Auto

auto — Intent Router

The front door to OrchestKit. You describe a goal in plain English; the router classifies it and hands off to the right specialist. One entry point, many execution paths.

Why this exists: OrchestKit has 110 skills, but usage telemetry shows users fire only the handful they can name by memory (10 distinct skills across thousands of sessions). The dominant cause of "dead" skills is no front door — not low quality. This router turns "you must know the exact /ork:<name>" into "describe what you want."

Core principle: routing is a deterministic workflow, not an autonomous agent (Anthropic, Building Effective Agents). Classify → confirm → hand off. The router never does the work itself — it picks who does.

When to use

By default, for any goal-shaped request. An unambiguous goal is a 1-step route: auto classifies, confirms in one line, and hands off — no extra hops, so there is no "too obvious for auto".

Design note (2026-07-12): this table previously said "Go direct when you already know the skill / the request maps unambiguously to one". That inverted instruction made the front door structurally unreachable — a competent model always believes it knows the target, so the router recorded near-zero invocations across thousands of sessions (the exact dead-skill problem the "Why this exists" note above describes). Routers must be framed as the default path, not an escape hatch for confusion.

Intent categories → OrchestKit skill

Full per-category parameter extraction + edge cases: references/routing-rules.md.

Model weight (orthogonal second dimension)

Intent picks who does the work. Weight picks how expensive that worker should be. A route is {intent} @ {weight}. Weight never changes the intent and never replaces it. The taxonomy above and the 7 disambiguation rules are untouched by it.

Tiers are the ones already declared in src/agents/*.md frontmatter (haiku 3 · sonnet 23 · opus 6 · inherit 4). No parallel taxonomy.

Resolution is asymmetric: ANY heavy signal ⇒ Heavy; Light requires ALL light signals; everything else is Standard. Under-powering a security review yields a confident wrong answer nobody catches; over-powering a rename only wastes money.

Weight is per leg, not per route. A PR review can be a Heavy security leg plus a Light lint leg. Full signal table, per-intent defaults, and the honest limits of this lever: references/routing-rules.md.

This is the selector, not the cap. src/hooks/src/pretool/task/team-size-gate.ts is an ex-post, per-session counter keyed on ORK_TEAM_OPUS_MAX (default 8). Its default posture is advisory (outputWarning); with ORK_TEAM_SIZE_HARD=1 it escalates to outputDeny and refuses the spawn outright. Either way it reads the model read-only: it can refuse a premium spawn, but it cannot choose a cheaper one for you. Routing is what chooses. The two compose, cap as backstop and routing as selector; never duplicate the cap's counting here.

The flow

Copy code
  CLASSIFY  ->  CONFIRM  ->  HAND OFF
     |            |             |
  intent       show the     invoke the
  + weight     route        target skill;
  one line     + nod        follow ITS phases

1. Classify

Name the route and the one signal that decided it in a single short line. Example: "'get latency under 200ms' names a metric + a direction, so optimize, not fix."

Apply the disambiguation rules (most specific wins; explicit verb beats inferred intent). The load-bearing one: explicit verb wins — "Fix the slow query" → fix, not optimize. For the full ordered ruleset (all 7, including the truly-ambiguous fallback), references/routing-rules.md is canonical.

Then classify weight in the same pass, naming the signal that decided it: "touches auth and models an attacker → Heavy." Intent first, weight second; a weight call never rewrites the intent you just committed to.

A route: line is a prior, not a verdict. When the prompt context carries one line of the shape route: <class> -> /ork:<skill> (conf 0.xx) or route: <class> -> Agent(<name>) (conf 0.xx), it came from the Jev routing seam (ORK_ROUTE_JEV=steer, off by default, #4233): one typed judgment of the capability class, taken before you read this table. Treat it as the prior for this step, state in one sentence whether you agree and why, and then classify as above. Your reasoning still decides; a confident prior you disagree with is worth a sentence, never a silent override in either direction. The confirm step is unchanged; the hand-off tool follows the target: a /ork: skill goes through the Skill tool, an Agent(...) target through the Agent tool.

2. Confirm (low ceremony)

Show the chosen route in one line and get a nod before handing off:

Copy code
Goal:   "{original goal}"
Intent: {category}
Weight: {Light|Standard|Heavy} ({tier}), decided by: {signal}
Route:  {skill or loop} {extracted args}
        [run] · [adjust] · [cancel]

For low-risk single-pass routes (verify, review), an inline "routing you to verify — ok?" is enough. Never hand off without a nod.

Premium spend is never silent. Routing down (Light/Standard) needs no approval, because spending less is not a decision the user has to make. Routing up to Heavy is premium spend and gets its own line the user must accept:

Copy code
⚠️  Heavy route: {N} opus-tier leg(s), triggered by: {heavy signal}
    [approve premium] · [run Standard instead] · [cancel]

If they decline, run Standard and say plainly which check is weakened. Never upgrade mid-handoff or inside a spawned agent the user did not see.

3. Hand off

Invoke the target skill with the extracted parameters and follow that skill's own phases and guardrails — do not override them. The router's job ends at the handoff; the specialist owns execution and its own report.

A hand-off is a Skill-tool invocation, not a recommendation. The failure mode that motivated M170 (#3127): telemetry traced 43 router hand-offs and found ZERO reached an executor skill (implement, cover, fix-issue, review-pr) — the route was named in chat, then the work happened inline in the main loop, so the executors' specialist wiring (implement → backend-system-architect, cover → test-generator) never activated. Therefore:

• Once the user nods, the SAME turn must contain the Skill-tool call for the routed skill. Never end the routing turn with only a description of what will run.
• Doing the routed work inline "because it's faster" is a routing failure, not a shortcut — the executor's parallel specialists and guardrails are the point of routing.
• If the routed skill genuinely cannot run (missing prerequisite, wrong repo state), say exactly that and stop; do not silently absorb the work into the main loop.

Fallback + honest gaps

• Fallback category. If no category clears a confident threshold, ask exactly ONE clarifying question rather than guessing. A rising fallback rate is the leading indicator that the taxonomy needs work — surface it, don't bury it.
• optimize has no dedicated skill (yet). OrchestKit's metric-driven optimization runs as a /goal loop using the loop recipe library (prd-to-goal → references/recipe-library.md). Route optimize there and say so plainly — don't pretend a experiment skill exists.
• improve-skill routes to the evolution gate. Self-optimizing a SKILL.md goes through the champion/challenger holdout-promotion gate (assess evals + evolution-engine), not a one-shot edit. It requires a benchmark + holdout set first.

Stacked invocation (CC 2.1.199+)

/skill-a /skill-b <goal> loads all leading skills (up to 5) into context at once; the trailing args belong to the whole stack. So auto brainstorm <goal> pre-loads the specialist alongside the router — useful when the user already knows part of the route. The router still owns classification and handoff; a pre-loaded specialist does not bypass the confirm step.

Guardrails

• No recursion. auto must not route to itself, directly or via a spawned agent.
• No bypass. Routing does not skip the target skill's guardrails, readonly enforcement, or confirmation steps.
• A pronoun target is yours to resolve. Pass a concrete target, not this or them: read it back from the conversation and announce it in one line. If a routed skill returns TARGET_UNRESOLVED: <word>, resolve <word> the same way and invoke the skill again; never relay the sentinel to the user (Read("../../shared/rules/target-resolution.md")).
• Classification quality is the whole job. A misroute that fails silently is worse than a fallback question. When two categories are equally plausible, ask — don't gamble.
• No silent upgrade. A Heavy (opus/fable) leg is premium spend and requires an explicit nod on its own line. Downgrades stay silent.
• The cap is not the router's to move. Never read, set, or suggest raising ORK_TEAM_OPUS_MAX. That is the user's budget.
• A gate denial is not a re-route trigger. If team-size-gate denies a spawn, do NOT relabel a Heavy leg as Standard to slip under the cap. Report the denial and let the user decide.
• Weight is a hint, not an override. The router cannot change an agent's declared model:. It selects skills and agents, and sets the caller's model that the 4 inherit agents adopt. It never overrides a target skill's own agent selection.

Validation

Routing accuracy is gateable, not vibes. routing-benchmark.json holds 50 labeled goal → category pairs (easy + genuinely ambiguous). Validate after any change to the category table or disambiguation rules:

bashCopy code
# isolated classification check via the bare-eval harness
bare-eval   # grade router output against routing-benchmark.json

Target ≥95% category accuracy; track the fallback rate as a degradation alarm as the skill library grows.

References

• references/routing-rules.md — per-category parameter extraction, edge cases, disambiguation
• routing-benchmark.json — 50 labeled goal→category pairs for accuracy validation

Quality Bar

Done means all of these hold:

• Classification reasoning is stated out loud BEFORE a route is committed, naming the chosen intent category and the signal words that triggered it.
• Weight is classified in the same pass as intent, naming the signal that decided it, and resolves asymmetrically (ANY heavy signal ⇒ Heavy; Light needs ALL light signals; else Standard).
• The confirm block names one of the taxonomy's intent categories, its target skill or /goal loop, and the extracted args — on one line.
• Every Heavy (opus/fable) leg is surfaced for explicit approval before handoff; no premium spend is silent, and declining it runs Standard with the weakened check named.
• Adversarial, security, safety, architecture, and cross-cutting work is never routed below opus tier; mechanical single-file IO-bound work is not routed above haiku tier.
• The router neither counts spawns nor touches ORK_TEAM_OPUS_MAX, and never downgrades a Heavy leg to evade a team-size-gate denial.
• No target skill is invoked without an explicit nod (or -y); handoff never precedes confirmation.
• When two categories are equally plausible, exactly ONE clarifying question is asked — the fallback is never silently guessed.
• optimize routes to a /goal loop and improve-skill to the evolution gate; neither claims a dedicated skill that does not exist.
• The router does none of the target work itself and never routes to auto (no recursion).

Related skills

• help — static categorized directory (browse, don't route)
• prd-to-goal — decompose a spec into a /goal line (the optimize route's engine)
• fix-issue · cover · expect · brainstorm · implement · review-pr · verify: the route targets
• assess — champion/challenger holdout gate (the improve-skill route)