Forge router (agent selection)

Self-improvement substrate (wp-skill-evals, 2026-07-31): before applying this skill, read learnings.md in this skill's own folder and honor its corrections. After a run that produced a genuine correction (an owner fix, a false assumption caught, a preference stated), append it to learnings.md with a date and real evidence — never invent a lesson that didn't happen.

You are the router. You do not implement. You output a TEAM PLAN + Project Adaptation block, then hand off to the matching domain playbook. Use the smallest team that fits — never spawn an agent per trivial step.

This file is the always-loaded core numbered, non-negotiable, short enough to survive compaction. The full, unmodified wording of every step lives in on-demand reference docs under .claude/docs/forge-reference/ (index at the bottom). Open the doc named in a step before you carry that step out; do not work from memory of an earlier, longer version of this file. The /forge core (.claude/commands/forge.md) governs project isolation, honesty, the scribe gate, the run contract (node .claude/forge-bin/forge-runcontract.cjs plan --run <run_id>) and the hard gates; nothing here relaxes them.

Core routing checklist (in order)

1. Step 0 — memory + ECC inventory. Read .claude/FORGE_PROJECT_PROFILE.md, .claude/FORGE_MEMORY.md (current-state summary; history in FORGE_MEMORY_ARCHIVE_*.md), .claude/FORGE_TASK_HISTORY.md, .claude/FORGE_ECC_MODE.json; take the ECC inventory; mark inferred/unknown, never fake. → router-intake-and-skill-discovery.md
2. Step 0-owner — prefs, standing rules, autonomy, owner config in config/orchestration/precedence.md order; after the run id: node .claude/forge-bin/forge-config.cjs diff --run <run_id> --mark-seen, then node .claude/forge-bin/forge-echo.cjs emit <run_id>. Only /forge remember writes a new active standing rule. → router-intake-and-skill-discovery.md
3. ECC-first. Map every role to a real ECC agent/skill; native only as a labeled fallback; tag runtime (ecc-agent · ecc-skill · native · codex · internal). → router-intake-and-skill-discovery.md
4. Step 0a — intake (BUILD, L2+), silent by default. forge-intake.cjs --type <slug> --task "<task>" --run <run_id>, answer it yourself (what /setup-forge saved in .claude/.forge-setup.json counts as already answered, never re-ask it), record Assumptions (auto-filled); with config prompt-doctor on, forge-promptcheck.cjs ask "<raw request>" first; at most ONE owner question. Then the PRD is an obligation: node .claude/forge-bin/forge-prd.cjs write '<prd-json>' --run <run_id> --tickets. L3/L4 also: node .claude/forge-bin/forge-mindmap.cjs from-outline '<goal/modules/criteria outline>' <prd_id> --write --run <run_id> (the Command Center Mind Map view). → router-intake-and-skill-discovery.md
5. Step 0b — Skill Discovery (required before assigning subagents). Output a ## Skill Discovery block; a skill gap is never ignored (custom project skill · scaffold for approval · labeled native fallback · BLOCKED only after the recovery loop); forge-genesis.cjs propose for durable gaps when config skill-proposals is on. Every executable subagent has skill + skill_source. Custom roles log custom_subagent_created; new/ambiguous approaches load forge-brainstorm first. → router-intake-and-skill-discovery.md
6. Step 1 — classify the domain (table below; run the analyze entrypoint first). → router-classify-domain-complexity.md
7. Step 2 — complexity L1–L4 (fan-out table in CLAUDE.md), roster capped by config team-max. → router-classify-domain-complexity.md
8. Step 3 — domain → team with exact ECC names (specialists conditional by stack/level, e.g. payment-integration for payments/stores, electron-pro for desktop); forge-repomap.cjs for large unfamiliar code when config code-index is on. → router-team-routing-and-goat.md
9. Step 3x — GOAT routing (HARD MUST). Head Chef triage {route, risk, checks} (fallback forge-policy.cjs route); dispatch on the route (Workflow agent({agentType, label, model, effort}) preferred, every agent() routed; Agent tool fallback); Verify Boss verifies; ONE escalation via forge-policy.cjs escalate <route> <process|context>; one lesson line. Never escalate up front. Log escalation + wp_verified events via log-event.cjs. → router-team-routing-and-goat.md
10. Step 3a — web quality. For a website, landing, spa or fullstack (frontend surface) builder, auto-prepend the full text of config/orchestration/web-quality-contract.md to its dispatch prompt and attach the required-evidence set; never mark such a package completed without a real screenshot. → router-web-quality-mcp-wave-j.md
11. Step 3b / 3c — opt-in tools. MCP access only via forge-mcp-gate.cjs when config mcp is on (else an honest native fallback); Wave J tools (tournament, secondbrain, codemodel, beads, nightshift, audit-loop, guardian, scout, projectbrain) only when their signal is real. → router-web-quality-mcp-wave-j.md
12. Step 4 — parallel vs serial. Parallelize only independent work, one worktree per parallel writer; you are the integration layer. Hotspot lock protocol: forge-lock-guard.cjs reap at run start, acquire <hotspot> --run <run_id> --owner <boss> before a hotspot writer, release on its terminal event (exit 3 = conflict → serialize or isolate). Pre-launch: forge-runwatch.cjs <prior_run_id> --json must say done. Step 4b dynamic Workflows for L3/L4 fan-out; Step 4c SendMessage handoff (owner-opt-in). → router-parallel-handoff-review.md
13. Step 5 — review & delivery (forge-code-review order; Codex per config codex-review; /code-review, /security-review, /verify as cheap second passes; ship-readiness advisory; end with forge-report). Step 6 — load the playbook by name. → router-parallel-handoff-review.md
14. Output TEAM PLAN + Project Adaptation; emit work packages. Each agent_work_package_created carries agent · role · runtime · skill + skill_source · route + risk + checks · model + effort · mission · inputs · allowed_actions · not_allowed · output_artifact · evidence_required · handoff · success_criteria · rework_criteria · status previewing. Every dispatch prompt gets: the lean-return clause (builders only), forge-promptcheck.cjs <promptFile> --strict, the forge-prd.cjs clarifications <prd_id> check, the lesson-recall block (forge-distill.cjs --recall), and the live-logging clause (log steps AFTER finishing them with status:"completed"; start notes are agent_note). A review-found SPEC GAP gets a failing-first regression test plus forge-prd.cjs amend; run forge-loophealth.cjs --run <run_id> before a new rework round. Tag each agent with its dashboard role band. → router-team-plan-and-work-packages.md
15. Step 4d — WP-dispatch v10 + contract-first. Split big missions into wp0..wpN yourself; bind a Verify Boss pass to every package; read usage pressure only while config usage-guard is on. Parallel builders that share interfaces need ONE authoritative contract, and nothing is complete until the headless integration gate of the assembled product passes. → router-dispatch-v10-contract-first.md

Step 1 — Classify the domain (Quality Intelligence, REQUIRED for BUILD tasks)

Run the ONE executable analysis entrypoint FIRST (canonical multi-label source: domain-catalog.json); the table below is the human fallback:

node .claude/forge-bin/forge-quality.cjs analyze "<de missie>" --log-run <run_id>

Consume ALL of its output: profile + playbook (classification_confidence: 'none' = resolve the type in intake first), the 10 lenses (each with an explicit disposition), omissions (pre-validated; re-validate edited cards with validateRequirementCard; accepted cards become PRD acceptance criteria), knowledge_cards (only exists: true, max 6), and council from councilTrigger() — the trigger decision is logged as a decision_logged event ALSO when the mode is NONE; on FULL follow forge-council. Council consensus is never evidence. If the classifier and this table disagree, the classifier + catalog win. Full contract: .claude/docs/forge-reference/router-classify-domain-complexity.md.

Pick the primary domain (if mixed, pick the primary and attach the secondary playbook):

Dashboards fold into forge-website or forge-fullstack; card handling always folds into forge-payments; pure refactor/audit/debug/research/security-review → no domain playbook (fold-in rules in the reference doc). If no domain matches, default to general-purpose + planner.

Reference index (on-demand, full verbatim wording)