get-work-done
Skill9 starsChanged 7 months agoMITView on GitHub
Central work dispatcher ("mother hub" front door) for the GetWorkDone fleet. EVERY task gets a contract + T-id; owner-present tasks run in-session in a worktree of the target repo (default), else a background worker on the cheapest-correct model lands via CI-gated PR (v0.10). Grills to >95% confidence at intake, one question per turn. Use for "/get-work-done fix X in IPODhan", intake mode, status, or "cancel T-042". SSOT: plans/get-work-done-dispatcher.md.
In this repository44 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- anthropic-multi-agent-research-system-skillSkill
- apply-selectionsSkill
- bootstrap-dogfood-projectSkill
- brainstormSkill
- branch-choiceSkill
- cc-adoption-scoutSkill
- claude-guardianSkill
- continueSkill
- end-sessionSkill
- executing-plansSkill
- five-advisorsSkill
- git-branch-lifecycleSkill
- githubSkill
- grill-meSkill
- hook-transcript-fixture-testSkill
- instagramSkill
- linkedinSkill
- modern-data-pipeline-diagramSkill
- pattern-qualitySkill
- platform-event-live-probeSkill
- plugin-lifecycleSkill
- promote-to-coreSkill
- prompt-auto-enhanceSkill
- provision-reportSkill
- redditSkill
- review-new-claude-featuresSkill
- save-sessionSkill
- scan-discovery-reportSkill
- scan-repoSkill
- scan-urlSkill
- self-improveSkill
- skill-evaluatorSkill
- skill-masterSkill
- ssot-workflow-auditSkill
- start-sessionSkill
- synthesize-hubSkill
- synthesize-projectSkill
- twitter-xSkill
- verify-effect-at-destinationSkill
- workflow-doc-reviewerSkill
- writing-plansSkill
- writing-skillsSkill
- youtube-transcriptSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
34 rules. About 0 are common, 16 are distinctive, and 18 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈16
Distinctive · estimate
This author's own choices. Worth a look.
18
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 42
Verify every prerequisite NOW, report ALL missing items in one stop - never discover a missing tool mid-run, never improvise a fallback (an …
- Very specific: line 116
FAIL AT INTAKE, NOT AT MIDNIGHT: run every abort-capable check now - registry/remote identity, branch protection + secret-scan gate (gh …
- Very specific: line 345
MUST resolve every repo through GWD/settings.json repo_registry and assert git remote get-url origin matches before any edit.
- Distinctive: line 347
MUST give EVERY task a contract + T-id - owner-present tasks fast-lane under it in a worktree of the target repo (SCOPE), everything else …
- Very specific: line 350
MUST keep the lanes separate: a lane: fast contract is session-executed and checked by fast-lane-check.py, never handed to a worker.
- Very specific: line 352
MUST NOT dispatch into a host whose claim reconciler has not ticked inside 2x settings.fleet_keeper_tick_minutes.
- Very specific: line 354
MUST NOT dispatch past host pressure (commit charge over the limit, or live workers at settings.fleet.max_concurrent_workers), and MUST …
- Very specific: line 357
MUST serialize same-repo work fleet-wide (any machine) unless related:-linked with disjoint file scopes.
- Very specific: line 359
MUST copy the registry's context_docs into every contract and open the worker prompt with the read-these-first mandate.
- Distinctive: line 361
MUST route models cheapest-correct (sonnet default, haiku mechanical, opus deep + security), never Fable as a worker, editing tier before …
- Very specific: line 363
MUST carry a routing rationale on model:, a data_source: on a data-reading task, evidence: required, and a dod: of checkable predicates.
- Very specific: line 365
MUST size max_turns from task shape and auto-resume an error_max_turns death into the SAME worktree at double budget, lifetime THREE runs, …
- Distinctive: line 368
MUST open every fix-round prompt with its honesty / PR-body / STATUS.md duties.
- Very specific: line 370
MUST apply the ROOT-CAUSE GATE: a SECOND occurrence is fixed at the MECHANISM or the impossibility recorded.
- Distinctive: line 373
MUST keep this skill conformant with the live fleet - paths resolve, settings keys exist, every preflight exit code documented, one launch …
- Very specific: line 376
MUST carry a gate:<id> on every rule here.
- Distinctive: line 378
MUST NOT write a worker prompt naming a path that belongs to another machine.
- Very specific: line 380
MUST inject the three standing mandates from GWD/worker-mandates.txt - never hand-copied, never paraphrased - into every worker, checker …
- Very specific: line 383
MUST add the HOLD-LABEL line whenever a dod: needs the PR to stay open.
- Very specific: line 385
MUST land ALL work via a PR gated on the repo's CI, and the WORKER never merges or closes any PR on any repo: PUBLIC repos arm auto-merge, …
- Distinctive: line 389
MUST verify PR-state post-run for every task - a PR merged inside the worker's run window is a task FAILURE independent of work quality, …
- Distinctive: line 392
MUST run the pre-queue dedup gate over ALL open contracts for the repo before queueing.
- Distinctive: line 394
MUST batch same-repo trivia from one intake into ONE contract (one T-id, one PR).
- Distinctive: line 396
MUST pass the PORTFOLIO REGISTRATION GATE before queueing, and create project folders only under the work root.
- Distinctive: line 398
MUST run every abort-capable check at INTAKE while the owner is present, and grill to >95% confidence of WHAT is asked.
- Distinctive: line 401
MUST run no fleet actor (worker or checker) unattended under interactive permissions.
- Distinctive: line 403
MUST keep maker != checker: evidence + LEDGER are checker-written only, a self-reported pass is never proof, a verdict with no …
- Very specific: line 406
MUST branch on stop_reason - a refusal is not success.
- Distinctive: line 408
MUST honor origin-session reporting affinity: only the origin session, or a labeled keeper rescue of a provably-dead origin, reports or …
- Distinctive: line 410
MUST show the owner a queue ack at dispatch AND the terminal status card when the watcher fires, and MUST apply the OWNER STATUS CADENCE: a …
- Very specific: line 413
MUST send a terminal-state card as FALLBACK for session-origin tasks and PRIMARY for origin-less ones - and on a non-VPS machine push the …
- Distinctive: line 416
MUST reprioritize by REORDERING only - a running worker is never preempted.
- Very specific: line 417
MUST honor the deploy tier computed from the ACTUAL merged diff - auth/payment/DNS/migration force-upgrades to HOLD.
- Very specific: line 419
MUST append the INVOCATIONS.log line at the end of every intake turn using the defined tids= tokens - only tids=none is the inline defect.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
/get-work-done - central work dispatcher (v0.10, 2026-08-27)
v0.10 = procedure only. Incident narratives live VERBATIM in references/incident-log.md
(I-01 ... I-37); a rule's [log: I-nn] cites the WHY, never a step to run. references/ also
holds the fast-lane runbook + routing table. GWD = this machine's state root
(GWD/settings.json -> fleet_home): D:/Abhay/GetWorkDone/ here, C:/Abhay/GetWorkDone on
the VPS. Contract format = the goal-creator shape (plugins/loop-engineering/skills/goal-creator/SKILL.md) + STEP 5 fields.
Mode router
| Invocation | Mode |
/get-work-done <task> | INTAKE (steps 1-7) |
/get-work-done intake | Standing intake loop: next task, steps 1-6, return to "ready" IMMEDIATELY. A fast-lane task (STEP 3) session-executes after queueing; else dispatches. Never freelances outside a T-id |
/get-work-done status | Fleet state from GWD/queue/ + GWD/heartbeats/ + LEDGER tail, plus host pressure (commit % + live workers via preflight-guard.ps1 -HostMemoryStatus); re-rendered every 15 min while live [log: I-22] |
/get-work-done cancel <T-id> | Read GWD/heartbeats/<id>.hb -> kill PID tree; rename <id>.cancelled.md with a reason; gh pr close --delete-branch any PR it opened; ledger it. cancel all = all claimed |
/get-work-done sweep | Promote/reject GWD/inbox/ per the P17 table, run INTAKE 4-7 on promoted items, then claim + dispatch unclaimed *.queued.md. DISPATCH-FIRST: claiming is every tick's first duty, never satisfied by watching claimed tasks |
SCOPE (owner 2026-08-27)
Owner-present work runs IN-SESSION: a worktree of the target repo, T-id + contract still
required, Opus/Sonnet Agent implements, a fresh Opus reviews, Fable lands. /get-work-done
is ONLY for unattended/hands-off app-repo work - overnight, VPS batches, clear-spec app tasks
the owner is not watching. Fleet-core (bus scripts, this skill/tests) is FROZEN - never a fleet
worker, in-session only when an app task is blocked. Weekly measure app vs fleet landed
(fleet-ratio.py on the bus); under ~2:1 by 2026-09-10 -> GWD becomes a thin launcher. Every
task keeps its T-id + contract + LEDGER line + registry flips regardless of mode.
Prerequisites
git, an authenticated gh (PR/label/merge-state calls), python and PowerShell on PATH, a
writable bus checkout at GWD (queue, heartbeats, settings.json, guard scripts). Fast-lane also
needs GWD/fast-lane-gate.py + GWD/fast-lane-check.py.
STEP 0: Preflight
Verify every prerequisite NOW, report ALL missing items in one stop - never discover a missing tool mid-run, never improvise a fallback (an unreachable bus means no T-id, no safe degraded mode).
STEP 1 - INTAKE: parse the ask
Split the input into tasks (may span repos). Resolve each target against GWD/settings.json ->
repo_registry - the ONLY source of repo paths [log: I-04]. A repo not in the registry, or
ambiguous, is an intake question, never a guess.
PORTFOLIO REGISTRATION GATE [log: I-04]: also check the target against PORTFOLIO.yml on
5wealths main. A repo absent from it, or a task creating a NEW project folder, cannot queue
until a registration row (name, pillar, machines) exists. Carry it into STEP 4; never skip it
silently.
STEP 2 - SCOUT (~1 min per task)
In the target repo: git remote get-url origin (must equal the registry's remote - a mismatch
aborts intake with a question), branch state, does the named area exist, do tests exist, which
paths the task will touch. Ground the gate in looked-at reality, never the prompt alone.
STEP 3 - GATE: every task gets a contract + T-id; SIZE picks the lane
A get-work-done task with no T-id is a defect [log: I-01]. Every task is contracted (STEP 5),
then FAST-LANED or DISPATCHED (STEP 6); old carve-outs are subsumed by those two paths. Blast
radius (auth, payments, config, migrations, deploy) informs model tier, budget, deploy_tier
and lane eligibility.
Repo identity is by REGISTRY KEY, never path or folder name: resolve cwd and target via
git remote get-url origin -> key; no key is foreign, same-key tasks still get their own
worktree. An artifact whose source of truth lives elsewhere (template copy, map cards, Deluge
specs) resolves its repo by ssot_artifacts; the session relays owner approvals, never authors
the text.
FAST LANE (owner decision A [log: I-01]; DEFAULT when the owner is present, per SCOPE)
Session-executed under a T-id in a worktree of the TARGET repo, with context_docs and a
checker - the DEFAULT whenever the owner is present; STEP 6's background worker is the
unattended path.
ELIGIBILITY = ALL of: deliverable: content|mechanical * <=5 files in files: * <=300
changed lines at PR time * no path matching the sensitive-path denylist in
GWD/fast-lane-gate.py * no unknowns after scout. code is NOT eligible in v1; revisit after
10 clean runs.
FLOW: gate -> stage-stamp -> worktree edit -> PR -> diff gate -> fast-lane-check.py -> merge
on green -> LEDGER line, detailed in references/fast-lane-runbook.md. SLO <=20 min
launched->merged (settings.fast_lane_slo_minutes), a miss = FAST-LANE-SLO-MISS via
lesson.py status. A lane: fast contract at WORKER dispatch is preflight exit 14.
STEP 3.5 - ROOT-CAUSE GATE: the SECOND occurrence fixes the MECHANISM
[log: I-02] - eight workers died at their turn cap; the fix was always a firmer prose mandate, and it failed every time.
A first occurrence may be fixed as an instance. From the SECOND occurrence of the same failure SHAPE
(same symptom class, not the same file or repo) an instance fix alone is NOT acceptable on its own -
fix the MECHANISM that permits it or record in status_log, with the reason, why a mechanism fix is
impossible - silence is not an exemption.
PROSE IS NOT A MECHANISM: a mechanism is CODE, a GUARD, a HOOK, a SCHEMA CONSTRAINT or a TEST
THAT FAILS when the defect returns - the eight-times-failed prose mandate is why. Before
calling it done, SWEEP the repo (and the estate, where the shape travels) and REPORT the count - "fixed 1 of N found" / "swept, no other instances" - and
append a LESSON(CODIFIED -> <where>) line to GWD/PATTERNS-SEEN.md plus a row in
GWD/MECHANISM-DUE.md (python GWD/lesson.py); the learning-debt gate (exit 11) reads it.
STEP 4 - CLARIFY: resolve everything at intake, one question per turn
NON-SKIPPABLE - "run it" / "go ahead" means PROCEED, not skip [log: I-03]. (a) DETERMINE,
don't ask anything you can scout ("compare IPODhan's data" -> the app's OWN DB/API, not its
public site); asking what you can determine is itself a defect. (b) For a GENUINE material unknown
(2+ valid answers changing the OUTCOME, unscoutable) -> the 95%-gate: ONE question per turn,
*Sync-check:* opener, recommended answer + one-line justification, until confidence exceeds
95%. "Decide-don't-ask" governs reversible EXECUTION detail only.
FAIL AT INTAKE, NOT AT MIDNIGHT: run every abort-capable check now - registry/remote identity,
branch protection + secret-scan gate (gh api), credentials/tools present, deploy tier, and the
permission surface (any unattended worker/checker on a no-prompt path, verified now - never
discovered mid-run with the owner away).
| Deploy situation | Tier |
| Re-deploy of a live app, same domain/infra | AUTO (Tier 1) - no question |
New app onto the owner-granted sandbox domain (settings.json _sandbox_note) | AUTO (Tier 1) - pre-authorized |
| First deploy / DNS / new domain / auth or payment surface | HOLD (Tier 2) - collect the yes/no NOW |
A portfolio-registration flag from STEP 1 IS a material unknown - resolve it here; determine the
pillar yourself when knowable. AUTO-tier deploys run via GWD/deploy-site.sh: capture
KNOWN-GOOD -> niced build -> swap -> nginx -t gate -> probe live URL -> restore known-good on
failure [log: I-05]. Nothing answerable from GLOBAL.md, the repo, or the registry may be asked.
STEP 5 - CONTRACT
Build the goal-creator contract from the task + answers, add the dispatcher fields, write it via
bus-write.ps1 to GWD/queue/T-<nnn>-<slug>.queued.md (T-384, the one bus write path). Each rule
below is normative in CRITICAL RULES; full v0.9 text + incident at [log: I-06 ... I-14].
- T-ID [log: I-06]:
git pullthe bus ->python GWD/next-task-id.py <GWD>-> on a non-fast-forward push rejection, pull and RE-RUN the allocator. Never invent an id; never reuse one seen historically. - DEDUP [log: I-07]: read EVERY open contract in
GWD/queue/for the same repo. Duplicate -> don't queue, name the existing T-id. Overlapping ->related: [T-xxx]. Supersedes -> rename the old one.superseded-by-T-<new>.md. - BATCH [log: I-08]: same-repo trivia from one intake -> ONE contract, one
dod:item each, samedeliverable:only. WAVE-CHAIN [log: I-09]: a wave contract'srelated:names the earlier wave T-id it depends on. - CONTEXT DOCS [log: I-10]: copy
repo_registry.<key>.context_docs(repo-relative) into the contract; prompt opens "Before ANY work, read these files at the repo root:<list>. If any is missing, STOP and report it." CLAUDE.md is never listed; the repo wins over a stale doc. - MANDATES: the three standing mandates live VERBATIM in
GWD/worker-mandates.txt- never hand-copied or paraphrased [log: I-11, I-12, I-14]. The DISPATCHER prepends that file to every prompt today; T-372 makesworker-wrapper.ps1do it instead (not landed -grep -c worker-mandates GWD/worker-wrapper.ps1= 0 until it does). Mandate 3 (T-209, PR #580):[skip ci]matches ANYWHERE in a commit message, headline or body; a REQUIRED check that never reports leaves the PR blocked forever, per PRs #577/#579 under T-191. Add the HOLD-LABEL line (gh pr edit <n> --add-label hold) whenever thedod:needs the PR to stay open [log: I-13] - not merging is not enough to stop this repo's auto-landing hooks. - LANDING BATCH [log: I-14]: same-repo same-day contracts default to ONE branch/PR/CI-run; exceptions are P1 break-fixes, conflicting file scopes, and a checker FAIL on separable hunks.
repo: <registry key> # path + remote resolved from the registry
origin: <session-id>@<machine> <project-dir> # owns all reporting
related: [] # T-ids sharing scope in the same repo
context_docs: [] # copied VERBATIM from the registry
model: haiku|sonnet|opus # + MANDATORY one-line rationale (lint-enforced)
deliverable: code|deploy|content|claude-resource|data|mechanical
lane: normal|fast # fast needs content|mechanical + files:/checks: (STEP 3)
priority: P1|P2|P3 # dispatcher-assigned - never ask the owner to rank
deploy_tier: none|auto|hold-approved|hold-denied
approvals: [<granted at intake>]
budget: {max_turns: <sized floor, 6.8c>, wall_clock_hours: 4}
evidence: required # checker-written, never worker-written
dod: # CHECKABLE predicates, dod-verbs discipline
- <ACTION + COMPLETENESS BAR, verifiable by a stranger>
status_log: []dod: items state the ACTION and the completeness bar ("documents the per-partner pattern
INCLUDING the two fallbacks", not "update the doc") - a worker satisfies the weakest literal
reading, the checker verifies only what is written as a predicate.
Routing: haiku = mechanical/classification, sonnet = DEFAULT for any briefed machine-checkable
job, opus = deep design + preemptively all security work, Fable/Mythos = NEVER a worker (exit 4).
Full table: references/model-routing-table.md. When torn pick the cheaper tier - escalation
recovers a wrong cheap pick; a wrong expensive one is never detected.
STEP 6 - DISPATCH: one background worker, guarded
- Same-repo scan [log: I-15]:
git pullthe bus, scan for ANY*.claimed.*.mdon the same registry key from ANY machine; one exists -> do not dispatch unlessrelated:-linked with declared-disjoint file scopes. Preflight exit 8 is the backstop; keep them in sync. - Atomic claim: rename
...queued.md->...claimed.<session-id>.md; failed -> skip. - Contract lint FIRST:
python GWD/contract-lint.py <claimed contract>- non-zero -> park with its stderr as the reason. Runs before preflight so nothing is cloned for a bad contract. Preflight gate:
powershell -NoProfile -ExecutionPolicy Bypass -File GWD/preflight-guard.ps1 -ContractPath <c> -RepoPath <workspace> -ExpectedRemote <remote> -TaskId <T-NNN|T-NNNC*|T-NNNF*> -PromptPath GWD/heartbeats/<id>.prompt.txt -MaxTurns <cap>- non-zero BLOCKS; park with the reason.
-TaskIdsizes exit 12's kind (build/checker/fix_round); a checker/fix round launches against its PARENT's contract, so omitting it prices as build. The script header table is the SSOT:
Code Gate exit 0 OK to dispatch exit 1 -SelfTestfailure only; never from a gate runexit 2 contract missing exit 3 no model:fieldexit 4 model not an allowed worker tier (Fable/Mythos) exit 5 workspace path missing exit 6 repo identity mismatch exit 7 a context_docs:entry is absolute or missing from the workspaceexit 8 same-repo contract already claimed, not related:exit 9 prompt file references a foreign-machine path exit 10 HOST-MEMORY (T-312): commit charge over 75%, or at settings.fleet.max_concurrent_workersexit 11 LEARNING-DEBT (T-323): MECHANISM-DUE.mdrow occurrences >= 2, not claimed/done, >48h, nolesson_class:; malformed FAILS CLOSEDexit 12 TURN-BUDGET (T-345): -MaxTurnsundersettings.worker_defaults.max_turns_by_deliverableexit 13 FIX-ROUND PROMPT-LINT (T-336): T-<n>F*prompt's first step omits honesty/PR-body/STATUS.md dutiesexit 14 FAST-LANE contract at worker dispatch (T-351): set lane: normal, re-lintexit 15 KEEPER-LIVENESS (T-363): no reconciler tick on THIS host in 2x settings.fleet_keeper_tick_minutesexit 16 HOST-MEMORY-UNKNOWN (T-364): commit/worker query returned nothing - fails CLOSED exit 17 SETTINGS-READ-FAILURE (T-364): settings.json unreadable/malformed/non-integer - blocks exit 18 ROUND-CAP-EXCEEDED (T-367): round suffix over settings.max_rounds_by_kind- park QUALITY/SCOPE-SUSPECTexit 19 DISPATCH-CEILING (T-369): today's heartbeats/*.prompt.txt>=settings.max_dispatches_per_day- non-zero BLOCKS; park with the reason.
- Workspace [log: I-16]: clone FRESH (
git clone --filter=blob:none) unless one insidesettings.workspaces.retention_daysexists AND is clean. Janitor deletes idle workspaces only when provably clean; dirty ones escalate. - LAUNCH - the one recipe [log: I-21]. Write
GWD/heartbeats/T-<id>.prompt.txtFIRST, then:
The wrapper passes the prompt as an argv pointer (STDIN abandoned - invisible to console-less processes), writespowershell -NoProfile -ExecutionPolicy Bypass -File GWD/worker-wrapper.ps1 ` -TaskId T-<id> -RepoPath <workspace> -Model <tier> -MaxTurns <cap> -StateRoot <GWD>T-<id>.hb(PID + tick) and the result JSON; the dispatcher, not yet the wrapper, prependsGWD/worker-mandates.txtto that prompt (T-372).-StateRootis MANDATORY off the VPS. A same-repo-as-dispatcher contract orders the worker into its own worktree. Workers run tests/builds in the FOREGROUND only, neverrun_in_background(T-393). - Terminal state [log: I-18]: parse
stop_reason- a refusal is NOT success. Refusal on haiku/sonnet -> reroute to opus (editmodel:, appendstatus_log, re-lint, re-preflight, relaunch); on opus -> PARK, no second reroute. Error -> one retry same tier; a 2nd failure: environmental -> park, capability/quality -> escalate ONE tier (max one). Keeper deaths are environmental: re-queue once, park on the second. - Post-exit PR-state check [log: I-19]:
gh pr view <PR> --json state,mergedBy,mergedAton every PR named. Merged inside the run window = a task FAILURE line instatus_log+GWD/LEDGER.md+ an owner card. - Lanes + ceilings [log: I-20]: up to
settings.fleet.max_concurrent_workersconcurrently (exit 10 enforces it); same-repo serializes, others parallel; P1 > P2 > P3, a P1 always admitted; reprioritize by REORDERING only, never preempting.settings.max_dispatches_per_daysettings.daily_token_ceiling-> pause + ping once, never retry-storm.GWD/state/ci-hold.flag=> only P1 proceeds.
- Queue ack + watcher, same turn as the launch [log: I-22]. Ack: render T-id, repo, tier,
priority, budget, one-line DoD on screen in the origin session. Watcher: a
Monitoruntil-loop (or sizedScheduleWakeup) breaking on every terminal state - a NON-EMPTY result JSON (the wrapper pre-creates it at 0 bytes), anEXITEDheartbeat, or one older thansettings.heartbeat_stale_after_seconds- after which the origin session renders the status card (outcome, PR, verdict, evidence path, tier) automatically. OWNER STATUS CADENCE [log: I-22]: while any session-origin task is live, run a PERSISTENT 15-minute ticker (never a fixed timeout - the 2026-08-20 20:30 lapse: a timeout-armed ticker stopped and nobody noticed until the owner asked twice). Every tick opens with the CURRENT TIME IN IST, e.g.[20:45 IST]- a tick with no timestamp does not satisfy the cadence. CONTENT FLOOR per tick: what changed since the last tick, what is running, the estimate - an explicit "NOTHING changed" tick, never a skipped one, and never fabricated progress. - Budget from task shape [log: I-23]:
max_turnsfromsettings.worker_defaults.max_turns_by_deliverable, floored bysettings.worker_defaults.max_turns_by_tier. A DoD running a full suite, rebuilding assets, writing docs AND driving a PR is >=70 turns however small the diff. A turn-cap death resumes into the SAME worktree at a raised budget - never redo finished work from origin/main. - Reconcile (keeper duty, every tick) [log: I-24]: per claimed contract read its
.hb- fresh tick -> leave alone;EXITED-> CHECK; stale/dead PID -> kill remnant, appendstatus_log, re-queue ONCE, park + card on a second death. Foreign-machine claims have no local.hbby design - require one before calling a task dead, else wait outbudget.wall_clock_hours. Run the janitor here too.error_max_turnsis NOT a death: raisebudget.max_turnsto min(2x current, 200), relaunch into the SAME worktree; lifetime THREE runs, then parkQUALITY/SCOPE-SUSPECT. Reconcile must RUN on this host or exit 15 blocks the next dispatch.
SWEEP RECONCILIATION [log: I-25] - newest signal wins (status_log + bus git log); a live
origin's activity blocks reclaim except the provably-dead-origin rescue; an unmerged PR blocks
.done.
BUDGET PRE-APPROVAL [log: I-26]: retries/resumes/raises/reroutes are pre-approved within plan + daily-ceiling limits, which pause + notify rather than ask. New recurring 3rd-party spend stays owner-gated.
INVOCATION LOG [log: I-27]: end EVERY intake turn with one GWD/INVOCATIONS.log line:
<UTC ISO> | <session-id>@<machine> | cwd-repo=<key|none> | tasks_parsed=<n> | tids=<value>;
only tids=none is the inline defect.
STEP 7 - CHECK + REPORT: maker != checker
- Reporting affinity [log: I-28]: the contract's
origin:session owns all in-session reporting; another session may LIST a foreign-origin task but never check, report, retry or park it. Exceptions: a keeper/sweep rescue of a provably-dead origin and an explicit owner ask elsewhere (answer, change nothing). - Checker launch [log: I-29, I-30]: mandatory, auto-chained, HEADLESS -
worker-wrapper.ps1,bypassPermissions, heartbeatT-<id>C.hb, watcher armed same turn. Tier = opus when the contract is opus, else sonnet - never weaker than the maker. In-sessionAgent()only with the owner present and said so. The worker itself runs in the TARGET repo's dir, never the hub's, or hub governance leaks into its machine-parsed JSON. - Tier receipt [log: I-31]:
python GWD/verify-model-tier.py <contract> GWD/heartbeats/<id>.result.json- tier-as-run must equal tier-as-contracted; non-zero = a task FAILURE line, never a silent pass.
- Merge-guard predicate: re-derive item 8 of STEP 6 independently, before scoring any dod.
- Verify every
dod:predicate by thedeliverable:procedure:
deliverable: | Checker procedure (re-derivation, never review-and-agree) |
code | Re-run the project's OWN full test gate from scratch; CI green on the PR; diff inspected against the dod predicates |
deploy | verify-effect-at-destination: probe the LIVE URL, capture the screenshot, config-validity gate |
content | Trace EVERY claim to its source (code, PR diff, captured data) - no source row = a FINDING; check placement, structure, each dod predicate. Sample floor: ALL claims <=10, else 10 + every number. lane: fast -> fast-lane-check.py instead |
mechanical | fast-lane-check.py (deterministic) or the repo's own lint - never review-and-agree |
claude-resource | Run /skill-evaluator (output mode min; full for a new skill) from the hub; if the target repo can't, execute the resource's own trigger + one scenario end-to-end, capture the transcript |
data | Re-pull a sample from data_source: independently and re-derive the headline numbers |
- Evidence = re-derivation artifacts, not attestations [log: I-32]:
GWD/evidence/<date>-T-<id>/holds the raw proof (test output / probe screenshot / claim-source table / eval report or transcript / re-pulled sample + recomputation) plus SHA + PR URL, AND theGWD/LEDGER.mdline with tier + costUSD from the receipt. An opinion with no artifact is no verdict; the WORKER NEVER writes evidence. - ROOT-CAUSE CLOSE-OUT [log: I-33]: not done until the verdict carries the SWEEP result for
every defect fixed and, on a second-occurrence shape, names the MECHANISM installed plus its
PATTERNS-SEEN.mdlesson line. New prose alone is INCOMPLETE. - Report + lessons [log: I-34]: per task - outcome, PR link, evidence path, cost tier, plus
anything parked and why. Append the shape signature to
GWD/PATTERNS-SEEN.md(3rd occurrence -> a PROPOSED codify card) and every failure/park/reroute/refutation asLESSON(OPEN): <mistake> -> <root cause> -> <rule>; lifecycle OPEN -> CODIFIED -> ARCHIVED; intake reads only the newest 20 OPEN lines. Fleet-mechanics lessons only - hub lessons stay in.claude/tasks/lessons.md. - Cards [log: I-35]: the origin session's on-screen render is PRIMARY;
GWD/notify-owner.ps1fires only when that session is dead at terminal time. Origin-less tasks use the card as primary; silence is never success. Non-VPS: the ping reaches the relay only after a bus push. Parked digest [log: I-36]:GWD/parked-digest.ps1(weekly) cards every*.parked.mdwith age + reason; owner replies<T-id> retry/drop- parked is never lost.
ARTIFACT PLACEMENT
[log: I-37] By WHAT the artifact is, not where the fleet runs: project-SPECIFIC -> that repo via a PR (never bus or hub); fleet-GENERIC machinery -> hub/bus scripts; fleet runtime STATE -> bus.
CRITICAL RULES
Each bullet carries gate:<id> resolving in config/gwd-gates.yml (hub). gate:PROSE-ONLY
means no machine enforces it yet.
- MUST resolve every repo through
GWD/settings.jsonrepo_registryand assertgit remote get-url originmatches before any edit. gate:PREFLIGHT-REPO-IDENTITY-MISMATCH - MUST give EVERY task a contract + T-id - owner-present tasks fast-lane under it in a worktree of the target repo (SCOPE), everything else dispatches; never the cwd checkout, and authoring SSOT-owned artifact text outside a T-id is the same defect. gate:PREFLIGHT-CONTRACT-MISSING
- MUST keep the lanes separate: a
lane: fastcontract is session-executed and checked byfast-lane-check.py, never handed to a worker. gate:PREFLIGHT-FAST-LANE-AT-WORKER-DISPATCH - MUST NOT dispatch into a host whose claim reconciler has not ticked inside 2x
settings.fleet_keeper_tick_minutes. gate:PREFLIGHT-KEEPER-LIVENESS - MUST NOT dispatch past host pressure (commit charge over the limit, or live workers at
settings.fleet.max_concurrent_workers), and MUST fail CLOSED when the host or settings read fails rather than reading 0.0%/0. gate:PREFLIGHT-HOST-MEMORY-GATE - MUST serialize same-repo work fleet-wide (any machine) unless
related:-linked with disjoint file scopes. gate:PREFLIGHT-SAME-REPO-ALREADY-CLAIMED - MUST copy the registry's
context_docsinto every contract and open the worker prompt with the read-these-first mandate. gate:PREFLIGHT-CONTEXT-DOCS-MISSING - MUST route models cheapest-correct (sonnet default, haiku mechanical, opus deep + security), never Fable as a worker, editing tier before any reroute. gate:PREFLIGHT-MODEL-TIER-NOT-ALLOWED
- MUST carry a routing rationale on
model:, adata_source:on a data-reading task,evidence: required, and adod:of checkable predicates. gate:CONTRACT-LINT-BLOCK - MUST size
max_turnsfrom task shape and auto-resume anerror_max_turnsdeath into the SAME worktree at double budget, lifetime THREE runs, then parkQUALITY/SCOPE-SUSPECT; budget is standing pre-approved, never an owner question. gate:PREFLIGHT-TURN-BUDGET-GATE - MUST open every fix-round prompt with its honesty / PR-body / STATUS.md duties. gate:PREFLIGHT-FIX-ROUND-PROMPT-LINT
- MUST apply the ROOT-CAUSE GATE: a SECOND occurrence is fixed at the MECHANISM or the
impossibility recorded; every fix swept with the count reported; every class fix appends its
LESSON(CODIFIED -> ...)line andMECHANISM-DUE.mdrow. gate:PREFLIGHT-LEARNING-DEBT-GATE - MUST keep this skill conformant with the live fleet - paths resolve, settings keys exist, every preflight exit code documented, one launch recipe, shrink-only byte ratchet. gate:GWD-SKILL-CONFORMANCE-TEST
- MUST carry a
gate:<id>on every rule here; an unmapped MUST is red in CI. gate:GWD-SKILL-MUSTS-HAVE-GATES-TEST - MUST NOT write a worker prompt naming a path that belongs to another machine. gate:PREFLIGHT-FOREIGN-MACHINE-PROMPT-PATH
- MUST inject the three standing mandates from
GWD/worker-mandates.txt- never hand-copied, never paraphrased - into every worker, checker and fix-round prompt. gate:PROSE-ONLY (mechanism in flight: T-372) - MUST add the HOLD-LABEL line whenever a
dod:needs the PR to stay open. gate:PROSE-ONLY (mechanism queued: T-372) - MUST land ALL work via a PR gated on the repo's CI, and the WORKER never merges or closes any
PR on any repo: PUBLIC repos arm auto-merge, PRIVATE ones (free plan, no branch protection)
never do - the worker ends at
gh pr checks --watch, dispatcher/checker merges. Dod boilerplate telling a worker to merge is a DEFECT. gate:PROSE-ONLY (mechanism queued: T-372) - MUST verify PR-state post-run for every task - a PR merged inside the worker's run window is a task FAILURE independent of work quality, checked by dispatcher AND checker. gate:PROSE-ONLY (mechanism queued: post-run-pr-check.ps1, T-372)
- MUST run the pre-queue dedup gate over ALL open contracts for the repo before queueing. gate:PROSE-ONLY
- MUST batch same-repo trivia from one intake into ONE contract (one T-id, one PR). gate:PROSE-ONLY
- MUST pass the PORTFOLIO REGISTRATION GATE before queueing, and create project folders only under the work root. gate:PROSE-ONLY
- MUST run every abort-capable check at INTAKE while the owner is present, and grill to >95% confidence of WHAT is asked; MUST NOT ask what GLOBAL.md, the repo or the registry answers. gate:PROSE-ONLY
- MUST run no fleet actor (worker or checker) unattended under interactive permissions. gate:PROSE-ONLY
- MUST keep maker != checker: evidence + LEDGER are checker-written only, a self-reported pass is never proof, a verdict with no re-derivation artifact is no verdict, and a missing evidence folder is a task FAILURE. gate:PROSE-ONLY
- MUST branch on
stop_reason- a refusal is not success; reroute to opus once, then park. gate:PROSE-ONLY - MUST honor origin-session reporting affinity: only the origin session, or a labeled keeper rescue of a provably-dead origin, reports or acts on a task. gate:PROSE-ONLY
- MUST show the owner a queue ack at dispatch AND the terminal status card when the watcher fires, and MUST apply the OWNER STATUS CADENCE: a PERSISTENT IST-stamped 15-minute ticker while work is live. gate:PROSE-ONLY
- MUST send a terminal-state card as FALLBACK for session-origin tasks and PRIMARY for origin-less
ones - and on a non-VPS machine push the bus (via
bus-write.ps1, T-384) after writing the ping. gate:PROSE-ONLY - MUST reprioritize by REORDERING only - a running worker is never preempted. gate:PROSE-ONLY
- MUST honor the deploy tier computed from the ACTUAL merged diff - auth/payment/DNS/migration force-upgrades to HOLD. gate:PROSE-ONLY
- MUST append the INVOCATIONS.log line at the end of every intake turn using the defined
tids=tokens - onlytids=noneis the inline defect. gate:PROSE-ONLY
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

