Repository navigation
[claude-code-user-docs-review] 🔍 Claude Code User Documentation Review - 2026-10-10 #67432
Closed
Replies: 1 comment
|
This discussion was automatically closed because it expired on 2026-10-11T12:49:09.630Z.
|
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.
Executive Summary
Reviewed gh-aw docs as a Claude Code user who does not use Copilot. The
CLAUDE_CODE_OAUTH_TOKENtrap remains resolved (documented, not silent) for a 9th consecutive run. No critical blockers remain, but Claude Code users still hit real friction:gh aw initscaffolds a custom agent file and MCP wiring only for Copilot, example/smoke-test coverage for Claude trails Copilot by ~1.9x (especially enterprise-auth and ARM variants), and Claude's full setup is link-only rather than inline. Key finding: core engine parity (conformance tests, dynamic workflows, fallback-models) is solid — the gap is concentrated in advanced-config depth and first-run scaffolding, not basic functionality.Severity Findings
Critical / Major / Minor (click to expand)
Critical Blockers: None this run.
Major Obstacles:
gh aw initscaffolding parity gap — only--engine copilotauto-generates a custom agent file and MCP wiring; other engines get a prose pointer with no template (docs/src/content/docs/setup/cli.md:107,:127-132,:131)..github/workflows/*.md. Copilot has dedicated Azure/OIDC auth-variant tests (smoke-copilot-aoai-entra.md,smoke-copilot-aoai-apikey.md) and an ARM runner test (smoke-copilot-arm.md) with no Claude analog.docs/src/content/docs/setup/quick-start.mdx:153defers "complete setup" entirely to an external/gh-aw/engines/claude/page, while Copilot's two auth paths are fully inline in the same document (quick-start.mdx:131-141).engine.copilot-sdk: true,docs/src/content/docs/reference/tools.md:130); Claude's restriction (tools.md:128) is one sentence with no workaround guidance.Minor Confusion:
tools.timeoutdefault stated only for Claude/Codex (60s); Copilot/Gemini/Pi defaults unstated (reference/tools.md:239)."agy (experimental)"engine named incli.md:244with no definition anywhere in the reviewed docs.architecture.mdxis 833 lines but only ~5 lines reference specific engines — per-engine sandbox/isolation differences (e.g. does Claude Code run the same sandbox as Copilot?) are unaddressed.Engine & Tool Matrix
quick-start.mdx:131-141)COPILOT_GITHUB_TOKENquick-start.mdx:143-153)ANTHROPIC_API_KEYor Anthropic WIF;claude logintoken explicitly rejectedOPENAI_API_KEY/CODEX_API_KEYshared/genaiscript.md)Tool classification (from
reference/tools.md): most tools (edit,github,bash,cache-memory,playwright,mcp-servers, etc. — 11 of ~16 reviewed) are engine-generic. Engine-differentiated tools areweb-fetch/web-search(Copilot rejects outright with a proxy workaround; Claude/Codex support natively with lighter docs) andtools.timeout(default only specified for Claude/Codex). Two entries (drive-memory, preview-gated) and one ambiguous (agy) round out the remainder.Parity observation (engine-example-counter): Base-level conformance is solid — every engine, including Claude, has its own
engine-conformance-<id>.mdsmoke test, and dynamic-workflow coverage is at parity (smoke-claude-dynamic.mdvssmoke-copilot-dynamic-workflow.md). The gap is concentrated at the advanced-config layer: enterprise/OIDC auth, ARM runners, and SDK sub-agent orchestration examples exist for Copilot but not Claude, so a Claude Code user searching for an edge-case config example is statistically more likely to land on a Copilot example.Auth Gaps
quick-start.mdx:153— Claude's full setup is entirely link-only (/gh-aw/engines/claude/), unlike Copilot's fully inline PAT-creation steps in the same file.cli.md:132— MCP server registration for Claude/Codex/Gemini/Pi is generic and link-only ("registergh aw mcp-serverin your own MCP host configuration"); no Claude-specific config snippet (e.g. no sample.mcp.jsonorclaude mcp addinvocation) appears in any of the six reviewed files.cli.md:131— "Claude Code subagents" is named but never shown; zero inline example of what such a subagent file looks like.quick-start.mdx:151/cli.md:232—CLAUDE_CODE_OAUTH_TOKENfromclaude loginis explicitly rejected (now documented, resolving the prior critical finding), but no file explains what a Claude Code user coming from localclaudeCLI usage should do differently beyond "create an API key or use WIF."tools.md:130vs:128— Copilot's web-tool restriction gets a workaround; Claude's does not.Recommended Actions
Priority 1:
.mcp.jsonsnippet) directly incli.mdnear line 131, closing the biggest "I have no idea what to write" gap for Claude Code users runninggh aw init.Priority 2:
smoke-copilot-aoai-entra.md, and inline the Claude "complete setup" steps inquick-start.mdxinstead of deferring entirely to the engines page.Priority 3:
tools.timeoutdefaults for all engines, not just Claude/Codex, and define"agy"on first mention incli.md.All reactions