| name | mu-route |
| description | Use at the start of any unprefixed user message within DevMuse's domain (dev or product work) to route to the right skill. Confidence-based: high confidence invokes silently, medium gives a one-line proposal, low gives full proposal with override options. Bypassed by direct slash invocations (`/mu-<skill>`). |
Route
Pick the correct opening move for the user's task by pattern-matching intent + cheap repo signals against a routing table. Confidence determines behavior: high confidence invokes silently, medium gives a one-line check, low gives a full proposal. Not a HARD-GATE — a smart router.
Consumes @../../knowledge/principles/stance-detection.md heuristics indirectly (target creative skill runs its own Phase 0). Does NOT run stance detection itself.
When to Use
- First message in a conversation within DevMuse's domain
- Task transition: user's intent shifts to a different skill category (e.g., debug→fix, explore→implement). Bootstrap detects this via the Continuation vs Transition table.
- Bootstrap pre-filters: messages outside DevMuse's domain never reach mu-route
- NOT invoked when user prefixes their message with
/mu-<skill> — those are direct-invocation escape hatches
When NOT to Use
- User typed
/mu-arch, /mu-biz, etc. → honor direct invocation, skip routing
- Cadence work (e.g., user explicitly asks for weekly retro) → direct
/mu-retro
- User's CLAUDE.md or AGENTS.md pins a specific skill for a repo → Instruction Priority honors that
- Not a dev/product task → bootstrap filters this out before mu-route is invoked
- Continuation of the same task type → same-type follow-ups during an active skill (e.g., "查下另一个日志" during mu-debug) don't need re-routing
Confidence Levels
| Confidence | Criteria | Behavior |
|---|
| High | Single verb match, unambiguous intent, sufficient context | Silent invoke — no proposal |
| Medium | Two possible moves, one clearly dominant | One-line check — "→ Skill, ok?" |
| Low | Three+ possible moves, or two equally plausible | Full proposal — axes + override options |
Default to medium when unsure.
Process Flow
digraph mu_route {
"User message arrives" [shape=box];
"Slash prefix?" [shape=diamond];
"Direct skill invocation\n(bypass mu-route)" [shape=doublecircle];
"Gather cheap signals\n(verbs, artifacts, git, CODEOWNERS)" [shape=box];
"Apply routing table R1..R7" [shape=box];
"Assess confidence" [shape=diamond];
"Silent invoke\n(no proposal)" [shape=doublecircle];
"One-line check\n('→ Skill, ok?')" [shape=box];
"Full proposal\n(axes + overrides)" [shape=box];
"User confirms?" [shape=diamond];
"Invoke target skill" [shape=doublecircle];
"Accept override" [shape=box];
"User message arrives" -> "Slash prefix?";
"Slash prefix?" -> "Direct skill invocation\n(bypass mu-route)" [label="yes (/mu-*)"];
"Slash prefix?" -> "Gather cheap signals\n(verbs, artifacts, git, CODEOWNERS)" [label="no"];
"Gather cheap signals\n(verbs, artifacts, git, CODEOWNERS)" -> "Apply routing table R1..R7";
"Apply routing table R1..R7" -> "Assess confidence";
"Assess confidence" -> "Silent invoke\n(no proposal)" [label="high"];
"Assess confidence" -> "One-line check\n('→ Skill, ok?')" [label="medium"];
"Assess confidence" -> "Full proposal\n(axes + overrides)" [label="low"];
"One-line check\n('→ Skill, ok?')" -> "User confirms?";
"Full proposal\n(axes + overrides)" -> "User confirms?";
"User confirms?" -> "Invoke target skill" [label="ok"];
"User confirms?" -> "Accept override" [label="override"];
"Accept override" -> "Invoke target skill";
}
Checklist
Create tasks for each and complete in order:
- Check slash prefix — if user message starts with
/mu-<skill>, bypass mu-route entirely; invoke that skill directly
- Gather cheap signals (all in <5s):
- Axis-Intent: parse user message verbs against lexicon (table below)
- Axis-Familiarity:
git log --author="$USER" --since="30 days ago" -- <area> if target area inferable
- Axis-Missing-artifact: file-exists on
docs/biz/, docs/prd/, docs/specs/. Inline conversation content (payload examples, pseudocode, verbal descriptions) never counts as "specs exist" — only on-disk artifact files do.
- Axis-Stakeholder:
test -f .github/CODEOWNERS || test -f CODEOWNERS + git log multi-author check (feeds sign-off gate later; informational to user here)
- Axis-Plugin: scan the available skills list (from system-reminder) for non-DevMuse skills (i.e. skills whose name does NOT start with
devmuse:). Check if the user's message plausibly matches any such skill's description or triggers. Record matched skill name(s), if any.
- Apply routing decision table (below) top-to-bottom, first match wins → one opening move
- Assess confidence and act accordingly:
- High (single unambiguous verb match + sufficient context) → silently invoke target skill, no proposal
- Medium (clear direction but some ambiguity) → one-line check: "→
<Skill>, ok?"
- Low (multiple possible moves, vague intent) → full proposal with axes + override options
- If proposing, accept user reply:
ok / bare confirm → proceed; one-word override → use overridden move; anything else → ask user to clarify (non-blocking)
- Invoke target skill (via Skill tool) with optional hint (e.g.,
stance=create for downstream creative skills)
Trigger Signal Table
| Verb / phrase in user message | Axis-Intent | Default opening move |
|---|
| "understand", "figure out", "read", "take over", "evaluate", "what does this do" | understand | Explore |
| "add feature", "build feature" | create-feature | Design-tech (or Explore if unfamiliar) |
| "refactor", "clean up", "rename", "restructure" | reshape | Design-tech (or Explore if unfamiliar) |
| "fix", "broken", "error", "bug", "test failing", "crash" | fix | Reproduce |
| "implement", "write this", "build this", "code it up" | implement | Implement (if design exists; else Design-tech) |
On-demand only (not auto-routed): "validate idea", "business model", "product requirements", "user flows", "competitive analysis" → respond with a pointer to /mu-biz or /mu-prd. "retro", "look back", "how did X go" → respond with a pointer to /mu-retro. "wiki", "architecture docs", "generate documentation", "project documentation" → respond with a pointer to /mu-wiki. | | |
When multiple verbs fire, Axis-Intent prefers the primary action — fix > reshape > create-feature > implement > understand (most-specific wins).
Routing Decision Table
Rows evaluated top-to-bottom; first match wins.
| # | Slash prefix | Axis-Intent | Axis-Missing-artifact | Axis-Familiarity | → Opening Move | Hint to target |
|---|
| R1 | /mu-<skill> | — | — | — | bypass direct invocation | — (user owns intent) |
| R2 | none | understand | — | — | Explore | — |
| R3 | none | fix | — | — | Reproduce (via mu-scope 1 UC repro) | — |
| R4 | none | reshape | — | unfamiliar | Explore (pre-change variant) → then Design-tech | — |
| R5 | none | reshape / create-feature | no specs | familiar | Design-tech | stance=auto |
| R5.5 | none | implement | no specs | — | Design-tech | stance=auto |
| R6 | none | implement | specs exist | — | Implement | — |
| R6.5 | none | Axis-Plugin matched | — | — | Delegate to plugin (invoke matched skill via Skill tool) | — |
| R7 | none (no verb match) | — | — | — | Explore (safe default) | — |
On-demand skills (mu-biz, mu-prd, mu-retro, mu-wiki) are not auto-routed. mu-route responds with a pointer to the appropriate slash command instead of invoking the skill.
Hint semantics: when the target is mu-arch, mu-route MAY pass a stance=auto hint indicating Phase 0 should run its own detection without further pre-confirmation.
Proposal Wording
High confidence (silent invoke)
No proposal. Directly invoke the target skill. The user sees the skill's output, not a routing question.
Medium confidence (one-line check)
"→ , ok?"
Example:
"→ Design-tech, ok?"
Low confidence (full proposal)
"Looks like . Axes: Intent=<verb>, Familiarity=<familiar|unfamiliar|n/a>, Missing=<specs|none>. Confirm (ok) or override (one word: explore / design-tech / reproduce / implement)?"
On-demand skill pointer
When on-demand skill language is detected:
"This sounds like <biz/product/retro> work. Use /mu-biz, /mu-prd, or /mu-retro to start."
Plugin delegation
"Detected installed skill matching your request. Route to /<skill-name>? Confirm (ok) or override."
Slash-Command Escape Hatch
Users can bypass mu-route entirely:
/mu-explore → directly invoke mu-explore (skip routing)/mu-arch → directly invoke mu-arch- etc.
This matches industry convention (Aider /ask//code//architect, Roo Code /architect//debug//code/etc.) and preserves muscle memory. mu-route is the default for unprefixed messages where classification adds value.
Ambiguity Handling
- 2+ moves tie on routing rules → propose the one that fires first in R1..R7 ordering; note in the proposal sentence:
"(tied with <other move>)". User overrides with one word.
- No verb matches Axis-Intent → R7 fires (default Explore). Safe default: understand before acting.
- Repo state is pathological (empty repo, shallow clone, submodule root, repo outside git) → skip the routing table and ask user directly: "Can't confidently route — repo state is unusual (
<detected anomaly>). Which opening move? (explore / design-tech / reproduce / implement)".
Failure Handling
- ER-R1 Heuristics computation error (git command fails, file read fails, regex error, etc.) → do NOT fabricate a signal. Surface the error briefly and ask user: "Couldn't compute routing signals (
<specific failure>). Which opening move?". Non-blocking.
- ER-R2 User reply cannot be parsed (multi-word / off-topic / typo) → ask user to restate with one word from the accepted list. Non-blocking.
All paths non-blocking — mu-route always produces a proposal or a single clarifying ask.
Interaction with Sign-off Gate
mu-route does NOT run the sign-off-gate protocol itself. It surfaces Axis-Stakeholder as context (e.g., "CODEOWNERS detected; sign-off will be required at artifact exit") but the gate itself fires inside the target creative skill's exit step, per @../../knowledge/principles/sign-off-gate.md.
Key Principles
- Confidence adapts friction — clear intent gets silent routing, ambiguity gets proposals
- Cheap signals only — detection must complete in <5s (file exists, git log --format)
- Slash hints bypass — power users never see mu-route unless they want to
- R7 safe default — when ambiguous, propose Explore (understand before acting)
- No HARD-GATE — mu-route is a router, target skills enforce their own gates
Integration
- Invoked by:
rules/bootstrap.md for any unprefixed user message
- Produces: a skill invocation (Skill tool call to the target move's skill)
- Terminal state: invoking the target skill; mu-route exits
- Consumed principle references:
@../../knowledge/principles/sign-off-gate.md (surfaced to user, not executed)
- Bypassed by: any
/mu-<skill> slash prefix
