artifact-librarian

name	artifact-librarian
description	Operate the local Agent Librarian workflow using the synthetic sample collection, approval-gated runtime wrapper, deterministic CLI outputs, and human-review summary.

Artifact Librarian

Use this skill to run a public-safe Codex demo of agent-librarian. Lead with the functional portable-agent workflow, not taxonomy.

Workflow

Codex reads AGENTS.md
-> Codex explains safe scope
-> Codex proposes runtime-wrapper commands
-> user approves exact command
-> wrapper runs deterministic CLI backend
-> Codex summarizes CLI evidence
-> human reviews generated outputs

Steps

1. Explain the agent in plain language.

   • Codex is the interface layer.
   • The deterministic CLI and generated outputs are the source of truth.
   • The demo uses only the synthetic examples/sample-collection.

2. Inspect only allowed public demo files.

   • Allowed source: examples/sample-collection.
   • Allowed generated catalog: examples/generated-catalog.
   • Allowed package instructions: packages/openai/codex/.
   • Do not scan private/work files, work-internal folders, credentials, secrets, private prompts, private traces, logs, memory snapshots, state snapshots, employer/client data, internal URLs, or private generated catalogs.
   • Do not use non-demo paths.

3. Propose commands before running anything.

   • Prefer runtime-wrapper propose commands.
   • Show exact command, read scope, write scope, generated files, and sensitivity note.

4. Require exact approval.

   • Do not run on vague approval.
   • Do not run if the approval string is different from the command shown.
   • Changed command, path, argument, sensitivity, or retry requires fresh approval.

5. Run only approved wrapper commands.

   • Do not run arbitrary shell.
   • Do not chain commands.
   • Do not execute, edit, delete, merge, publish, or rewrite source files.
   • Do not create git commits, pushes, tags, releases, pull requests, or GitHub/yeet publishing flows.
   • Do not edit repo files during the demo unless the user explicitly asks for development work.

6. Summarize deterministic outputs.

   • Use runtime-wrapper output, CLI output, and generated files as evidence.
   • Preserve warnings, diagnostics, validation failures, and overlap candidates.
   • Do not invent counts, files, findings, or status.

7. Explain what the demo proves.

   • Codex can scope a synthetic public demo.
   • Codex can propose bounded wrapper commands.
   • Exact approval gates local execution.
   • The deterministic backend produces evidence for human review.

8. Explain what the demo does not prove.

   • It does not certify safety, privacy, correctness, completeness, approval, compliance, or publication readiness.
   • It does not scan private or work-internal material.
   • It does not add OpenAI API integration, network behavior, MCP server code, arbitrary shell execution, or autonomous publication.

Allowed Demo Commands

python -m agent_librarian.runtime_wrapper propose catalog examples/sample-collection --out examples/generated-catalog
python -m agent_librarian.runtime_wrapper propose validate examples/generated-catalog
python -m agent_librarian.runtime_wrapper propose report examples/generated-catalog
python -m agent_librarian.runtime_wrapper run report examples/generated-catalog --approve-exact "agent-librarian report examples/generated-catalog"

Wrong approval demonstration:

python -m agent_librarian.runtime_wrapper run report examples/generated-catalog --approve-exact "wrong command"

The wrong approval should fail with a nonzero status and must not run the backend.