Resolve runtime paths before file or script operations.

Iron Law

Check existing tools before installing new ones.

Red Flags

• "This package looks useful, let me install it" → STOP, because installation requires explicit user approval after scoring — never auto-install.
• "It has 5000 stars so it must be safe" → STOP, because popularity alone is not safety — check for lifecycle scripts, license, recency, and dependency weight before recommending.
• "I need an external tool for this" → STOP, because built-in capabilities (Read, WebFetch, Bash) must be checked first — install only when built-ins genuinely cannot handle the task.
• "The README looks fine, skip the inspection" → STOP, because candidate inspection of the actual SKILL.md/README is mandatory for the top 3 — metadata alone is insufficient.
• "It scored 2.5 but it's the only option" → STOP, because packages below 3.0 must not be recommended — suggest building a custom pipeline instead.

Discover

Find existing skills first, then search external marketplaces only when needed and safe. Discovery is registry-first: prefer the official/local/verified registry path before broader community search.

When to Use

• The current skills do not cover the task
• A pipeline references a missing capability
• The user wants to browse or install additional skills

Workflow

0. Check built-in capabilities: Before searching, check if the requested capability is already available through built-in tools (Read for PDFs, WebFetch for URLs, Bash for shell tasks, etc.). If built-in tools suffice, report that and skip external search.
1. Scan local skills by name and description.
2. If no strong match, search external sources in priority order (see Source Prioritization). Treat official/local/verified registry sources as the default adoption path.
3. Candidate Inspection: Fetch and read README/SKILL.md for the top 3 candidates. If inspection is blocked (private repo, rate limit): (a) search package name + "review"/"tutorial", (b) check npm page for README, (c) note "inspection blocked" and apply a -1 score penalty. See references/discover-scoring.md for the full workflow.
4. Score candidates on relevance, popularity, recency, dependency weight, and source trust. Show the full weighted breakdown (see references/discover-scoring.md).
5. Apply the Build vs Install threshold. If no candidate scores above 3.0, recommend a custom pipeline instead.
6. Present ranked options with pinned versions and wait for explicit approval.
7. Install only with approval and only when the environment supports it.

Source Prioritization

Trust tier scores: Tier 1 = 5, Tier 2 = 4, Tier 3 = 3, Tier 4 = 2. Bumps apply for verified orgs or scoped packages.

Evaluation Weights

Each criterion is scored 1-5. The weighted sum produces a final score (1.0-5.0).

Thresholds

Safety

• Never auto-install. Pin exact versions (see references/discover-scoring.md).
• Postinstall check: Before recommending any npm package, inspect its package.json for preinstall, install, or postinstall scripts. Flag packages with lifecycle scripts and apply a -1 score penalty. Packages with obfuscated or network-calling lifecycle scripts must be rejected entirely.
• Flag heavy or stale packages.
• Degrade gracefully to local-scan-only mode if marketplace tooling is missing.
• Flag repos with no LICENSE file or suspicious content.

Output Format

Return a ranked list AND save results to <plugin-data>/discovers/{query-slug}.json:

{
  "query": "...",
  "searched_at": "ISO-8601",
  "built_in_check": { "capable": false, "reason": "Read tool handles PDFs up to 20 pages but cannot extract tables" },
  "candidates": [{ "rank": 1, "name": "...", "source": "...", "score": 3.8, "install_cmd": "...", "inspection_notes": "..." }],
  "recommendation": "install|build-custom|use-builtin"
}

Gotchas

• Never invent package names.
• Never recommend a package without checking recency and source.
• Never skip the approval step.
• Never recommend a bare repo URL without a pinned version or commit hash.
• Never rely solely on metadata -- always attempt Candidate Inspection.
• If marketplace CLI does not exist, skip silently and note the gap.
• Never search externally when built-in tools already handle the capability.
• Always persist discover results to JSON for future reference.

Subagents

skill-searcher: { model: haiku, tools: [Bash, WebSearch], constraint: "return only real results; query sources in priority order" }
skill-inspector: { model: sonnet, tools: [Bash, Read], constraint: "fetch and read README/SKILL.md for top 3; verify claims" }
skill-evaluator: { model: sonnet, tools: [Read], constraint: "score consistently; show full breakdown with rationale" }