[workspace-split] Phase 0 + 1 (partial): Workspace skeleton, leaf extractions, CI adaptations

[yaohaowei0914]

Closes #69. Advances #70 (partial — cc-types deferred, see issue body). Part of #68.

Summary

Pure path reshuffle — converts the single-crate layout into a Cargo
workspace without changing source code semantics.

• Root Cargo.toml becomes a virtual [workspace] with
  members = ["crates/*"] and one [workspace.dependencies] block
  holding all shared dep versions.
• All existing sources move under crates/claude-code-rs/ via
  git mv (rename tracking preserved).
• crates/claude-code-rs/Cargo.toml references shared deps with
  dep = { workspace = true }, keeping crate-local feature layering.
• build.rs moved into the package and now derives the web-ui/
  path via CARGO_MANIFEST_DIR + ../.. (cargo has no
  workspace-level build script, so keeping it at root was not
  viable — deviated from the design doc here, with a note).
• rust-embed's #[folder] attribute and the settings-schema
  snapshot test pick up the same two-level offset to the workspace
  root.
• .cargo/config.toml, Cargo.lock, web-ui/, ui/, docs/, and
  architecture/ stay at the workspace root.

Baseline measurements

Recorded in docs/workspace-split-measurements.md:

Scenario	Time
touch src/main.rs (incremental)	6.74s
touch src/tools/file_read.rs (incremental)	0.42s
Cold cargo build --release	deferred to Phase 8

Cold release is skipped intentionally — the shared target-dir
(F:/cargo-target/cc-rust) is also used by other in-flight
worktrees, and cargo clean would churn caches unrelated to this
measurement. Phase 8 already requires a full rebuild, so the cold
figure lands there alongside the "after" number.

CI adaptations (commits 3c6c176, e36deb4)

Two follow-up commits after the Phase 0 move landed to get CI
actually running (rust-lite HEAD has been red on every push for
days — the Phase 0 move did not introduce that redness, it only
moved the failure from "build blocker" to "test fragility"):

3c6c176 — CI config fixes

• .cargo/config.toml: removed target-dir = "F:/cargo-target/cc-rust".
  That is a developer-local absolute path; on GitHub Actions it caused
  failed to create directory F:/cargo-target before cargo even
  started building. Dev-side override now documented in the same file
  pointing to ~/.cargo/config.toml or CARGO_TARGET_DIR.
• Dockerfile.ci and ui/Dockerfile.test: replaced COPY src/ /
  COPY tests/ with COPY crates/ to match the new workspace layout.
  The "empty-shell cache warmup" trick (only works for single-crate
  layouts) was dropped; we rely on BuildKit layer caching instead.

e36deb4 — Long-standing Linux-only bugs

These three errors reproduce on rust-lite@ffd38b7 (run
24679262954
timestamp 2026-04-20T16:59). All predate Phase 0. They were hidden
because the CI config problems above short-circuited the Docker build
before cargo build even started.

• sandbox/availability.rs:89 — which::which used in the Linux code
  path but which was declared only in [dev-dependencies]. Promoted
  to a regular dependency.
• computer_use/screenshot/linux.rs:13,18 — or_else on an async
  future can't short-circuit across awaits (E0277 / E0308). Replaced
  with a flat match .await dispatch.
• computer_use/input/linux.rs:132 — E0515. p.to_lowercase().as_str()
  returned an &str tied to a temporary String. Rebound to a named
  local; arms now return owned String.

Phase 1 progress (commits 9fd7229, ef45df3)

Two of the three leaf crates in the #70 checklist are extracted here.
Build + test pass at every commit boundary.

cc-keybindings (9fd7229) — 2,362 LOC, 7 files

• Pure mechanical move to crates/cc-keybindings/. Direct deps: serde,
  serde_json, parking_lot, crossterm, tracing (+ tempfile dev).
• Root crate keeps crate::keybindings::… path unchanged via
  pub use cc_keybindings as keybindings; at crate root.
• cargo test -p cc-keybindings: 51/51 pass.
• cargo test -p claude-code-rs --bins --test-threads=1: 1804/1805
  (single pre-existing off_mode_skips_redaction flake, unrelated).

cc-observability (ef45df3) — 990 LOC, 4 files

• Design doc labelled observability as a "pure leaf" but grep shows one
  tie: sink.rs:122 → crate::config::paths::runs_dir. Broken by
  dependency injection: AuditSink::init now takes a runs_dir: PathBuf
  parameter; main.rs:605 resolves the path at the call site before
  constructing the sink.
• Root crate keeps crate::observability::… path via
  pub use cc_observability as observability; alias.
• cargo test -p claude-code-rs --bins --test-threads=1: 1794/1794
  (parallel runs hit pre-existing HOME/CC_RUST_HOME env races in 9
  unrelated tests; single-threaded is clean).
• cargo test -p cc-observability: 10/11 (same pre-existing
  off_mode_skips_redaction flake, now resident in the extracted crate).

cc-types — deferred

Grep shows types::app_state::AppState references teams::types,
keybindings (now cc_keybindings), ui::status_line,
config::settings, and types::tool::ToolUseContext references
ipc::agent_channel. Not a true leaf. Per the routine's analysis in
issue #70, cc-types extraction is deferred until enough of
teams/ui/config/ipc have migrated (likely during Phase 5+ when the hub
cycles break). Design doc's "leaf" claim for cc-types should be
revisited as a small follow-up doc amendment.

Verification

• cargo build (dev) : ok, 5 warnings (identical to pre-split).
• cargo build --release : ok, 5 warnings (identical to pre-split).
• cargo test --bin claude-code-rs --offline -- --test-threads=1
  : 1856 / 1857 pass. The single failure is
  observability::sink::tests::off_mode_skips_redaction — a
  pre-existing bug in redact_value (always redacts regardless of
  RedactionMode::Off). File last modified in 856dc41, so this is
  not something the workspace move introduced.
• Without --test-threads=1, up to ten additional tests flake on
  env/cwd races — also pre-existing.
• claude-code-rs --version and --help produce identical output
  to rust-lite.

Known failing CI tests (not blockers for this PR)

rust-lite CI has never successfully run any test — it was red on
compile since at least ffd38b7. Peeling back the build blockers in
this PR exposed the underlying test state for the first time.

test-windows — 1794 / 1802 passing (99.6%)

Test	Likely cause	Category
observability::sink::tests::off_mode_skips_redaction	redact_value bug, pre-existing (pre-Phase-0)	pre-existing
plugins::tests::test_paths	Asserts path contains .cc-rust; CI runner HOME lacks the dir	CI-env fragile
session::storage::tests::test_stable_workspace_path_uses_git_root	Relies on a particular git config	CI-env fragile
utils::git::tests::test_find_git_root	Same	CI-env fragile
teams::helpers::tests::test_add_member	Temp-dir read fails on windows-latest	CI-env fragile
daemon::memory_log::tests::test_today_log_path_format	Timezone / path format	CI-env fragile
tools::hooks::execution::tests::test_execute_command_hook_plain_text	Shell not bash on runner	CI-env fragile
tools::hooks::execution::tests::test_execute_command_hook_echo	Same	CI-env fragile

Locally with the same codebase the suite runs 1856/1857 (the routine
reported 1804/1805 earlier on a slightly older SHA). The extra 7 CI
failures are all tests that pass locally — classic "CI environment
≠ developer machine" fragility, not regressions from this PR.

test-rust-offline — 4 / 8 e2e suites green, the rest have 1–3 assertion failures each

Failures observed include:

• print_mode_bash_tool_no_crash_without_api
• print_mode_read_tool_no_crash_without_api
• system_prompt_omits_send_message_by_default

These assert on CLI stdout with no API key configured. Need inspection
to determine whether the Phase 0 path move shifted any path-dependent
string they compare against. Not blocking, but worth triage as
follow-up work during Phase 1+ crate extractions (file-path references
inside tests will naturally get revisited as tests move with their
crates).

test-ui — pre-existing, not introduced by this PR

error: Cannot find package 'wrap-ansi' from '/app/ui/ink-terminal/src/core/wrapAnsi.ts'
error: Cannot find module '@alcalzone/ansi-tokenize' from '/app/ui/ink-terminal/src/core/screen.ts'

Reproduces on rust-lite@ffd38b7
(run 24679262954).
The ui/ink-terminal/ submodule's package.json is missing two
transitive deps that bun install can't resolve. Unrelated to the
workspace split — should be tracked separately.

Test plan

• CI: cargo build --release on the workspace layout.
• CI: cargo test -- --test-threads=1 compiles + runs (1794/1802).
• Local smoke: cargo run -- --version, --help.
• Tool edit re-detects: touch crates/claude-code-rs/src/tools/file_read.rs triggers rebuild.
• cc-keybindings extracted (commit 9fd7229) — workspace crate builds + tests green.
• cc-observability extracted (commit ef45df3) — AuditSink::init takes injected runs_dir.
• cc-types extraction — deferred until teams/ui/config/ipc move (Phase 5+).
• Follow-up: triage 8 CI-env-fragile Windows tests (separate PR).
• Follow-up: triage 3 print-mode e2e assertions (likely during Phase 1 test migration).
• Follow-up: fix ui/ink-terminal missing Bun deps (separate concern).
• Follow-up: amend design doc to reflect cc-types non-leaf reality.

🤖 Generated with Claude Code