Intent Routing v2

The always-on rule that maps plain-language user messages to the right specialist agent — no use <agent> syntax required.

Lives at rules/common/agent-intent-routing.md (832 lines, auto-loaded by every AI-tool session). Under 5% of routing decisions require the user to know an agent's name. The rest is inference.

The 10-tier signal system

Ten priority tiers cover the domain. Higher tiers win when signals overlap.

Priority	Family	Example agents
1	Crisis & emotional	`kodelyth-advisor`, `pair-programmer`
2	Active pain (broken)	`debug-detective`, `silent-failure-hunter`, `build-error-resolver`
3	Quality & review	`code-reviewer`, `security-reviewer`, `api-guardian`, `ux-reviewer`, 8 devil-mode adversarial agents
4	Performance & scale	`incident-commander`, `load-tester`, `performance-optimizer`
5	Planning & architecture	`planner`, `architect`, `code-architect`, `migration-guide`
6	Testing	`tdd-guide`, `e2e-runner`, `pr-test-analyzer`
7	Code hygiene	`refactor-cleaner`, `code-simplifier`, `type-design-analyzer`
8	Documentation	`doc-updater`, `docs-lookup`, `comment-analyzer`
9	Specialised workflows	`chief-of-staff`, `seo-specialist`, opensource pipeline
10	Multi-agent patterns	Parallel commands (`/team-review`, `/debug-blitz`, `/devil-mode`, ...)

Each tier has 4-12 signal tables of real human phrasing. The AI reads intent — not keywords.

Routing v2 — 8 new dimensions layered on top

1. Confidence tiers

Tier	Trigger	AI behaviour
High	2+ signals from one table, OR 1 signal + emotion, OR explicit `use X` / `@X`	Route immediately, one-line announcement
Medium	1 clean signal, no counter-signals	Route + "not X? say so, I'll switch" tail
Low	1 weak signal, or signals from 2+ tables of similar priority	Do NOT auto-route. Name the 2 best candidates in one line, ask which fits
None	No signal, or explicit anti-routing	Answer directly, no routing announcement

2. Session-state awareness (sticky routing)

Once you've routed to an agent this session, stay in its voice for follow-ups on the same thread unless:

The next message contains a stronger signal for a different agent
The user says "back to normal", "stop routing", "just answer me"
The task has clearly concluded

Never re-announce the same routing. Sticky = quiet. Only re-announce when the agent actually changes.

3. Anti-routing whitelist

Skip Priority 1-10 entirely when:

Message is under 5 words with no error/code paste ("hi", "ok", "thanks", "cool")
Message is a factual question with a one-line answer ("what does git stash pop do?")
User already invoked with use X, @X, run <agent> — respect explicit choice
User said "don't route" / "just answer" / "stop routing" this session
User is mid-workflow with a previous agent
The message is a bug report about ECC itself — answer as maintainer, not specialist

4. New signal families (v2.4+)

User says	Routes to
"compress this memory file" / "shrink CLAUDE.md"	`/terse-compress`
"talk shorter" / "be brief" / "caveman mode"	`/terse` (skill activation)
"who calls X" / "trace call chain" / "impact analysis"	`codebase-memory-mcp` — `search_graph` / `trace_path`
"index this project" / "build the graph"	`codebase-memory-mcp` — `index_repository`
"remove ECC" / "uninstall kodelythecc"	`kodelythecc uninstall --dry-run`
"how much did I save" / "RTK stats"	`kodelythecc dashboard` OR `kodelythecc rtk gain`
"install for another IDE" / "add to Cursor"	`kodelythecc --target <ide>`

5. Compound intent → parallel commands

When a message matches two priority tables at once, fire the parallel command instead of both agents sequentially.

Signals	Fires
security question + review question	`/security-audit`
bug + been-stuck-for-hours + multi-layer	`/debug-blitz`
architecture + security + a11y (new project)	`/project-launch`
pre-release + full audit	`/pre-release`
refactor + types + tests	`/refactor-sprint`
attacker mindset + multi-vector	`/devil-mode`
new codebase + orientation	`/onboard`
general audit + multiple angles	`/team-review`

6. Announcement style adapts to terse mode

If the user has /terse active this session:

Drop the Tip: next time you can type "use <agent>" line
Use one-token announcement: → debug-detective instead of the full form
No trailing decoration

7. Cultural + multi-language cues

Emotional markers in any language route the same way: "arre yaar broken hai" = frustration = debug-detective
Filler like "bro", "yaar", "man", "please" is not a signal — strip and read the substance
Respond in the user's this-turn language
Never translate code identifiers, commands, or error text (NullPointerException stays)

8. Evolve integration

Routing misses and weak matches feed ~/.kodelythecc/evolve/routing-misses.jsonl. The evolve pipeline aggregates them into proposals for new routing triggers.

You never write this file yourself. Just be honest in the routing announcement — if the match is weak, say "medium confidence — routing to X, not Y, because Z" so the evolve pipeline has a clean signal.

Output formats

Default (verbose)

→ Routing to debug-detective (stack trace + frustration signals match the bug-tracking pattern)
 
[response in debug-detective's style — methodical, hypothesis-driven, asks for repro]
 
Tip: next time you can type "use debug-detective" to invoke me directly.

Terse mode active

→ debug-detective
 
[response — no tip line, no explanation of why]

Medium-confidence single match

→ Routing to code-reviewer (single file paste, no other context — treating as review)
 
[response as code-reviewer]
 
Not what you wanted? Say "use debug-detective" or "just answer" to switch.

Low-confidence, two candidates

Two agents fit this: debug-detective (bug + error) or silent-failure-hunter (no error thrown).
Which one? Or should I answer directly?

Sticky routing (already in an agent)

[response continues in the same agent's voice — no re-announcement]

Compound intent → parallel command

→ Firing /debug-blitz — 3 agents in parallel (debug-detective + silent-failure-hunter + env-debugger)
 
[synthesised result]

Live examples

User writes	Routes to	Why
"I'm getting a TypeError on line 42"	`debug-detective`	Specific error
"no error but the data is wrong"	`silent-failure-hunter`	Silent failure pattern
"should I use React Context or Zustand here?"	`pair-programmer`	Pre-implementation approach
"Review my login component"	`typescript-reviewer`	Single file + likely TS
"I have no idea where to start"	`kodelyth-advisor`	Lost / overwhelmed
"make this faster"	`performance-optimizer`	Direct perf question
"is my JWT signing secure?"	`security-reviewer`	Auth + security keyword
"build failed on Vercel"	`build-error-resolver`	Build failure
"production is down, getting 500s"	`incident-commander`	Active production incident
"will this hold under 10k users?"	`load-tester`	Capacity question
"generate a hero image"	`image-architect`	Explicit image request
"help me build a todo app"	`/project-launch`	New build
"review my code before I deploy"	`/team-review`	Pre-release, full scope
"cutting v2 today, last check"	`/pre-release`	Release + confidence combo
"been stuck on this bug for 2 days"	`/debug-blitz`	Persistent frustration
"who calls `ProcessOrder`?"	`codebase-memory-mcp trace_path`	Structural query
"remember we always use pnpm"	`/lessons`	Preference to encode
[paste code with no text]	`code-reviewer`	Implicit review
[paste stack trace with no text]	`debug-detective`	Implicit debug

Counter-patterns (do NOT route)

User already invoked explicitly (use <agent>, @agent)
Message is one-liner factual ("what does git stash pop do?")
Purely conversational ("hi", "thanks", "ok")
User said "just answer" / "don't route"
Trivially simple (rename a variable)

Where it lives

Rule file: rules/common/agent-intent-routing.md (832 lines)
Loaded: automatically at every AI-tool session start via ~/.claude/rules/ or platform-equivalent
Companion rule: cost-aware-model-routing — picks the model tier (trivial / standard / hard) once the agent is picked

Intent Routing v2 — Plain-Language to Specialist Agent in Claude Code