Analysis Date: 2026-10-10
Repository: github/gh-aw
Scope: 334 total workflows, ~119 using Copilot engine (21 use simple engine: copilot)
📊 Executive Summary
Key Findings
- Basic hygiene is strong: 322/334 workflows set
timeout-minutes, 282 use safe-outputs, 181 define network.
- Several Copilot features remain nearly unused:
--share (1), manual --add-dir (1), --disable-builtin-mcps (2), plugins (0), model overrides (3).
- Sandbox is declared in 140 workflows (down from 230 in the prior run's grep; method differences may account for part of this).
max-turns/continuations used in 39 workflows, with no consistent policy.
Primary recommendation: define repo-wide guardrails (sandbox, network, timeout, turn limits) through shared imports, and add conversation sharing/tracking for complex workflows.
Note: counts come from grep over .github/workflows/*.md and may undercount imports/shared configs.
🔴 High Priority
- Sandbox/firewall consistency – only 140 of 334 workflows declare
sandbox:. Add it to Copilot workflows via a shared import so the default is explicit.
- Turn-limit policy – only 39 workflows set
max-turns; long-running Copilot workflows lack runaway protection. Add engine.max-turns or autopilot max-continuations to the heavy ones.
🟡 Medium Priority
--share conversation tracking (1 use) – add via engine.args: ["--share"] on complex triage/research workflows to ease auditing.- Plugins (0 uses) – evaluate the
plugins config for reusable tool bundles instead of repeating tool configs.
- Model selection (3 overrides, all
copilot/gpt-5.3-codex) – cheap/simple workflows could use a lighter model; use the model env var overrides for experiments.
- Narrower GitHub toolsets – 185 use
toolsets:; audit the remaining GitHub-tool users for default toolset overreach.
Full Analysis
Capability inventory (from pkg/workflow/copilot_engine_execution.go)
Flags: --disable-builtin-mcps, --no-ask-user, --agent, --autopilot --max-autopilot-continues, --add-dir (auto: /tmp/gh-aw/, cache/drive memory dirs), --log-level all --log-dir, --allow-tool/--deny-tool, --allow-all-paths, --no-custom-instructions; SDK headless mode (--headless --no-auto-update --port); model via env var with BYOK fallback; model routing.
Usage matrix
| Feature |
Workflows |
| timeout-minutes |
322 |
| safe-outputs |
282 |
| network |
181 |
| toolsets |
185 |
custom agent (agent:) |
151 |
| sandbox |
140 |
| cache-memory |
110 |
| max-turns |
39 |
| repo-memory |
36 |
| version pin |
26 |
| model override |
3 |
| --share / --add-dir / disable-builtin-mcps |
1 / 1 / 2 |
| plugins |
0 |
Trends vs. 2026-09-08
Workflows 299→334; Copilot extended 88→~119 total; repo-memory 38→36; max-turns 30→39; agent 232→151 and sandbox 230→140 (likely counting method differences); plugins still 0; share still 1.
Best practices
- Centralize sandbox/network/timeout defaults in shared imports.
- Pin versions only where reproducibility matters; otherwise use defaults.
- Bound agent runs with turn/continuation limits.
Action Items
Methodology
Grep counts over workflow markdown, review of pkg/workflow/copilot_engine_execution.go, comparison with repo-memory latest.json.
Generated by Copilot CLI Deep Research (Run: 38022056720)
Generated by 🔬 Copilot CLI Deep Research Agent · copilot · auto · 13.1 AIC · ⌖ 0.677 AIC · ⊞ 10.7K · ◷
Analysis Date: 2026-10-10
Repository: github/gh-aw
Scope: 334 total workflows, ~119 using Copilot engine (21 use simple
engine: copilot)📊 Executive Summary
Key Findings
timeout-minutes, 282 usesafe-outputs, 181 definenetwork.--share(1), manual--add-dir(1),--disable-builtin-mcps(2),plugins(0),modeloverrides (3).max-turns/continuations used in 39 workflows, with no consistent policy.Primary recommendation: define repo-wide guardrails (sandbox, network, timeout, turn limits) through shared imports, and add conversation sharing/tracking for complex workflows.
🔴 High Priority
sandbox:. Add it to Copilot workflows via a shared import so the default is explicit.max-turns; long-running Copilot workflows lack runaway protection. Addengine.max-turnsor autopilotmax-continuationsto the heavy ones.🟡 Medium Priority
--shareconversation tracking (1 use) – add viaengine.args: ["--share"]on complex triage/research workflows to ease auditing.pluginsconfig for reusable tool bundles instead of repeating tool configs.copilot/gpt-5.3-codex) – cheap/simple workflows could use a lighter model; use the model env var overrides for experiments.toolsets:; audit the remaining GitHub-tool users fordefaulttoolset overreach.Full Analysis
Capability inventory (from
pkg/workflow/copilot_engine_execution.go)Flags:
--disable-builtin-mcps,--no-ask-user,--agent,--autopilot --max-autopilot-continues,--add-dir(auto:/tmp/gh-aw/, cache/drive memory dirs),--log-level all --log-dir,--allow-tool/--deny-tool,--allow-all-paths,--no-custom-instructions; SDK headless mode (--headless --no-auto-update --port); model via env var with BYOK fallback; model routing.Usage matrix
agent:)Trends vs. 2026-09-08
Workflows 299→334; Copilot extended 88→~119 total; repo-memory 38→36; max-turns 30→39; agent 232→151 and sandbox 230→140 (likely counting method differences); plugins still 0; share still 1.
Best practices
Action Items
--share.Methodology
Grep counts over workflow markdown, review of
pkg/workflow/copilot_engine_execution.go, comparison with repo-memorylatest.json.Generated by Copilot CLI Deep Research (Run: 38022056720)