forge-router

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.

Step 0 — Read project memory + ECC inventory first (required)

Before routing, read (if present): .claude/FORGE_PROJECT_PROFILE.md, .claude/FORGE_MEMORY.md, .claude/FORGE_TASK_HISTORY.md, .claude/FORGE_ECC_MODE.json. Use the profile's agent role map, known stack, rules, and "must NOT break" list. Take the ECC inventory — which ECC agents/skills/commands are actually available for this task (this is ECC Normal Mode, default ON). If no profile exists, infer one from the files and create it. Never fake project knowledge or ECC availability — mark inferred/unknown.

Step 0-owner — Owner Prefs, Standing Rules & Autonomy (required, before intake) — WAVE B / B4

Read the owner-governance stack in config/orchestration/precedence.md order (hard-gates > current owner instruction > standing-rules + owner config > owner-profile defaults > auto-defaults — never re-derive this ordering here, that file is the single source of truth): (1) forge-bin/forge-prefs.cjs::resolve()/list() — the owner-profile defaults (.claude/FORGE_OWNER_PROFILE.json + the owner-write-only global copy + an optional env override; merge order project < global < env); (2) forge-bin/forge-standing.cjs::match({type, paths}) — the active, evidence-backed standing rules for this run (advisory, injected into dispatch prompts as ADVISORY OWNER CONSTRAINTS, never a silent block); (3) forge-bin/forge-autonomy.cjs::decide() — which continue-within-mission mode applies (opts.mode > the owner config autonomy > the FORGE_AUTONOMY.json default), always deferring to the live hard-gate classifier first; at a phase boundary use decideLive() / decide "<text>" --phase --live, which also stops on a real usage-guard pause; (4) the owner config (/forge config, v2.7.0) — owner config = tier 3, next to the standing rules: right after the run id exists run node .claude/forge-bin/forge-config.cjs diff --run <run_id> --mark-seen; exit 3 means a setting changed since the last run (it logged ONE config_changed event) — repeat its lines to the owner in the first message ("Je hebt usage-guard.pause-at op 97 gezet — toegepast.") and continue. Every later step reads its own key with node .claude/forge-bin/forge-config.cjs get <key> (or the module's get(key)), never by opening FORGE_CONFIG.json by hand. Then compose and log the applied-prefs ECHO — node .claude/forge-bin/forge-echo.cjs emit <run_id> (or the module API forge-echo.cjs::emitEcho()) — a single owner_prefs_loaded event summarizing which prefs/rules actually applied, so precedence.md's ordering is VISIBLE on the dashboard/run log, not merely documented. Re-run forge-echo.cjs compose --type <domain> after Step 1's domain classification below if a domain-scoped standing rule should be re-checked with the now-known type. Never fabricate a pref/rule — an unresolved layer degrades honestly (see forge-prefs.cjs's per-layer notes); nothing here promotes a staged FORGE_PREF_CANDIDATES.json entry to active — that requires a real owner /forge remember (forge-standing.cjs::remember()).