Executive Summary

Analysis of 233 compiled workflow lockfiles (.github/workflows/*.lock.yml) on 2026-05-21 — total 22.55 MB, no files skipped.

File Size Distribution

Sizes are tightly clustered: every lockfile is at least 62 KB, and the largest is only ~2.8× the smallest — suggesting a substantial shared compiled scaffold (engine setup, safe-output plumbing, MCP server boilerplate) that dominates the per-file footprint regardless of the user-authored content.

Trigger Analysis

Top trigger combinations:

Nearly two-thirds of all workflows are scheduled, almost universally paired with workflow_dispatch for manual reruns. Only 8 workflows omit workflow_dispatch entirely — a strong convention across the repo.

Safe Outputs Analysis

⚠️ Limitation: In compiled .lock.yml files, safe-outputs configuration is embedded as a JSON env var (GITHUB_AW_SAFE_OUTPUTS_*), not as YAML keys, so the regex-based extractor returned no key-level counts in this run. The lockfiles do reference the safeoutputs MCP server / CLI throughout. A future bump (lockfile_stats_v2.py) should parse the embedded JSON env var to recover counts.

What we can observe:

• Every lockfile loads the safeoutputs MCP server (it is the standard write channel for these agentic workflows).
• Discussion categories were not captured directly; the next analyzer revision should pull category: from the embedded safe-outputs JSON.

Structural Characteristics

Compiled workflows are large but uniform: every workflow has at least 5 jobs and 67 steps, suggesting a fixed compiled skeleton (preflight, engine setup, MCP setup, agent run, safe-output dispatch, teardown) plus per-workflow tail.

Timeout distribution (across all jobs):

The 6–30 minute band covers 91% of all timeouts — consistent with single-turn agent runs sized to fit comfortably within GitHub-hosted runner billing increments.

Permission Patterns

All 233 lockfiles set permissions: {} at the top level, then declare fine-grained permissions per job. This is the safest pattern: no implicit token scope at the workflow level, with each job's needs declared explicitly. (Per-job permission counts could not be aggregated in this run because the YAML parser was unavailable in the runtime; the regex pass over inline permissions: {} confirms the empty-top-level convention is universal.)

Tool & MCP Patterns

MCP server frequency (count of references across all lockfiles):

Engine distribution:

Interesting Findings

1. Tight size band. Min 62.8 KB, max 177.4 KB — a 2.8× spread across 233 workflows. The compiled scaffold is the dominant contributor to lockfile size, not workflow-author content.
2. schedule + workflow_dispatch is the dominant pattern (66.5% of all workflows). Only 3 workflows are pull_request-only; the repo is overwhelmingly oriented toward autonomous scheduled agents, not PR gatekeepers.
3. All 5 smoke-*-claude/copilot/codex lockfiles cluster at the top of the size list (4 of the top 5 largest), reflecting the broader test/setup matrix smoke tests carry.
4. The github MCP server is referenced 6,448 times across lockfiles — roughly 27.7 references per workflow on average, dwarfing every other MCP server combined (≈300). Reducing this footprint (selective tool enabling instead of full toolset opt-in) would meaningfully shrink lockfiles.
5. 124-of-233 workflows enable the full GitHub MCP read-toolset together (every top-30 GitHub tool shows the same count of 124). This is the single biggest target for tool-narrowing across the repo.
6. All 233 lockfiles use permissions: {} at the top level — a strict, opt-in security posture is universal.
7. No cron at :00 or :30 outside a handful of cases. The scheduler-spreading guidance is being followed broadly: most off-zero minutes (e.g. :23, :38, :49) are in active use.

Historical Trends

No reliable prior-day comparison: the only history entry (2026-05-20.json) was generated by an earlier broken analyzer version that reported 0 lockfiles. Today's run (schema v1) writes a clean baseline at /tmp/gh-aw/cache-memory/history/2026-05-21.json — future runs will be able to diff against it.

Recommendations

1. Bump analyzer to lockfile_stats_v2.py to parse the embedded GITHUB_AW_SAFE_OUTPUTS_* JSON env vars so safe-output type counts and discussion categories can be recovered. Today's report has a gap there.
2. Audit the 124 workflows pulling the full GitHub MCP read-toolset. If those workflows only use 2–3 GitHub tools each, narrowing the enabled set would noticeably shrink compiled lockfiles (top contributor to the 22.5 MB total).
3. Investigate the smoke- engine workflows topping the size chart.* If smoke matrices include unused engine setup blocks, trimming them would reduce repo-wide compiled bytes.
4. Install PyYAML in the analyzer runtime (or vendor a small parser) so per-job permission and trigger details can be aggregated precisely instead of via regex fallback.
5. Continue the :00/:30 cron-avoidance convention — it is already nearly universal, worth codifying in a linter check.

Methodology

• Approach: single-script compact JSON analysis. One bash invocation wrote and executed /tmp/gh-aw/cache-memory/scripts/lockfile_stats_v1.py, which scanned all 233 .lock.yml files and emitted a single ≤50 KB JSON summary at /tmp/gh-aw/agent/lockfile-stats-summary.json. All discussion text below was derived from that summary, not from re-reading lockfiles.
• YAML parser: unavailable in this runner — analyzer fell back to regex extractors. Counts for triggers, schedules, jobs/steps, timeouts, engines, and MCP usage remain reliable; per-key safe-output and per-job permission breakdowns are partial (see Limitation note above).
• Caching: an earlier cached version of lockfile_stats_v1.py was discovered to be broken (skipped all 233 files); it was overwritten in-place and rerun. Today's run is the new baseline.
• Skipped files: 0.

References:

• §26252187909

Generated by 📊 Lockfile Statistics Analysis Agent · ● 5.7M · ◷

• expires on May 22, 2026, 8:55 PM UTC