[lockfile-stats] Agentic Workflow Lock File Statistics — 2026-04-12 #25867
Replies: 1 comment
-
|
💥 KAPOW! 🦸♂️ WHOOSH! The smoke test agent has ARRIVED! Meanwhile, at the Claude engine... 🤖: "By the power of all neural networks combined... SMOKE TEST ACTIVATED!" 💥 ZAP! All systems: NOMINAL! The villain: flaky CI — DEFEATED!
|
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
-
Overview
Analysis of all 187
.lock.ymlfiles in.github/workflows/as of 2026-04-12. Total corpus size reached 13.5 MB — a +17.4% increase from 11.5 MB two weeks ago (2026-03-30), while file count grew from 178 to 187 (+5.1%). Today's run shows notable growth: 3 additional files crossed the 100 KB threshold (now 6 files > 100 KB, up from 3 yesterday), and average file size jumped from 71.1 KB to 74.0 KB (+2.9 KB) in a single day.The repository has 4 AI engines represented: Copilot (66%), Claude (28%), Codex (6%), and Gemini (<1%). The dominant workflow pattern is
schedule + workflow_dispatch(66% of all workflows), reflecting a fleet primarily designed for automated, recurring tasks.Executive Summary
codex-github-remote-mcp-test.lock.yml(30.5 KB)smoke-claude.lock.yml(155.9 KB)File Size Distribution
The 50–100 KB bucket dominates (93.6%), indicating a highly standardized lock file structure. The 6 files above 100 KB are likely more complex multi-job workflows.
Trigger Analysis
Most Popular Triggers
workflow_dispatchschedulepull_requestissue_commentissuespull_request_review_commentdiscussiondiscussion_commentworkflow_callworkflow_runpushCommon Trigger Combinations
schedule + workflow_dispatchworkflow_dispatchonlypull_request + workflow_dispatchpull_request + schedule + workflow_dispatchissue_commentonlyissue_comment + issues + pull_requestworkflow_call + workflow_dispatchThe
schedule + workflow_dispatchcombo is used by nearly two-thirds of all workflows, making it the canonical agentic pattern: automated scheduling with optional manual trigger.Schedule Pattern Analysis (top 30 unique cron expressions)
37 2 * * *48 12 * * *23 3 * * *7 4 * * *6 11 * * 1-527 10 * * *52 11 * * *19 10 * * *37 3 * * *27 */6 * * *25 */6 * * *49 */4 * * *7 8 * * 133 5 * * 026 11 * * 3Frequency breakdown:
Schedules use deliberately spread-out times (off :00/:30) to avoid thundering-herd effects on the CI/CD platform.
Safe Outputs Analysis
Action Safe Output Types (top tools, excluding baseline
noop/missing_tool/missing_data)create_discussioncreate_issue(all variants)create_pull_request(all variants)add_comment(all variants)upload_assetadd_labels(all variants)push_to_pull_request_branchsubmit_pull_request_reviewcreate_pull_request_review_commentremove_labelscreate_code_scanning_alertdispatch_workflowassign_to_agentsend_slack_messagenotion_add_commentlink_sub_issuecreate_project_status_updateBaseline tools (
noop,missing_tool,missing_data) appear in 181 of 187 workflows (96.8%) — standard safe-output hygiene.Most Common Safe Output Combinations
create_discussiononlycreate_issueonlycreate_pull_requestonlycreate_discussion + upload_assetadd_comment(max:2)onlyadd_commentonlyadd_comment + create_pull_requestcreate_issue + create_pull_requestcreate_code_scanning_alertonlycreate_discussion + create_issue(max:3)Discussion Category Distribution
auditsannouncementsreportsartifactsdevresearchagent-researchdaily-newsThe
auditscategory accounts for 71% of all discussion-targeting workflows, making it the primary output channel for reporting agents. 5 workflows reference the discussion category via an object (dynamic lookup) rather than a static string.Engine / Agent Distribution
The Copilot fleet is the largest, but Claude represents over a quarter of all workflows. Codex is a smaller but distinct segment. Gemini represents an experimental single workflow.
MCP Server Usage
github-mcp-servergh-awserena-mcp-servermcp(generic)markitdownbrave-searchast-greparxiv-mcp-servernotionsemgrepcontext7pythonmemorynodegithub-mcp-serveris nearly universal (97.3%).gh-aw(16%) andserena-mcp-server(12.3%) are secondary servers for specialized capabilities. The long tail includes domain-specific tools: academic (arxiv-mcp-server), code analysis (ast-grep,semgrep), documentation (markitdown), and external integrations (notion,brave-search).Permission Patterns
Read Permissions
contentsactionspull-requestsissuesdiscussionssecurity-eventscheckspackagesWrite Permissions
issuesdiscussionspull-requestscontentscopilot-requestssecurity-eventsactionsattestationsid-tokenpackagesstatusescontents: read(945 instances) dominates — virtually every job needs to read the repository. Write permissions are spread acrossissues,discussions, andpull-requests, reflecting the three primary GitHub output surfaces. Thecopilot-requests: writepermission (95 instances) is Copilot-engine-specific.Structural Characteristics
Step & Job Complexity
copilot-token-audit.lock.yml)codex-github-remote-mcp-test.lock.yml)Timeout Distribution
Average timeout: 19.1 minutes. The bimodal peak at 15 and 20 minutes reflects two standard step timeout tiers used across the fleet. Only 6 workflows use timeouts ≥ 60 minutes, likely for long-running analysis tasks.
Historical Trends (2026-03-30 → 2026-04-12)
Key trends:
Interesting Findings
Lock files grow over time. The +17.4% total size growth over 14 days while file count grew only 5.1% suggests lock files accumulate run history or context. Three files crossed the 100 KB mark today alone — at this trajectory, more will follow.
schedule + workflow_dispatchis the canonical pattern. 65.8% of workflows use this exact combination — automated recurring execution with a manual escape hatch. This is the dominant architectural idiom of the agentic fleet.github-mcp-serveris near-universal. 97.3% of workflows use it, making it effectively a required dependency of the harness — only 5 workflows go without it.Safe output diversity is expanding. 70+ distinct safe output tool names exist in the corpus (including
max:Nvariants), with exotic tools likeassign_to_agent,create_agent_session,notion_add_comment,close_discussion, andcreate_project_status_updateappearing in niche workflows.Copilot dominates by count, but Claude has the largest single workflow. The largest file is
smoke-claude.lock.ymlat 155.9 KB vs. the smallest being a Codex test file at 30.5 KB — Claude workflows tend to be larger in structure.5 workflows use dynamic discussion category references. Identified as "object - Category info with name field" — these workflows look up the category at runtime rather than hardcoding a string, which is a more flexible but potentially fragile pattern.
Recommendations
Monitor size growth trajectory. At +2.9 KB/day average, files will begin exceeding 100 KB more frequently. Consider whether lock file history rotation or pruning is needed to prevent runaway growth.
Standardize
add_commentmax limits. There are 8 distinctadd_comment(max:N)variants (2, 3, 5, 10, 15, 20, 50, plus unlimited). A fleet-wide convention would reduce complexity.Investigate the 6 workflows missing baseline safe-output tools. The 6 files without
noop/missing_tool/missing_datamay be outdated configurations that predate the safe-output hygiene standard.Track the 5 dynamic discussion category references. These "object - Category info" patterns deserve attention to ensure they resolve correctly at runtime and don't silently fail.
The Gemini workflow is unique. With only 1 Gemini-engine workflow in the fleet, it warrants monitoring to understand if it's experimental, deprecated, or a planned expansion.
Methodology
.lock.ymlin.github/workflows/)/tmp/gh-aw/cache-memory/scripts/analyze_lockfiles.py(reused from cache)/tmp/gh-aw/cache-memory/history/(14-day trend)References:
Beta Was this translation helpful? Give feedback.
All reactions