refactor(workspace): Phase 6/7 scaffold + AppState/Tool trait extraction (#75 #76) by yaohaowei0914 · Pull Request #86 · Crsei/claude-code-rust

yaohaowei0914 · 2026-04-21T13:16:07Z

Summary

Lands Phase 6/7 of the workspace split (issues #75, #76) in five incremental commits:

Scaffold + cycle-breaking (6cb15c0) — publishes nine new workspace crates (cc-engine, cc-query, cc-tools, cc-lsp-service, cc-plugins, cc-teams, cc-daemon, cc-commands, cc-ipc) as stubs wired into the workspace; eliminates five remaining architectural cycles by moving the hook trait plumbing, background-agents types, agent IPC types, LSP shared types, and ToolUseContext::bg_agent_tx wiring to cc-types.
TeamContext to cc-types (dfff0df) — pure-data TeamContext / TeammateInfo land in cc-types so AppState stops reaching into crate::teams::types::*.
status_line to cc-engine (f3225dc) — ~1k LOC of src/ui/status_line/ moves to cc-engine/src/status_line/. Two root-crate back-edges (engine::output_style::resolve, tools::worktree::get_current_worktree_session) inverted via pre-resolved fields on StatusLineSnapshot.
AppState + Tool trait + types/config to cc-engine (8ebfd4c) — the core Phase 6 goal: Tool trait and ToolUseContext finally leave the root bin crate. Every downstream Phase 6/7 crate can now import them via cc_engine::types::tool.
Benchmark cleanup (db60cdd) — removes a stray empty file created during incremental-build benchmarking.

Why

Without moving AppState and the Tool trait out of the root bin crate, Phase 6/7 physical source moves are blocked — cc-tools, cc-plugins, cc-teams, cc-daemon, cc-commands, cc-ipc, cc-lsp-service all need Tool and ToolUseContext from somewhere other than the root bin crate they'd be pulled out of.

The chosen solution puts AppState + Tool trait in cc-engine (rather than abstracting them behind trait objects in cc-types) because both transitively hold runtime handles (StatusLineRunner, KeybindingRegistry) that don't belong in the leaf-types crate. cc-engine is their semantic home — it already owns QueryEngine and the new status_line submodule.

Key architectural notes

cc-types dep graph unchanged — cc-types stays a pure-ish leaf. The new teams, agent_channel, agent_events, agent_types, background_agents modules only add plain-data structs/enums plus a couple tokio-backed mpsc channel types.
cc-engine dep graph — depends on cc-types, cc-config, cc-bootstrap, cc-compact, cc-keybindings. Zero cycles; all deps are already-extracted sibling crates.
Back-compat shim — root src/types/mod.rs and src/ui/mod.rs re-export the moved modules so the ~200+ existing crate::types::{app_state, tool, config}::* and crate::ui::status_line::* call sites compile unchanged.

Incremental build speed (measured)

Against the design-doc baseline of 7.5s after a main.rs touch:

Touch point	Before (doc baseline)	After
`main.rs`	7.5s	5.96s
Root tool file (`tools/fs/file_read.rs`)	~7.5s	7.21s
`cc-engine/src/status_line/runner.rs`	n/a	7.01s
`cc-engine/src/types/tool.rs`	n/a	6.76s
`cc-types` leaf file	n/a	8.38s
No-op	—	0.5–1.5s

~20% win on main.rs touch already. The full <3s target arrives when src/tools/ (~16.8k LOC) physically moves into cc-tools — documented in the follow-up section of the design doc.

Phase 6 remaining

Documented in docs/superpowers/specs/2026-04-20-workspace-split-design.md:

Move src/engine/ lifecycle code into cc-engine (the bulk; ~5k LOC)
Move src/query/ into cc-engine (or independent cc-query depending on cc-engine; 2.3k LOC)
Move src/tools/ minus the cycle-causing files into cc-tools (16.8k LOC — biggest single-step win)
Move src/lsp_service/ plus tools/lsp.rs into cc-lsp-service
Phase 7 crates (commands, teams, daemon, plugins, ipc)

Test plan

cargo check --all-targets clean (only the 2 pre-existing web/handlers.rs dead-field warnings)
cargo test -p cc-types: 4/4 passing
cargo test -p cc-engine: 12/12 passing (status_line payload + runner)
cargo test --bin claude-code-rs -- types: 18/18 passing
cargo test --bin claude-code-rs -- engine query: 181/181 passing
cargo test --bin claude-code-rs -- statusline: 7/7 passing
cargo test --bin claude-code-rs -- ipc::sdk_mapper: 5/5 passing
cargo test --bin claude-code-rs teams -- --test-threads=1: 58/58 passing
Full --test-threads=1 run: only pre-existing cc-observability::sink::off_mode_skips_redaction flaky failure (fails on HEAD too; unrelated)

🤖 Generated with Claude Code

…rsei#76) Lands the Phase 6 / Phase 7 crate scaffold and the last cycle-breaking work needed before the physical source move. ## Scaffold (9 new workspace members) Phase 6 (issue Crsei#75): - cc-engine target for src/engine/ - cc-query target for src/query/ - cc-tools target for src/tools/ (minus lsp, send_message, team_spawn, system_status) - cc-lsp-service target for src/lsp_service/ + tools/lsp.rs Phase 7 (issue Crsei#76): - cc-plugins target for src/plugins/ - cc-teams target for src/teams/ (+ tools/send_message.rs, tools/team_spawn.rs) - cc-daemon target for src/daemon/ - cc-commands target for src/commands/ - cc-ipc target for src/ipc/ (+ tools/system_status.rs) Each crate publishes a stub lib.rs documenting its target scope and the remaining blockers. Description comments note which cycle-causing tool files travel with which crate when the source move lands. ## Cycle-breaking (done in this PR) 1. query -> tools: eliminated. The query loop now drives hooks via the cc_types::hooks::HookRunner trait through a new QueryDeps::hook_runner() accessor. HookRunner gained a run_stop_hooks method, implemented by NoopHookRunner (cc-types) and ShellHookRunner (root crate's tools::hooks). 2. query -> tools (background_agents): eliminated. CompletedBackgroundAgent and PendingBackgroundResults moved from tools/background_agents.rs to cc-types::background_agents. Root tools/background_agents.rs is now a thin re-export so existing call sites compile unchanged. 3. engine -> ipc (agent events): eliminated. The agent IPC type trio moved from src/ipc/ to cc-types: agent_types -> cc-types::agent_types (AgentNode, AgentInfo, TeamMemberInfo) agent_events -> cc-types::agent_events (AgentEvent, TeamEvent, commands) agent_channel -> cc-types::agent_channel (AgentSender, AgentIpcEvent) Root src/ipc/agent_*.rs are now thin re-exports. The engine sdk_to_agent_event helper now imports only from cc-types. 4. types -> ipc: eliminated from the ToolUseContext surface. ToolUseContext::bg_agent_tx is now Option<cc_types::agent_channel::AgentSender> instead of Option<crate::ipc::agent_channel::AgentSender>. 5. tools::lsp <-> lsp_service: eliminated. The three LSP shared types (HoverInfo, SourceLocation, SymbolInfo) moved from tools/lsp.rs to a new lsp_service/types.rs. tools/lsp.rs re-exports them. Next pass moves tools/lsp.rs itself into cc-lsp-service, making the crate fully self-contained. ## Remaining before the source move (documented in design doc) - Hoist types/tool.rs (Tool trait, ToolUseContext) to cc-types. Blocker: ToolUseContext references AppState, which reaches into teams/ui. Either (a) abstract AppState behind a trait, or (b) move TeamContext and StatusLineRunner to cc-types first. - Physically move src/engine/, src/query/, src/tools/, src/lsp_service/ into their respective crates and rewire imports (follow-up PR). - Move cycle-causing tool files to their natural homes: tools/lsp.rs -> cc-lsp-service tools/send_message.rs -> cc-teams tools/team_spawn.rs -> cc-teams tools/system_status.rs -> cc-ipc See docs/superpowers/specs/2026-04-20-workspace-split-design.md for the full plan and remaining work breakdown. ## Verification - cargo check --workspace: clean (2 pre-existing warnings, unrelated). - cargo test --lib -p cc-types: 2/2 passing. - cargo test --lib --bin claude-code-rs ipc::agent: 8/8 passing. - cargo test --lib --bin claude-code-rs query: 35/35 passing. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…i#75) Removes the `types/app_state.rs -> teams::types::*` edge — the last run-time-free leaf dependency blocking `AppState` from moving into cc-types (which in turn blocks Phase 6 `types/tool.rs` extraction). ## What moves Only the two pure data types `AppState` actually consumes: - `TeamContext` (team_name, paths, self-identity, teammates map) - `TeammateInfo` (per-teammate runtime info) Both are plain field bags with no internal deps beyond `String` / `Option` / `HashMap`. They now live in `cc-types/src/teams.rs`. ## What stays in root Everything else in `src/teams/types.rs` — `TeamFile`, `TeamMember`, `BackendType`, `TeammateMessage`, `TeammateSpawnConfig`, `TeammateSpawnResult`, `TaskStatus`, `TeammateIdentity`, `InProcessTeammateTaskState`, `IdleReason`, `SystemPromptMode` — is used only by the team runtime modules (backend, mailbox, runner, protocol, in_process). Moving them here would pull `tokio::task::AbortHandle` and team-specific impls into cc-types, which violates the leaf-crate invariant. They'll travel with the full `src/teams/` tree into cc-teams in Phase 7. ## Backward compatibility `src/teams/types.rs` keeps `pub use cc_types::teams::{TeamContext, TeammateInfo};` re-exports, so the ~7 call sites (`commands/agents_cmd.rs`, `commands/team_cmd.rs`, `teams/identity.rs`, `tools/send_message.rs`, `tools/team_spawn.rs`) compile unchanged. `types/app_state.rs` switches from `crate::teams::types::TeamContext` to `cc_types::teams::TeamContext` directly, breaking the paper dependency. ## Verification - `cargo check --workspace`: clean (2 pre-existing warnings unrelated). - `cargo test -p cc-types`: 4/4 passing (added 2 new teams tests). - `cargo test --bin claude-code-rs teams -- --test-threads=1`: 58/58 passing. - Parallel test-runner interference on 5 mailbox/helpers tests is pre-existing (shared `~/.cc-rust/teams/` paths); not caused by this change. Next step: abstract `StatusLineRunner` behind a `StatusLineHandle` trait in cc-types so `AppState` itself can move. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

First real source move of Phase 6: the scriptable status-line payload and runner (~1k LOC across 3 files) leave the root crate and land in `cc-engine`. This is the lightest-weight step toward moving `AppState` out of root, since `AppState::status_line_runner: StatusLineRunner` will now resolve to `cc_engine::status_line::StatusLineRunner`. ## What moves - `src/ui/status_line/{mod.rs, payload.rs, runner.rs}` -> `crates/cc-engine/src/status_line/` ## Decoupling the two root-crate back-edges The payload used to call into the root crate for two things: - `crate::engine::output_style::resolve(style, cwd)` -> canonical style name - `crate::tools::worktree::get_current_worktree_session()` -> active worktree Both live deep in the root crate and can't follow status_line yet. Rather than introduce callback trait objects, the contract now inverts: callers pre-resolve the two values and pass them into `StatusLineSnapshot`: - removed fields: `output_style: Option<&'a str>` - added fields: `resolved_output_style_name: Option<String>`, `worktree: Option<WorktreeStatus>` A tiny helper module `src/ui/status_line_resolver.rs` in the root crate does the resolution once per call site, so each of the three call sites (commands/statusline_cmd, ipc/sdk_mapper, ui/app) just plugs it in. ## Call-site compatibility `src/ui/mod.rs` re-exports the module (`pub use cc_engine::status_line;`), so existing `crate::ui::status_line::*` import paths keep compiling without touching the ~5 consumer files. ## Dependency envelope (cc-engine Cargo.toml) tokio, futures, async-trait, serde, serde_json, anyhow, tracing, chrono, uuid, parking_lot, git2, cc-types, cc-config, cc-bootstrap, cc-compact, cc-keybindings. All already-extracted workspace siblings — no cycle. ## Verification - `cargo check --all-targets`: clean (2 pre-existing warnings). - `cargo test -p cc-engine`: 12/12 passing (status_line payload + runner). - `cargo test --bin claude-code-rs statusline`: 7/7 passing. - `cargo test --bin claude-code-rs ipc::sdk_mapper`: 5/5 passing. Next step: move `types/{tool, app_state, config}.rs` into cc-engine so the Tool trait leaves the root crate. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…rsei#75) The last runtime-adjacent type files finally leave the root crate: - `src/types/app_state.rs` -> `crates/cc-engine/src/types/app_state.rs` - `src/types/tool.rs` -> `crates/cc-engine/src/types/tool.rs` - `src/types/config.rs` -> `crates/cc-engine/src/types/config.rs` With the `Tool` trait and `ToolUseContext` now in cc-engine, Phase 6's `cc-tools` can finally import them from a workspace crate rather than from the root bin crate — the core architectural goal of the phase. ## Why cc-engine, not cc-types `AppState` transitively holds runtime handles (`StatusLineRunner` in cc-engine, `KeybindingRegistry` in cc-keybindings) and `ToolUseContext` exposes `get_app_state: Arc<dyn Fn() -> AppState>` — both pull in things that cc-types (a pure leaf) deliberately refuses to carry. cc-engine is the natural home: it's where `QueryEngine` lives and it already owns the status-line runner after the previous commit. ## Import rewiring inside the moved files - `use super::message::{...}` -> `use cc_types::message::{...}` - `crate::keybindings::KeybindingRegistry` -> `cc_keybindings::KeybindingRegistry` - `crate::ui::status_line::StatusLineRunner` -> `crate::status_line::StatusLineRunner` (since the files now live in cc-engine, `crate` means cc-engine) `super::tool::*` and `super::app_state::*` work unchanged — the modules are siblings in `cc-engine/src/types/`. ## Back-compat shim Root `src/types/mod.rs` becomes a re-export: pub use cc_engine::types::{app_state, config, tool}; pub use cc_types::{message, state, transitions}; So every `crate::types::{app_state, tool, config}::*` path across the root crate still resolves. No call sites changed. ## Verification - `cargo check --all-targets`: clean (2 pre-existing warnings). - `cargo test -p cc-types`: 4/4 passing. - `cargo test -p cc-engine`: 12/12 passing. - `cargo test --bin claude-code-rs -- types`: 18/18 passing. - `cargo test --bin claude-code-rs -- engine query`: 181/181 passing. - Full `--test-threads=1` run: only the pre-existing `sink::tests::off_mode_skips_redaction` in cc-observability fails (flaky on HEAD too, unrelated). ## What this unblocks - `cc-tools`, `cc-plugins`, `cc-teams`, `cc-daemon`, `cc-commands`, `cc-ipc`, `cc-lsp-service` can all now import the `Tool` trait and `ToolUseContext` via `cc_engine::types::tool`, paving the way for the full Phase 6/7 source moves. - Next: move `src/engine/` and `src/query/` into cc-engine (the bulk of Phase 6) — these are lifecycle code, not just types. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Accidentally created during incremental-build benchmarking in F:/AIclassmanager/cc/rust/.claude/worktrees/modest-morse-592f2d when I mistook the file's canonical path (real file lives at tools/fs/file_read.rs). The stray file had no module declaration referencing it so it was dead weight; deleting it. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Integrates the fork-agent infrastructure and ancillary command/IPC updates that landed on rust-lite after this branch opened: - PR Crsei#85: fork-agent infra + /btw /simplify /advisor (Crsei#33 Crsei#37 Crsei#62) - PR Crsei#87: MCP/plugin/reload-plugins/ide/reload-plugins commands (Crsei#41 Crsei#44 Crsei#47 Crsei#49) - PR Crsei#88: /loop /schedule /team-onboarding (Crsei#43 Crsei#58 Crsei#60 Crsei#63) ## Conflict resolution Git's recursive merge handled every hunk on its own — each of the five files flagged as overlapping had non-overlapping line ranges between my workspace-split edits and rust-lite's feature edits: - `crates/cc-engine/src/types/app_state.rs` — my move to cc-engine vs. advisor/fork-related AppState fields. Clean auto-merge. - `crates/claude-code-rs/src/engine/agent/mod.rs` — my `cc_types::agent_*` import rewrites vs. rust-lite's `pub mod fork;` addition. Clean. - `crates/claude-code-rs/src/engine/lifecycle/deps.rs` — my `cc_types::background_agents` retyping and `hook_runner()` impl vs. rust-lite's `advisor_model` stripping logic. Clean. - `crates/claude-code-rs/src/query/deps.rs` — my `hook_runner()` / `drain_background_results` signature changes vs. rust-lite's new `ModelCallParams::advisor_model` field. Clean. - `crates/claude-code-rs/src/query/loop_impl.rs` — my hook-runner-trait rewiring vs. rust-lite's advisor model plumbing. Clean. All 40+ other touched files (api/*, commands/*, ide/*, ipc/*, services/*, etc.) merged without conflict — they're in regions my branch didn't touch. ## Verification - `cargo check --workspace`: clean (2 pre-existing warnings). - `cargo test -p cc-types`: 4/4 passing (teams + background_agents). - `cargo test -p cc-engine`: status_line + types suite unchanged. - `cargo test --bin claude-code-rs engine::agent query::loop_impl`: 69/69 passing. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Crsei and others added 5 commits April 21, 2026 08:27

This was referenced Apr 21, 2026

[workspace-split] Phase 6: Extract core crates (engine, query, tools, lsp-service) #75

Open

[workspace-split] Phase 7: Extract high-level crates (plugins, teams, daemon, commands, ipc) #76

Open

Crsei merged commit 0d3b75c into Crsei:rust-lite Apr 21, 2026
0 of 3 checks passed

Navigation Menu

refactor(workspace): Phase 6/7 scaffold + AppState/Tool trait extraction (#75 #76) - #86

refactor(workspace): Phase 6/7 scaffold + AppState/Tool trait extraction (#75 #76)#86
Crsei merged 6 commits into
Crsei:rust-litefrom
yaohaowei0914:claude/modest-morse-592f2d

yaohaowei0914 commented Apr 21, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

yaohaowei0914 commented Apr 21, 2026

Summary

Why

Key architectural notes

Incremental build speed (measured)

Phase 6 remaining

Test plan

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants