Layered PowerContext Skills

The Skill entry is a small intent router. It explains when to search, list, save, hand off, or inspect candidates; workflow references carry detailed arguments, result interpretation, and approval boundaries.

Ordinary coding and summaries with sufficient current context need neither a Skill load nor a PowerContext call. An explicit operation still needs its real tool and observed result. An agent may call a self-contained tool directly, or load the relevant domain when it needs detail. Reading the router first, loading all domains, and loading before every response are not prerequisites. English and Chinese intent phrases live in actual Skill descriptions.

Host layout

File-backed hosts follow MiniMax's existing local-reference organization. Keep references beside the installed entry. OpenCode's npm archive includes the complete Skill directory. WorkBuddy's installer resolves Python and Scope-helper placeholders in all installed Markdown resources. OpenClaw uses the SDK's manifest skills directory mechanism. DSH uses its existing runtime Skill service, so the host can discover domain descriptions and load one domain directly.

All maintained Skill entries use powercontext-project-context, including runtime registration and installation paths. There are no alternate entry names. DSH's separately loadable domain Skills keep their domain-specific names. Framework adapters and Bub do not expose these Skill entries.

Workflow boundaries

Memory search finds relevant history; inventory lists a requested collection. An empty search does not authorize listing. Explicit save uses the Memory write; automatic Source capture is not saved Memory. Preserve exact citations on revisions.

Temporary Handoff returns the complete prepared carrier, including required nullable fields and generation receipts. Use each host's actual high-level or capture/activate/finalize workflow. Only an explicit durable-milestone request permits commit. Inspected facts without exact PowerContext citations remain declared claims. Acknowledgement checks current evidence, capability, and authority; it does not prove that work ran.

Review inspection and generation do not grant approval, publication, installation, or execution. Each host retains its actual exposed operations and approval channel. OpenClaw retains private-session and tool-availability gates.

Skill absence does not disable independently sufficient tool guidance. A missing resource must identify its exact Skill/path; a failed operation must identify its name and safe returned reason. Distinguish empty, rejected, unavailable, and unknown outcomes. Do not invent a diagnosis, repeat an unconfirmed write blindly, or claim success without a result.

Validation

Run uv run pytest tests/test_layered_skills.py to read each shipped entry and reachable reference, check missing-resource diagnostics, and inspect OpenCode/OpenClaw npm archive contents. Installer regressions exercise actual OpenCode reference copying and WorkBuddy placeholder expansion. Pi's package test uses its SDK Skill loader. OpenClaw's package test invokes skills list --json with an isolated profile. DSH's runtime suite verifies model-visible domain metadata and loading via the real host skill tool. These tests are separate from live-model routing evidence.

Export catalogs as described in tool-routing validation, then opt into resource reads:

uv run python scripts/evaluate_integration_guidance.py \
  --catalog /tmp/pc-guidance/dsh.json --layered-skills --env-file .env \
  --output /tmp/pc-guidance/layered.json --skill-modes unloaded unavailable

Repeat --catalog for each host. The evaluator offers an optional reader of the actual shipped resource text and records every requested resource. It does not force a Skill load or a data tool. Unavailable mode exposes no reader. This controlled reader is evaluation infrastructure, not a native host loader or distribution source. MCP-backed catalogs, including MiniMax, do not prove each product's native automatic discovery. Routing and argument validation remain separate from transcript-bound reporting review. Preserve failed observations and evaluation limits.

Explicit skill_search and skill_handoff probes must successfully read their respective Memory or Handoff workflow before the data operation. Reading only the router, a different domain, or reading after the operation does not qualify. The report records the expected resource and resources read before the operation, preserving any more specific failure. Ordinary requests do not require a Skill read, and unavailable mode still evaluates independently usable tools.

Run-specific configuration, transcripts and findings belong in an evaluation report; the PR records delivery status.