agentleFS
hushh-labshushh-labs/hushh-research/.agents/skills/puppy-one-harness/SKILL.md

puppy-one-harness

Skill40 starsChanged 5 days agoApache-2.0View on GitHub

Use when running, grading, or extending the on-device model harness for Puppy One, including the review-queue judge handoff, capability probing, and the evolution ledger.

The file

We split this skill into the rules it gives an agent

6 rules. About 0 are common, 1 is distinctive, and 5 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 5

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Avoid overlap with quality-contracts and backend-agents-operons.

    line 13
  2. Very specific:

    Every wrong verdict cites the offending value verbatim.

    line 63
  3. Distinctive:

    Before comparing two runs, check the capability profile.

    line 70
  4. Very specific:

    If the request is still broad or ambiguous, route it back to backend.

    line 75
  5. Very specific:

    If the finding is a provider-routing or gate issue, use backend-runtime-governance.

    line 76
  6. Very specific:

    If it touches the vault or the PKM write path, use security-audit.

    line 78

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Puppy One Harness Skill

Purpose and Trigger

  • Primary scope: puppy-one-harness
  • Trigger on benchmarking a local model, grading its output, adding a model to the ladder, or reading the evolution ledger.
  • Avoid overlap with quality-contracts and backend-agents-operons.

The harness answers two different questions and must never conflate them. The benchmark asks well-formed and fast; the judge asks correct. A model can emit a perfectly shaped save_to_pkm call that files a dietary restriction under finance.accounts, and the shape check passes it.

Coverage and Ownership

  • Role: spoke
  • Owner family: backend

Owned repo surfaces:

  1. .agents/subagents/local_model_judge.md
  2. .agents/skills/puppy-one-harness/references/judging-contract.md

Non-owned surfaces:

  1. consent-protocol
  2. hushh-webapp
  3. backend

Do Use

  1. Running the benchmark or the judge over a local model.
  2. Grading a review queue, or reviewing someone else's verdicts.
  3. Adding a model to the ladder, or reading the ledger for a trend.
  4. Grading the device's real daily jobs (the cron quality suite, see .agents/skills/puppy-one-harness/references/cron-quality-suite.md).

Do Not Use

  1. Changing the PKM write path or the vault (security-audit).
  2. Provider routing or the on-device gate itself (backend-runtime-governance).
  3. Frontend surfaces for Puppy One (frontend-surface-placement).

Read First

  1. .agents/skills/puppy-one-harness/references/judging-contract.md
  2. .agents/subagents/local_model_judge.md

Workflow

  1. Collect outputs from the local model, then write a review queue. Controls are planted, shuffled per run, and unmarked; scoring reads them from the seal.
  2. Grade in a session that did not write the queue. This cannot be enforced from a script, so it is a discipline the report states rather than assumes.
  3. The answer key is the seal (rows, rules, seed) and memory_judge_controls in the harness. Never open or hunt for either; the manifest locates no control.
  4. Every wrong verdict cites the offending value verbatim. Ingest checks the citation against the output and discards it if absent, because an uncited failure is indistinguishable from a hallucinated one.
  5. Grade every row. Ungraded rows void the run: skipping the hard ones would raise accuracy for free.
  6. Ingest, then append to the ledger. Record void runs too — dropping them makes the ledger an unbroken record of successes.
  7. Before comparing two runs, check the capability profile. A model tested via tool calling and one tested via JSON mode were not asked the same question.

Handoff Rules

  1. If the request is still broad or ambiguous, route it back to backend.
  2. If the finding is a provider-routing or gate issue, use backend-runtime-governance.
  3. If it touches the vault or the PKM write path, use security-audit.

Required Checks

python3 .agents/skills/agent-orchestration-governance/scripts/agent_orchestration_check.py
python3 .agents/skills/codex-skill-authoring/scripts/skill_lint.py

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.