[prompt-clustering] Copilot Agent Prompt Clustering — 2026-06-15 (1000 PRs, 80.2% merge, 10 clusters)

[github-actions[bot]]

Summary

Window: 2026-05-27 → 2026-06-15 (30d) · PRs: 1000 · Clusters: 10 (sil 0.0391) · Merge: 80.2% (727M/180C/93O)

Merge rate holds in its stable 78–81% band (today 80.2%, ~13-run mean ~79%) — no regression. TF-IDF (1–2 grams) on PR title+task-intro; commits = iteration proxy (per-PR turn logs unjoinable: only 20 aw_info.json exist, none linked).

Key findings

1. Infra/tooling dominates — top 4 clusters (CLI runtime 23%, docs/schema 16%, Go refactors 11%, lockfile/pin bumps 11%) ≈61%. The fleet mostly maintains gh-aw's own plumbing.
2. Highest merge = mechanical — Lockfile/pin bumps 87% at fewest commits (3.0); regenerated lock/compiled files land first-pass.
3. Lowest merge = open-ended/reactive — PR safe-output handling 67%, WIP fix-failing-Actions 73%; small, often closed/superseded.
4. Most iterations = SDK harness — Copilot SDK driver 5.2 commits/PR, hard integration vs a moving SDK.
5. Cost accounting is a standing theme — AI-credits (8.6%@86%) + AIC reporting (9.7%@76%) ≈18% of PRs.

Clusters

Theme	Size	Merge	Commits	Files	Top terms
CLI runtime / toolcache / setup	234 (23.4%)	84%	4.18	25	cjs, setup, behavior
Docs, skills & frontmatter schema	157 (15.7%)	80%	3.72	14	docs, md, reference
Go package refactors (pkg/cli, linters, helpers)	114 (11.4%)	76%	3.54	76	pkg, pkg workflow, linters
Lock files / compiled workflows / CLI pin bumps	105 (10.5%)	87%	3.01	52	lock, workflows, lock yml
AIC usage & cost reporting / artifacts	97 (9.7%)	76%	3.93	55	aic, usage, conclusion
Safe-outputs feature work	95 (9.5%)	75%	3.64	22	safe, safe outputs, outputs
AI credits guardrails & renaming	86 (8.6%)	86%	4.0	52	credits, ai, ai credits
Copilot SDK driver harness	67 (6.7%)	80%	5.21	29	sdk, copilot, copilot sdk
WIP: fix failing GitHub Actions jobs	23 (2.3%)	73%	2.43	15	failing actions, actions job, job
PR / pull-request safe-output handling	22 (2.2%)	67%	2.68	19	pull request, pull, request

Representative PRs per cluster

• CLI runtime / toolcache / setup — #35802 · #35992 · #38725
• Docs, skills & frontmatter schema — #35694 · #36748 · #36936
• Go package refactors (pkg/cli, linters, helpers) — #37162 · #36327 · #36010
• Lock files / compiled workflows / CLI pin bumps — #38706 · #36704 · #37097
• AIC usage & cost reporting / artifacts — #38856 · #37004 · #37472
• Safe-outputs feature work — #36353 · #36701 · #38237
• AI credits guardrails & renaming — #38611 · #38197 · #37101
• Copilot SDK driver harness — #36358 · #36549 · #36307
• WIP: fix failing GitHub Actions jobs — #38397 · #37886 · #38265
• PR / pull-request safe-output handling — #39118 · #36702 · #36238

Trend

Date	PRs	k	Merge
2026-06-05	999	8	75.4%
2026-06-06	999	9	78.4%
2026-06-07	999	9	79.0%
2026-06-08	999	10	79.7%
2026-06-12	1000	9	78.8%
2026-06-13	991	10	78.5%
2026-06-14	1682	10	80.7%
2026-06-15	1000	10	80.2%

Recommendations

1. Separate auto-fix tracking — WIP fix-failing-Actions (73%) is reactive firefighting; isolate it so its merge rate isn't read as a fleet regression.
2. Tighten PR-plumbing prompts — lowest cluster (67%) is open-ended; pin concrete acceptance criteria (field names, target behavior) up front.
3. Raise iteration ceilings for SDK work — Copilot SDK driver averages 5.2 commits; don't cut it off mid-integration.
4. Consolidate cost-accounting logic — ~18% of PRs touch AI-credits/AIC across two overlapping clusters; a recurring maintenance tax.

References: §27543489324

Generated by 📊 Copilot Agent Prompt Clustering Analysis · 217.2 AIC · ⌖ 17.9 AIC · ⊞ 13.3K · ◷

• expires on Jun 16, 2026, 4:01 AM UTC-08:00

[github-actions[bot]]

This discussion has been marked as outdated by Copilot Agent Prompt Clustering Analysis.

A newer discussion is available at Discussion #39536.