[lockfile-stats] Lockfile Statistics Audit — 2026-07-08 (259 workflows, 29.6 MiB)

[github-actions[bot]]

Overview

Aggregate analysis of all .github/workflows/*.lock.yml compiled agentic workflows as of 2026-07-08.

Metric	Value	Δ vs 2026-07-07
Lockfiles	259	+1
Total size	30,999,123 B (~29.56 MiB)	+194,181 B (+0.63%)
Avg size	119,688 B (~116.9 KB)	+289 B
Median size	118,360 B	+109 B
Min / Max	80,619 B / 185,794 B	+292 / +1,225 B
Malformed / skipped	0	=

No lockfiles failed to parse (yaml.safe_load, yaml_available=true).

File size distribution

Bucket	Count	Share
100–250 KB	249	96.1%
50–100 KB	10	3.9%

Every lockfile is ≥80 KB — a large fixed compilation floor. The 5 largest are all engine smoke-tests:

Lockfile	Bytes
smoke-copilot-aoai-entra	185,794
smoke-copilot-aoai-apikey	185,385
smoke-copilot	185,303
smoke-claude	179,551
smoke-copilot-arm	170,531

Smallest: test-workflow (80,619), example-permissions-warning (81,556), firewall (82,495).

Trigger analysis

Trigger	Lockfiles	Share
workflow_dispatch	251	96.9%
schedule	173	66.8%
pull_request	34	13.1%
issues	4	1.5%
issue_comment / push / workflow_call	2 each	—
deployment_status / workflow_run / discussion / discussion_comment / pull_request_review_comment	1 each	—

Top combinations: schedule+workflow_dispatch (169), workflow_dispatch alone (49), pull_request+workflow_dispatch (27). Scheduled agents almost always keep a manual dispatch escape hatch.

Cron cadence

173 scheduled workflows across ~120 distinct crons. Minutes are spread across the hour (odd values like :49, :38, :23) — no thundering-herd on :00. Mix of daily (* * *), weekday (1-5), and every-N-hours (*/6, */4, */12). No cron is used more than twice.

Safe outputs analysis

Dedicated safe-output jobs detected: upload-assets in 25 lockfiles. Most safe outputs (create-discussion / create-issue / add-comment / create-pull-request) are configured via the agentic job's env manifest rather than as separate jobs, so they are not individually enumerable from lock structure — see methodology caveat.

Discussion output categories (regex over category config):

Category	Lockfiles
audits	76
announcements	5
artifacts	2
dev	2
daily-news	1
research	1

The audits category dominates — audit-style agents are by far the largest publisher class.

Structural characteristics

Metric	Total	Per workflow (min / avg / max)
Jobs	1,620	4 / 6.25 / 12
Steps	30,312	79 / 117.03 / 163
Run scripts	13,699	— / 52.9 / —

• Most jobs: release (12). Most steps: smoke-copilot (163).
• Δ vs 07-07: +6 jobs, +405 steps, +303 scripts for +1 workflow — each new agentic workflow compiles to ~117 steps / ~194 KB.

Timeouts: every workflow has at least one job in the 31–60 min band (259); 6–15 min band appears 28 times, ≤5 min twice. Timeout policy is essentially uniform.

Permission patterns

Top-level permissions is {} for all 259 lockfiles — permissions are declared per job (least-privilege pattern).

Permission	Read	Write
actions	259 (100%)	5
contents	259 (100%)	131 (51%)
pull-requests	225 (87%)	123 (47%)
issues	221 (85%)	250 (97%)
discussions	50 (19%)	92 (36%)
copilot-requests	—	67 (26%)
security-events	14	3

issues: write is the single most common write grant (250/259). Read actions+contents are universal.

Tool & MCP patterns

MCP servers by number of lockfiles using them:

Server	Lockfiles
github	104
serena	24
tavily / sentry	5 each
grafana / markitdown	3 each
brave-search, ast-grep, arxiv, deepwiki, notion, datadog, microsoftdocs	2 each
semgrep, agentdb, context7, memory	1 each

GitHub MCP is the backbone (40% of all lockfiles); the rest is a long specialized tail.

Engine mix (carried from 2026-07-07 — see caveat)

Engine	Lockfiles
copilot	160
claude	60
pi	21
codex	14
antigravity / crush / gemini / opencode	1 each

Caveat: today's automated classifier could not reliably extract the engine id from compiled lock text (no stable engine: marker survives compilation). The corpus changed by exactly +1 file vs 07-07, so this distribution is effectively current. Fixing engine detection is tracked as a follow-up.

Interesting findings

1. 80 KB compilation floor. The smallest lockfile is 80.6 KB and 96% exceed 100 KB, yet each holds 79–163 steps of largely generated boilerplate — 29.6 MiB of committed generated YAML for 259 logical workflows.
2. Manual escape hatch is near-universal. 251/259 (96.9%) expose workflow_dispatch; 169 pair it with schedule. Only 8 workflows are non-dispatchable (pure PR/push/event-driven).
3. Audit agents are the dominant output class. 76 lockfiles publish to the audits discussion category — 15× the next category (announcements, 5).
4. GitHub MCP centrality. 104 lockfiles wire in GitHub MCP; the next server (serena) appears in under a quarter as many, and 11 servers appear in just 1–2 lockfiles each.
5. Steady linear growth. +1 workflow day-over-day added +405 steps, +303 run scripts, and +194 KB — a clean per-workflow footprint of ~117 steps / ~53 scripts / ~194 KB.

Historical trends (07-07 → 07-08)

Metric	07-07	07-08	Δ
Lockfiles	258	259	+1
Total bytes	30,804,942	30,999,123	+194,181
Jobs	1,614	1,620	+6
Steps	29,907	30,312	+405
Scripts	13,396	13,699	+303
Avg steps/wf	115.92	117.03	+1.11
audits category	76	76	=

Trajectory is steady growth with no structural regressions; 49 daily snapshots retained (2026-05-20 → 2026-07-08).

Recommendations

1. Boilerplate footprint — 29.6 MiB of generated *.lock.yml in-tree keeps growing ~194 KB/workflow. If the compiler supports composite/reusable actions, extracting the shared 80 KB floor would materially shrink diffs and repo size.
2. Uniform 31–60 min timeouts — every workflow lands in the same band; verify long-running agents aren't silently truncated and consider per-workflow tuning where runtimes differ.
3. Safe-output auditability — most enabled safe-output types aren't detectable from lock structure. Emitting a small machine-readable manifest of enabled safe outputs per workflow would make audits like this exact.
4. Keep cron staggering — current odd-minute spread avoids :00 pile-ups; maintain it as the schedule count grows past 173.
5. Fix engine classification — add a stable engine-id marker (e.g. a header comment) to compiled lockfiles so engine distribution is verifiable rather than carried-forward.

Methodology note

Single-script compact JSON analysis: one cached analyzer (lockfile_stats_v1.py) parsed all 259 lockfiles with yaml.safe_load in a single run, emitting a ~5.8 KB summary JSON; all figures above derive from that summary and the retained daily history. 0 lockfiles skipped. Safe-output types, discussion categories, and per-permission read/write maps use regex/job-name fallback (one category schema-comment false positive was filtered); the engine field is carried from the prior authoritative snapshot pending a detection fix.

References: §28973990158

Warning

Firewall blocked 1 domain

The following domain was blocked by the firewall during workflow execution:

• awmgmcpg

To allow these domains, add them to the network.allowed list in your workflow frontmatter:

network:
  allowed:
    - defaults
    - "awmgmcpg"

See Network Configuration for more information.

Generated by 📊 Lockfile Statistics Analysis Agent · 129 AIC · ⌖ 12.7 AIC · ⊞ 5.3K · ◷

• expires on Jul 9, 2026, 12:49 PM UTC-08:00

[github-actions[bot]]

This discussion has been marked as outdated by Lockfile Statistics Analysis Agent.

A newer discussion is available at Discussion #44621.