Analysis Date: 2026-10-05
Repository: github/gh-aw
Scope: 320 workflow sources, 112 explicitly using the Copilot engine (about 35%). A further 18 declare no engine and fall back to the default.
📊 Executive Summary
Key findings
- Copilot engine adoption is broad (112 workflows).
engine.args is rare (10 files), and max-continuations appears in 10.
timeout-minutes is set almost everywhere (308 files). toolsets appears in 188 files, so GitHub toolset scoping is well adopted.cache-memory (108) is used much more than repo-memory (38). mcp-scripts (5) and mcp-servers (2) are barely used.- The CLI flags emitted by
copilot_engine_execution.go are --add-dir, --agent, --allow-all-paths, --autopilot, --disable-builtin-mcps, --headless, --log-dir, --log-level, --max-autopilot-continues, --no-ask-user, --no-auto-update and --no-custom-instructions.
- The count of 320 sources is up from 299 on 2026-09-08. The previous snapshot (run 34184590907) counted 112 Copilot workflows (24 simple, 88 extended), so the total is unchanged.
Primary recommendation: Pin the engine version consistently and define the model centrally, because model overrides are scattered across 185 files.
Caveat: these counts come from grep over frontmatter patterns, so they are approximate.
Critical Findings
🔴 High Priority
- Version pinning gap:
version: appears in 42 files, and not all of them are for the engine. Most Copilot workflows float on the default CLI version, so a CLI regression can hit many workflows at once. Pin the version through a shared import or default and bump it deliberately.
- Untyped engine selection: 18 workflows declare no engine. Make the engine explicit so a default change cannot silently change behaviour.
🟡 Medium Priority
- Scattered model overrides (185 files with
model:). Move them to shared imports or a central default so model upgrades are one change.
max-continuations / autopilot is used in only 10 workflows. Long multi-step workflows could use bounded autopilot rather than relying on high max-turns.repo-memory is under-used for trend workflows. Research and audit workflows that rely on cache-memory lose state when the cache expires. Persist durable trend data to repo-memory.mcp-scripts / mcp-servers are rarely used. Custom, deterministic tool wrappers would reduce token use compared with free-form shell tool calls.
View Full Analysis
Feature Usage Matrix (approximate)
| Category |
Observation |
| CLI flags |
Engine-managed flags are applied automatically. engine.args is used in 10 files. |
| Engine config |
model in 185 files, version in 42, agent in 178 (includes non-engine uses) |
| MCP |
toolsets in 188 files, playwright in 30, mcp-scripts in 5, mcp-servers in 2 |
| Sandbox |
sandbox: in 138 files |
| Memory |
cache-memory in 108 files, repo-memory in 38 |
| Timeouts |
timeout-minutes in 308 files |
Trends
Compared with the 2026-09-08 snapshot: total sources went from 299 to 320 and Copilot workflows stayed at 112. Several metrics were counted differently this time (custom_agent_usage was 232 before, and agent: is 178 now), so the earlier snapshot does not allow a like-for-like trend on those.
Best Practices
- Pin the engine version and manage it centrally.
- Declare the engine and model explicitly or via shared imports.
- Use
repo-memory for durable state and cache-memory for scratch data.
- Use
mcp-scripts for deterministic helper tools.
Methodology
Counted frontmatter patterns with grep across .github/workflows/*.md, listed the CLI flags in pkg/workflow/copilot_engine_execution.go, and compared the result with /tmp/gh-aw/repo-memory/default/copilot-cli-research/latest.json from the previous run. I did not review individual workflow bodies or the docs in depth.
Action Items
Immediate
Short-term
Long-term
Generated by Copilot CLI Deep Research (Run: 37261579083)
Generated by 🔬 Copilot CLI Deep Research Agent · copilot · auto · 11.9 AIC · ⌖ 6.8 AIC · ⊞ 10.7K · ◷
Analysis Date: 2026-10-05
Repository: github/gh-aw
Scope: 320 workflow sources, 112 explicitly using the Copilot engine (about 35%). A further 18 declare no engine and fall back to the default.
📊 Executive Summary
Key findings
engine.argsis rare (10 files), andmax-continuationsappears in 10.timeout-minutesis set almost everywhere (308 files).toolsetsappears in 188 files, so GitHub toolset scoping is well adopted.cache-memory(108) is used much more thanrepo-memory(38).mcp-scripts(5) andmcp-servers(2) are barely used.copilot_engine_execution.goare--add-dir,--agent,--allow-all-paths,--autopilot,--disable-builtin-mcps,--headless,--log-dir,--log-level,--max-autopilot-continues,--no-ask-user,--no-auto-updateand--no-custom-instructions.Primary recommendation: Pin the engine version consistently and define the model centrally, because model overrides are scattered across 185 files.
Caveat: these counts come from grep over frontmatter patterns, so they are approximate.
Critical Findings
🔴 High Priority
version:appears in 42 files, and not all of them are for the engine. Most Copilot workflows float on the default CLI version, so a CLI regression can hit many workflows at once. Pin the version through a shared import or default and bump it deliberately.🟡 Medium Priority
model:). Move them to shared imports or a central default so model upgrades are one change.max-continuations/ autopilot is used in only 10 workflows. Long multi-step workflows could use bounded autopilot rather than relying on highmax-turns.repo-memoryis under-used for trend workflows. Research and audit workflows that rely oncache-memorylose state when the cache expires. Persist durable trend data torepo-memory.mcp-scripts/mcp-serversare rarely used. Custom, deterministic tool wrappers would reduce token use compared with free-form shell tool calls.View Full Analysis
Feature Usage Matrix (approximate)
engine.argsis used in 10 files.modelin 185 files,versionin 42,agentin 178 (includes non-engine uses)toolsetsin 188 files,playwrightin 30,mcp-scriptsin 5,mcp-serversin 2sandbox:in 138 filescache-memoryin 108 files,repo-memoryin 38timeout-minutesin 308 filesTrends
Compared with the 2026-09-08 snapshot: total sources went from 299 to 320 and Copilot workflows stayed at 112. Several metrics were counted differently this time (
custom_agent_usagewas 232 before, andagent:is 178 now), so the earlier snapshot does not allow a like-for-like trend on those.Best Practices
repo-memoryfor durable state andcache-memoryfor scratch data.mcp-scriptsfor deterministic helper tools.Methodology
Counted frontmatter patterns with grep across
.github/workflows/*.md, listed the CLI flags inpkg/workflow/copilot_engine_execution.go, and compared the result with/tmp/gh-aw/repo-memory/default/copilot-cli-research/latest.jsonfrom the previous run. I did not review individual workflow bodies or the docs in depth.Action Items
Immediate
Short-term
cache-memorytorepo-memoryin audit workflows.Long-term
mcp-scriptsfor the most token-heavy shell patterns.Generated by Copilot CLI Deep Research (Run: 37261579083)