fix(acp): emit [Workspace] section for Windows drive-letter cwd

[kcao-gss]

Summary

workspace_section() gated the [Workspace] system-prompt block on cwd.starts_with('/'), so it was silently dropped for every Windows agent — cwd comes from std::env::current_dir() (lib.rs:1547-1550), which on Windows is always drive-letter form (C:\Users\<user>\.buzz).

Replaces the string-prefix test with Path::is_absolute() && Path::parent().is_some(), which preserves both invariants the original guard existed for — reject relative paths, reject the / sentinel current_dir() falls back to on failure — and generalizes the second to reject C:\, \\?\C:\, and UNC share roots uniformly.

Fixes #3124

Scope, stated honestly

This is not "Windows agents don't know about their workspace." base_prompt.md:84-100 already ships the Workspace Layout table (REPOS/, OUTBOX/) and the "never run find over $HOME" instruction unconditionally on every platform, and the agent's OS cwd is the workspace (the child inherits it). What Windows was missing is the absolute-path anchor — one reinforcing sentence. Filed and fixed as a correctness bug, not a user-facing breakage.

Why the platform-dependent predicate is deliberate

Path::is_absolute is platform-dependent, and that is correct here rather than a hazard: cwd only ever originates from current_dir() in the same process on the same host, so the compile-time platform always matches the path's dialect. Audited — PromptContext (pool.rs:482) has no derives (so no deserialization path), its only production constructor is lib.rs:1530, no crate outside buzz-acp references it, and there is no CLI flag or env override for cwd. The reasoning is now recorded in the function's doc comment so a future maintainer doesn't "fix" it back to a string prefix test.

Deliberately not done: normalizing the cwd string upstream. The same string is the ACP session/new cwd (pool.rs:831), so rewriting it would change protocol behavior for every harness.

Tests

• #[cfg(windows)]: drive-letter cwd → Some; C:\ → None.
• #[cfg(unix)]: /Users/me/.buzz → Some.
• Ungated (platform-neutral): / → None, relative/path → None, "" → None.

Two POSIX-path assertions needed #[cfg(unix)] gating as a direct consequence of this change — including the pre-existing test_framed_system_prompt_absolute_cwd_prepends_workspace_before_base, which was green on Windows before, because Path::new("/Users/me/.buzz").is_absolute() is false on Windows (root, no drive prefix). Called out explicitly since it's the one non-obvious side effect of the fix.

Reviewer note: the #[cfg(windows)] tests do not currently execute in CI. .github/workflows/ci.yml:945-947 runs cargo clippy and cargo check --workspace --all-targets on windows-latest (so they compile), but the only Windows cargo test invocations are -p buzz-dev-mcp (:953) and the Tauri crate (:988). Adding -p buzz-acp to that step would make them enforcing — happy to include it here if you want it, or leave it separate.

Test plan

• cargo test -p buzz-acp workspace_section — 3 passed
• cargo test -p buzz-acp framed_system_prompt — 7 passed
• cargo fmt --check -p buzz-acp — clean
• cargo clippy -p buzz-acp — clean
• cargo test -p buzz-acp on a Windows target (no Windows host available to me — this is the one gap, and it's exactly what the two #[cfg(windows)] tests are for)

Out of scope, noted for follow-up

Now that Windows agents receive this text:

• The section says "do not search $HOME" — POSIX phrasing that reads oddly on Windows (base_prompt.md:100 has the same issue).
• {cwd}/REPOS/ renders C:\Users\me\.buzz/REPOS/ — mixed separators, functional but untidy.

Related

#3126 — the same [Workspace] section never reaches protocol-v1 harnesses (Claude Code, Codex) on any platform, because the legacy prompt-framing path never calls workspace_section at all. Independent of this predicate bug: a Windows Claude Code agent is affected by both, a macOS one only by #3126. This PR does not address it, and should not close it.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FnfD8YHcZF5KbBTcLxc1P1