hm-l3-opencode-project-audit

Files (20)

• phase-1-skills.md · .harness-audit/compiled/
• phase-2-commands.md · .harness-audit/compiled/
• phase-3-tools.md · .harness-audit/compiled/
• phase-4-permissions.md · .harness-audit/compiled/
• phase-5-agents.md · .harness-audit/compiled/
• phase-6-subagents.md · .harness-audit/compiled/
• phase-7-synthesis.md · .harness-audit/compiled/
• SKILL.md
• phase-1-skills.md · assets/profiles/
• phase-2-commands.md · assets/profiles/
• phase-3-tools.md · assets/profiles/
• phase-4-permissions.md · assets/profiles/
• phase-5-agents.md · assets/profiles/
• phase-6-subagents.md · assets/profiles/
• phase-7-synthesis.md · assets/profiles/
• evals.json · evals/
• rich-gate-scorecard.md · metrics/
• pointers.md · references/
• compile-bundle.sh · scripts/
• validate-skill.sh · scripts/

SKILL.md (first lines)

## Overview

Comprehensive audit of OpenCode project configuration across skills, agents, commands, tools, and permissions. Use when onboarding to a new project, after structural changes, or when verifying ecosystem integrity. Produces a fact-based audit report covering boundaries, governance, and configuration health.

## The Iron Law

```
AUDIT REPORTS FACTS. NEVER BLOCKS. NEVER FIXES. NEVER DOES THE WORK ITSELF.
```

<files_to_read>
.opencode/skills/harness-audit/references/pointers.md
.opencode/skills/harness-audit/scripts/compile-bundle.sh
.opencode/skills/harness-audit/scripts/validate-skill.sh
</files_to_read>

# harness-audit

Comprehensive audit orchestrator for ANY OpenCode project. Reports facts, leaves judgment to the agent.
## Architecture

```
harness-audit/
├── SKILL.md                    # Thin orchestrator (this file) — no YAML agent config
├── assets/
│   └── profiles/               # 7 subagent profile templates
│       ├── phase-1-skills.md
│       ├── phase-2-commands.md
│       ├── phase-3-tools.md
│       ├── phase-4-permissions.md
│       ├── phase-5-agents.md
│       ├── phase-6-subagents.md
│       └── phase-7-synthesis.md
├── references/                 # Skill pointers for execution_context
│   └── pointers.md
└── scripts/
    ├── compile-bundle.sh
    └── validate-skill.sh
```

## On Load

1. Run `bash scripts/compile-bundle.sh` — compiles all 7 subagent profiles
2. Run `bash scripts/validate-skill.sh` — validates structure before dispatch
3. Read project context: `opencode.json`, `AGENTS.md`
4. Apply the official OpenCode scope matrix before deciding what exists or is missing: agents in `.opencode/agents` or `opencode.json.agent`, commands in `.opencode/commands` or `opencode.json.command`, config overrides from `OPENCODE_CONFIG`/`OPENCODE_CONFIG_DIR`, and rules precedence from `AGENTS.md`/`CLAUDE.md` plus configured instructions.

## Execution Flow

### Phase 0: Bootstrap (FIRST RUN ONLY)
If `assets/profiles/` is empty, inform user:
> "Subagent bundle compiled. Please restart session and re-run audit."

### Phases 1-6: Parallel Dispatch (run simultaneously)

| Phase | Target | Profile | Execution Context |
|-------|--------|---------|-------------------|
| 1 | Skills | `assets/profiles/phase-1-skills.md` | `use-authoring-skills` |
| 2 | Commands | `assets/profiles/phase-2-commands.md` | `command-dev` |
| 3 | Tools | `assets/profiles/phase-3-tools.md` | `custom-tools-dev` |
| 4 | Permissions | `assets/profiles/phase-4-permissions.md` | `opencode-platform-reference` |
| 5 | Agents | `assets/profiles/phase-5-agents.md` | `agents-and-subagents-dev` |
| 6 | Subagents | `assets/profiles/phase-6-subagents.md` | `agents-and-subagents-dev` |
| 7 | Verification | spec-verifier agent | Independent validation pass |

Dispatch each via Task tool with `run_in_background: true`.

### Phase 7: Sequential Synthesis (after 1-6 complete)

| Profile | Focus |
|---------|-------|
| `assets/profiles/phase-7-synthesis.md` | Aggregate all findings, write audit-report-YYYY-MM-DD.md |

## Subagent Profile Envelope

Each profile contains:

```
role: <specialist-auditor>
core_principle: <audit focus>
verification_dimensions:
  - <dimension-name>: <what to verify>
templates:
  - name: <template-name>
    description: <what-to-check>
    command: <bash-command-if-applicable>
forbidden_files:
  - <paths-to-avoid>
critical_rules:
  - <rule-1>
  - <rule-2>
structured_returns:
  findings: <json-schema>
  evidence: <file-references>
  risk_level: <none|low|medium|high|critical>
success_criteria:
  - <criterion-1>
  - <criterion-2>
```

## Dispatch Protocol

```
delegate-task (run_in_background: true)
  └── spawns: 6 parallel subagents (Phases 1-6)
      └── each MAY spawn multiple children for deep investigation
  └── after all complete → Phase 7 (sequential)
```

- NO YAML agent configuration files
- All routing controlled via this SKILL.md
- One subagent can spawn multiple children
- Parallel: 4-6 simultaneous subagents maximum

## Inputs

| Input | Description | Required |
|-------|-------------|----------|
| `target_repo` | Path to OpenCode project | Yes (defaults to cwd) |

From shynlee04/hivemind-plugin, MIT.

More from shynlee04/hivemind-plugin