retrieval-system-audit
Skill17 starsChanged 2 days agoApache-2.0View on GitHub
Audit and improve stellar-raven-codemode retrieval across every exposed service operation and skill. Use for comprehensive endpoint, response-shape, catalog-description, search-ranking, execute-grounding, evaluation-coverage, golden-truth, or cross-source weighting reviews that require live probes and measured A/B evidence.
In this repository9 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- audit-reviewabilitySkill
- cloudflare-observability-reviewSkill
- golden-truthSkill
- improvements-pipelineSkill
- live-drift-resolutionSkill
- run-evalsSkill
- truth-maintenanceSkill
The file
We split this skill into the rules it gives an agent
7 rules. About 0 are common, 6 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈6
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 26
Keep data, soft-empty, and error responses distinct.
- Distinctive: line 27
Keep secrets and service traffic outside model-authored code.
- Distinctive: line 28
Never call a paid or side-effecting operation during inventory work.
- Distinctive: line 29
Never tune production behavior for one question.
- Distinctive: line 30
Prefer deletion, simpler descriptions, and existing interfaces to new code.
- Very specific: line 48
Trace retrieval decisions: Follow search from catalog generation through scoring, tiering, quotas, descriptions, profiles, and recovery …
- Distinctive: line 64
Implement the smallest proven change: Keep measurement changes separate from product changes, with disjoint write ownership.
≈ “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.
Files
Retrieval system audit
Measure the retrieval system before changing it. Ship only general improvements that survive real endpoint calls and reviewed golden questions.
Scope first
A full audit is the largest task in this repo. Before you start, write in the round ledger the
question the audit answers, the surfaces and steps below that it needs, and the expected time and
cost. Run only those steps. Probes and offline instruments are free and run here against the local
dev server; every paid run follows the run-evals spend rule. Stop when the question is answered.
Companion skills: run-evals for every evaluation and spend decision, golden-truth before any
golden change, improvements-pipeline for verified upstream defects, and live-drift-resolution
first when the committed catalog differs from the live surfaces. Read PLAN.md,
ARCHITECTURE.md, eval/EVALS.md, and eval/qa/README.md for the parts in scope.
Fixed boundaries
- The manifest is the exposed surface. Keep exact operation and skill IDs.
- Keep data, soft-empty, and error responses distinct.
- Keep secrets and service traffic outside model-authored code.
- Never call a paid or side-effecting operation during inventory work.
- Never tune production behavior for one question.
- Prefer deletion, simpler descriptions, and existing interfaces to new code.
- Raw paid results stay local and gitignored.
- The coordinator holds Git, spending, deployment, and external filing authority.
Steps
- Pin. Work in a dedicated worktree with a round ledger. Record the revision and the manifest,
runner, and corpus digests. Run the free baseline from
AGENTS.mdand confirm that regeneration leaves the tree clean. - Build one surface ledger. For each operation and skill in scope, compare the committed
manifest, generated spec, adapters, live catalog, and fresh-client MCP descriptions. Use the
fields in
references/evidence-contract.md. Separate deployed drift from a client that cached an older MCP schema. - Probe live response contracts. Design the probe matrix first and mark every probe read-only, paid, side-effecting, secret-bearing, or unsafe. Call each safe read-only operation with a representative input, then add bounded soft-empty and invalid-input probes by response class. Compare returned keys and nested shapes with the model-visible schema. Keep hashes and compact projections, not raw bodies.
- Trace retrieval decisions. Follow
searchfrom catalog generation through scoring, tiering, quotas, descriptions, profiles, and recovery guidance. Followexecutethrough adapters, envelopes, truncation, artifacts, and evidence packing. For each suspected defect, name the first failed boundary: discovery, schema, retrieval, execution, upstream data, synthesis, judge, or golden truth. When the evidence is absent upstream, file a finding and make no runtime change. - Audit evaluation coverage. Map each operation and skill to its routing, holdout, discovery, plan, QA, live, temporal, and agentic coverage. Separate direct case labels from operations that agents called. Review every consequential verdict change against the transcript and golden. Claim no gain below the documented noise floor, and compare different memberships only on common cases.
- Pre-register each experiment. One hypothesis per candidate: mechanism, target cases, symmetric controls, one mechanism metric, expected time and cost, hard cap, stop rule, and acceptance threshold. Size the controls to the decision, and state and justify their coverage before any outcome is read. Use identical case IDs, runner code, judge tuple, and arm order across arms. Collect answers before judging when the runner supports it.
- Implement the smallest proven change. Keep measurement changes separate from product changes, with disjoint write ownership. Use generation scripts. Remove rejected experimental code and dormant flags. Advance a candidate only after two unrelated targets show the intended mechanism. Stop after a verified control, security, exposure, or exact-match regression.
- Close. Goldens change only through
golden-truth; service defects go toimprovements/, never behind a local answer patch. A reviewer who did not author the change reviews it, and every finding is reconciled. Run the tests, smoke tests, build, routing gate, relevant evals, improvements lint, and secret scan that the change touches. Record artifacts, costs, failures, and rejected hypotheses. Stop only the lanes you started.
Evidence rule
An aggregate score never overrides a blocking regression. A shipped change needs reviewed
correctness, causal evidence, stable retrieval, complete accounting, and no security or exposure
regression. Cost counts as run-evals states: a much cheaper path that is only slightly less
correct can be the better product.
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
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.

