plan-guardian-skill
A mandatory 7-step plan, then verification by independent subagents — as a
DeepSeek Harness plugin. The skill ships inside the bundle: install the plugin
and it is there, with the reference documents it cites.
A Codex skill that enforces verifiable planning, then validates completion with
independent memoryless subagents.
Install
DeepSeek Harness Desktop — open Plugins in the sidebar, choose Add
plugin, and enter:
https://github.com/bauerelizabeth07139/plan-guardian-skill
Then switch the new dsh-plan-guardian bundle on. The Desktop app boots the
reserved desktop profile, so that is where it has to be enabled.
dsh CLI — install it into the profile you actually boot:
dsh plugin --profile web add bauerelizabeth07139/plan-guardian-skill
No git on the machine? pnpm resolves a git shorthand with git ls-remote,
which fails with 'git' is not recognized when git is missing. Use the tarball
instead — that path is plain HTTPS:
dsh plugin --profile web add https://codeload.github.com/bauerelizabeth07139/plan-guardian-skill/tar.gz/master
The same address works in the Desktop Add plugin dialog. Replace master
with a commit SHA to pin an exact revision (/tar.gz/<sha>).
Uninstall with dsh plugin --profile web remove dsh-plan-guardian.
The workflow
| Step | What happens |
|---|
| 0 | Establish verification capability — once per conversation |
| 1 | Restate the intent; name the core deliverable and the implicit requirements |
| 2 | Draft exactly 7 concrete steps, implementation and verification kept apart |
| 3 | Give every step at least one binary acceptance criterion |
| 4 | Execute through worker subagents — they never see the criteria |
| 5 | Verify every deliverable through a memoryless verifier subagent |
| 6 | On FAIL, revise the plan, re-execute, re-verify |
| 7 | Report every step as PASS or FAIL, with evidence |
Hard rules the skill will not bend: never skip a step, never self-verify, never
produce fewer than 7 steps, never report a step as passed without verifier
evidence.
Full instructions: skills/plan-guardian/SKILL.md.
Supporting protocols: plan_protocol.md
and memoryless_review.md.
What changed for DeepSeek Harness
The 7 steps, the hard rules and the acceptance-criteria discipline are
unchanged. Two things are genuinely different:
- Step 0 is no longer a model probe. In Codex the skill had to discover
whether the session model could see images, by reading
~/.codex and
~/.mimo2codex and probing a model API. On DSH, every agent has read_image,
so visual verification is always available and there is nothing to detect.
The original probe is still in scripts/detect_multimodal.py
for Codex users; the DSH skill does not call it.
- Subagent calls are the harness's.
subagent (memoryless) replaces
multi_agent_v1__spawn_agent(..., fork_context=false), and a forked
subagent is explicitly forbidden for verifiers because it would inherit the
conversation's assumptions. Workers report back on their own instead of a
separate wait_agent call.
Repository layout
| Path | Purpose |
|---|
index.js | the DSH plugin: registers the bundled skill |
cordis.patch.yml | the loader row that activates the plugin |
skills/plan-guardian/ | the skill in Harness tool names, with its references |
locale/{en,zh}.json | card title and description for the plugin lists |
assets/icon.svg | card artwork |
test/plugin.test.mjs | static composition checks |
SKILL.md, references/, scripts/, agents/, AGENTS*.md | the original Codex skill, unchanged |
Development
No build step and no runtime dependencies — @deepseek-ai/cordis and
@deepseek-ai/dsh-skill are peers supplied by the Harness.
npm test # node >= 22: manifest, card metadata, loader patch, skill, syntax
Codex usage (unchanged)
The original files are still here, byte for byte:
git clone https://github.com/bauerelizabeth07139/plan-guardian-skill.git
cd plan-guardian-skill
cp -r . ~/.codex/skills/plan-guardian
cp AGENTS.global.md ~/.codex/AGENTS.md # optional: the global instructions
python scripts/detect_multimodal.py # Codex-only capability probe
python scripts/validate_skill.py . # Codex-only skill validator
DSH also reads AGENTS.md — including a user-global $DSH_HOME/AGENTS.md — so
AGENTS.md / AGENTS.global.md remain useful if you want the rules applied
outside a skill invocation, with the subagent tool names swapped for the ones in
the mapping table above.
License
MIT
