agent-detector · diff
git:20260610.9689a23 to git:20260812.7cc066c
6 added, 23 removed. Audit A to A.
---
name: agent-detector
description: "CRITICAL: MUST run for EVERY message. Detects agent, complexity, AND model automatically. Without this, tasks route to wrong agents and use wrong models, degrading quality and wasting tokens."
autoInvoke: true
priority: highest
model: haiku
triggers:
- "every message"
- "always first"
allowed-tools: NONE
user-invocable: false
---
- > **AI-consumed reference.** Optimized for Claude to read during execution.
- > Human-readable explanation: see [docs/architecture/HIERARCHICAL_PLANNING.md](../../../docs/architecture/HIERARCHICAL_PLANNING.md)
- > or [docs/getting-started/](../../../docs/getting-started/) depending on topic.
-
-
# Agent Detector
**Runs FIRST for every message.**
## Complexity
```toon
complexity[4]{level,criteria,approach}:
Quick,"Single file / simple fix / clear scope","Direct implementation"
Standard,"2-5 files / feature / some unknowns","Scout then implement"
Deep,"6+ files / architecture / vague scope","run-orchestrator"
Project,"Multi-feature / multi-session / weight ≥ 3 on bridge heuristic AND no active plan","/aura-frog:plan bootstrap then per-task /run anchored"
```
- **Project (v3.7.2+):** Emitted when `rules/workflow/run-plan-bridge.md` triggers sum to weight ≥ 3 AND `.claude/plans/active.json` is absent. `run-orchestrator` Step 0 owns the user prompt (`plan` / `deep` / `details`) and the scratch-file handoff. Otherwise Quick/Standard/Deep classification is unchanged.
-
## Model Selection
- **Prefer the session model — inherit it.** The model the user launched with is the signal of their intent and budget; use it for all substantive work (Standard / Deep / architecture / planning / build / review). Do NOT force-upgrade to a named model (e.g. Opus) for "complex" — if the user wanted more capability they'd run it, and naming a model means a newer/stronger one is ignored.
-
- The only deliberate override is **down-shifting to `haiku`** for trivial mechanical work (classification, detection, state bookkeeping) where a wrong answer costs little. So: Quick/classification → `haiku`; everything else → **inherit (session model)**. See `rules/core/small-to-large-routing.md`.
-
- ## Detection (Priority Order)
-
- 1. **Task content** (highest): Analyze task keywords — backend repo may have frontend tasks. Score ≥50 overrides repo detection.
- 2. **Explicit tech** (+60): User mentions react-native/flutter/angular/vue/react/next/node/python/go/laravel → agent.
- 3. **Intent** (+50): Action keywords: implement/fix/test/design/database/security/deploy → agent.
- 4. **Project context** (+40): Package files/configs. Use cached detection when valid (<24h).
- 5. **File patterns** (+20): Recent file naming: *.vue→frontend, *.go→architect, etc.
+ Inherit the session model for all substantive work. Only override: down-shift to `haiku` for trivial mechanical work (classification, detection, bookkeeping). Never force-upgrade to a named model. See `rules/core/small-to-large-routing.md`.
- ## Scoring
+ ## Detection (priority order)
- Primary ≥80 (leads), Secondary 50-79 (supports), Optional 30-49, Skip <30.
- **tester:** Always secondary unless explicit test request.
+ Task content > explicit tech (+60) > intent keywords (+50) > project context (+40) > file patterns (+20).
+ Scoring: Primary ≥80 (leads), Secondary 50-79, Optional 30-49, Skip <30. **tester:** always secondary unless explicit test request.
- ## Cache
+ Cache: `.claude/cache/agent-detection-cache.json` (reuse within workflow; invalidate on new workflow / phase 1 / user override).
- `.claude/cache/agent-detection-cache.json` — reuse within workflow (phase >1). Invalidate on new workflow, phase 1, or user override.
+ **Detail (Project-level heuristic, signal weights, model rationale):** `skills/agent-detector/reference.md` — load on demand only.