Lite router

Lite Router - Task Router (No Opus)

Performs task classification and routing WITHOUT spawning the CEO (Opus) agent. All classification happens in the current session context (Sonnet-level), and agents are dispatched using only Sonnet and Haiku tiers.

When to Activate

• User runs /crew lite "description"
• User runs /crew task --lite "description" (backward compatible)

Prerequisites

These files MUST exist (run /crew init first):

• .claude/crew-profile.md — project context
• .claude/crew-team.json — agent configuration
• .claude/crew-skills.json — skill catalog
• .claude/crew-index.json — codebase index (Layer 1)
• .claude/crew-symbols.json — symbol map (Layer 2)

If any are missing, tell the user to run /crew init first.

Lite Mode Advantage

In LITE mode:

• 0 Opus calls — CEO is bypassed entirely
• Classification is done by this skill (runs in user's Sonnet session)
• vp-engineering (Sonnet) takes over planning/architecture duties that CEO would normally handle

Routing Algorithm

Step 1: Read Context

1. Read .claude/crew-profile.md for project context
2. Read .claude/crew-team.json for available agents
3. Read .claude/crew-index.json for codebase awareness
4. Check for .claude/.index-stale — if present, invoke index-updater for listed files, then delete the marker

Step 2: Classify Task

Task Type (pick one):

Type	Signals
bug-fix	"fix", "bug", "broken", "error", "crash", "not working", "issue"
feature	"add", "implement", "create", "build", "new", "feature"
refactor	"refactor", "clean up", "reorganize", "restructure", "improve"
docs	"document", "readme", "comment", "explain", "docs"
test	"test", "coverage", "spec", "TDD", "unit test", "e2e"
security	"security", "vulnerability", "auth", "permission", "XSS", "injection"
devops	"deploy", "docker", "CI/CD", "pipeline", "kubernetes", "build"
review	"review", "check", "audit", "quality"
performance	"slow", "optimize", "performance", "speed", "memory", "cache"

Complexity (assess based on scope):

Level	Signals
trivial	Single word change, typo, formatting, one-line fix
simple	Single file, clear fix, under 20 lines changed
moderate	2-4 files, some design decisions, 20-100 lines
complex	5+ files, architecture decisions, new patterns, 100+ lines
architectural	System-wide changes, new subsystems, major refactors

Step 3: Select Agent Chain (LITE — No Opus)

LITE ROUTING TABLE (Sonnet + Haiku only):

bug-fix:
  trivial    → junior-dev (haiku)
  simple     → debugger (sonnet) investigates → junior-dev (haiku) fixes
  moderate   → debugger (sonnet) investigates → senior-dev (sonnet) fixes
  complex    → vp-engineering (sonnet) analyzes → debugger investigates → senior-dev fixes → code-reviewer reviews
  architectural → vp-engineering (sonnet) plans → debugger + senior-dev execute → vp-quality reviews

feature:
  trivial    → junior-dev (haiku)
  simple     → senior-dev (sonnet)
  moderate   → vp-engineering (sonnet) plans → senior-dev implements → test-engineer tests
  complex    → vp-engineering (sonnet) architects → [parallel: senior-dev implements, test-engineer writes tests] → code-reviewer reviews
  architectural → vp-engineering (sonnet) designs + plans → [parallel: multiple agents] → vp-quality reviews all

refactor:
  any        → vp-engineering (sonnet) plans → senior-dev executes → code-reviewer reviews

docs:
  any        → doc-writer (haiku)

test:
  trivial/simple → test-engineer (sonnet)
  moderate+  → vp-quality (sonnet) strategizes → test-engineer implements

security:
  any        → security-analyst (sonnet) → senior-dev fixes findings

devops:
  trivial/simple → devops-engineer (haiku)
  moderate+  → senior-dev (sonnet) handles complex infra

review:
  any        → code-reviewer (sonnet) → report to user

performance:
  any        → vp-engineering (sonnet) analyzes → senior-dev optimizes → test-engineer verifies

Key difference from full mode: vp-engineering (Sonnet) takes over the CEO's planning and architecture role. No Opus agent is ever spawned.

Step 4: Decide Execution Strategy

PARALLEL when:
- Subtasks operate on DIFFERENT files (check via crew-index.json)
- Implementation + test writing (test-engineer on test files, senior-dev on source files)
- Multiple independent bug fixes
- Documentation for different components

SEQUENTIAL when:
- Investigation must complete before fix begins
- Planning must complete before implementation
- Implementation must complete before review
- Subtasks modify the SAME files

HYBRID (most common for moderate+ tasks):
1. Sequential: investigation/planning phase
2. Parallel: independent implementation subtasks
3. Sequential: integration review

Step 5: Inject Relevant Skills

Read .claude/crew-skills.json and select skills to inject into agent prompts:

1. Always include: "always" category skills
2. Add stack-specific skills based on detectedStack
3. Add task-type skills based on classified type
4. Each agent gets ONLY the skills relevant to their subtask
5. vp-engineering gets expanded budget (2500 tokens) when acting as planner in lite mode

Step 6: Generate Execution Plan

Create a structured plan before executing:

[LITE] {type}/{complexity}: "{task}"
  1. {agent} → {subtask}
  2. {agent} + {agent} → {subtask A} | {subtask B}
  3. {agent} → {review}

Display this plan to the user, then execute.

Step 7: Execute

Use the execution-orchestrator skill to dispatch agents according to the plan.

For each agent dispatch, construct a self-contained prompt that includes:

1. The subtask description
2. Relevant project context from crew-profile.md
3. The Index-First Protocol instructions
4. Injected skill guidance
5. Files the agent should focus on (from index analysis)
6. Expected output format
7. Note: In lite mode, if an agent needs to escalate, escalate to vp-engineering (not CEO)

Step 8: Collect Results & Report

After all agents complete:

1. Collect each agent's output
2. Check for conflicts (multiple agents editing same lines)
3. If conflicts: escalate to vp-engineering for resolution (NOT CEO)
4. Auto-update index for all FILES_MODIFIED (handled automatically by execution-orchestrator)
5. Log task to .claude/crew-history.json (with "mode": "lite" flag, including skillsInjected and agent details)
6. Report final status to user (including index update status and skills injected)
7. Include skills used in the report — execution-orchestrator's Result Collection template handles the format

Error Handling

• If agent fails → retry once with more context
• If agent reports index is stale → trigger incremental update, then retry
• If parallel agents conflict → fall back to sequential for conflicting files
• If task is unclear → ask user for clarification instead of guessing
• If task complexity exceeds expectations → escalate to vp-engineering (NOT CEO)