forgeplan_route

Classifies an incoming task description into a depth level and recommends an artifact pipeline. This is the first MCP call in every non-trivial workflow — it tells the agent whether to just code (Tactical), create a PRD → RFC (Standard), the full PRD → Spec → RFC → ADR chain (Deep), or escalate to an Epic with cross-team review (Critical). Getting routing right saves hours of overthinking OR underthinking.

Category: Reasoning & AI

When an agent calls this

User says “please add X to the project” — agent routes before deciding whether to create artifacts.
Before forgeplan_new: the depth from the route determines which template sections are mandatory.
When the agent is uncertain whether a refactor is Tactical or Standard — route settles it.

Input parameters

Name	Type	Required	Description
`description`	`string`	yes	Task description in natural language.

Schema source: crates/forgeplan-mcp/src/server.rs::RouteParams

Returns

A recommendation with depth, pipeline (list of artifact kinds in order), confidence (0-1), rationale, and alternatives the agent can offer if the user disagrees. Level 1 (LLM) returns richer rationale; Level 0 (keyword fallback) returns a terser shape but is still actionable.

Example response shape:

{
  "depth": "Standard",
  "pipeline": ["prd", "rfc"],
  "confidence": 0.85,
  "rationale": "Multi-file feature touching auth flow; reversible within a sprint.",
  "alternatives": [
    { "depth": "Deep", "pipeline": ["prd", "spec", "rfc", "adr"], "when": "if crypto primitives change" }
  ]
}

Example invocation

{ "description": "add rate limiting to API" }

With typical agent context:

User asks for a feature. Agent routes first so the rest of the flow aligns with the correct depth.

{ "description": "rewrite the LanceDB storage layer to use SQLite" }

Typical sequence

forgeplan_route → (if Tactical) jump straight to code → (if Standard+) forgeplan_search (dup check) → forgeplan_new → forgeplan_update → forgeplan_validate → forgeplan_reason. Routing is cheap; call it whenever the task feels ambiguous.

CLI equivalent

forgeplan route — same classifier, human output