[lockfile-stats] Lockfile Statistics — 2026-08-23 (290 lockfiles, 41.1MB)

[github-actions[bot]]

Lockfile Statistics — 2026-08-23

Analysis of .github/workflows/*.lock.yml in github/gh-aw. Single-script compact JSON analysis (see Methodology). Compared against prior snapshot 2026-08-22.json where schema overlaps.

Executive summary

Metric	Value	Δ vs 2026-08-22
Lockfiles parsed	290	+1 (289)
Lockfiles skipped/malformed	0	—
Total size	43,090,530 bytes (~41.1 MB)	+423,285 B (+1.0%)
Avg size	148,588 B (~145.1 KB)	+950 B
Min / Max size	89,104 B / 238,455 B	~flat
Engine unknown	0 / 290	—
Permissions unknown (agent job)	0 / 290	—

File size distribution

• Range 89.1 KB – 238.5 KB, avg 145.1 KB. Largest: api-consumption-report.lock.yml (176.1 KB), agent-performance-analyzer.lock.yml (163.6 KB), agentic-token-audit.lock.yml (160.3 KB).
• Size correlates with job count: the 5 largest files are also among the highest job-count workflows (7–9 jobs), consistent with more jobs → more embedded step/script YAML.

Trigger analysis

Trigger	Count	% of 290
workflow_dispatch	283	97.6%
schedule	195	67.2%
pull_request	42	14.5%
issues	5	1.7%
issue_comment	3	1.0%
workflow_call	2	0.7%
discussion / discussion_comment / push / workflow_run / deployment_status / pull_request_review_comment	1 each	0.3% each

Top combinations: schedule + workflow_dispatch (192, 66.2%), workflow_dispatch only (48, 16.6%), pull_request + workflow_dispatch (37, 12.8%).

Cron scheduling is heavily staggered — of 195 scheduled workflows, cron strings are almost all unique. Only a handful repeat exactly: 7 5 * * * (×3), and 38 3 * * *, 40 3 * * *, 41 4 * * *, 26 5 * * *, 9 3 * * *, 23 11 * * *, 49 14 * * 1-5, 37 5 * * * (×2 each). This indicates deliberate per-workflow minute/hour jitter rather than round-number scheduling, minimizing thundering-herd load on runners.

Safe outputs analysis

Type	Text-presence count	% of 290
create-discussion	290	100%
create-issue	145	50.0%
add-comment	145	50.0%
create-pull-request	81	27.9%
create-pull-request-review-comment	18	6.2%
update-issue	12	4.1%

Caveat: this counts textual occurrence of each safe-output key anywhere in the lockfile (including shared handler boilerplate), so it's an upper bound, not actual usage. The MCP-tool-invocation signal (below) is tighter: only 91/290 (31%) workflows actually have create_discussion wired as an invokable tool, 142 have create_issue, 76 have add_comment, 62 have create_pull_request. Treat the tool-usage numbers as the more reliable indicator of real safe-output configuration.

Discussion category extraction returned 0 matches this run (regex against discussion_category did not match current lockfile structure) — flagged as a detection gap rather than "no categories configured"; do not read this as zero discussion-producing workflows.

Structural characteristics

Metric	Avg	Max
Jobs / workflow	7.33	14
Steps / workflow	142.77	202
Scripts (run:) / workflow	61.44	83

Structural averages are essentially flat vs. 2026-08-22 (jobs 7.32→7.33, steps 142.13→142.77, scripts 61.04→61.44), consistent with the +1 lockfile / +1.0% byte delta being incremental additions rather than a broad rewrite.

Timeout-minutes (per job, summed across all jobs in all workflows): 10min ×308, 45min ×291, 60min ×284, 90min ×3, 120min ×2, 180min ×1, 5min ×2.

Permission patterns (agent job)

The top-level permissions: {} in every lockfile carries no signal — these figures come from jobs.agent.permissions, which mirrors the workflow frontmatter:

Scope	read	write	none
contents	290	0	0
pull-requests	250	0	40
issues	244	0	46
actions	108	0	182
discussions	51	0	239
security-events	14	0	276
checks / packages / deployments	1	0	289
id-token	0	2	288
statuses / pages / attestations	0	0	290

Zero write scopes appear at the agent-job level across all 290 workflows (aside from 2 id-token: write). All real write capability lives in the union of jobs per workflow (safe-output handler jobs):

Scope	Workflows granting write (union)
issues	290 (100%)
contents	201 (69.3%)
pull-requests	142 (49.0%)
discussions	94 (32.4%)
copilot-requests	93 (32.1%)
actions	104 (35.9%)
checks	16
id-token	3
security-events	3
code-quality	2
attestations / packages / drives / statuses	1 each

All 290 workflows (100%) grant at least one write scope somewhere in their job set — confirming the safe-outputs isolation pattern: the agent job itself is read-only/none, writes are delegated to dedicated jobs.

Engine distribution

(from gh-aw-metadata agent_id, 0 unresolved)

Engine	Count	%
copilot	145	50.0%
claude	61	21.0%
codex	46	15.9%
pi	21	7.2%
crush / aider / opencode / goose	3 each	1.0% each
cursor / deepseek-harness / gemini / kiro / pydantic-ai	1 each	0.3% each

Model distribution (top): copilot/mai-code-1-flash-picker (32), copilot/gpt-5.4 (21), copilot/claude-sonnet-4.5 (12), gpt-5.4 (7), claude-sonnet-4.6 (5), claude-haiku-4.5 (4), plus 4 workflows resolving model dynamically via ${{ needs.activation.outputs.model_size }}.

Tool & MCP patterns

MCP servers by frequency: safeoutputs 290 (100%, always present), github 119 (41.0%), agenticworkflows 42 (14.5%), serena 25 (8.6%), mcpscripts 12, tavily 5, sentry/datadog/ast-grep/deepwiki/microsoftdocs/grafana 2–3 each, long tail of 1-off integrations (mempalace, graft, agentdb, ruflo). Manifest-based detection succeeded for all 290 lockfiles (0 fallback to comment-scraping).

Most common individual tools: create_issue (142), github read-tool cluster (get_commit/get_file_contents/get_latest_release/get_release_by_tag/get_tag/list_branches/list_commits/list_releases/list_starred_repositories/list_tags/search_code/search_repositories, 117 each), PR-read cluster (~111 each), wildcard * tool grant (77), create_discussion (91), create_pull_request (62).

Interesting findings

1. Cron jitter is near-universal: of 195 scheduled workflows, cron strings are almost all distinct — only 9 exact strings repeat, each just 2–3×. This looks like deliberate per-workflow offset assignment to avoid synchronized runner load, not coincidence.
2. Copilot is the plurality engine (50%) despite this being "Claude"-branded tooling context; Claude is second (21%), Codex third (16%), with 9 other engines in single digits — a wide agent-backend experimentation surface.
3. Agent jobs carry essentially zero write permission (0/290 write any scope except 2 stray id-token writes) — all mutation capability is delegated to separate jobs, a strict enforcement of the safe-outputs isolation pattern across the entire fleet.
4. Safe-output "type presence" overstates real usage by ~3× for discussions: 290/290 lockfiles textually contain create-discussion, but only 91 (31%) actually wire it as an invokable tool — a reminder that naive text search on compiled lockfiles is not a reliable usage proxy.
5. Day-over-day drift is minimal: +1 lockfile, +1.0% total bytes, flat structural averages — today's change set looks incremental (added/modified 1 workflow) rather than a fleet-wide regeneration.

Historical trends (vs 2026-08-22)

Metric	2026-08-22	2026-08-23	Δ
Lockfile count	289	290	+1
Total bytes	42,667,245	43,090,530	+423,285 (+1.0%)
Avg size	147,637.5	148,588.0	+950.5
Avg jobs/workflow	7.32	7.33	+0.01
Avg steps/workflow	142.13	142.77	+0.64
Avg scripts/workflow	61.04	61.44	+0.40
create_discussion (safe-output, prior text-key method)	91	91 (tool-usage) / 290 (text)	flat (tool)
create_issue	140	145	+5

Note: the 2026-08-22 snapshot used a different detection methodology for triggers, engines, and per-scope permissions (those fields were empty in that run), so this comparison is limited to size/structural metrics and the safe-output counts that both methods captured comparably.

Recommendations

• Audit the 46 workflows without issues: read and 40 without pull-requests: read at the agent-job level to confirm they genuinely don't need that context (vs. an oversight).
• Fix discussion-category extraction at generation time (or document the field) — this run could not recover any category values from the compiled lockfiles.
• Consider spacing out the 9 duplicate-cron-slot pairs/triples (e.g. 7 5 * * * ×3) since they currently compete for runners at the same instant.
• Track copilot-requests: write (93 workflows) alongside AI-credit usage — it's a distinct quota surface from the engine-level token accounting.

Methodology

Single-script compact JSON analysis: one Python script (cached at /tmp/gh-aw/cache-memory/scripts/lockfile_stats_v3.py) parsed all 290 .lock.yml files with PyYAML.safe_load in one pass, emitting a compact summary JSON (11.9 KB) that this report was written from — no individual lockfile was read outside that script.

Generated by 📊 Lockfile Statistics Analysis Agent · agent · 45.6 AIC · ⌖ 8.12 AIC · ⊞ 7.6K · ◷

• expires on Aug 24, 2026, 12:22 PM UTC-08:00