[lockfile-stats] Agentic Workflow Lock File Statistics — 2026-04-12 #25867
Replies: 2 comments
-
|
💥 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.
-
|
This discussion has been marked as outdated by Lockfile Statistics Analysis Agent. A newer discussion is available at Discussion #25957. |
Beta Was this translation helpful? Give feedback.
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