Skill v1.0.1
currentAutomated scan100/100+1 new
name: harness-initializer description: "Conducts a non-destructive audit of harness readiness. Use when assessing a repo before harness setup to inspect conventions. Do NOT use for installing files (use harness-bootstrap)." metadata: version: "1.0.1"
Harness Initializer
Audit any project before writing. Propose before creating. Write only with explicit user approval.
When to Use
Use for /harness-init, new projects, missing harness artifacts, or before declaring harness maturity complete.
Audit Sequence
Run in order from the current project root. Record each result. Do not create files during audit.
- Repo root — locate
.git, package files, or current directory. If unknown, say unknown. - Entry —
AGENTS.md,CLAUDE.md,README.md, or no root instructions? - Context —
docs/ARCHITECTURE.md,docs/PRODUCT.md,docs/RELIABILITY.mdpresent or missing? - State — find
feature_list.json,.specs/features/*/tasks.md,.specs/project/STATE.md,docs/harness/progress.md, oragent-progress.md. Also check foragent-os/learned-rules.json(learned rules ledger),agent-os/prevc-autonomy.json(PREVC autonomy policy: verificar se existe e seautoConfirméfalse; aviso setrue),agent-os/project-profile.json(verificar sefixLoop.maxAttempts<= 3), anddocs/harness/session-handoff.md. - Feedback & Review — find
init.ps1,init.sh,make check, package scripts, test/build/lint commands, or none. Check ifagent-os/quality-thresholds.jsonexists and which commands are wired (bypass_scan,skill_lint,redteam). Check ifdocs/review.mdexists and contains real critical failures. Check if.github/pull_request_template.md(ordocs/pull-request.md) exists and contains the 4 required sections (O que muda e por quê,Tier,Testing,Como reverter). Check if thereviewrole inconfig.ymluses a different model family fromimplement. - Evaluation — find
agent-os/judges/, project rubrics, CI checks, or none. - Skills & Evolution — find project
.opencode/skills/harness-*and global harness skills. Also check foragent-os/skills-proposals/(project-local skill proposals mined from failure traces). - Fresh-session test — answer from files only:
- What does this repo do?
- How is it organized? (architecture)
- How do I start and verify it?
- What is currently unfinished?
- What is the next task?
- What decisions were made and why?
Output
Produce a table. Then stop. Do not create files.
| Layer | Artifact | Status | Gap | Recommended Fix | |
|---|---|---|---|---|---|
| Review | .github/pull_request_template.md | Present/Missing | Missing 4 required sections | Install PR template with required sections | |
| Review | docs/review.md | Present/Missing | Generic checklist, no real critical failures | Populate real area critical failures | |
| Autonomy | agent-os/prevc-autonomy.json | Present/Missing | Missing or autoConfirm: true | Ensure autoConfirm: false | |
| Profile | agent-os/project-profile.json | Present/Missing | fixLoop.maxAttempts > 3 | Set fixLoop.maxAttempts <= 3 | |
| Roles | config.yml | Checked | review uses same model family as implement | Configure distinct model family for review |
Then propose a minimal bootstrap plan and ask approval before writing anything.
Rule
Do not assume init.ps1, feature_list.json, .specs/project/STATE.md, or zharnessengineering/ exist. If a discovered startup command fails, make that the first fix unless the user asks for an audit-only report. Context layer gaps are non-blocking — flag them but continue the audit.