forge-report

Produces the standard end-of-task Forge report — project adaptation, agent ledger, memory update. Use when delivering the result of any non-trivial /forge build, fix, or review.

SKILL.md

GitHub

Forge delivery report

Be concrete and honest; reflect only checks that actually ran. Never claim agents/Codex/tests/preview that did not really happen. Do not say "done" if it isn't. Include all sections below.

0. Chat reply vs the file on disk (usage-minimisation)

The 14 sections below are the CONTRACT for .claude/forge-runs/<run_id>/final-report.md — never trimmed, never skipped, always complete on disk. What actually goes in the chat reply is a SHORT summary only: verdict (Section 13's one-line pick) · what changed (2-4 lines, from Sections 3-4) · checks run (Section 5's real commands, one line each) · blockers (Section 11, or "none") · next step (Section 14) — then the exact path to the full report for anyone who wants the rest. Never re-paste the whole file into chat by default; the owner can always open the path. A small L1 fix may compress this into 3-5 plain lines. A large/client deliverable still gets this short chat form plus the complete file — never a full re-paste in chat instead of writing the file.

1. Classification

task type · complexity (L#) · START NEW or CONTINUE · active project folder · fan-out level used.

2. Project Adaptation

detected project type (incl. any custom taxonomy for unknown project types) · detected stack · selected agents · custom subagents created (+ why) · optional agents · agents not used and why · playbooks loaded · why this team fits the project.

2b. Mission Blueprint + Skill Discovery

  • Blueprint: one-line expanded mission · quality target · do-not-touch list · definition of done.
  • Skill Discovery: required capability areas · skills selected · skills assigned to subagents · custom skills created (path) · skills not selected + why · missing skills · fallback plan. Confirm every executable subagent had an assigned skill/method (or honest BLOCKED).

3. What was done

Plain summary.

4. Files changed

Paths (all inside the project), one-line note each.

5. Checks/tests run

Actual commands + results only. No fabricated output. Before declaring a run complete, run node .claude/forge-bin/forge-verify.cjs <run_id> and report its result; a mismatch means rework, not a pass.

6. Agent Activity Ledger (required table — evidence, no fake claims)

Each row has a Runtime (ECC agent · ECC skill · native/main · Codex) and a Status. Statuses ONLY: ECC REAL INVOKED · ECC SKILL LOADED · NATIVE AGENT INVOKED · INTERNAL ROLE ONLY · NOT USED · FAILED · BLOCKED.

Agent Custom? Runtime Skill / source Status Project role Task Files read Files changed Evidence Result

If a real ECC subagent was dispatched → ECC REAL INVOKED (Runtime = ECC agent). If an ECC skill/playbook ran → ECC SKILL LOADED (Runtime = ECC skill). If a native/main-session agent did the work as a labeled fallback → NATIVE AGENT INVOKED (Runtime = native). A thinking role → INTERNAL ROLE ONLY. ECC tried but unavailable/blocked → BLOCKED. Mark Custom? = yes for Lead-created custom roles and give the assigned Skill / source (ecc-skill/forge-skill/project-local/native/internal). Never label a native role as ECC, or claim a custom subagent/skill that wasn't really created. Don't claim parallel agents unless they really ran separately. Evidence = command output / file diff / tool result / commit hash / log line / created file — or an explicit note that evidence is unavailable and why.

6c. Rework loop (if any)

rework tasks created · how many fixed · retests run · loop iterations (max 2 / 3 high-end) · anything left blocked. No fabricated fixes.

6b. ECC Status block (required — ECC-first honesty)

  • ECC Normal Mode: ON / OFF / BLOCKED (from FORGE_ECC_MODE.json)
  • ECC Full Test Mode: OFF / ON (opt-in only)
  • ECC attempted: yes / no
  • ECC agents selected / invoked: <list> / <list>
  • ECC skills loaded: <list>
  • ECC succeeded: yes / no / partial
  • Native fallback used: yes / no — and which agents
  • ECC blocked reason (if any): ECC unavailable · ECC skill not found · ECC agent failed · task too small (no ECC needed) · … Be honest: if Forge used native roles because ECC was unavailable/blocked/too-small, say exactly that — never present native work as ECC.

7. Memory Update

  • FORGE_PROJECT_PROFILE.md: created / updated / unchanged
  • FORGE_MEMORY.md: created / updated / unchanged
  • FORGE_DECISIONS.md: created / updated / unchanged
  • FORGE_TASK_HISTORY.md: created / updated / unchanged
  • FORGE_AGENT_LEDGER.md: created / updated / unchanged
  • Summary of memory changes · any uncertain/inferred memory.

8. Dashboard Update (no fake starts)

The dashboard is the Command Center (http://127.0.0.1:4100, one dashboard for every project); the old per-project Control Center was ** ** (not merely retired, the file and its UI no longer exist) and only its log-event.cjs stays in service as the run-event writer. Report the Command Center first:

  • Command Center health check: passed / failed / not run (GET http://127.0.0.1:4100/api/health — the report may only say the dashboard runs if this passed)
  • Command Center URL: http://127.0.0.1:4100 (only when the health check passed)
  • latest run id: <id> · events written: <n> (via .claude/forge-dashboard/log-event.cjs)
  • Forge Session Mode: on / paused / off (from FORGE_SESSION_STATE.json) If the owner explicitly said legacy dashboard this run: there is nothing to start or report on any more — say in one line that the old Control Center was and the dashboard is the Command Center above. Allowed wording: "Command Center health check passed (4100)." · "Command Center not reachable — Forge tried to start it itself (supervisor) and it is still not answering." · "Command Center not installed in this project (optional) — continuing without it." · "Events written via log-event.cjs; dashboard not started (not requested)." Never print a start/build command for the user to run, and never claim a URL is live unless tested/confirmed.

8b. Command Pack Update (include on install / when it changed)

  • forge-bin installed: yes/no · PowerShell scripts: installed/not · CMD scripts: installed/not · Bash scripts: installed/not
  • package.json scripts: added / skipped (no package.json) / conflict
  • VS Code tasks: added / example file / skipped
  • commands tested: <list> · commands not tested: <list>
  • how to start dashboard — PowerShell: .\.claude\forge-bin\forge-dashboard.ps1 · CMD: .claude\forge-bin\forge-dashboard.cmd · Bash: bash .claude/forge-bin/forge-dashboard.sh (Forge defines no npm scripts — never document any)

8c. Project governance (CLAUDE.md + custom skills)

  • CLAUDE.md: created / updated (safe-merge) / existed·no change / conflict — and what existing instructions were preserved (never silently deleted).
  • Custom project-local skills: .claude/skills/<name>/SKILL.md created / updated / used (by which subagent) / skipped / conflict. Each is real, documented, and linked to a work package.
  • Four things maintained: CLAUDE.md (brain) · .claude/skills/* (capabilities) · Forge memory · dashboard.
  • Conflicts (if any) are listed here and were owner-approved before any deletion.

9. Issues found

10. Issues fixed

11. Remaining risks

12. Codex block (exact state, never faked)

  • Codex considered / invoked: · Codex state (one of): CODEX REAL INVOKED · CODEX BLOCKED: TRUST/TTY · CODEX BLOCKED: NO GIT · CODEX BLOCKED: NO OUTPUT · CODEX NOT AVAILABLE · CODEX NOT INVOKED · CODEX FALLBACK USED.
  • If blocked: cite artifacts/codex-unlock-diagnosis.md (CLI/version/auth/git/TTY/trust/exit-code) and state in one plain line that Codex was not run and why — never hand the user a manual command to run; the diagnosis artifact keeps the attempted command only as an internal record. Never claim Codex proof without real output.

12b. Browser proof + Skill Registry + Fresh-install

  • Browser proof: tool used · dashboard URL · run id · lenses checked · screenshot paths · layout observations · pass/fail — or PARTIAL with the exact reason (no Playwright/Chrome). Cite artifacts/dashboard-browser-proof.md. Never claim a screenshot you didn't take.
  • Skill Registry: .claude/FORGE_SKILL_REGISTRY.md created/updated/checked + entries touched.
  • Fresh-install verification (if run): CLAUDE.md · skills dir · registry · memory · session · dashboard starts · ECC Normal ON/Full Test OFF · project-local only → artifacts/fresh-install-verification.md.

13. Final verdict (strict)

Pick ONE overall level + give the per-layer verdicts. Do not use FULL PASS if a required Codex/browser proof was blocked.

  • FULL PASS — all critical requirements proven; Codex proof passed or not required; browser proof passed if dashboard changes were tested; no critical PARTIALs.
  • PASS CORE / PARTIAL PROOF — core Forge behavior passed; one or more proof layers missing; no safety/isolation failure.
  • PARTIAL — an important required feature unproven/blocked; fallbacks used; manual follow-up needed.
  • BLOCKED — cannot safely/honestly complete. A BLOCKED status requires a real recovery ledger, not just an assertion: node .claude/forge-bin/forge-recovery.cjs check-block <record.json> must report ok:true (≥3 alternatives attempted, ≥5 for a high-value item, a Verify-Agent verdict, non-empty queries/tools). While check-block reports ok:false, BLOCKED is not yet a valid final status — keep attempting safe alternatives first (GLOBAL_RESEARCH_RECOVERY_POLICY.md).
  • FAIL — critical requirement failed · safety/isolation violated · fake claim detected · required artifact missing.

Per-layer (state each): Core · Codex · Browser proof · Dashboard · Project isolation · Overall.

14. Next step

Single most useful next action.

The FILE on disk keeps every section above in full. The CHAT reply stays short — see Section 0.