메뉴
End-to-end auto-factory driver — picks up beads + GH issues tagged factory, dispatches coders to drive worldai PRs to /green + /advice + /er, runs verifier ticks, repeats until done. Designed to work even when GH API is rate-limited (falls back to beads-only).
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Run the Dark Factory DOT pipeline runner against a goal. Slash command: /factory. Implements StrongDM's Attractor pattern as an external Python runner — .dot files are the versioned artifact, sealed holdouts live in a separate repo, every step is recorded to CXDB, and the Healer clusters failures into diagnoses. Use when you want the goal_harness idea executed as a reproducible external pipeline instead of in-Claude subagent dispatch.
Dark Factory two-phase spec workflow (/factory-spec, /fs): create a main spec (spec.md) AND an attractor spec (attractor_spec.md) via the spec_gen pipeline, review an existing spec of either kind, or display the pipeline node graphs — spec-review pipelines, factory gates, node types, edge conditions, handler mappings. Create mode runs dark-factory --pipeline slim/spec_gen.dot; both the main and attractor specs must be codex-cold-reviewed and pass before exit. Review and show modes are in-session only.
Build a developer reproduction bundle for agent-review failures, missed-plan retrospectives, Claude/Codex transcript comparisons, or cases where Anthropic/OpenAI engineers need raw conversation JSONL, subagent state, repo evidence, hashes, and a concise replay guide.
Run matched A/B reviewer calibration for /f: compare factory reviewer, raw codex exec, and delegated subagent reviews on the same frozen PR/work-item envelope.
| name | auto-factory |
| description | End-to-end auto-factory driver — picks up beads + GH issues tagged factory, dispatches coders to drive worldai PRs to /green + /advice + /er, runs verifier ticks, repeats until done. Designed to work even when GH API is rate-limited (falls back to beads-only). |
The auto-factory is the agent-orchestrator-style system that drives worldai PRs to merge. This skill is its orchestrator: it picks up work (beads + GH issues), dispatches coder subagents, runs verifier ticks, and iterates until gates pass.
H=daemon/factory-overlay.sh
CONFIG=config/daemon.toml
[ -f "$CONFIG" ] || CONFIG=daemon/contracts/daemon.toml.example
$H init # idempotent
The overlay harness daemon/factory-overlay.sh is the canonical executable spec for the auto-factory state machine (restored from e60b5a31b~1:daemon/factory-lite-harness.sh and extended in PR #167). All sqlite3 mutations to ~/.dark-factory/daemon-cxdb.sqlite flow through it. Subcommands: init, intake-upsert, route-record, capacity, dispatch-record, pr-opened, autonomy-tick, gate-assessment, prev-gate-assessment, ready, reroll-verdict, park, park-duplicate, bead-closed-check, tick-summary, recover-held, unstick-dispatching, redrive-pr, list.
Historical note: the original
daemon/factory-lite-harness.shanddaemon/run-factory-lite.shwere removed in commite60b5a31b(2026-07-05, jleechan-xrdx). The decommissioned factory-lite-coder / factory-lite-verifier skills are gone; their protocols are now inline in this skill + factory-af-tick.sh + factory-overlay.sh.
Pick up work from BOTH sources (bead store + GitHub):
br list --status open --label factory --json
For each bead: read body, detect drive-existing-pr mode (fields existing_pr, existing_branch, target_repo). If present, this bead drives an existing PR — coder must push to existing branch via git push wa <existing_branch>. Otherwise default to new-work (create factory/<bead>-r<attempt> branch).
gh issue list --repo "$TARGET_REPO" --label factory --state open --json number,title,body --limit 20
If GH API rate-limited (returns error), skip this step — beads-only mode. Log the fallback: [intake] GH API rate-limited, beads-only mode.
For each GH issue, treat as a bead: read body, file a local bead via br create "<title>" --body "<body>" --label factory --label drive-existing-pr, then continue with bead pickup.
A bead/issue has drive-existing-pr mode if body contains ALL of:
existing_pr: <N>existing_branch: <name>target_repo: <owner>/<name>The coder MUST push to the existing branch via git push <remote> <existing_branch> (NOT create a new factory/* branch). This is the general-purpose pattern.
$H list QUEUED → for each bead, use LLM judgment to pick SMALL_PATH (single-file fix) vs STANDARD_PATH (multi-file/architecture). Record via $H route-record <bead_id> <VERDICT> "<note>".
free=$($H capacity) — min(max_workers - active, max_batch). If 0, skip to step 7.
Select up to $free routed QUEUED beads (file-overlap rule: serialize if any share a mutable file). For each selected bead:
$H dispatch-record <bead_id> <existing_branch> # e.g., fix/quota-banner-modal-cta-7945
$H dispatch-record <bead_id> factory/<bead_id>-r<attempt>
Spawn ALL selected coders in ONE message — multiple Agent tool calls, each subagent_type: minimax-pair-coder, run_in_background: true. Parallel dispatch is the point.
Each coder prompt MUST include:
git push wa <branch> for worldai, git push origin <branch> for dark-factory)/tmp/<bead>-wt; remove after push; never check out branches in the shared repo working tree)Step number kept stable.
For every row in $H list DISPATCHED:
$H bead-closed-check <bead_id> # guards against direct bead closure
For drive-existing-pr beads, check the existing PR branch for new commits:
gh pr view <existing_pr> --repo <target_repo> --json headRefOid,mergeable,statusCheckRollup,reviewDecision,comments
If new commits detected → $H pr-opened <bead_id> <existing_pr> <url>.
For new-work beads, check for new PR:
gh pr list --repo "$TARGET_REPO" --head "factory/<bead_id>-r<attempt>" --state open --json number,url
daemon/src/verifier.rs::GateName (Ci, NoConflicts, CodeRabbitApproved, BugbotClean, <<<<<<< HEAD CommentsResolved, EvidenceFloor, Skeptic, VacuousRedGreen). The code_standards and zfc checks are
optional advisory reviews — they are NOT required keys in the gate-assessment JSON and their
absence never blocks all_green. If you DO record them, they participate like any recorded
gate: a fail verdict blocks all_green and routes through the same reroll-verdict fix
loop. See bead jleechan-1gft for promoting them to required GateName gates in the Rust verifier.CommentsResolved, EvidenceFloor, Skeptic, VacuousRedGreen). The code_standardsandzfcchecks are optional advisory reviews — they are NOT required keys in the gate-assessment JSON and their absence never blocksall_green. If you DO record them, they participate like any recorded gate: a failverdict blocksall_greenand routes through the samereroll-verdictfix loop. See bead jleechan-1gft for promoting them to requiredGateName gates in the Rust verifier. Gate 8 (VacuousRedGreen`, issue #387 / bead jleechan-ijod) was added in
PR #389 / r5 commit 175c6ad — runtime vacuous-test detector verdict.
53a3999 (claude/antig: feat(daemon): runtime vacuous-red-green detector + gate-contract updates (#387))
| Verdict | Meaning | Gate result |
|---|---|---|
pass | Reviewer returned positive evidence; ready to advance | counts toward all_green |
warn | Reviewer returned mixed evidence; non-blocking, surface in next tick | counts toward all_green |
fail | Reviewer returned negative evidence; reroll required | forces reroll-verdict |
unknown | Reviewer could not gather evidence yet; wait for the next tick | blocks all_green |
all_green=true iff every gate returned pass or warn. fail routes through
reroll-verdict → HUMAN_HELD → recover-held → QUEUED — the bounded fix loop
<<<<<<< HEAD
shared with the original 8 gates (7 canonical + VacuousRedGreen); no parallel implementation. unknownshared with the original 8 gates; no parallel implementation. unknown
53a3999 (claude/antig: feat(daemon): runtime vacuous-red-green detector + gate-contract updates (#387)) defers to the next tick rather than racing to READY.
gh pr view <pr> --repo "$TARGET_REPO" --json headRefOid,mergeable,reviewDecision,statusCheckRollup
gh pr checks <pr> --repo "$TARGET_REPO" --json name,state,conclusion
Each gate is a model-delegated review (NOT keyword routing); the verifier
dispatches them as subagents / claude --print /<slash> / codex exec --yolo
against the PR diff. The overlay only records verdicts — the pass/warn/fail
decision is the model's, not ours.
Assess each gate:
conclusion=success (or state=SUCCESS)mergeable=MERGEABLE, mergeStateStatus not DIRTYcoderabbitai[bot] review APPROVEDcursor[bot] commentsisResolved=truecode-standards review (overrides ~/.claude/skills/code-standards/SKILL.md or repo-local .claude/skills/code-standards/SKILL.md) — pass/warn only when every claimed pass carries file/line evidence per the lane's Iron Law; fail on bare assertions without file/line proofzfc review against .claude/skills/zero-framework-cognition/SKILL.md + .claude/skills/root-cause-first/SKILL.md — pass/warn only when no banned-pattern scan finds a violation; fail otherwiseEach gate's verdict value in gates_json is either a plain string OR a
structured object so reviewers can audit the verdict without re-running the
model:
"zfc": "pass" /* shorthand */
"zfc": {"verdict": "fail", /* structured */
"evidence": [{"path": "daemon/factory-overlay.sh",
"line": 201,
"msg": "keyword blacklist enforced in app code"}]}
Record via $H gate-assessment <bead_id> <pr> '<gates_json>'. The 7-key JSON
schema (strict, matching daemon/src/verifier.rs::GateName) is enforced by factory-overlay.sh:
{"ci_green":"pass","no_conflicts":"pass","coderabbit":"pass",
"bugbot":"pass","comments_resolved":"pass","evidence_review":"pass",
"skeptic":"pass"}
// optional advisory keys — not required; if present, a fail still blocks:
// "code_standards":"pass","zfc":"pass"
Legacy aliases "green" → "pass", "red" → "fail" are still accepted so
existing test fixtures don't have to migrate in lockstep, but new callers
should use the pass/warn/fail vocabulary.
All-green (all_green=true on stdout line 1) → $H ready <bead_id> <pr>
(terminal state; verifier stops driving). Any-fail → $H reroll-verdict <bead_id> <pr> <in_place_fixable|reroll_worthy> "<rationale>". The cooldown_ready
line indicates whether the prior GATE_ASSESSMENT for this PR was a false
result (cooldown handling is unchanged from the original 7-gate design).
$H autonomy-tick $ELAPSED_SECS — increment actives, warn at 80%, park over-box.
$H tick-summary coder (or verifier if you ran verifier steps).
$H.base_branch.gh pr merge — dispatch is not the merge authority.dispatch-record.factory/* branch when the bead body has drive-existing-pr fields — push to the existing branch.bash daemon/factory-af-tick.sh for one tick; for a 24/7 poll loop, install the launchd plist from daemon/launchd/ai.dark-factory.af-tick.plist (bead jleechan-57h0) and launchctl bootstrap gui/$UID <plist>.factory-af-tick.sh already calls $H recover-held every tick, which requeues any HUMAN_HELD bead with attempt < 10 back to QUEUED (incrementing attempt, resetting autonomy_secs) automatically. To force it immediately: $H recover-held (no bead-id argument — it processes every eligible HUMAN_HELD row). Never mutate bead_overlay with a raw sqlite3 command.