godmode
Autonomous product factory. A finite programme delivers a bounded roadmap. A perpetual programme continuously pairs independent models to discover evidence-backed value, create and execute goals, run full regression, repair confirmed fallout, and create the next progression until a hard safety, budget, or deadline boundary fires.
- License 许可
- MIT
- 最近更新
- 2026-11-12
- Stars
- ★ 1
- Forks
- 1
01 前言 Front Matter
- name
- godmode
- description
- Autonomous product factory. A finite programme delivers a bounded roadmap. A perpetual programme continuously pairs independent models to discover evidence-backed value, create and execute goals, run full regression, repair confirmed fallout, and create the next progression until a hard safety, budget, or deadline boundary fires.
02 资源文件 7
- 资产文件assets/programme-dashboard.html4.7 KB
- 参考文档references/charter-template.md5.8 KB
- 参考文档references/lessons.md5.2 KB
- 参考文档references/stage-workflows.md12.2 KB
- 参考文档references/state-and-beads.md8.2 KB
- 参考文档references/verify-lanes.md4.8 KB
- 脚本scripts/beads_remote.sh3.1 KB
03 SKILL.md 正文
godmode — the autonomous product factory
A runtime orchestrator that OWNS the loop. It composes existing machinery —
/loop (driver), Workflow (staged ultracode), beads (state), /here-now
(dashboard hosting), tmux-verify/browser-harness/Playwright (verify lanes) —
into one self-driving programme. You are the driver each tick; these files are
your constitution and playbooks.
Entry modes
| Invocation | Action |
|---|---|
/godmode <north-star> --mode finite|perpetual [--approval none|roadmap] [--no-court] [--budget <tokens>] [--deadline <ISO>] (alias: /godmode init <north-star> ...) |
INIT: generate charter + state + dashboard + beads, then begin finite delivery or perpetual evolution |
/godmode run [--take-over] |
Resume/continue the loop from state (any session, incl. post-crash). Claims the driver lease first: refuses while another session's lease is fresh, auto-claims a stale one, --take-over forces after an AskUserQuestion confirm. On a fresh machine: sync.sh discover, pick the slug, claim, then sync.sh adopt <slug> reconstructs local state |
/godmode status |
Standup: read state + beads + dashboard + lease, report, change nothing. NEVER adopts (observer-safe: no scratch state means hooks stay inert). Preflight prints "status publishing DISABLED: missing <X>" when here-now/credentials are absent |
/godmode pause |
Stop re-arming the loop; state stays resumable; release the lease (lease.sh release) |
Provider parity
| Capability | Claude | Codex | Copilot |
|---|---|---|---|
| init / run (drive the loop) | native scheduler | native Codex automation required | external scheduler adapter required |
| status / discover | yes | yes | yes |
| dashboard publish + sidecar sync + lease guard | all hooks | hooks.json (shared format) | sessionStart pull + staleness nudge |
Never claim a provider is driving when its scheduler or peer-model adapter is missing. Record the capability result in state and defer that lane visibly.
The three-layer machine
LAYER 1 CHARTER .agents/goals/<slug>-charter.md — constitution, re-read EVERY tick
LAYER 2 DRIVER /loop dynamic mode — ScheduleWakeup ~600s, same prompt re-entered
LAYER 3 WORKFLOWS one ultracode Workflow per stage — resumable via resumeFromRunId
STATE beads (epic/feature nodes) + .agents/scratch/<slug>-state.json (working
copy) + hook-maintained sidecar mirror {state,charter,lease} on the
dedicated ref refs/godmode/<slug> (cross-machine; GODMODE_SYNC=local opts out)
Pipeline (stage names are the primary vocabulary; W-numbers are aliases)
SIGNAL TTL ONE model gathers CITED external evidence: comparable
│ products, user-reported gaps, constraining standards.
│ Reads the append-only ledger first and returns only what is
│ NEW; parked rows keep their reason and are never re-proposed.
│ /research if available, else WebSearch+WebFetch. Charter-gated;
│ unavailable degrades to repo-only discovery, never a failure.
▼
DISCOVER once BRAIN brainstorms the nirvana landscape from the
│ north-star + a repo scan + the SIGNAL rows + any existing
│ backlog → writes the feature REGISTRY
│ (.agents/plans/<slug>-registry.md), provenance per row
│ (repo | signal | user). User lists appended, not replaced.
▼
FEASIBILITY COURT (W0) once every REGISTRY idea → verdict: feasible | downgrade | park+blocker
│ skipped only by explicit --no-court (registry taken verbatim)
▼
ROADMAP (W1) once cluster survivors → ordered epics + parallel-pair candidates,
│ seed beads (one per epic + per feature)
▼
[OPTIONAL ROADMAP GATE] only when approval_policy is roadmap
▼
per epic, serial on stacked branches:
PLAN (W2) planner → adversarial review → revise → VERIFY THE REVISE EDITED THE
│ PLAN FILES (mtime/content) — see references/lessons.md
▼
EXECUTE (W3) build (model pair per workstream) → pair review → adversarial epic
│ review (+fix) → VERIFY (surface lane, references/verify-lanes.md)
▼ → bounded fix loop (≤2) → build gate (unpiped!)
SHIP stacked PR (labelled for review) → close epic+feature beads with
│ evidence notes → dashboard update
▼
PERPETUAL after every ship: full regression + refresh SIGNAL if stale +
next discovery/plan. Backlog-dry refreshes SIGNAL first and
becomes adaptive evidence research, never DONE
TERMINATION finite backlog-dry | --budget exhausted | --deadline reached |
security, production-safety, authority, or lease loss
Perpetual progression: after each epic ships, start cumulative regression in an isolated lane. At the same time, the completeness critic re-enters Discover and creates evidence-backed candidates for the next generation, refreshing the external Signal first when it is stale, so a dry backlog is replenished from new outside evidence rather than from re-reading the same repo. The refresh appends to the signal ledger and dedupes against every row ever recorded, parked ones included, so the programme does not pay to rediscover a dead end it already rejected. A confirmed defect preempts the single mutation lane. Discovery and planning continue in parallel. An empty Court result enters adaptive-backoff research, never DONE.
Creative quorum
Every creative decision has two independent model perspectives: Discover, Feasibility Court, Roadmap, completeness critic, incident root-cause analysis, and strategic reprioritisation. Signal is not one of them: gathering cited external evidence is retrieval, so it runs on one model and is judged by its citations, not by quorum. Resolve the route before launching work with:
${CLAUDE_PLUGIN_ROOT}/scripts/programme_policy.py creative-route <availability.json>
- Claude primary pairs a Claude creative model with
codex:codex-rescue. - Codex primary pairs a Codex model with Claude Fable.
- Copilot primary uses two distinct Copilot models, then enriches with Claude or Codex when available.
- Fallback uses any two distinct available models. One model means creative work is deferred, while regression and grounded work continue.
- Persist both proposals, model identifiers, evidence, disagreement, and the synthesis. A third model synthesises when available; otherwise use the evidence rubric and preserve dissent.
Parallel epics: allowed ONLY if blessed AND file-sets provably disjoint AND worktrees available; otherwise serialise (single worktree = single checkout).
INIT procedure
- Parse the north-star + flags. Derive
<slug>. - Generate the charter from
references/charter-template.md— fill outcome, model policy (resolve fable toggle), verify doctrine, termination, stop rules, dashboard slug. Write.agents/goals/<slug>-charter.md. - Write
.agents/scratch/<slug>-state.json(schema:references/state-and-beads.md) withdriver_session_id: nulland aphase_sincetimestamp. You cannot know your own session_id: the PostToolUse hook fillsdriver_session_idfrom the hook event on this very write (first-writer-wins), which is why the write must precede the claim. Then claim the driver lease:bash ${CLAUDE_PLUGIN_ROOT}/scripts/lease.sh claim <repo> <slug>. - Dashboard publishing is HOOK-OWNED: every state.json write triggers
render + publish + sidecar push (PostToolUse). Your duties: keep
state.json truthful, write
current_note, pick a stabledashboard_slug. Publish failure never blocks (pending marker + dashboard banner; hooks retry on the next state write). - Run DISCOVER (template:
references/stage-workflows.md) → writes the registry. Then, unless--no-court, the Feasibility Court workflow, then Roadmap. With--no-court: registry taken verbatim, straight to Roadmap. - If
approval_policyisroadmap, present the roadmap, sethuman_gate: "pending", and wait. Otherwise sethuman_gate: "none"and start the loop immediately. Perpetual mode normally usesnone.
LOOP protocol (every wake — this is the driver contract)
- Read the charter + state.json. They outrank your memory of last tick. Run
programme_policy.py validate-state <state>before any side effect. If an optionalhuman_gateis pending, do not re-arm. - Check the running workflow (task output file / TaskList). If done: read its
result AND its journal if the result looks off; persist artifacts; verify
commits landed per the commit policy; advance the state machine (consult
bd readyfor dependency order when picking the next epic); launch the next stage workflow. - Heartbeat + lease: run
${CLAUDE_PLUGIN_ROOT}/scripts/lease.sh refreshEVERY wake, state changed or not (quiet multi-tick workflows must not starve the lease). Exit 5 or a<slug>-lease-lostmarker: post a handoff note, downgrade to read-only, do NOT re-arm. The dashboard republishes itself on every state write (hook-owned); a pending-publish marker means degraded publishing, mention it and keep going. - Update beads (recipes in
references/state-and-beads.md), state.json, and the token ledger (accounting rules in the same file — fail CLOSED if spend is unmeasurable under a --budget). state.json writes go through the Write/Edit TOOL only (Non-negotiables). On a phase flip also stampphase_since; after SHIP / HUMAN_GATE / DONE write the phase explainer (/explain-to-me) and publish it via${CLAUDE_PLUGIN_ROOT}/scripts/explainer-publish.sh(the Stop gate blocks your stop until its receipt exists). - Ask
programme_policy.py next-action <state>which lane owns the next action. Confirmed defects own mutation, shipped epics queue regression, and perpetual empty backlogs refresh Signal then queue research. Re-arm: ScheduleWakeup ~600s, reason = current phase, prompt = the DRIVER RE-ENTRY PROMPT verbatim (below). Honour STOP RULES first.
Driver re-entry prompt (the exact string for every ScheduleWakeup)
/godmode run — programme charter: .agents/goals/<slug>-charter.md, state:
.agents/scratch/<slug>-state.json. Follow the charter's LOOP PROTOCOL exactly.
Never improvise a different prompt; this string is what makes every future wake re-enter this skill and re-read the constitution.
STOP RULES (halt the loop + page the human)
Page = PushNotification + a red banner note on the dashboard, then stop re-arming. Triggers:
- Budget or deadline is exhausted, or budget spend is unmeasurable under a cap.
- Security, production-safety, lost lease, or missing required authority.
- A declared production deployment breaches its health policy after rollback.
Repeated workflow or validation failure is not a global stop. Use bounded retries, create a quarantined incident, release the mutation lane, and keep independent research and planning moving. Any confirmed defect preempts the mutation lane until repaired and fully re-verified.
Model policy (defaults; charter may override any line)
| Role | Default | Fallback |
|---|---|---|
| BRAIN | host primary plus mandatory creative peer | availability route from Creative quorum |
| BUILD | opus + codex pair (codex:codex-rescue agentType), disagreements surfaced not silently resolved |
charter override |
| TEST / VALIDATE | sonnet | charter override |
| SCAFFOLD / mechanical | sonnet | charter override |
VERIFY (≠ validate — the point of this skill)
Auto-detect the epic's surface from touched files + project manifest, then run
the mandated lane from references/verify-lanes.md:
Web UI → real-browser drive · TUI → tmux+VHS frame-truth · API → real requests
- side-effect asserts · library → unit+property(+mutation). Cross-cutting: mock ONLY the human at the input boundary; READ the artefact (never blank-check); evidence uploaded to here.now and linked from the dashboard Evidence tab AND the PR body. In perpetual mode, every shipped epic runs the entire cumulative regression spine before auto-merge. Regression uses an isolated worktree. A confirmed failure pauses unrelated mutation, creates a repair incident, then resumes paused work from a rebased baseline.
References (load when you reach that step)
references/charter-template.md— the constitution scaffold (INIT step 2)references/stage-workflows.md— Court/Roadmap/Plan/Execute Workflow script templates (each stage launch)references/verify-lanes.md— surface detection + the four proof playbooks (Execute's VERIFY step)references/state-and-beads.md— state.json schema, beads JSONL plumbing, resume procedurereferences/lessons.md— hard-won operational catches; consult BEFORE authoring any stage workflow and whenever a stage result looks wrong
Scripts
scripts/beads_remote.sh(skill-local) — close/annotate beads on origin/main via git plumbing without touching any checkout (safe under concurrent worktrees)- Plugin scripts at
${CLAUDE_PLUGIN_ROOT}/scripts/:render_dashboard.py(deterministic dashboard) ·on-state-write.sh(PostToolUse pipeline) ·explainer-publish.sh(receipt writer, the ONLY sanctioned explainer publish path) ·explainer-gate.sh(Stop gate) ·sync.sh(pull|push|adopt|discover) ·lease.sh(claim|refresh|check|release) ·sidecar_remote.sh(refs/godmode/<slug> plumbing) ·staleness-note.sh(Copilot nudge)
Non-negotiables
- Commit policy: atomic single-concern commits, conventional, named paths only
(never
git add -A), no AI attribution, never commit charters/state/ dashboards/scratch/env files, EXCEPT the sidecar mirror onrefs/godmode/<slug>, which hooks maintain via git plumbing. Sign only when the gpg cache is warm; never spawn GUI pinentry headless. - state.json is written ONLY via the Write/Edit tool, never Bash redirection/jq/mv: the status and heartbeat hooks key on the PostToolUse event (belt: the Stop-hook heartbeat entry; a Bash-written state file means a silently stale dashboard and a starving lease).
- Plugin version policy: ANY change under
plugins/godmode/**bumps at least the patch version, all three provider manifests together. - The dashboard is not optional and never goes stale past one tick.
- Stale-work check: before fixing any backlog item, verify it isn't already
fixed in-tree (
git log/show+ live probe). Close-with-citation beats re-fix. - Build gates run UNPIPED to a log file, backgrounded, gated on
EXIT=0.
04 同仓库更多 Skills
well-formed
Fixture skill that satisfies every mechanical rule, used to prove the conformance check passes a clean tree instead of failing everything.
underscore-tools
Fixture skill that spells the tool allowlist key with an underscore instead of the hyphen the spec requires.
no-frontmatter
暂无简介
long-description
Fixture skill with a deliberately oversized description used to prove the conformance check rejects anything past the 1024 character ceiling. Fixture skill with a deliberately oversized description used to prove the conformance check rejects anything past the 1024 character ceiling. Fixture skill with a deliberately oversized description used to prove the conformance check rejects anything past the 1024 character ceiling. Fixture skill with a deliberately oversized description used to prove the conformance check rejects anything past the 1024 character ceiling. Fixture skill with a deliberately oversized description used to prove the conformance check rejects anything past the 1024 character ceiling. Fixture skill with a deliberately oversized description used to prove the conformance check rejects anything past the 1024 character ceiling. Fixture skill with a deliberately oversized description used to prove the conformance check rejects anything past the 1024 character ceiling. Fixture skill with a deliberately oversized description used to prove the conformance check rejects anything past the 1024 character ceiling. Fixture skill with a deliberately oversized description used to prove the conformance check rejects anything past the 1024 character ceiling. Fixture skill with a deliberately oversized description used to prove the conformance check rejects anything past the 1024 character ceiling. Fixture skill with a deliberately oversized description used to prove the conformance check rejects anything past the 1024 character ceiling. Fixture skill with a deliberately oversized description used to prove the conformance check rejects anything past the 1024 character ceiling.
some-other-name
Fixture skill whose frontmatter name does not match its parent directory name.
bad_Name
Fixture skill whose name uses an underscore and a capital letter, so it violates the agentskills.io name pattern.