Deep-evaluate GitHub repositories across quantitative metrics (API-based) and qualitative signals (code-level). Dispatches parallel subagents for efficient per-repo assessment.
Trigger boundary
Use when: evaluating known repos for quality, comparing candidates, scoring maintainer credibility
Do NOT use when: searching for repos (use run-github-repo-search instead)
Workflow
Phase 1: Quick Screen (main agent, NO subagents)
CEILING: Maximum 30 repos. If input list is >30, take top 30 by stars.
Batch repos into groups of 5. For each batch, run ONE GraphQL call:
gh api graphql -f query='
fragment R on Repository {
nameWithOwner stargazerCount forkCount
createdAt pushedAt primaryLanguage{name} licenseInfo{spdxId}
defaultBranchRef{target{...on Commit{history(first:0){totalCount}}}}
openIssues:issues(states:OPEN){totalCount}
closedIssues:issues(states:CLOSED){totalCount}
mergedPRs:pullRequests(states:MERGED){totalCount}
releases{totalCount}
}
{
r0:repository(owner:"O0",name:"N0"){...R}
r1:repository(owner:"O1",name:"N1"){...R}
...
}' --jq '.data | del(.rateLimit) | to_entries[] | .value |
"\(.nameWithOwner)\t\(.stargazerCount)★\t\(.primaryLanguage.name // "?")\t\(.defaultBranchRef.target.history.totalCount) commits\tI:\(.openIssues.totalCount)/\(.closedIssues.totalCount)\tPR:\(.mergedPRs.totalCount)\tR:\(.releases.totalCount)\t\(.pushedAt[:10])"'
Read references/graphql-batch-repos.md for the exact pattern.
Output: overview table for user to review before committing to deep evaluation.
Phase 2: Deep Evaluation (subagents, 1 per repo)
For each repo, dispatch a subagent (Sonnet model) with the full evaluation prompt.
CEILING: Maximum 15 parallel subagents. If >15 repos, batch in waves. THRESHOLD: If evaluating <=3 repos, skip subagents — do it inline.
Each subagent runs THREE query types:
-
GraphQL repo deep dive (1 API call, 1 rate limit point) Read
references/graphql-repo-deep-dive.md— copy the exact query template. -
REST-only signals (~8 API calls, 8 rate limit points) Read
references/rest-unique-signals.md— these 6 endpoints have NO GraphQL equivalent. -
Author profile (1 API call, 1 rate limit point) Read
references/graphql-user-profile.md— deep author assessment with 6-year contribution history. -
Code-level analysis (3-5 API calls) Read
references/code-level-analysis.md— README quality, key file presence, source code sampling.
Each subagent returns a structured metrics report. See references/subagent-dispatch-guide.md for the exact prompt template.
Phase 3: Score & Report (main agent)
After all subagents complete:
- Read all metrics reports
- Apply scoring rubric from
references/scoring-rubric.md - Calculate composite score per repo
- Rank by score descending
Output format:
Summary table:
| # | Repo | Score | Stars | Health | Author | Code | Flags | Key Strength | Key Risk |
Per-repo analysis (top 10): 2-3 sentences covering standout metrics, concerns, and recommendation.
Landscape insights: Common patterns, gaps, trends across evaluated repos.
Scoring formula
Score = (Health × health_weight) + (Author × author_weight) + (Code × code_weight) + RedFlagPenalties
Default weights: Health 40%, Author 30%, Code 30%. User can customize.
Read references/scoring-rubric.md for all 29 metrics with 0-10 scales.
Decision rules
- <=3 repos → evaluate inline (no subagents)
- 4-15 repos → one wave of parallel subagents
- 16-30 repos → two waves of subagents
- User only wants author assessment → skip repo metrics, author-only path
- Feature Fit metrics are CUSTOM per use case — user defines what features matter
- If a repo doesn't exist or API errors → score 0, note in report
- If author is an org → use org profile query, score all Author metrics at 5/10 (neutral). Note: org repos lack personal contribution history — this is an information gap, not a red flag.
Reference routing
| File | Read when |
|---|---|
references/graphql-repo-deep-dive.md |
Phase 2 — single repo full query |
references/graphql-batch-repos.md |
Phase 1 — batch screening query |
references/graphql-user-profile.md |
Phase 2 — author assessment |
references/graphql-user-profile.md (batch section) |
Phase 1 — batch author comparison |
references/rest-unique-signals.md |
Phase 2 — the 6 REST-only endpoints |
references/code-level-analysis.md |
Phase 2 — README, files, source sampling |
references/scoring-rubric.md |
Phase 3 — all metrics with scoring scales |
references/author-red-flags.md |
Phase 2/3 — detecting red flags |
references/api-gotchas.md |
Any phase — known API pitfalls |
references/subagent-dispatch-guide.md |
Phase 2 — exact subagent prompt template |