[lockfile-stats] Lockfile Statistics Audit — 2026-10-04

[github-actions[bot]]

Analysis of 320 compiled .github/workflows/*.lock.yml files (single-script compact JSON analysis, 0 malformed/skipped).

Executive summary

Metric	Value	Δ vs 2026-10-03
Lockfiles	320	+1
Total size	47.5 MB (49,835,058 B)	+241 KB
Avg size	152.1 KB	+0.3 KB
Min / Max size	83.8 KB / 250.1 KB	+186 B / +2.3 KB

File size distribution

Range 83.8 KB – 250.1 KB, tight clustering around the 152 KB average — lockfiles are dominated by boilerplate (safe-output tool defs, MCP tool lists) rather than per-workflow logic.

Trigger analysis

Trigger	Count
workflow_dispatch	312 (+1)
schedule	250 (+1)
pull_request	43
issues	5
issue_comment	3
workflow_call	3
discussion / discussion_comment / push / workflow_run / deployment_status / pull_request_review_comment	1 each

Top combinations: schedule+workflow_dispatch (217, +1), solo workflow_dispatch (51), pull_request+schedule+workflow_dispatch (29). Most common cron: 0 0 */2 * * (46 workflows, every 2 days). Notably 7 5 * * * grew from 3→4 workflows day-over-day.

Safe outputs analysis

Every workflow includes the standard safe-output control set (missing_data/missing_tool/noop/report_incomplete/create_report_incomplete_issue, ~313 each). Beyond those:

Type	Count	Δ
create_issue	164	+1
create_discussion	92	—
add_comment	75	—
create_pull_request	66	+1
push_repo_memory	35	—
add_labels	32	—
upload_asset	26	—
upload_artifact	15	+1

Discussion categories (92 create_discussion workflows, 100% resolved, 0 unresolved): audits (79), announcements (5), artifacts/dev/research (2 each), general/daily-news (1 each). safe_outputs_config_missing = 1 (consistent with prior day).

Structural characteristics

Metric	Avg	Max	Δ max
Jobs/workflow	7.18	14	—
Steps/workflow	144.0	213	+1
Script steps/workflow	99.9	143	+1

Timeout-minutes distribution: 10min (359 jobs), 45min (320), 60min (313) dominate; long tail at 90/120/180/15/5/3 min.

Permission patterns (agent job)

Scope	read	write	none
contents	320	0	0
issues	263	0	57
pull-requests	257	0	63
actions	115	0	205
discussions	53	0	267
security-events	14	0	306
id-token	0	2	318

Agent jobs are read-only almost everywhere — writes happen via the safeoutputs MCP server, not direct job permissions. Union-across-jobs tells the real write story: all 320 workflows grant issues: write somewhere in the job graph (safe-output processing jobs), 211 grant contents: write, 147 grant pull-requests: write, 97 grant discussions: write.

Engine distribution

Engine	Count	Δ
copilot	130	+1
codex	78	—
claude	57	—
pi	32	—
aider / goose	4 each	—
opencode	3	—
crush / cursor / deepseek-harness / gemini / kiro / pydantic-ai	2 each	—

engine_unknown = 0 — every lockfile's agent_id resolved cleanly from gh-aw-metadata. copilot/auto (16) is the single most common model config; the model field is highly fragmented (49+ distinct values) across date-coded model names (gpt-6.1-sol, gpt-5.6-terra, etc.) and provider-prefixed variants.

Tool & MCP patterns

Server	Workflows	Δ
safeoutputs	320	+1
github	192	—
agenticworkflows	44	—
serena / mcpscripts	25 each	—
tavily	5	—
sentry	4	—

mcp_fallback_used_count = 0 — all manifests parsed via gh-aw-manifest JSON, no legacy scraping needed. Most-used individual tools: safeoutputs:missing_data/missing_tool/noop (314 each), 15 distinct github:* read tools at 174–184 each (commits/releases/tags/PRs), safeoutputs:create_issue (165).

Interesting findings

1. Permission model is bifurcated by design: agent jobs are near-universally read-only; all mutation capability is funneled through the safeoutputs MCP server and a dedicated write-permission job, visible only in the union-of-jobs view — auditing the top-level/agent permissions alone would badly understate write exposure.
2. copilot just overtook the field at 130/320 (40.6%) as the single largest engine share, ahead of codex (78) and claude (57) combined with second place.
3. Discussion-category coverage is perfect: 92/92 create_discussion workflows resolved a category (79 to audits alone), with zero fallback-regex parses needed — the structured GH_AW_SAFE_OUTPUTS_CONFIG extraction path is fully reliable across the fleet.
4. Model naming entropy: 49+ distinct agent_model strings for only 13 distinct engines, including a literal unresolved template token (${{ needs.activation.outputs.model_size }}, 3 occurrences) — suggests some lockfiles compile with dynamic model selection left unexpanded.
5. Nearly every workflow is schedule-driven: 250/320 (78%) have a schedule trigger, and 0 0 */2 * * alone accounts for 46 — a strong clustering that could be spread out to reduce simultaneous Actions-minute contention.

Historical trends (vs 2026-10-03)

Growth was incremental and uniform: +1 lockfile, +1 workflow_dispatch/schedule trigger, +1 create_issue/create_pull_request/upload_artifact/upload_asset(unchanged)/copilot engine — consistent with a single new workflow added using the copilot engine with create_issue+create_pull_request+upload_artifact safe outputs and schedule+workflow_dispatch triggers. No regressions in discussion-category resolution or engine detection (both still 0 unresolved/unknown).

Recommendations

• Spread cron schedules currently stacked on 0 0 */2 * * (46 workflows) across more offsets to reduce runner contention.
• Standardize agent_model naming (49+ variants for 13 engines) and ensure ${{ needs.activation.outputs.model_size }}-style expressions are resolved before compilation, or confirm they're intentionally dynamic.
• Continue relying on union-of-jobs permission view (not agent-job-only) for any future security audit of this fleet, since that's where real write capability lives.

Methodology note

Single-script compact JSON analysis: .github/workflows/*.lock.yml parsed once via a cached PyYAML-based analyzer (lockfile_stats_v4.py), producing a ~9 KB compact JSON summary subsequently reasoned over without reopening individual lockfiles.

Generated by 📊 Lockfile Statistics Analysis Agent · claude · agent · 76.6 AIC · ⌖ 6.51 AIC · ⊞ 6.4K · ◷

• expires on Oct 5, 2026, 1:08 PM UTC-08:00

[github-actions[bot]]

This discussion was automatically closed because it expired on 2026-10-05T21:08:27.641Z.

Closed by Workflow