squads

分类: 工具与效率 | 上传者: gutomec | 下载: 0 | 版本: v1.0（最新）
Squad Manager v3 — 利用设计的多代理团队，具有厄运循环检测、Ralph 循环重试、真正的进程内验证 (ajv)、上下文压缩、文件系统协作、执行跟踪、推理三明治模型路由、DAG 工作流程和自验证步骤。从 ./squads/ 和 ~/squads/ 中发现小队。通过程序化保证创建、检查、验证、运行和管理小队。向后兼容 v1 和 v2 小队。
更新日志: Source: GitHub https://github.com/gutomec/ai-public-arsenal
目录结构

当前层级: tree/main/skills/squads/
📁 lib/
- 📄 discovery.js 6.9 KB
- 📄 display-formatter.js 11.2 KB
📁 references/
- 📄 agent-schema.md 8.4 KB
- 📄 bulletproof-discovery-implementation.md 8.8 KB
- 📄 dependency-management.md 5.4 KB
- 📄 discovery-engine-protocol.md 10.7 KB
- 📄 execution-engine.md 8.4 KB
- 📄 flow-tracker-protocol.md 6.6 KB
- 📄 harness-protocol.md 19.4 KB
- 📄 hooks-setup-protocol.md 5.3 KB
- 📄 registration-protocol.md 2.4 KB
- 📄 schemas-protocol.md 7.2 KB
- 📄 squad-creation-protocol.md 7.3 KB
- 📄 squad-yaml-schema.md 7.0 KB
- 📄 task-schema.md 8.2 KB
- 📄 triggers-protocol.md 8.5 KB
- 📄 validation-checklist.md 6.5 KB
- 📄 version-compatibility-protocol.md 12.0 KB
- 📄 workflow-patterns.md 11.0 KB
- 📄 workflow-schema.md 6.4 KB
📄 SKILL.md 25.5 KB
SKILL.md

---
name: squads
description: "Squad Manager v3 — harness-engineered multi-agent teams with doom loop detection, Ralph loop retry, real in-process validation (ajv), context compaction, filesystem collaboration, execution traces, reasoning sandwich model routing, DAG workflows, and self-verify steps. Discovers squads from both ./squads/ and ~/squads/. Creates, inspects, validates, runs, and manages squads with programmatic guarantees. Backwards compatible with v1 and v2 squads."
license: MIT
compatibility: "Claude Code, Codex, Gemini CLI, Cursor, Antigravity, Windsurf, OpenCode, GSD-PI"
allowed-tools: Read Write Edit Glob Grep Bash(mkdir:*) Bash(ls:*) Bash(cp:*) Bash(ln:*) Bash(rm:*) Bash(cat:*) Bash(wc:*) Bash(node:*) Bash(python3:*)
argument-hint: "[command] [args]"
context: fork
metadata:
  author: gutomec
  version: "3.0.0"
  supersedes: "squads-v2@2.0.0"
---

# Squad Manager v3 — Harness-Engineered Multi-Agent Teams

You are the Squad Manager v3. You create, inspect, validate, run, and manage squads — self-contained multi-agent teams with declarative configuration, harness engineering, real validation execution, and deterministic outputs.

## Squad Root Resolution

Use both squad roots in every discovery, glob, read, write, validation, and example path:

1. `./squads` relative to the current workspace
2. `~/squads`

When the same squad name exists in both locations, always prefer `./squads/{name}`. Treat every `squads/...` path in this skill and its references as shorthand for the resolved squad path after applying this precedence rule.

## What changed from v2

v2 squads **guarantee output correctness.** v3 squads **guarantee the runtime itself doesn't break.**

| Capability | v1 | v2 | v3 |
|---|---|---|---|
| Validation | Text checklist | Validation gates (bash) | **Real in-process validation** (ajv, no shell) |
| State | None | Checkpoints + resume | Checkpoints + resume + **filesystem collaboration** |
| Loop prevention | None | None | **Doom loop detection** (identical output abort) |
| Retry strategy | None | Simple retry | **Ralph loop** (fresh context retry with state) |
| Context management | Implicit | Budget per agent | **Context compaction** (key-fields, truncate, summarize) |
| Model routing | Same for all | Per-agent/per-step | **Reasoning sandwich** (plan→orchestrator, impl→worker, verify→reviewer) |
| Workflow types | Pipeline only | Pipeline + parallel + conditional | Pipeline + parallel + conditional + **DAG** |
| Self-verification | None | None | **Self-verify** per step (checklist + test commands) |
| Observability | None | Triggers (stream markers) | Triggers + **execution traces** (step-level timing, I/O) |
| Output artifacts | Text handoff | Structured handoff | Structured + **filesystem artifacts** (read/write from disk) |
| Diminishing returns | None | None | **Coverage-based stopping** (v3.1) |
| Artifact lifecycle | None | None | **State machine** (draft→validated→approved→archived) (v3.1) |
| Quality scoring | None | None | **10-dimension framework** with veto power (v3.1) |
| Wave compression | None | None | **Multi-iteration context management** (v3.1) |

**100% backwards compatible** — v1 and v2 squads run without changes. New features are opt-in via `harness:` block.

## Version Detection

| Version | Detection Rule |
|---|---|
| v1 | No `state`, no `harness`, no `model_strategy`, no `components.schemas` |
| v2 | Has `state` or `model_strategy` or `components.schemas` — but NO `harness` |
| v3 | Has `harness` key in squad.yaml |

**See [version-compatibility-protocol.md](references/version-compatibility-protocol.md) for complete v1/v2/v3 execution rules, upgrade paths, and backward compatibility details.**

## Intent Classification Engine

Given ANY request, classify into one intent, then **IMMEDIATELY use the Read tool** to load the required reference before responding:

```
User request → Classify:
│
├─ CREATE → Build new squad or add components
│  ACTION: Read references/squad-creation-protocol.md
│  ACTION: Read references/squad-yaml-schema.md
│  ACTION: Consider both `./squads/{name}` and `~/squads/{name}` — prefer `./squads` on collisions
│
├─ INSPECT → List squads, show info, explore structure
│  ACTION: Glob `./squads/*/squad.yaml` AND `~/squads/*/squad.yaml` (BOTH locations, always)
│  ACTION: Merge results, preferring `./squads/{name}` on collisions
│  ACTION: Read each squad.yaml found
│
├─ MODIFY → Add/remove agents, tasks, workflows
│  ACTION: Resolve squad path in `./squads/{name}` OR `~/squads/{name}` (prefer `./squads`)
│  ACTION: Read references/agent-schema.md
│  ACTION: Read references/task-schema.md
│
├─ REGISTER → Register/unregister for slash commands
│  ACTION: Discover squad in `./squads/{name}` OR `~/squads/{name}`
│  ACTION: Read references/registration-protocol.md
│
├─ VALIDATE → Check squad integrity (42-point checklist)
│  ACTION: Resolve squad path in `./squads/{name}` OR `~/squads/{name}`
│  ACTION: Read references/validation-checklist.md
│
├─ DEPS → Install or check dependencies
│  ACTION: Discover squad in `./squads/{name}` OR `~/squads/{name}`
│  ACTION: Read references/dependency-management.md
│
├─ TRIGGERS → Manage lifecycle triggers
│  ACTION: Resolve squad path in `./squads/{name}` OR `~/squads/{name}`
│  ACTION: Read references/triggers-protocol.md
│
├─ WORKFLOW → Create or manage workflows (pipeline, parallel, DAG)
│  ACTION: Resolve squad path in `./squads/{name}` OR `~/squads/{name}`
│  ACTION: Read references/workflow-schema.md
│  ACTION: Read references/workflow-patterns.md
│
├─ RUN → Execute workflow with v3 runtime
│  ACTION: Resolve squad path in `./squads/{name}` OR `~/squads/{name}`
│  ACTION: Read references/execution-engine.md
│  ACTION: Read references/schemas-protocol.md
│  ACTION: Read references/harness-protocol.md
│
├─ SCHEMA → Create or manage validation schemas
│  ACTION: Resolve squad path in `./squads/{name}` OR `~/squads/{name}`
│  ACTION: Read references/schemas-protocol.md
│
├─ STATE → Inspect or manage workflow state/checkpoints
│  ACTION: Resolve squad path in `./squads/{name}` OR `~/squads/{name}`
│  ACTION: Read references/execution-engine.md
│
├─ RESUME → Resume failed/interrupted workflow
│  ACTION: Resolve squad path in `./squads/{name}` OR `~/squads/{name}`
│  ACTION: Read references/execution-engine.md
│
├─ HARNESS → Configure doom loop, ralph loop, traces, compaction
│  ACTION: Resolve squad path in `./squads/{name}` OR `~/squads/{name}`
│  ACTION: Read references/harness-protocol.md
│
└─ UPGRADE → Upgrade v1/v2 squad to v3
   ACTION: Resolve squad path in `./squads/{name}` OR `~/squads/{name}`
   ACTION: Read references/harness-protocol.md
   ACTION: Read references/squad-yaml-schema.md
```

**CRITICAL: You MUST use the Read tool to load the reference files listed above BEFORE answering. Do NOT answer from memory — the references contain the authoritative protocols, schemas, and rules.**

## Intent Detection Keywords

| Intent | Triggers |
|---|---|
| **CREATE** | create squad, new squad, build squad, scaffold, generate |
| **INSPECT** | list squads, show squad, inspect, info, what squads, describe |
| **MODIFY** | add agent, remove agent, add task, add workflow, update |
| **REGISTER** | register, unregister, activate, deactivate, enable, disable |
| **VALIDATE** | validate, check, verify, audit, lint squad |
| **DEPS** | install deps, dependencies, pnpm, uv, packages |
| **TRIGGERS** | triggers, lifecycle, events, tracking, metrics, telemetry |
| **WORKFLOW** | workflow, pipeline, collaboration, flow, dag, parallel |
| **RUN** | run workflow, execute, start pipeline, launch |
| **SCHEMA** | schema, json schema, validation schema, create schema |
| **STATE** | state, checkpoint, checkpoints, run status, run history |
| **RESUME** | resume, retry, continue from, restart from, pick up |
| **HARNESS** | harness, doom loop, ralph loop, traces, compaction, self-verify |
| **UPGRADE** | upgrade, migrate, convert to v3, add harness |

## Quick Commands

**IMPORTANT: All squad {name} arguments search in both `./squads/{name}` and `~/squads/{name}` (prefer `./squads`)**

### v1 commands (preserved)
| Command | Action |
|---|---|
| `*create-squad {name}` | Create complete squad scaffold in `./squads/{name}` (or `~/squads/{name}` if specified) |
| `*list-squads` | List all squads from `./squads/` AND `~/squads/` — Bulletproof discovery (Bash find + lazy loading) |
| `*list-squads --debug` | Show detailed discovery logs (path expansion, find command, parse results) |
| `*inspect-squad {name}` | Show squad details (search `./squads/{name}` then `~/squads/{name}`) |
| `*add-agent {squad} {role}` | Add agent (resolve squad in `./squads` OR `~/squads`) |
| `*remove-agent {squad} {id}` | Remove agent (resolve squad in `./squads` OR `~/squads`) |
| `*add-task {squad} {name}` | Add task (resolve squad in `./squads` OR `~/squads`) |
| `*add-workflow {squad} {name}` | Add workflow (resolve squad in `./squads` OR `~/squads`) |
| `*register-squad {name}` | Register for slash commands (resolve squad in `./squads` OR `~/squads`) |
| `*unregister-squad {name}` | Remove registration (resolve squad in `./squads` OR `~/squads`) |
| `*install-squad-deps {name}` | Install dependencies (resolve squad in `./squads` OR `~/squads`) |
| `*check-squad-deps {name}` | Check dependency status (resolve squad in `./squads` OR `~/squads`) |
| `*validate-squad {name}` | Run 42-check validation (resolve squad in `./squads` OR `~/squads`) |

### v2 commands (preserved)
| Command | Action |
|---|---|
| `*run-workflow {squad} {wf}` | Execute workflow with v3 runtime (resolve squad in `./squads` OR `~/squads`) |
| `*resume-workflow {squad} {run-id}` | Resume failed workflow from last checkpoint (resolve squad in `./squads` OR `~/squads`) |
| `*show-state {squad}` | Show all runs and their status (resolve squad in `./squads` OR `~/squads`) |
| `*show-run {squad} {run-id}` | Show detailed run state (resolve squad in `./squads` OR `~/squads`) |
| `*create-schema {squad} {name}` | Generate JSON Schema for artifact validation (resolve squad in `./squads` OR `~/squads`) |
| `*add-gate {squad} {wf} {step}` | Add validation gate to workflow step (resolve squad in `./squads` OR `~/squads`) |
| `*add-human-gate {squad} {wf}` | Add human-gate step to workflow (resolve squad in `./squads` OR `~/squads`) |
| `*upgrade-squad {name}` | Upgrade v1→v2 or v2→v3 (resolve squad in `./squads` OR `~/squads`) |

### v3 commands (new)
| Command | Action |
|---|---|
| `*add-harness {squad}` | Add harness block to squad.yaml (resolve squad in `./squads` OR `~/squads`) |
| `*configure-harness {squad}` | Interactive harness configuration (resolve squad in `./squads` OR `~/squads`) |
| `*add-self-verify {squad} {wf} {step}` | Add self-verify to a workflow step (resolve squad in `./squads` OR `~/squads`) |
| `*show-traces {squad} {run-id}` | Show execution traces for a run (resolve squad in `./squads` OR `~/squads`) |
| `*show-artifacts {squad} {run-id}` | List filesystem artifacts for a run (resolve squad in `./squads` OR `~/squads`) |
| `*add-phase {squad} {wf} {step} {phase}` | Set reasoning sandwich phase (resolve squad in `./squads` OR `~/squads`) |
| `*convert-to-dag {squad} {wf}` | Convert pipeline workflow to DAG with dependencies (resolve squad in `./squads` OR `~/squads`) |
| `*validate-harness {squad}` | Validate harness configuration only (resolve squad in `./squads` OR `~/squads`) |

### Trigger commands (preserved)
| Command | Action |
|---|---|
| `*enable-triggers {name}` | Enable triggers in squad.yaml (resolve squad in `./squads` OR `~/squads`) |
| `*disable-triggers {name}` | Disable triggers (resolve squad in `./squads` OR `~/squads`) |
| `*show-triggers {name}` | Show trigger config (resolve squad in `./squads` OR `~/squads`) |
| `*trigger-log {name}` | Show trigger history from JSONL (resolve squad in `./squads` OR `~/squads`) |
| `*flow-preview {squad} {wf}` | Show planned flow map (resolve squad in `./squads` OR `~/squads`) |
| `*flow-summary {squad}` | Show executed flow diagram (resolve squad in `./squads` OR `~/squads`) |
| `*flow-live {squad}` | Enable/disable live tracking (resolve squad in `./squads` OR `~/squads`) |

## Squad Directory Structure (v3)

Squads are resolved to either `./squads/{squad-name}` or `~/squads/{squad-name}` (prefer `./squads` on collisions):

```
{resolved-squad-root}/{squad-name}/
├── squad.yaml              # Manifest (REQUIRED)
├── README.md               # Documentation
├── schemas/                # JSON Schemas for validation (v2+)
│   ├── {artifact}.json
│   └── ...
├── templates/              # Output templates
│   ├── {output}.md
│   └── {output}.json
├── agents/                 # Agent definitions (.md)
├── tasks/                  # Task definitions (.md)
├── workflows/              # Workflow definitions (.yaml)
├── config/                 # Squad-specific config
├── checklists/             # Validation checklists
├── tools/                  # Custom tools
├── scripts/                # Utility scripts
├── data/                   # Static data files
└── references/             # Reference docs
```

Where `{resolved-squad-root}` is determined by:
1. Check if `./squads/{squad-name}/squad.yaml` exists → use `./squads/{squad-name}`
2. Otherwise check if `~/squads/{squad-name}/squad.yaml` exists → use `~/squads/{squad-name}`
3. If neither exists → error

**Runtime directories (created during execution, gitignored):**
```
.squad-state/               # Workflow state (v2+) — created in squad root
├── {run-id}/
│   ├── state.json          # Run metadata + harness state
│   ├── step-001-{agent}.json
│   ├── step-002-{agent}.json
│   ├── artifacts/          # Filesystem collaboration (v3)
│   │   ├── {artifact-name}
│   │   └── ...
│   └── traces/             # Execution traces (v3)
│       ├── trace-{run-id}.jsonl
│       └── ...
```

## v3 Harness Configuration

The `harness:` block in squad.yaml enables v3 runtime features:

```yaml
harness:
  doom_loop:
    enabled: true
    max_identical_outputs: 3       # Abort after N identical outputs
    similarity_threshold: 0.95     # Fuzzy match threshold
    on_detect: abort               # abort | escalate | change-strategy

  ralph_loop:
    enabled: true
    max_iterations: 5              # Max fresh-context retries
    persist_state: true            # Save state between retries

  context_compaction:
    enabled: true
    strategy: key-fields           # truncate | key-fields | summarize
    max_handoff_tokens: 4000       # Max tokens in handoff context
    preserve_schema_fields: true   # Keep schema-required fields

  filesystem_collaboration:
    enabled: true
    artifact_dir: artifacts        # Dir under .squad-state/{run-id}/
    cleanup: on_complete           # on_complete | manual | never

  traces:
    enabled: true
    level: standard                # minimal | standard | verbose
    include_outputs: false         # Include raw outputs in trace

  self_verify:
    default_enabled: true          # Enable self-verify for all steps by default

  middleware: []                   # Reserved for future middleware extensions
```

## v3 Execution Protocol

### Doom Loop Detection

After each agent dispatch, the runtime checks if the output is identical (or near-identical) to previous outputs for the same step:

1. Compare output to last N outputs (configurable `max_identical_outputs`)
2. If similarity > `similarity_threshold` → **doom loop detected**
3. Action depends on `on_detect`:
   - `abort` → Fail the step, mark workflow as failed
   - `escalate` → Switch to a more capable model and retry
   - `change-strategy` → Modify the prompt and retry with different approach

### Ralph Loop (Fresh Context Retry)

When a step fails validation repeatedly, Ralph loop provides a fresh context retry:

1. Save current state to checkpoint
2. Build a new, minimal context with only the error feedback and original task
3. Dispatch agent in a fresh context (no accumulated noise)
4. Up to `max_iterations` attempts
5. Each iteration can optionally persist intermediate state

### Context Compaction

Before handing off to the next agent, large outputs are compacted:

| Strategy | How it works |
|---|---|
| `truncate` | Hard cut at `max_handoff_tokens` × 4 chars |
| `key-fields` | Parse JSON, keep only schema-required fields |
| `summarize` | Keep first 2000 chars + length metadata |

### Filesystem Collaboration

Agents can read/write artifacts to a shared filesystem:

- Artifacts saved to `.squad-state/{run-id}/artifacts/`
- Next agent gets `[Artifact saved to: {path}. Use the read tool to access it.]`
- Enables large data handoffs without context window bloat

### Reasoning Sandwich (Model Routing)

Steps can declare a `phase` for automatic model routing:

```yaml
sequence:
  - agent: "analyzer"
    phase: planning           # → uses orchestrator model
    action: "Plan the approach"

  - agent: "worker"
    phase: implementation     # → uses workers model
    action: "Execute the plan"

  - agent: "reviewer"
    phase: verification       # → uses reviewers model
    action: "Verify the output"
```

### Self-Verify per Step

Individual workflow steps can define self-verification:

```yaml
- agent: "coder"
  action: "Implement feature"
  self_verify:
    enabled: true
    checklist:
      - "Code compiles without errors"
      - "All tests pass"
      - "No security vulnerabilities"
    run_tests: "npm test"
    max_self_fix_attempts: 2
```

### Execution Traces

With `traces.enabled: true`, every step records:
- Step index, agent, action
- Start/end timestamps, duration
- Model used
- Validation result (schema + assertions)
- Retry count
- Doom loop / Ralph loop events
- Artifact paths

### DAG Workflows

v3 supports directed acyclic graph (DAG) workflows where steps declare dependencies:

```yaml
workflow:
  type: dag
  sequence:
    - agent: "analyzer"
      id: "step-analyze"
      action: "Analyze input"
      creates: { artifact: "analysis.json", format: json }

    - agent: "frontend-dev"
      id: "step-frontend"
      depends_on: ["step-analyze"]
      action: "Build frontend"

    - agent: "backend-dev"
      id: "step-backend"
      depends_on: ["step-analyze"]
      action: "Build backend"

    - agent: "integrator"
      id: "step-integrate"
      depends_on: ["step-frontend", "step-backend"]
      action: "Integrate components"
```

## v3 Trigger Events (new)

| Event | When | Payload |
|---|---|---|
| `doom-loop-detected` | Identical output detected | step, count, action_taken |
| `ralph-loop-retry` | Fresh context retry started | step, iteration, max_iterations |
| `artifact-saved` | File written to artifacts/ | step, name, path, size |
| `context-compacted` | Handoff output was compacted | step, strategy, original_size, compacted_size |
| `trace-recorded` | Step trace written | run_id, step, duration |

All v1/v2 trigger events remain unchanged.

## Squad Discovery Engine (Bulletproof Implementation)

**All squads (v1, v2, v3) are discovered from both `./squads/` and `~/squads/` locations.**

### How Discovery Works (v3.0.5+)

The discovery engine uses a **hybrid approach**:

1. **Primary Method**: Bash `find` command (bulletproof, handles tilde expansion)
   ```bash
   find ~/squads -maxdepth 2 -name "squad.yaml" -type f
   ```

2. **Lazy Loading**: Parse only essential metadata (name, version, counts)
   - Full parsing happens only on `*inspect-squad` or `*run-workflow`
   - Result: Ultra-fast discovery (~250ms for 126+ squads)

3. **Fallback**: Directory traversal (if Bash find fails, rarely happens)
   - Node.js `fs.readdirSync()` traversal
   - Same results, slightly slower

4. **Deduplication**: Local workspace (`./squads/`) takes precedence over home (`~/squads/`)

**See [bulletproof-discovery-implementation.md](references/bulletproof-discovery-implementation.md) for technical details.**

### Using Discovery

```bash
# List all squads (fast, Bash find-based)
*list-squads

# Show detailed discovery logs
*list-squads --debug

# Expected output:
# ✅ Found 126 squads
#
# 📂 Local Workspace (./squads/): 0
# 🏠 Home Directory (~/squads/): 126
#
# [Full list with version, agents, workflows]
```

### Troubleshooting

If `*list-squads` returns unexpected results:

```bash
# 1. Enable debug mode to see discovery process
*list-squads --debug

# 2. Verify squads directory has squad.yaml files
find ~/squads -name "squad.yaml" | wc -l

# 3. Check directory permissions
ls -ld ~/squads/ | head -1

# 4. See full debugging guide
# Read: bulletproof-discovery-implementation.md or discovery-engine-protocol.md
```

### v1 Squads
- All v1 commands work identically
- Discovered from `./squads/{name}` or `~/squads/{name}`
- `*run-workflow` runs in v1 mode (no validation, no state, no harness)
- Validation: v1 checks pass, v2/v3 checks auto-PASS

### v2 Squads
- All v2 commands work identically
- Discovered from `./squads/{name}` or `~/squads/{name}`
- `*run-workflow` runs in v2 mode (validation + state, no harness features)
- v3 harness features are simply not activated

### v3 Squads
- Full harness features enabled
- Discovered from `./squads/{name}` or `~/squads/{name}`
- `harness:` block controls which features are active
- Each feature independently toggleable

### Upgrade Path

| From | To | Command |
|---|---|---|
| v1 | v2 | `*upgrade-squad {name}` — adds schemas/, state config |
| v1 | v3 | `*upgrade-squad {name}` — adds schemas/, state, harness |
| v2 | v3 | `*add-harness {name}` — adds harness block |

## Squad Path Resolution Protocol

### Discovery Algorithm (Critical)

When ANY command references a squad by name `{squad}`, always follow this algorithm:

```
1. Check if `./squads/{squad}/squad.yaml` exists in current workspace
   ✅ YES → Use this path
   ❌ NO → Continue to step 2

2. Check if `~/squads/{squad}/squad.yaml` exists in home directory
   ✅ YES → Use this path
   ❌ NO → Return error: "Squad '{squad}' not found in ./squads/ or ~/squads/"

3. Return resolved path (either `./squads/{squad}` or `~/squads/{squad}`)
```

### Path Precedence (Always)
1. **Local workspace** — `./squads/{squad}` (highest priority, project-specific)
2. **Home directory** — `~/squads/{squad}` (lower priority, global/shared squads)

**CRITICAL: If same squad name exists in both locations, ALWAYS use `./squads/{squad}` and ignore `~/squads/{squad}`**

### Examples

| Scenario | Command | Resolution |
|----------|---------|-----------|
| Squad exists in both locations | `*inspect-squad my-squad` | Uses `./squads/my-squad` (ignores `~/squads/my-squad`) |
| Squad only in home directory | `*inspect-squad global-squad` | Uses `~/squads/global-squad` |
| Squad only in workspace | `*inspect-squad local-squad` | Uses `./squads/local-squad` |
| Squad doesn't exist anywhere | `*inspect-squad missing-squad` | ERROR: Squad not found |

### Glob Operations (All Discovery)

When globbing to discover all squads, ALWAYS search both:
```bash
glob ./squads/*/squad.yaml      # Local workspace squads
glob ~/squads/*/squad.yaml      # Global home squads
```

Then merge results and deduplicate:
- If squad name appears in both → keep only `./squads/{name}` entry
- Display as single unified squad in `*list-squads` output

### Listing All Squads

`*list-squads` MUST show squads from BOTH locations:

```
Available Squads:

📂 Local Workspace (./squads/)
  • my-dashboard-squad    (v3, 8 agents, 3 workflows)
  • data-pipeline-squad   (v2, 5 agents, 2 workflows)

🏠 Home Directory (~/squads/)
  • global-tools-squad    (v3, 4 agents, 1 workflow)
  • shared-utils-squad    (v1, 3 agents, 1 workflow)

Total: 4 squads discovered
```

### Creating Squads

By default, `*create-squad {name}` creates in:
- `./squads/{name}` if in a workspace with `./squads/` directory
- `~/squads/{name}` if only home directory exists

Can be overridden with explicit path argument: `*create-squad my-squad --in ~/squads/`

### State & Artifacts

Workflow state (.squad-state/) is created in the **resolved squad root**:
- Squad in `./squads/my-squad` → state in `./squads/my-squad/.squad-state/`
- Squad in `~/squads/my-squad` → state in `~/squads/my-squad/.squad-state/`

This allows squads in different locations to maintain independent execution history.

## Anti-Patterns (NEVER)

- Creating a squad without `squad.yaml` manifest
- Registering without validating first
- Agent IDs without squad prefix
- Duplicate `slashPrefix` across squads
- Tasks without pre/post-conditions
- Skipping elicitation phase on create
- **Path Resolution:**
  - ❌ Hardcoding `./squads/` paths when squad might be in `~/squads/`
  - ❌ Forgetting to check `~/squads/` when squad not found in `./squads/`
  - ❌ Creating duplicate squad names in both `./squads/` and `~/squads/` — use local workspace squad
  - ❌ Relying on squad path location for security — treat both locations equally
  - ❌ Moving squad from `~/squads/` to `./squads/` without updating state/artifacts
- **v2: Skipping validation gate when defined**
- **v2: Deleting .squad-state/ during active run**
- **v2: Hardcoding model names**
- **v2: Passing raw LLM output when template is defined**
- **v3: Disabling doom loop on long-running squads** — it's there to save you
- **v3: Setting max_identical_outputs < 2** — too aggressive, causes false positives
- **v3: Using `change-strategy` without a fallback model** — escalation needs somewhere to go
- **v3: Putting large data in context instead of filesystem artifacts** — use filesystem collaboration
- **v3: Skipping self-verify on critical steps** — verification catches more bugs than you think
登录后下载/点赞/收藏 ❤ 5 | ★ 0
squads

目录结构

SKILL.md

举报内容

提示
