Skip to content
cccskills
FreePublished on GitHub

rulake-query

Query the witness-anchored ruLake cache — search, verify, explain, or refresh. Returns ranked results plus a decision_trace block with cost, latency, witness match, and substrates used.

The English summary is being prepared. Showing the original description.

See how to install

Files included (1)

  • SKILL.md2.2 KB

SKILL.md (original)

Review the instructions the agent will receive before you install.

/rulake-query

Wraps the rulake_query MCP tool with sensible defaults so a fresh operator can issue their first query without thinking about the intent / risk / freshness / target axes.

Inputs (all optional except the question)

  • question (required): natural-language query
  • intent: search | verify | explain | refresh (default: search)
  • freshness_ms: tolerance for stale cache hits (default: 5000)
  • risk: low | medium | high (default: low)
  • target_collection: limit to a specific backend.collection (default: federated across all)

What you get back

{
  "result": [ /* ranked retrieval */ ],
  "decision_trace": {
    "chosen_path": "deterministic-retrieval-path-v0.1",
    "intent": "search",
    "freshness": { "budget_ms": 5000, "actual_ms": 1031 },
    "cache":  { "hit": true, "hit_ratio_session": 0.87 },
    "substrates_used": ["gcs-backend", "rvdna-backend"],
    "kernel": { "id": "cpu-naive", "deterministic": true },
    "witness": { "match": true },
    "cost":    { "compute_kernel": 0.0, "backend_fetch": 0.0, "cache_hit_discount": -1.0 },
    "latency": { "total_ms": 1.02, "cache_ms": 0.4, "witness_ms": 0.6 },
    "refusals": []
  }
}

The decision_trace block is the new contract from ADR-009 — it lets a calling agent negotiate cost vs trust vs latency without reaching around the abstraction. Note: the trace shape ships in v2.4 — until then, responses carry the data only and the trace is in the JSONL audit log on the server side.

Examples

/rulake-query "what does ADR-157 commit to?"

/rulake-query --intent verify --freshness_ms 0 "is bundle b1f4...3d valid?"

/rulake-query --target_collection adrs.rulake "list all merged ADRs"

See also

  • ADR-004 §4b — the eight tool tier matrix
  • ADR-009 — the deterministic retrieval path + decision_trace shape

Reviews

No reviews yet. Tell others what you thought.

Comments

No login needed. Share how it worked for you, ask a question, or note where it did not run.

No comments yet. Be the first to write one.

Comments are published right away. Do not include personal information such as email addresses or phone numbers. Up to 2 URLs.

0 / 1,000 characters

This form is protected by reCAPTCHA, and the Google Privacy Policy and Terms of Service apply.

More skills from this repository

memory

Free

Use ruLake as the agent's working memory. Auto-invoke /rulake-memory:memory-recall when the agent needs to look up something it might already know; auto-invoke /rulake-memory:memory-remember after the agent learns or decides something worth pinning. Closes the self-learning loop without requiring the agent author to wire it explicitly.

The English summary is being prepared. Showing the original description.

ruvnet/RuLake15Updated Oct 3, 2026

See all skills by ruvnet

Report a problem with this skill