Back to home

oukeming64-tech

codex-skills

Evidence-first agent skills for handoff auditing and documentation sync, packaged for Codex and DeepSeek Harness.

Stars
0
Language
Python
Created
Jul 1, 2026
Updated
Aug 17, 2026

Introduction

Evidence-First Agent Skills

Portable agent skills distilled from real project work and maintained by oukeming64-tech. The current collection focuses on one boundary: an agent should not call work complete until its code, evidence, documentation, release state, and stated scope agree.

Skills

  • handoff-auditor: audit implementation handoffs before accepting, merging, or declaring them complete. It checks claims against current code, verification evidence, docs, release state, and human acceptance requirements.
  • docs-sync-guardian: keep repository documentation aligned with code, configuration, asset, release, workflow, or product changes. It treats docs as suspicious after changes, grounds updates in live implementation evidence, and uses a risk gate for human-judgment text.

Both skills are instruction-only. They add no MCP server, network client, credential handling, telemetry, background service, or install hook.

DeepSeek Harness

This repository includes a community, Skill-only bundle for the DeepSeek Harness developer preview. It is not an official DeepSeek product and does not imply DeepSeek endorsement.

Compatibility target: @deepseek-ai/dsh@0.1.0-rc.6 with Node.js ^22.19.0 || >=24.0.0.

Install the tagged GitHub package into the Web profile:

dsh plugin --profile web add 'github:oukeming64-tech/codex-skills#dsh-v0.1.0'

Invoke either skill explicitly:

Use the handoff-auditor skill to verify this completion claim before accepting it.
Use the docs-sync-guardian skill to check this change for documentation drift.

Uninstall:

dsh plugin --profile web remove @oukeming64-tech/codex-skills

The integration only registers this repository's skills/ directory through Harness's ordinary filesystem Skill provider. It does not add native tools or change the default permission model.

Other Agent Runtimes

Copy a skill folder from skills/ into the runtime's user or project Skill directory. For Codex, use ~/.codex/skills/ or a compatible .agents/skills/ location. For a Harness project-local installation, use .dsh/skills/.

Keep the folder name unchanged so it matches the name in SKILL.md.

Repository Validation

Run the fusion A/B coverage smoke check before claiming an absorbed skill update is complete:

python3 evals/fusion_ab_eval.py --old-ref 1d9949e --new-ref HEAD

To focus one skill:

python3 evals/fusion_ab_eval.py --old-ref 1d9949e --new-ref HEAD --skill docs-sync-guardian

This check catches missing fusion targets and baseline regressions. It does not replace independent review or scenario-based behavior testing.

The DeepSeek Harness package contract is checked separately:

node --test dsh/test/*.test.mjs

Ownership and Issues

Released under the MIT License.