Evaluate dependency

Evaluate Dependency

Dual-purpose: upfront selection when adding a new dep, and review when a PR adds or bumps one. Same evaluation criteria either way, with per-language addenda for ecosystem-specific quirks.

When to use

Mode	Triggers
Selection	Considering adding a dep, asks "should I use X", compares alternatives, before go get / npm install / pip install / cargo add
Review	PR diff touches a manifest, or review-security / review-code sees a new dep in scope

Language-specific quirks (read first)

The canonical package coordinate usually has ecosystem-specific gotchas that override the general criteria. The wrong coordinate is the wrong dep, no matter how good the package is.

• Go: references/go.md. Semantic import versioning (/v2, /v3 paths), +incompatible smell, pkg.go.dev / vuln.go.dev, govulncheck.
• Other ecosystems: not yet covered here. Run the general checklist and flag uncertainty about the canonical coordinate with unsure: (per terse-output) rather than guessing.

Future addenda: npm (scoped vs unscoped, deprecated packages, ESM/CJS), Python (package vs distribution name, wheel/sdist, ABI compatibility), Rust (pre-1.0 churn, crate renames), Ruby (transitive native-extension hazards). Add when a real evaluation surfaces the need.

Evaluation criteria

Run through these. Cite specific data: no hand-waving, no I think.

1. License. Detect via the ecosystem's registry or the repo's LICENSE / LICENSE.md / COPYING. Permissive (MIT, BSD, Apache-2.0, ISC, MPL-2.0) is generally fine. Copyleft (GPL, AGPL, LGPL) needs an explicit compatibility check against the user's project license. No license = no use. Note relicense history; relicenses are a red flag.

2. Internal precedent. Before weighing external popularity, check whether the candidate (or an equivalent) is already used or wrapped internally. Sources, in order:

   1. Project-local hints: CLAUDE.md, REVIEW.md, .claude/rules/, and the manifests of sibling services if mounted locally.
   2. Internal code search if available: Sourcegraph (mcp__sourcegraph__keyword_search / mcp__sourcegraph__list_repos), gh search code scoped to the user's orgs.
   3. Internal package registry / module index if the project documents one.

   A candidate with strong internal precedent beats a marginally more-popular external alternative — consistency, shared CVE response, shared tooling all compound. An internal wrapper or replacement (e.g. an internal/x package covering the same need) usually means don't add the external dep; use the internal one. If no internal-search tooling is reachable, mark this row unsure: no internal index available rather than skipping.

3. Popularity / community signal. Stars (rough proxy), the ecosystem's "depended by" / "imported by" count, Sourcegraph importer search if available. Popularity without recent maintenance is a trap, not validation.

4. Maintenance / activity. Last commit date: <12 months = active, 12–24 = stale, >24 = abandoned. Last release vs last commit (releases lagging commits = pre-release work or maintainer absent). Open issues / PRs ratio; unanswered bug reports older than 6 months; "looking for maintainers" signals.

5. Vulnerability history. Check the ecosystem's vuln DB (vuln.go.dev for Go, GitHub Advisory DB, OSV, npm audit, pip-audit, RustSec). Pattern of repeat CVEs in the same subsystem = avoid.

6. API stability. Semver discipline. Major-version bump cadence: every six months is chaotic. v0.x.y after years signals unstable surface. Look for a documented breaking-changes / release-notes practice.

7. Surface and transitive cost. Each dep brings its own deps. Count transitive deps; smaller is better when alternatives are comparable. Note pulled-in transitive risk (e.g. a Go module that quietly pulls in a deprecated logger).

8. Fit. Does it do what's needed with minimal surface? A library doing 10× what's needed brings 10× the risk. Check the language's standard library / built-ins first. Many ecosystems' stdlib covers what third-party libs once owned (e.g. Go's slices, maps, cmp, errors; Python's pathlib, dataclasses; JS's modern URL, fetch, structuredClone).

9. Footguns. Ecosystem-specific (init-time side effects, monkey-patching, peer-dep churn, global mutable state, unbounded goroutines/threads). Check the language reference for known patterns; surface anything weird in the README or top-level types.

10. Capabilities. What privileged operations can the dep (and its transitive closure) reach? A logging library that transitively pulls os/exec or net/http is doing more than it advertises. For Go, capslock (see references/go.md) classifies transitive calls into capability buckets (NETWORK, FILES, EXEC, REFLECT, CGO, UNSAFE_POINTER, ARBITRARY_EXECUTION, etc.). Two modes:

• Selection mode: snapshot capabilities of the candidate. Flag high-signal combinations — ARBITRARY_EXECUTION (any), REFLECT + EXEC (dynamic-code gadget), EXEC in a non-exec-purpose dep, CGO / UNSAFE_POINTER (analyzer blind spots downstream).
• Review mode (bump): diff capabilities against the prior version. A dep gaining NETWORK or EXEC it didn't have before is a strong supply-chain signal, regardless of release notes.
• For non-Go ecosystems, equivalents are immature; use unsure: no capability tool available rather than skipping.
• SSA / callgraph blind spots: capability analysis relies on SSA + call graph; reflection, cgo boundaries, unsafe.Pointer arithmetic, and inactive build tags are invisible. A clean capslock report doesn't prove safety — calibrate against the UNANALYZED count.

Workflow

1. Confirm scope (per confirm-before-implementing): selection (prospective add) or review (PR touched manifest). Identify the language and the candidate package.

2. Resolve canonical coordinate. Read the per-language reference if one exists; apply its rules first. This is non-negotiable. Examples: Go's /vN path probe, npm's scope check, Python's distribution-name lookup.

3. Survey alternatives (selection mode, optional). When multiple candidates are in the running, delegate per-candidate data lookups to Explore subagents (model: haiku per subagent-model-routing — mechanical lookup across 10 fixed criteria; per parallelize-subagents and delegate-investigation). Each subagent returns one verdict table prefixed with Status: per subagent-prompt-contract. Parent merges and applies the final verdict synthesis (model: opus per subagent-model-routing — hard reasoning combining multiple criteria across candidates).

4. Run the 10 criteria against the resolved coordinate. Cite specific numbers, dates, license names.

5. Produce verdict: GO / CAUTION / NO-GO. Always include the coordinate to use explicitly (so the user doesn't import the wrong major / wrong scope / wrong distribution).

6. Review mode: output as a finding row suitable for review-code or review-security finding tables, wrapped per the existing path:line deep-link convention if pr_url is set. Severity prefix per terse-comments: risk: (CAUTION) or bug: (NO-GO).

Output format

## Dependency Evaluation: `<canonical coordinate>`

**Verdict:** GO | CAUTION | NO-GO
**Coordinate to use:** `<canonical coordinate>` (note any path/scope/name gotcha the user might miss)

| Criterion | Finding |
|---|---|
| Language-specific path/coordinate | (e.g. v2 is current; root path is at v1.4.7, 2022) |
| License | (e.g. Apache-2.0) |
| Internal precedent | (e.g. used in 3 sibling repos as `internal/httpx` wrapper; or: no internal usage found) |
| Popularity | (e.g. 12.4k stars, ~3200 importers) |
| Last commit | (e.g. 2026-04-29, active) |
| Vulnerabilities | (e.g. None at v2.1.0; govulncheck clean) |
| API stability | (e.g. v2 released 2023-06, no major bumps since) |
| Surface | (e.g. 4 transitive deps, all common) |
| Fit | (e.g. Covers HTTP-client retry; no stdlib equivalent) |
| Capabilities | (e.g. NETWORK, FILES — expected; or: EXEC unexpectedly transitive via `findExternalDriver`; or: `unsure: no capability tool available`) |

**Mitigations** (CAUTION only): pin to specific minor; vendor; wrap behind an internal interface.
**Alternatives** (NO-GO only): `<alt-1>` — brief reason; `<alt-2>` — brief reason.

Cross-references

• delegate-investigation: surveying alternatives belongs in Explore subagents.
• subagent-prompt-contract: when fanning out per-candidate evaluation.
• terse-output: verdict tables; use unsure: when a criterion can't be confirmed.
• terse-comments: review-mode findings follow the review-comment shape with severity prefix.
• review-security and review-code: this skill is their natural delegation target when a new dep appears in scope.

Anti-patterns

• Skipping the per-language reference. The wrong coordinate is the wrong dep.
• Verdict without a recommended coordinate. Always name the exact import path / package name / scope to use.
• Recommending an older major because docs are better. Name a hard reason (missing feature, breaking change incompatible with caller) or recommend current.
• Accepting "X is popular" as sufficient. Popularity without recent maintenance is a trap.
• Hand-waving with hedge words instead of citing numbers ("seems active", "I think it's maintained"). Use unsure: when you don't know.
• Auto-running go get / npm install / pip install. Out of scope; the skill recommends, the user applies.
• Recommending an external dep when an internal package already covers the use case. Internal precedent overrides external popularity unless there's a hard reason (missing feature, abandoned, license incompatibility) — name it.
• Per ~/.claude/rules/probe-not-assume.md: confirm via tool/command before recommending; do not infer.