Task Router

Task Router

Intro

task-router is the primary routing entry point for processkit agents. Call route_task(task_description) once at the start of any domain task. It returns — in a single call, without an LLM — the matching skill, the project-specific process override (if any), and the recommended MCP tool. Use it instead of calling find_skill() directly.

Overview

When to use

Call route_task() before any create_*, transition_*, link_*, or record_* MCP tool. The 1% rule applies: if there is even a 1% chance a processkit skill covers this task, route first.

Routing algorithm

Two-phase heuristic, no LLM call required:

1. Phase 1 — domain group match (cheap keyword classifier). Scores the task against 13 domain groups by keyword overlap. Eliminates ~90% of candidates in one pass.

2. Phase 2 — tool ranking (token overlap or local embedding-style cosine scoring within the matched group). Scores the ~3–6 tools in the matched group. Returns the top tool with rationale and two runner-up candidates.

3. Fallback — if no group scores ≥ 0.3, falls back to the skill-finder trigger-phrase table for cross-domain or meta tasks.

Reading the result

When confidence < 0.5

routing_basis will be needs_llm_confirm. Show candidate_tools[] to the user or pass them to an LLM for disambiguation. Do NOT call a create_* tool without confirmation when confidence is low.

Callers may pass allow_llm_escalation=true; the router then returns a llm_escalation status with the configured fast model class hint. The stdio MCP server does not make provider network calls itself, so the caller owns the actual cheap-model confirmation step.

When process_override is present

Read the process override file before the generic skill. It contains project-specific step overrides, gates, and blockers that supersede the skill defaults. Example: context/processes/PROC-release.md overrides release-semver for this project.

Gotchas

• Do not call find_skill() directly in normal flow — route_task() calls it internally as a fallback. Direct find_skill() calls skip the process override lookup and the domain group fast path.
• Low confidence is a signal, not a failure. Confidence < 0.5 means the task straddles multiple domains or uses uncommon vocabulary. Use candidate_tools[] to present options.
• Task description quality matters. Verb-noun phrases score best: "create work item", "record decision", "log event". Single words ("task", "log") are ambiguous — add context.

Full reference

MCP server

See mcp/SERVER.md for tool schema and return-shape details.

Domain groups

The 13 registered groups: workitem, decision, discussion, event, actor, scope, index, skill, gate, model, id, role, binding.

Extending the router

To add a new domain group:

1. Add an entry to DOMAIN_GROUPS in mcp/server.py.
2. Add trigger phrases to the skill-finder trigger table.
3. Run the smoke tests: uv run scripts/smoke-test-servers.py.