--`; an engine kit supplies the COMPLETE object in its own fragment, replaced whole, never deep-merged): { \"argv\": [\"<runner>\", \"...\"], \"estSec\": <seconds>, \"summary\": { \"kind\": \"junit|nunit|regex|exit-only\", \"path\": \"<file, for junit/nunit>\", \"regex\": \"<named groups total, failed>\" }, \"rerunArgv\": [\"...\", \"{names}\"], \"hermeticHome\": false } is the command THIS project's merge-first root gate runs in the foreground; validated by hooks/lib/heavy-run-config.cjs, read by scripts/t1k-verify-tier.cjs (`classify --json` -> heavyRun). rerunArgv absent means a red unit cannot be re-run alone and is reported unclassified. hermeticHome true points HOME at a per-run dir (default false: a Unity-style run needs the real HOME).",
"generated": [],
"heavyGateClass": "long",
"heavyRunSkipGlobs": [
"docs/**",
"plans/**",
".claude/**",
".github/**",
"**/*.md",
"LICENSE*"
],
"pathClasses": {
"_comment": "Path classes for `t1k-verify-tier.cjs classify --classes` (plan 261022-model-adaptive-strictness phase 14): which review lanes and sync-back rounds a diff needs, chosen from `git diff --name-only`, never by the model. Anchored globs (same dialect as heavyRunSkipGlobs). First match wins in the order risk, tests-only, docs-only, config; a path matching none is `unknown` and means ALL rounds. risk is seeded from t1k-review/references/adversarial-review.md NEVER-skip list plus hooks, settings, module.json and workflows. laneMap: class -> review lanes (correctness, security, tests, contracts); the domain lane resolves separately. The security lane is never pruned beside a non-docs, non-tests path. Pruning also needs evidenceExit.enabled and loses to the sampled audit (evidenceExit.auditPerMille).",
"config": {
"globs": [
"**/*.json",
"**/*.yml",
"**/*.yaml",
"**/*.toml"
]
},
"docs-only": {
"globs": [
"docs/**",
"plans/**",
"**/*.md",
"**/*.mdx",
"LICENSE*"
]
},
"laneMap": {
"config": [
"correctness",
"security",
"tests",
"contracts"
],
"docs-only": [
"contracts"
],
"risk": [
"correctness",
"security",
"tests",
"contracts"
],
"tests-only": [
"tests",
"correctness"
]
},
"risk": {
"globs": [
"**/auth/**",
"**/middleware/**",
"**/security/**",
"**/crypto/**",
"**/package.json",
"**/package-lock.json",
"**/bun.lock",
"**/bun.lockb",
"**/yarn.lock",
"**/pnpm-lock.yaml",
"**/*.lock",
"**/go.sum",
"**/.env",
"**/.env.*",
"**/*.env",
"**/schema/**",
"**/schema.*",
"**/migrations/**",
"**/*.sql",
"**/*.prisma",
"**/routes/**",
"**/router/**",
".claude/hooks/**",
"**/settings.json",
"**/settings.*.json",
"**/module.json",
".claude/t1k-modules.json",
".claude/metadata.json",
".github/workflows/**"
]
},
"tests-only": {
"globs": [
"**/__tests__/**",
"**/*.test.*",
"**/*.spec.*",
"tests/**",
"**/tests/**"
]
}
}
},
"worktreeBehindNudge": {
"_comment": "UserPromptSubmit advisory (hooks/worktree-behind-nudge.cjs): in a LINKED worktree on a topic branch (never main/master/the default branch), emit [t1k:worktree-behind] when the branch is >= minBehind commits behind origin/<default>, or >= 1 behind with a file changed on both sides. At most once per throttleMinutes per worktree; a remote ref older than fetchStaleMinutes triggers a detached background fetch. Overlap is skipped when either side changed more than overlapMaxPaths files. Nudge only, never rebases. Kill switch: features.worktreeBehindNudge.",
"fetchStaleMinutes": 15,
"gitTimeoutMs": 2000,
"minBehind": 20,
"overlapMaxPaths": 2000,
"throttleMinutes": 30
},
"worktrees": {
"_bundleRealDirtMergedNote": "Issue #6491, default false. Only read with reapMode quarantine. true: a MERGED lane whose REAL dirt is captured in a verified recovery bundle (telemetry/lane-recovery) is removed once its landed proof holds, with its restore command printed. false: real dirt stays DIRTY, as it always did. Removing real work is a data-loss policy call the issue marks as needing design, so the owner opts in per install.",
"_comment": "Worktree-root hygiene (plans/261016-scratch-worktree-hygiene). DOCUMENTS the defaults read from DEFAULTS in hooks/lib/scratch-root.cjs. THE worktree location is `root` (central, e.g. ~/.cache/t1k/worktrees -> <root>/<repo>/<name>; `<repo>` placeholder; ~ expands) when set, else <parent-of-clone>/<repo>-worktrees (scripts/lib/worktree-root.cjs). staleReap: false turns off the done/abandoned/unity-lib reaper. watchRoots/repoIdleDays/watchDiskPct feed scripts/worktree-watch.cjs. legacyRootNames are MIGRATE-ONLY locations the census reports and the sweep still reaps; nothing creates there. reapMode: report | regenerable | quarantine.",
"archivePatterns": [
"worktrees-archive*",
"_worktree-archive",
"stale-*"
],
"bundleRealDirtMerged": false,
"cloneDiscoveryHours": 24,
"cloneDiscoveryMaxClones": 100,
"extraRoots": [],
"finisher": {
"_comment": "Finisher dispatch (plans/261028-worktree-tracking phase 4, hooks/lib/worktree-finisher-dispatch.cjs). When a dead owner left real unfinished work in a .claude/** lane with a recorded task and no PR, the SessionStart notice carries a standing order to adopt the worktree (adopt --as-finisher) and spawn a NATIVE t1k-lane-worker to finish it. ON by default (Core Requirement #13); enabled: false or T1K_SKIP_WORKTREE_FINISHER=1 turns the dispatch off (the notice still prints). maxConcurrent caps active finishers machine-wide (lock dir <claudeDir>/session-state/worktree-finishers). maxAttempts is per worktree for its lifetime. Any other lane type is surface-only (no-native-agent-for-lane). Never a herdr pane, a claude -p run or a routed agent.",
"enabled": true,
"maxAttempts": 1,
"maxConcurrent": 2
},
"idleDays": 14,
"legacyRootNames": [
"worktrees",
"wt",
"_worktrees",
".worktrees",
".adnet-worktrees",
".kit-lanes",
"worktrees-archive*"
],
"light": {
"_comment": "Light worktree profile (opt-in; scripts/lib/worktree-light-profile.cjs). profile is the default for `worktree.cjs create` (full = the pre-profile create: every tracked file, every submodule). light.initSubmodules: submodules a light create initialises (name or path; [] = none). light.sparseDirs: cone-mode sparse-checkout dirs for a light create ([] = the whole tree). light.seedDirs: regenerable gitignored folders reflink-copied by a seed (default empty: core names no engine; an engine kit sets e.g. [\"Library\"]). light.seedFrom: the seed source for a light create (\"main\" = the main worktree, another value = a worktree name, \"\" = no seed; --seed-from wins). light.engineProjectMarkers: paths whose presence marks an engine project a sparse tree cannot open (a light create warns; an engine kit sets e.g. [\"ProjectSettings/ProjectVersion.txt\"]). Flags --profile / --init-submodules / --sparse / --seed-from / --seed-dirs override these per create.",
"engineProjectMarkers": [],
"initSubmodules": [],
"seedDirs": [],
"seedFrom": "",
"sparseDirs": []
},
"mergedQuietMinutes": 30,
"ownerGone": {
"_comment": "Dead-owner detection (plans/261028-worktree-tracking phase 2, scripts/lib/worktree-owner-gone.cjs). A worktree is owner-gone when its owner session is dead, nobody is inside, real work is unfinished and it has been idle minIdleMin. An ADOPTED finisher is live only while its adopter session lives AND its last activity is within heartbeatMin. surfaceHours throttles the SessionStart notice per worktree. enabled: false turns the notice off; T1K_SKIP_WORKTREE_OWNER_GONE=1 does too.",
"enabled": true,
"heartbeatMin": 15,
"minIdleMin": 30,
"surfaceHours": 24
},
"profile": "full",
"reapMode": "quarantine",
"regenerable": {
"_comment": "Issue #6491. Regenerable dirt a lane sweep may drop WITHOUT refusing the lane; real work is never listed here. submodulePointers: a checked-out submodule whose HEAD is an ancestor-or-equal of base's gitlink for that path and whose own tree is clean. suffixes: TRACKED paths that are unstaged-modified (` M`) and whose name ends with one of these; EMPTY in core on purpose, core stays engine-neutral. An engine kit supplies its own in its t1k-config-<kit>.json (per-key override, highest priority wins), for example Unity adds \".meta\". bundleMaxMB caps the recovery bundle written before any removal that takes content (lane-recovery under telemetry/, restore command in its MANIFEST.txt). reapMode: report only lists; regenerable also removes regenerable-only lanes; quarantine also removes regenerable-only lanes.",
"bundleMaxMB": 512,
"submodulePointers": true,
"suffixes": []
},
"repoIdleDays": 7,
"root": null,
"staleReap": true,
"trashRetentionDays": 7,
"unity": {
"_comment": "Engine facts for the Unity lane reaper. Core ships the built-in defaults; the Unity kit's t1k-config-unity.json (priority 90) overrides them per key. marker \"\" = the built-in Assets/ + ProjectSettings/ pair.",
"marker": "",
"regenerableDirs": [
"Library",
"Temp",
"obj",
"Logs"
]
},
"warnCount": 40,
"watchDiskPct": 85,
"watchRoots": []
}
}