Do

Tier 3: LLM Complexity Classifier (Cost: ~500 tokens | Latency: ~1-2s)

For every non-exact request, classify intent across 6 dimensions. Use Tier 2 candidates as evidence when present; when none match, classify from the request:

SCOPE: single-file | single-domain | cross-domain | platform-wide
COMPLEXITY: 1 (trivial) | 2 (simple) | 3 (moderate) | 4 (complex) | 5 (campaign)
INTENT: fix | build | create | add | audit | redesign | research | improve | wire | prune
REQUIRES_PERSISTENCE: true | false (multi-session?)
REQUIRES_PARALLEL: true | false (independent sub-tasks?)
REQUIRES_TASTE: true | false (quality judgment beyond tests?)

Routing rules (first match wins):

Ambiguity gate: When multiple Tier 2 candidates survive or classifier confidence is below 0.7, and the AskUserQuestion tool is available: present the top 2-3 candidate routes as options, each with a label and a one-line tradeoff, and route to the user's pick. When unavailable, apply the rules above unchanged (tie-break, then /marshal default).

Important: A repeated pattern complaint ("I keep doing X manually", "the agent always makes this mistake") should route to /create-skill. A repeated pattern is a skill waiting to be extracted.

Step 3.5: Proportionality Check

After classification and before execution, verify the response is proportional to the input:

Downgrade triggers (apply in order):

Upgrade triggers:

Fleet auto-decomposition — 1/2/3 confirmation prompt:

When 2+ independent tasks detected (non-overlapping scopes, complexity >= 3, not already routed to full Fleet), read consent.fleetSpawn from harness.json:

• auto-allow → route directly to /fleet --quick
• always-ask or null → show prompt: "These look independent — run in parallel? [1=yes 2=always 3=no]"
  • 1: route to --quick, preference unchanged
  • 2: route to --quick, write writeConsent('fleetSpawn', 'auto-allow')
  • 3: run sequentially; if "don't ask again", write always-ask

readConsent/writeConsent are in hooks_src/harness-health-util.js. Trust level: Read from harness.json trust object. Levels: novice (0-4 sessions), familiar (5-19), trusted (20+ with 2+ campaigns). trust.override takes precedence.

Step 4: After Classification

1. Enforce product activation before invocation. Run: node {citadelRoot}/scripts/citadel-config.js check route {bare-skill-name} --runtime {claude-code|codex|opencode} --json. enabled and explicitly named degraded routes may continue. For disabled, unavailable, or blocked, show the returned reason and activation plan; do not invoke the target. Enabling a bundle is a separate plan-first config mutation and never happens silently. An explicit --route override changes only route selection. It never skips this activation preflight or any worktree, approval, or verification boundary.

2. Log routing decision (fire-and-forget): node .citadel/scripts/telemetry-log.cjs --event agent-complete --agent do-router --session routing --status success --meta '{"tier":N,"target":"[skill]","input_chars":M}'

   Use .citadel/scripts/telemetry-log.cjs (the project-local copy). If it doesn't exist, skip logging silently — never block routing on telemetry failure.

3. Announce the routing decision: "Routing to [target] because [one-sentence reason]"

4. Invoke the target skill or orchestrator

5. If the target fails or the user says "wrong tool", try the next tier up. If the target is already Tier 3 (marshal fails or user explicitly escalates from a failed marshal attempt): re-route to /archon with the original input as context, then repeat activation preflight for that route.

/do --list

Output a grouped skill list drawn from the system reminder's available skills. Group by category (Orchestration, App Creation, Code Quality, Research & Debugging, GitHub & CI, Infrastructure, Monitoring, Utilities, Observability). For each skill, show /name — one-line description plus its product bundle. Include a footer: "Direct invocation bypasses routing, but not product-bundle activation or safety gates."

Fringe Cases

• .planning/ does not exist: The router works without .planning/. Tiers 0, 2, and 3 are fully independent of it. Tier 1 (active-state short-circuit) reads .planning/campaigns/ and .planning/fleet/ — if those directories are absent, skip Tier 1 gracefully and fall through to Tier 2. Never crash on a missing .planning/ directory.
• harness.json missing: Use the bootstrap Standard + Core + Persistence activation preview. Core/Persistence routes may proceed; other bundles return an explicit activation plan. Do not write config during routing.
• Multiple skills match at Tier 2: Carry candidates to Tier 3 per Tier 2 disambiguation rule above.
• User input is empty or whitespace: Respond with the --list output and a prompt to provide a direction.
• Routed skill not found: Report "Skill not found" and fall back to Marshal as the safe default.

Contextual Gates

Disclosure: "Routing to [skill]. See that skill's contextual gates for reversibility." Reversibility: depends on routed skill — check the routed skill's reversibility Trust gates:

• Any: routing and dispatch; inherits trust gates from the routed skill.

Quality Gates

• Tier 0-2 must resolve in under 1 second
• Tier 3 classification must be transparent (announce reasoning)
• Never route a trivial task (complexity 1) to Archon or Fleet
• Never route a multi-session task to a bare skill
• If routing fails, default to Marshal (safe middle ground)
• A routed target must pass central activation preflight; fallback targets do not bypass this gate

Exit Protocol

After routing and execution complete:

• If the routed skill/orchestrator produces a HANDOFF, relay it to the user
• If the task was trivial (Tier 0), just show the result
• Do not add overhead to simple tasks
• Telemetry is fire-and-forget — never surface telemetry errors to the user