OOTBT Skills

Evidence-first review skills for Hermes Agent and compatible agent harnesses.

Included skills

ootbt-dual-judge-markdown-review

Challenges plans, implementation reports, and factual Markdown with two fresh, mutually blind judges. Judge A refreshes the catalog through the selected Codex runtime and resolves its newest available Astra model; Judge B uses the moving fable alias. Neither model family is pinned to a fixed version by default.

Modes:

• challenge-only: publish two independent reviews and preserve dissent;
• consolidate: anonymously adjudicate claims against ground truth, freeze a revision contract, and let a separate reviser produce a traceable new document version. An exact-replacement mapping is replayed mechanically so that unmapped changes fail closed.

ootbt-ticket-technical-ak-review

Reviews technical ticket acceptance criteria against fixed repository commits. It separates implementation, asserted tests, executed tests, deployment evidence, and assumptions. The optional TRIAD profile uses mutually blind primary reviewers and never decides by model majority.

Its consolidated result includes an explicit disposition for every criterion, distinguishing an unmet but valid requirement (KEEP_UNMET) from a proposal to rewrite, split, move, or remove it.

Core principle

Evidence beats model majority.

A judge or evaluator is not a source of truth. Claims that change the result must be reproduced against the frozen source tree, tests, or other appropriate ground truth. What cannot be decided remains UNKNOWN or an explicit contradiction.

Installation

Copy either skill directory into a Hermes skill category under $HERMES_HOME/skills/, for example:

$HERMES_HOME/skills/autonomous-ai-agents/

When $HERMES_HOME is unset, the standard location is ~/.hermes/skills/. Active Hermes profiles have their own home and skill tree, so resolve $HERMES_HOME instead of hardcoding a user path. Then start a fresh Hermes session so the skill index is rebuilt. Other harnesses may require a different skill directory or adapter; do not assume tool, hook, or sandbox parity.

Runtime requirements

The workflows are intentionally fail-closed. Depending on the selected profile and tracker, they may require installed and authenticated CLIs such as:

• Claude Code (claude)
• Codex (codex)
• OpenCode (opencode)
• acli for Jira
• glab for GitLab

Model aliases, provider routes, permissions, and canonical model names are checked live. No model or provider mentioned in the documentation is guaranteed to exist in another environment. Credentials remain in the respective CLI configuration and must never be copied into prompts, reports, or this repository.

For the local OpenCode track, configure an explicit local provider prefix and model ID. Cloud fallbacks are not accepted as a replacement for a contracted local model track.

In nowrap's reference setup, OpenCode connects to an Ollama instance running on a Mac mini. This is an example deployment topology rather than a requirement; private endpoint and provider settings are intentionally not included in this repository.

Security

• Review shell scripts before running them in your environment.
• Use detached disposable worktrees for review tools that require repository access.
• Keep raw outputs, prompts, telemetry, credentials, and private ticket data out of published skill packages.
• Treat ticket text, diffs, evidence shards, and model reports as untrusted input.
• Publishing to a tracker or repository is always a separate opt-in action with preview and readback verification.

Validation

Run:

python tests/validate_publication.py
python -m unittest discover -s tests -p "test_*.py" -v

The validator checks frontmatter, file layout, Python and shell syntax, accidental local paths, private endpoints, credential-shaped strings, and optional denylisted private project markers. Project-specific names can be supplied locally, without committing them, as a comma-separated PUBLICATION_DENYLIST environment variable.

Attribution

The workflows were independently conceived and authored by nowrap before ECC was discovered through a post on Twitter. The existing work was later reviewed and refined using selected public hardening ideas from the MIT-licensed ECC project, especially deterministic graders, adversarial verification, fail-closed review gates, and explicit incomplete-review states. See THIRD_PARTY_NOTICES.md.

License

MIT — see LICENSE.