[lockfile-stats] Lockfile Statistics Report — 2026-08-26 (294 workflows)

[github-actions[bot]]

Executive summary

Analyzed 294 compiled lockfiles (.github/workflows/*.lock.yml) as of 2026-08-26. 0 malformed/skipped. Total size 44,272,954 bytes (~42.2 MB), avg 150.6 KB/file (min 88,958 B, max 238,038 B). All self-checks passed (yaml parsing OK, safe-outputs config found in every file, engine resolved for every file, permissions resolved for every file).

File size distribution

Metric	Value
Count	294
Total	44,272,954 B (42.2 MB)
Avg	150,588 B
Min	88,958 B
Max	238,038 B

Trigger analysis

Top triggers: workflow_dispatch (286), schedule (238), pull_request (43), issues (5), issue_comment (3).

Top combinations: schedule+workflow_dispatch (205), workflow_dispatch only (38), pull_request+schedule+workflow_dispatch (29), pull_request+workflow_dispatch (9).

workflow_dispatch present in 286/294 (97.3%). Most common cron: 0 0 */2 * * (every 2 days) in 42 workflows; the rest are long-tail unique schedules.

Full cron frequency table (30 distinct patterns)

Cron	Count
0 0 */2 * *	42
7 5 * * *	3
49 14 * * 1-5	2
23 11 * * *	2
38 3 * * *	2
9 3 * * *	2
41 4 * * *	2
26 5 * * *	2
37 5 * * *	2
40 3 * * *	2
remaining 19 patterns	1 each

Safe outputs analysis

create_discussion configured in 92 workflows; category resolved for all 92 (100%, no fallback parsing needed).

Discussion category	Workflows
audits	79
announcements	5
artifacts	2
dev	2
research	2
general	1
daily-news	1

Other common safe-output types: create_issue (143), add_comment (77), create_pull_request (63), add_labels (32), upload_asset (26), create_pull_request_review_comment / submit_pull_request_review (16 each). Universal defaults (missing_data, missing_tool, noop, report_incomplete, create_report_incomplete_issue) each appear in 288/294 workflows — the 6 without them are likely older or minimal configs.

Structural characteristics

Metric	Avg	Max
Jobs/workflow	7.33	14
Steps/workflow	143.24	202
Script steps/workflow	98.46	133

Timeouts: 10 min (312 job-instances), 45 min (295), 60 min (288), 90 min (3), 5 min (2), 120 min (2), 180 min (1).

Permission patterns

From jobs.agent.permissions (the actual per-workflow scope, not the always-empty top-level block): permissions_unknown = 0 for all 294 files.

Scope	read	write	none
contents	294	0	0
issues	258	0	36
pull-requests	254	0	40
discussions	52	0	242
actions	110	0	184
security-events	14	0	280
id-token	0	2	292

The agent job itself is essentially read-only everywhere — writes happen through the safeoutputs MCP server instead. Looking at the union across all jobs in each workflow (agent + post-processing jobs), all 294 workflows (100%) grant write on at least one scope, with issues: write universal (294/294), contents: write in 204/294, and pull-requests: write in 144/294 — this is where the actual safe-output side effects get applied.

Engine distribution

Engine	Count
copilot	148
claude	61
codex	47
pi	21
crush	3
aider	3
opencode	3
goose	3
cursor	1
deepseek-harness	1
gemini	1
kiro	1
pydantic-ai	1

engine_unknown = 0 — every lockfile's agent_id resolved cleanly from its gh-aw-metadata header. Top models: copilot/mai-code-1-flash-picker (32), copilot/gpt-5.4 (23), copilot/claude-sonnet-4.5 (12).

Tool & MCP patterns

Every workflow uses the safeoutputs MCP server (294/294) — expected, since it's the safe-output write path. github MCP server appears in 121 workflows, agenticworkflows in 43, serena in 25, mcpscripts in 12. mcp_fallback_used_count = 0, meaning every file's manifest was parsed directly (no legacy comment-scraping needed).

Full MCP server counts

Server	Count
safeoutputs	294
github	121
agenticworkflows	43
serena	25
mcpscripts	12
tavily	5
sentry	3
ast-grep	2
datadog	2
deepwiki	2
microsoftdocs	2
grafana	2
mempalace	1
graft	1
agentdb	1
ruflo	1

Interesting findings

1. Copilot is now the majority engine (148/294, 50.3%), more than double Claude (61) and ahead of Codex (47) — a notable shift if this ecosystem was Claude-first historically.
2. Agent jobs are read-only by design: 0 write grants at the agent-job level across every scope except a rare 2 workflows using id-token: write (likely OIDC-based publishing). All mutation happens through the sandboxed safeoutputs MCP layer, not direct job permissions — a consistent security pattern across the whole fleet.
3. audits is the dominant discussion category (79/92, 86% of discussion-producing workflows) — this very workflow is part of the majority pattern.
4. Workflow bulk is heavily boilerplate-driven: avg 143 steps but 98 of those are script steps, and every workflow ships the same 5 fallback safe-output types (missing_data, missing_tool, noop, report_incomplete, create_report_incomplete_issue) — suggesting a shared compiler template rather than per-workflow authoring of these steps.
5. Discussion category extraction had zero fallback dependency: all 92 create_discussion configs resolved their category straight from GH_AW_SAFE_OUTPUTS_CONFIG JSON, with no regex fallback triggered — a good signal the config-embedding convention is applied uniformly.

Historical trends (vs. 2026-08-25)

Metric	2026-08-25	2026-08-26	Δ
Lockfile count	293	294	+1
Total bytes	44,118,118	44,272,954	+154,836
Avg size	150,573.8	150,588.3	+14.5
workflow_dispatch count	285	286	+1
create_discussion workflows	92	92	0
codex engine count	46	47	+1
create_issue safe-output count	142	143	+1
github MCP server count	120	121	+1
safeoutputs MCP server count	293	294	+1

All deltas are consistent with exactly one new lockfile being added (using codex engine, workflow_dispatch+schedule trigger, create_issue + github + safeoutputs tooling) with no other files changing structurally.

Recommendations

• Consider consolidating the 5 universal fallback safe-output types into a single compiler-level default rather than duplicating them per-lockfile, if compile size becomes a concern (current avg 150 KB/file).
• The discussions: read grant sits at only 52/294 in the agent job but discussions: write reaches 95/294 in the job union — confirm this gap is intentional (write happens only in a later job stage) rather than a missed grant.
• Given Copilot's growing share (148/294), consider surfacing engine-specific timeout/step-count patterns in a future run to check whether structural characteristics correlate with engine choice.

Methodology note

Single-script compact JSON analysis: one cached Python analyzer (lockfile_stats_v4.py) parses all lockfiles once via yaml.safe_load plus targeted regex for gh-aw-metadata/gh-aw-manifest header comments, emits one compact summary JSON, and this report reasons only from that JSON (no per-file cat/grep outside the analyzer).

References:

• §33013031915

Generated by 📊 Lockfile Statistics Analysis Agent · claude · agent · 58.6 AIC · ⌖ 6.69 AIC · ⊞ 8.5K · ◷

• expires on Aug 27, 2026, 1:06 PM UTC-08:00

[github-actions[bot]]

This discussion was automatically closed because it expired on 2026-08-27T21:06:05.931Z.

Closed by Workflow