[deep-research] Deep Research: Pattern Library Analysis (2026-09-28 snapshot)

これは patterns/manifest.json と patterns/archetypes/*.json の読み取り専用の概要であり、コード変更の依頼ではありません。まず、これらのコミット済みの入力を調べ、報告されている件数または結果をそれらと照合して確認してください。issue には、すべてのファイルの解析が成功したと記載されています。完了とは、リポジトリのファイルを変更せずに、分析結果を実行ログに記録することです。

Read-only analysis of the committed pattern library (patterns/manifest.json + patterns/archetypes/*.json), generated 2026-09-28 from 563 source repos / 345 active workflows / 929 total workflows scanned. No repository writes were made; this issue is purely a summary for the run log.

Archetype coverage (32 files, 31 user-facing + hidden custom)

Empirical archetypes (count > 0), by success rate:

Curated archetypes with no measured runs yet (count: 0, retained per policy): accessibility-expert, agent-cost-tracker, backlog-drip, batched-ci-doctor, ci-failure-triage, code-health-auditor, community-digest, contribution-guidelines-checker, issue-hierarchy-manager, link-checker, linter-applier, linter-miner, linter-refiner, linter-workflows, nitpick-reviewer, performance-nut, pr-fix-assistant, pr-iteration-loop, repo-qa-assistant, security-scanner, user-simulator (21 archetypes).

custom (count: 469, 49% baseline) is excluded from the wizard's HOW-step archetype cards but kept as a supporting profile/match target.

Configuration defaults

• No pinned default model (model: null) — premium models only selected deliberately.
• Timeout defaults by trigger: schedule 30 min; issues, push, slash_command, workflow_run, discussion, pull_request all 15 min (archetypes may override upward).
• Prompt size sweet spot: 3,000–8,000 bytes.

Trigger combinations (curated, 15 tracked)

All 15 tracked combos are marked "Recommended"; 13 of 15 hit 90–100% success. Notable:

• permissions+schedule and lone reaction — 100% (n=20 each).
• issues+reaction+stop-after — 100% (n=13).
• bots+permissions+skip-if-match, manual-approval+pull_request+roles, cron+env+name+schedule+steps, etc. — 100% (n=10 each).
• No sub-90% combo appears in this curated top-15 "Recommended" list.

Configuration profiles (archetype × trigger × safe-outputs detail, 20 entries)

Highlights of best/worst within archetype:

• custom: schedule+skip-if-match → create-pull-request = 93.0% (n=43, best in dataset); schedule → create-issue = 53.6% (n=69); pull_request → add-comment = 60.0% (n=30).
• status-report: schedule+stop-after → create-discussion = 97.5% (n=40, near-best overall) vs. plain schedule → create-issue = 47.1% (n=87) vs. adding mentions/allowed-github-references → 33.6% (n=143, degrades further).
• repo-maintainer: permissions+reaction+slash_command+steps (rich safe-outputs incl. push-to-pull-request-branch) = 90.0% (n=30) vs. reaction+schedule+slash_command (no permissions/steps scoping) = 41.0% (n=39).
• issue-triage: issues+reaction → add-comment+add-labels = 83.3% (n=24); issue_comment+issues → add-comment/messages/noop = 80.0% (n=30).
• pr-review: pull_request → add-comment or review-comment variants = 71.4% (n=21 each) vs. slash_command → review-comment/messages = 20.0% (n=20).
• dependency-monitor: schedule → create-pull-request = 36.4% (n=176, largest sample in the whole profile set, still underperforms).
• documentation-updater: schedule → create-pull-request = 56.7% (n=30) vs. push → create-pull-request = 26.7% (n=30).
• code-improvement: workflow_run → add-comment+create-issue = 21.7% (n=23, best code-improvement profile, still poor); schedule+skip-if-match → create-pull-request = 12.0% (n=83); stop-after+workflow_run (broad safe-output set) = 0.0% (n=20, confirms workflow_run chaining anti-pattern).

Anti-patterns (20 named, all 0% success, each seen in exactly 1 repo)

All are singleton-repo failures, clustered around: documentation/doc-sync variants (documentation-updater, documentation-audit, update-docs, doc-sync, docs-sync, daily-repo-status ×2, localize-pt, news-translate), broad unscoped mandates (dependabot-remediation, breaking-change-checker, ci-doctor, csharp-guideline-improver), issue-triage/pr-review singleton failures, a duplicate "update-github-info" pattern (seen in 2 different repos), and two Imperia360-specific patterns (imperia-market-prices, imperia-cross-auditor). degraded_workflows was an empty array in this manifest snapshot.

Research findings

• Bimodal outcomes: 38% of workflows always succeed, 21% always fail, 41% mixed — averages are misleading.
• DO NOT constraints: workflows with explicit boundary instructions are 61% more likely to be healthy (p=0.009).
• Slash commands: act as dispatchers routing to target workflows via workflow_dispatch; their own execution metrics are recorded under the dispatched run, not the dispatcher.
• workflow_run risk: chaining via workflow_run has only 13–16% success rate — prefer pre-steps or schedule, confirmed by the 0% code-improvement workflow_run+stop-after profile above.
• Pre-steps help: workflows with pre-steps show +13pp internal and +5pp community activity.
• Prompt size matters: active workflows have 35–48% larger prompts than inactive ones.
• Template clones are fragile: ~32% of workflows are unmodified template copies, which underperform customized workflows.

Data integrity note

patterns/manifest.json is present and valid, and references all 32 archetype files under patterns/archetypes/, all of which parsed successfully and contained analyzable pattern data (either empirical success_rate/count data or curated defaults with count: 0). No missing-or-invalid-input condition was encountered.

This is a read-only reporting workflow; no code, workflow, or pattern files were modified.

Generated by 🔬 Deep Research · copilot · auto · 26.7 AIC · ⌖ 5.45 AIC · ⊞ 5.9K · ◷