maicongambini/omp-harness-guide/harness-redteam-check
Claude Code and agent skill by maicongambini/omp-harness-guide
- community
Checking that the SKILL.md file is still in the repository.
Run bounded adversarial scans on reconciled diffs. Use when verifying lane diffs before gates or commits. Do NOT use for full harness secret audits (use harness-security-scan).
Install
npx vskill@latest install maicongambini/omp-harness-guide/harness-redteam-checkSKILL.md
The same SKILL.md was published earlier as maicongambini/omp-harness-guide/harness-redteam-check.
Harness red-team check
Every reconciled lane runs one bounded adversarial check against that lane's changed files or explicit diff. The runner owns a 120-second scan deadline. The complete red-team ceremony for a simple or medium task has a 10-minute wall-clock budget. Deep active exploitation is not part of this loop.
Procedure
- Identify the changed files or explicit diff for one reconciled lane. Do not pass the whole
repository when a lane target is available.
- Call the bounded runner:
``text node scripts/harness-redteam-runner.mjs --lane-target <changed-files-or-diff> --deadline-ms 120000 --json ``
On OMP, use the installed runner path. The runner delegates detection to harness-redteam-scan.mjs, reports result, elapsedMs, scannedScope, and findings, and never dispatches an agent.
- Branch on the typed result:
pass: at tierauto, finish the check without a reviewer. At tiersamplingorfull,
escalate once to the read-only security-analyst for the reconciled lane. A wave may use one aggregated review instead. Never dispatch once per file.
fail: block lane reconciliation. Fix every finding, run the same lane target again, and
keep reconciliation blocked until a subsequent result is pass or an approved reviewer decision changes the lane state.
unavailable: block lane reconciliation and record the reason. A timeout, crash, invalid
target, unreadable target, or malformed scanner response is never a pass.
- Measure the aggregate wall-clock ceremony from runner start through any required review. If it
exceeds 10 minutes for a simple or medium task, record a blocking unavailable result. Do not convert a budget overrun into pass.
- Store the runner result, exact scanned scope, elapsed time, and any review decision in the lane
evidence. Findings remain tied to the lane that produced them.
Boundaries
- Authorized targets are the operator's own repository tree and explicit fixtures only.
- The scan is read-only and deterministic. It makes no network calls, mutates no target, and does
not perform active exploitation.
- Escalation is a scheduler decision. The runner never dispatches an agent.
- Image-bearing or multi-runtime payloads follow
harness-capability-check.mjs, not this skill.
Time budgets
| Step | Budget | |
|---|---|---|
| Per-lane deterministic runner | 120 seconds maximum | |
| Total ceremony for a simple or medium task | 10 minutes wall-clock maximum | |
| Security-analyst escalation | Once per reconciled lane, or once per aggregated wave |
Security
- Security scanPassed, 90 of 100Pattern checks on the skill's files.
- Model reviewPassed
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingomp/agent/skills/harness-redteam-check/SKILL.md
Popularity
- 7-day trend score
- 5
Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/maicongambini/omp-harness-guide/harness-redteam-check)