Documentation corrections and memories additions

.codescout/memories/architecture.md

@@ -33,7 +33,7 @@ src/
 │   ├── git.rs             #   git_blame, file_log (not registered; used by dashboard)
 │   ├── semantic.rs        #   semantic_search, index_project, index_status
 │   ├── github.rs          #   github_identity, github_issue, github_pr, github_file, github_repo
-│   ├── library.rs         #   list_libraries
+│   ├── library.rs         #   list_libraries, register_library
 │   ├── memory.rs          #   memory (action: read/write/list/delete/remember/recall/forget/refresh_anchors)
 │   ├── usage.rs           #   GetUsageStats (dashboard API; not an MCP tool)
 │   ├── ast.rs             #   list_functions, list_docs (not registered; tree-sitter offline tools)
@@ -48,7 +48,7 @@ src/
 1. `find_tool(name)` — linear scan over `Vec<Arc<dyn Tool>>`
 2. `check_tool_access(name, &security)` — denormalized match-arm gate in
    `src/util/path_security.rs`. Missing a write tool here bypasses access controls.
-3. Build `ToolContext { agent, lsp, output_buffer, progress }`
+3. Build `ToolContext { agent, lsp, output_buffer, progress, peer, section_coverage }`
 4. Apply `tool_timeout_secs` from `project.toml` (skipped for `index_project`, `onboarding`)
 5. `UsageRecorder::record_content` wraps `tool.call_content()`
 6. `route_tool_error`: `RecoverableError` → `isError:false` + JSON error/hint;
@@ -81,7 +81,8 @@ src/
 `build_index(root, force)` in `src/embed/index.rs`:
 1. `find_changed_files()`: git diff → mtime → SHA-256 fallback
 2. `ast_chunker::split_file()`: AST-aware chunking per language, respects chunk size config
-3. Concurrent embedding with semaphore cap=4 (`JoinSet` over `Embedder::embed()`)\n4. Single SQLite transaction: delete old chunks, insert new, upsert file hash
+3. Concurrent embedding with semaphore cap=4 (`JoinSet` over `Embedder::embed()`)
+4. Single SQLite transaction: delete old chunks, insert new, upsert file hash
 5. Drift detection (if enabled): cosine distance old→new embeddings → `drift_report` table
 6. High-drift files → mark memory anchors stale
 
@@ -98,6 +99,38 @@ chunk embeddings into memory for each query — known perf issue for large index
   referenced in content. `project_status` checks SHA-256 of anchored files to surface
   stale memories. Regenerated on each `write`; cleared via `refresh_anchors` action.
 
+## LSP Multiplexer (Kotlin-specific)
+
+The `mux` subcommand (`--hidden` in CLI) spawns a Unix-socket-based multiplexer for
+languages that support only one server instance per workspace (Kotlin's kotlin-lsp).
+
+Flow:
+- `LspManager::do_start(language, root)` for mux-enabled languages calls
+  `ensure_mux_running(language, root)` in `src/lsp/mux/process.rs`
+- Mux socket path: `/tmp/codescout-<lang>-mux-<workspace_hash>.sock`
+- Lock path: `/tmp/codescout-<lang>-mux-<workspace_hash>.lock`
+- Auto-terminates after `idle_timeout` (default 300s) with no connected clients
+- Multiple codescout instances share one kotlin-lsp process via the mux socket
+
+Kotlin-specific: passes `--system-path=/tmp/codescout-mux-kotlin-lsp` and
+`GRADLE_USER_HOME=/tmp/codescout-mux-gradle` to isolate from per-PID paths.
+
+## ToolContext Fields
+
+`ToolContext` (src/tools/mod.rs):
+| Field | Type | Description |
+|-------|------|-------------|
+| `agent` | `Agent` | Shared state: project config, memories, registry |
+| `lsp` | `Arc<dyn LspProvider>` | Mock-swappable LSP access |
+| `output_buffer` | `Arc<OutputBuffer>` | Session LRU store for @tool_* refs |
+| `progress` | `Option<Arc<ProgressReporter>>` | MCP progress notifications |
+| `peer` | `Option<Peer<RoleServer>>` | MCP peer for elicitation; None in tests |
+| `section_coverage` | `Arc<Mutex<SectionCoverage>>` | Tracks markdown sections read in session |
+
+`ctx.elicit::<T>(message)` sends an MCP elicitation request to the user. Returns
+`Ok(Some(T))` on response, `Ok(None)` if client doesn't support it, or
+`RecoverableError` if user declined/cancelled.
+
 ## Unregistered Tool Structs
 
 Several tool structs exist in code but are NOT registered in `from_parts`:

.codescout/memories/project-overview.md

@@ -9,7 +9,7 @@ rather than raw shell reads on source files.
 
 ## Tech Stack
 - **Language:** Rust 1.75+ (MSRV enforced in CI)
-- **MCP SDK:** `rmcp 0.1` (stdio + SSE transports)
+- **MCP SDK:** `rmcp 1.x` (stdio + HTTP transports; builder pattern for `StreamableHttpServerConfig`)
 - **LSP:** JSON-RPC clients for 9 languages via `lsp-types 0.97`
 - **AST:** `tree-sitter` with grammars for Rust/Python/TypeScript/Go/Java/Kotlin
 - **Embeddings:** SQLite via `rusqlite` (bundled); `remote-embed` (reqwest, Ollama/OpenAI-compatible) or `local-embed` (fastembed ONNX) — feature flags
@@ -29,4 +29,4 @@ rather than raw shell reads on source files.
 - `e2e-*`: integration tests requiring real LSP servers installed
 
 ## Current Version
-v0.2.2 (see `Cargo.toml`). Published to crates.io from `master` branch only.
+v0.7.2 (see `Cargo.toml`). Published to crates.io from `master` branch only.

CLAUDE.md

@@ -219,7 +219,7 @@ See `did_change_refreshes_stale_symbol_positions` in `src/lsp/client.rs` for the
 
 ## Key Patterns
 
-**Tool trait** (`src/tools/mod.rs`): Each tool is a struct implementing `name()`, `description()`, `input_schema()`, `async call(Value, &ToolContext) -> Result<Value>`. 29 tools registered. All use `#[async_trait]`.
+**Tool trait** (`src/tools/mod.rs`): Each tool is a struct implementing `name()`, `description()`, `input_schema()`, `async call(Value, &ToolContext) -> Result<Value>`. 27 tools registered. All use `#[async_trait]`.
 
 **Tool↔MCP bridge** (`src/server.rs`): Tools registered as `Vec<Arc<dyn Tool>>`, dispatched dynamically in `call_tool`. Errors are routed through `route_tool_error`:
 - `RecoverableError` (`src/tools/mod.rs`) → `isError: false` with JSON `{"error":"…","hint":"…"}` — LLM sees the problem and a corrective hint, **sibling parallel calls are not aborted by Claude Code**.

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "codescout"
-version = "0.7.1"
+version = "0.7.2"
 edition = "2021"
 description = "High-performance coding agent toolkit MCP server"
 repository = "https://github.com/mareurs/codescout"

README.md

@@ -29,7 +29,7 @@ cargo build
 ./target/debug/codescout start --project /path/to/code
 ```
 
-Add codescout as an MCP server in Claude Code `settings.json`:
+Add codescout as an MCP server  `~/.claude.json`:
 
 ```json
 {

docs/agents/claude-code.md

@@ -1,24 +1,53 @@
 # Claude Code
 
-## One-Time Setup
+## Setup
 
-Prerequisites: Rust toolchain, `cargo install codescout`. The binary lands at `~/.cargo/bin/codescout`.
+Prerequisites: Rust toolchain
 
-Register codescout as an MCP server. The recommended approach is user-level registration — edit `~/.claude/settings.json`:
+### Using `cargo install` (recommended)
+
+Install the binary:
+
+```bash
+cargo install codescout
+# lands at ~/.cargo/bin/codescout
+```
+
+For this route, we can register as a user-level MCP server via the CLI:
+
+```bash
+claude mcp add --scope user codescout ~/.cargo/bin/codescout -- start
+```
+
+This writes the entry into `~/.claude.json` (user-scoped).
+
+
+For a project-scoped alternative, place a `.mcp.json` file at the project root:
 
 ```json
 {
   "mcpServers": {
     "codescout": {
-      "command": "codescout",
-      "args": ["start"],
+      "command": "/absolute/path/to/target/release/codescout",
+      "args": ["start", "--project", "/absolute/path/to/your/project"],
       "type": "stdio"
     }
   }
 }
 ```
 
-For a project-scoped alternative, place a `.mcp.json` file at the project root with the same block.
+Claude Code detects `.mcp.json` automatically and prompts for approval on first use. 
+
+### Using a local build
+
+If you're working from source or want to pin to a specific local build:
+
+```bash
+cargo build --release
+# binary at target/release/codescout
+```
+
+If you opt for the local build, manually add to either json based on desired scope. Use absolute paths — relative paths are not reliable here.
 
 ## Workflow Skills
 

src/server.rs

@@ -41,7 +41,7 @@ pub struct CodeScoutServer {
     agent: Agent,
     lsp: Arc<dyn LspProvider>,
     output_buffer: Arc<crate::tools::output_buffer::OutputBuffer>,
-    // Arc<dyn Tool>: heterogeneous collection of 23+ different tool types dispatched by name at runtime.
+    // Arc<dyn Tool>: heterogeneous collection of 27 registered tool types dispatched by name at runtime.
     tools: Vec<Arc<dyn Tool>>,
     /// Pre-computed at construction because `get_info()` is sync.
     /// Becomes stale if project state changes mid-session (e.g. after onboarding or indexing).