forge

/forge

Input: $ARGUMENTS. Obey CLAUDE.md (project isolation + honesty + no fake claims). Work only in this project folder.

Sub-commands & natural-language triggers (check $ARGUMENTS first)

Also trigger on plain language: "use Forge system", "gebruik Forge systeem", "start Forge dashboard", "open Forge dashboard", "gebruik forge voor dit project".

• dashboard / start Forge dashboard → confirm the active project folder, then find the Command Center — the ONE dashboard for every project (owner decision 2026-07-31; it auto-discovers projects, so it does not need to live in THIS folder): (1) if command-center/gateway/bin.mjs exists in this project, that IS the Command Center home — if command-center/dashboard/dist is missing (never built), print the build command (cd command-center/dashboard && npm install && npm run build) instead of failing silently; prefer the supervisor for long sessions (node command-center/gateway/supervisor.mjs — auto-restarts the gateway) or start the gateway directly (node command-center/gateway/bin.mjs, port 4100); (2) if it does not live here, health-check the central instance anyway — http://127.0.0.1:4100/api/health answering means this project is already covered. Either way: health-check http://127.0.0.1:4100/api/health and report http://127.0.0.1:4100 as the dashboard URL — never claim it is running unless that check passed; never fake a start. Do NOT auto-start the retired per-project Control Center as a fallback — .claude/forge-dashboard/server.cjs is RETIRED (never started automatically in any project); only log-event.cjs in that directory stays in active service as the per-project run-event writer, completely unchanged. If no Command Center is reachable anywhere, say so honestly and print the exact command + expected URL. Also start the usage guard: node .claude/forge-bin/usage-guard.cjs start (idempotent, single global instance).
• legacy dashboard / old Forge dashboard → the RETIRED per-project Control Center, available ONLY on this explicit request (it never starts automatically): check .claude/forge-dashboard/server.cjs exists, start it, health-check http://localhost:<port>/api/health (port from .claude/forge-dashboard/PORT), report that URL, and remind that the Command Center on 4100 is the current dashboard.
• start → same as dashboard, then begin a new run (steps 0–1 below).
• use / use Forge system / gebruik Forge systeem → load project memory/profile; if Forge isn't installed, offer the installer; start (or show how to start) THIS project's dashboard; then continue the task via the flow below.
• status → read .claude/FORGE_MEMORY.md + .claude/forge-dashboard/DASHBOARD_STATE.json + newest .claude/forge-runs/*/run.json; summarize project status, latest run, and dashboard URL.
• runs → list THIS project's .claude/forge-runs/*/ newest-first.
• resume <run_id> (WAVE D) → a session/usage limit killed a swarm mid-run; reconcile that run's manifest purely from its logged events.jsonl content (never inferred from side effects, never a fabricated "done") and re-dispatch only the still-unfinished work packages with their original narrowed scope. Requires the run to have been ARMED at dispatch time — that ARM step is now arm-manifest in .claude/config/orchestration/run-checklist.json (advisory, sits between route and dispatch): node .claude/forge-bin/forge-manifest.cjs arm --run <run_id> --wps <file.json> --log-event — one record per WP: wp_id, agent, narrowed_prompt, deps?. --log-event logs the manifest_armed proof event in the SAME act as the write (2026-08-01), so arming and its proof can no longer disagree; without it you must log manifest_armed yourself or forge-orchestrate.cjs audit will report the step skipped. Once a run is armed, node .claude/forge-bin/forge-manifest.cjs ready --run <run_id> gives the honest frontier and ... waves --run <run_id> gives the full dispatch order in which everything inside one wave is mutually independent — that is the mechanical way to satisfy the owner's "one writer per hotspot at a time" HARD MUST instead of eyeballing the order. Then: node .claude/forge-bin/forge-swarm-resume.cjs --run <run_id> [--json] (or forge.cmd resume --run <run_id> / forge.ps1 resume --run <run_id> / bash forge.sh resume --run <run_id>) — exit 0 = complete (nothing to resume) / 3 = resumable / 2 = usage error or no manifest was ever armed. Re-dispatch each returned unfinished WP exactly as originally scoped and log wp_resumed per WP as it goes back out. A done WP is never re-dispatched. This module has zero side effects of its own beyond persisting the reconciled manifest — a WP whose own work is side-effecting (email/deploy/payment/etc.) still needs its own forge-checkpoint.cjs idempotency guard before repeating that side effect.
• open-report → print THIS project's newest .claude/forge-runs/*/final-report.md.
• learn (WAVE F, opt-in, read-only) → cross-project learning harvest: reads OTHER Forge projects' .claude/FORGE_*.md memory files (never writes to them) and stores real, evidenced lines from them into THIS project's reserved global lesson namespace — the same namespace forge-recall.cjs always blends into every dispatch, so a lesson from project A can now surface as advisory guidance in project B. node .claude/forge-bin/forge-harvest.cjs --scan <portfolio-dir> [--global-store <file>] [--dry-run] [--json] (or --projects <a,b,...> for an explicit list; or forge.cmd learn / forge.ps1 learn / bash forge.sh learn, passthrough args). Explicit discovery only — neither flag given means nothing is scanned; --scan <dir> only checks that dir's IMMEDIATE child folders for a .claude/FORGE_* marker (one level, never recursive, never a whole-disk walk). Secrets/PII excluded: .env/.env.*/*.key/*.pem/*secret*/*credential*/id_rsa* are never opened, and every line read is redacted (reusing the same secret-pattern set the leak scan uses) before it can become a lesson — a line that still looks secret-shaped after redaction is dropped outright, never stored. Evidenced-only: a stored lesson is a verbatim quote from a real logged line (a decision-log row, a "what worked" note, a recurring owner-ask) tagged with its source project/file/line — never synthesised or summarised into a new claim; un-recorded history yields no lesson. Always read-only on other projects — its only write is to this project's own global lesson store. Use --dry-run first to see what WOULD be learned without writing anything. Log lessons_harvested (projects scanned, lessons stored, skipped-secret, skipped-non-canonical) after a real (non-dry-run) harvest.
• remember <text> / "onthoud dat …" (WAVE B / B4) → promote an explicit owner instruction into a real, ACTIVE standing rule: node .claude/forge-bin/forge-standing.cjs remember "<text>" [--scope <scope>] [--trigger always|domain|glob|on-request] [--domain <domain>] [--glob <glob>] [--topic <topic>]. This is the ONLY sanctioned path that writes a new status:"active" rule — it stamps source:"owner /forge remember" (never a fabricated/inferred source) and can never mint a new cannot_override_core rule. Report the written rule id + text back to the owner.
• snapshot (owner request 2026-07-29, .claude/skills/forge-snapshot/SKILL.md) → manual/phase-boundary context-continuity snapshot: node .claude/forge-bin/forge-snapshot.cjs write --reason phase regenerates .claude/FORGE_SNAPSHOT.md from real sources (memory/task-history/work-packages/decisions/git/latest run events/doctor) — Mission preserved verbatim, every other section re-derived fresh, never a fabricated context percentage. This is also fired automatically by the wired PreCompact (matchers manual+auto) hook and re-injected on the following SessionStart (matcher compact) — see HOOKS_OPT_IN.md section 4 for the live wiring record. node .claude/forge-bin/forge-snapshot.cjs check --max-age-hours <n> reports staleness.