Design/Plan/Claude.md from Claude.

This commit is contained in:
Drew 2026-02-23 21:39:31 -08:00
commit 42e3ddacc2
5 changed files with 310 additions and 0 deletions

1
.gitignore vendored Normal file
View file

@ -0,0 +1 @@
target/

63
CLAUDE.md Normal file
View file

@ -0,0 +1,63 @@
# CLAUDE.md
Rust TUI coding agent. Ratatui + Crossterm + Tokio. See DESIGN.md for architecture decisions and PLAN.md for implementation phases.
## Commands
- `cargo build`: Build the project
- `cargo test`: Run all unit and integration tests
- `cargo test --lib`: Unit tests only
- `cargo test --test '*'`: Integration tests only
- `cargo clippy -- -D warnings`: Lint (must pass with zero warnings)
- `cargo fmt --check`: Format check
- `cargo run -- --project-dir <path>`: Run against a project directory
## Architecture
Six modules with strict boundaries:
- `src/app/` — Wiring, lifecycle, tokio runtime setup
- `src/tui/` — Ratatui rendering, input handling, vim modes. Communicates with core ONLY via channels (`UserAction` → core, `UIEvent` ← core). Never touches conversation state directly.
- `src/core/` — Conversation tree, orchestrator loop, sub-agent lifecycle
- `src/provider/``ModelProvider` trait + Claude implementation. Leaf module, no internal dependencies.
- `src/tools/``Tool` trait, registry, built-in tools. Depends only on `sandbox`.
- `src/sandbox/` — Landlock policy, path validation, command execution. Leaf module.
- `src/session/` — JSONL logging, session read/write. Leaf module.
The channel boundary between `tui` and `core` is critical — never bypass it. The TUI is a frontend; core is the engine. This separation enables headless mode for benchmarking.
## Code Style
- Use `thiserror` for error types, not `anyhow` in library code (`anyhow` only in `main.rs`/`app`)
- Prefer `impl Trait` return types over boxing when possible
- All public types need doc comments
- No `unwrap()` in non-test code — use `?` or explicit error handling
- Async functions should be cancel-safe where possible
- Use `tracing` for structured logging, not `println!` or `log`
## Conversation Data Model
Events use parent IDs forming a tree (not a flat list). This enables future branching. Every event has: id, parent_id, timestamp, event_type, token_usage. A "turn" is all events between two user messages — this is the unit for token tracking.
## Testing
- Unit tests go in the same file as the code (`#[cfg(test)] mod tests`)
- Integration tests go in `tests/`
- TUI widget tests use `ratatui::backend::TestBackend` + `insta` snapshots
- Provider tests replay recorded SSE fixtures from `tests/fixtures/`
- Sandbox tests use `tempdir` and skip Landlock-specific assertions if kernel < 5.13
- Run `cargo test` before every commit
## Key Constraints
- All file I/O and process spawning in tools MUST go through `Sandbox` — never use `std::fs` or `std::process::Command` directly in tool implementations
- The `ModelProvider` trait must remain provider-agnostic — no Claude-specific types in the trait interface
- Session JSONL is append-only. Never rewrite history. Branching works by writing new events with different parent IDs.
- Token usage must be tracked per-event and aggregatable per-turn
## Do Not
- Add MCP support (deferred, but keep tool trait compatible)
- Use `unsafe` without discussion
- Add dependencies without checking if an existing dep covers the use case
- Modify test fixtures without re-recording from a real API session

111
DESIGN.md Normal file
View file

@ -0,0 +1,111 @@
# Design Decisions
## Stack
- **Language:** Rust
- **TUI Framework:** Ratatui + Crossterm
- **Async Runtime:** Tokio
## Architecture
- Channel boundary between TUI and core (fully decoupled)
- Module decomposition: `app`, `tui`, `core`, `provider`, `tools`, `sandbox`, `session`
- Headless mode: core without TUI, driven by script (enables benchmarking and CI)
## Model Integration
- Claude-first, multi-model via `ModelProvider` trait
- Common `StreamEvent` internal representation across providers
- Prompt caching-aware message construction
## UI
- **Agent view:** Tree-based hierarchy (not flat tabs) for sub-agent inspection
- **Modes:** Normal, Insert, Command (`:` prefix from Normal mode)
- **Activity modes:** Plan and Execute are visually distinct activities in the TUI
- **Streaming:** Barebones styled text initially, full markdown rendering deferred
- **Token usage:** Per-turn display (between user inputs), cumulative in status bar
- **Status bar:** Mode indicator, current activity (Plan/Execute), token totals, network policy state
## Planning Mode
- Distinct activity from execution — planner agent produces a plan file, does not execute
- Plan file is structured markdown: steps with descriptions, files involved, acceptance criteria
- Plan is reviewable and editable before execution (`:edit-plan` opens `$EDITOR`)
- User explicitly approves plan before execution begins
- Executor agent receives the plan file + project context, not the planning conversation
- Plan-step progress tracked during execution (complete/in-progress/failed)
## Sub-Agents
- Independent context windows with summary passed back to parent
- Fully autonomous once spawned
- Hard deny on unpermitted actions
- Plan executor is a specialized sub-agent where the plan replaces the summary
- Direct user interaction with sub-agents deferred
## Tool System
- Built-in tool system with `Tool` trait
- Core tools: `read_file`, `write_file`, `edit_file`, `shell_exec`, `list_directory`, `search_files`
- Approval gates by risk level: auto-approve (reads), confirm (writes/shell), deny
- MCP not implemented but interface designed to allow future adapter
## Sandboxing
- **Landlock** (Linux kernel-level):
- Read-only: system-wide (`/`)
- Read-write: project directory, temp directory
- Network: blocked by default, toggleable via `:net on/off`
- Graceful degradation on older kernels
- All tool execution goes through `Sandbox` — tools never touch filesystem directly
## Session Logging
- JSONL format, one event per line
- Events: user message, assistant message, tool call, tool result, sub-agent spawn/result, plan created, plan step status
- Tree-addressable via parent IDs (enables conversation branching later)
- Token usage stored per event
- Linear UX for now, branching deferred
## Testing Strategy
### Unit Tests
- **`provider`:** SSE stream parsing from byte fixtures, message/tool serialization, `StreamEvent` variant correctness
- **`tools`:** Path canonicalization, traversal prevention, risk level classification, registry dispatch
- **`sandbox`:** Landlock policy construction, path validation logic (without applying kernel rules)
- **`core`:** Conversation tree operations (insert, query by parent, turn computation, token totals), orchestrator state machine transitions against mock `StreamEvent` sequences
- **`session`:** JSONL serialization roundtrips, parent ID chain reconstruction
- **`tui`:** Widget rendering via Ratatui `TestBackend`, snapshot tests with `insta` crate for layout/mode indicator/token display
### Integration Tests — Component Boundaries
- **Core ↔ Provider:** Mock `ModelProvider` replaying recorded API sessions (full SSE streams with tool use). Tests the complete orchestration loop deterministically without network.
- **Core ↔ TUI (channel boundary):** Orchestrator with mock provider connected to channels. Assert correct `UIEvent` sequence, inject `UserAction` messages, verify approval/denial flow.
- **Tools ↔ Sandbox:** Real file operations and shell commands in temp directories. Verify write confinement, path traversal rejection, network blocking. Skip Landlock-specific tests on older kernels in CI.
### Integration Tests — End to End
- **Recorded session replay:** Capture real Claude API HTTP request/response pairs, replay deterministically. Exercises full stack (core + channel + mock TUI) without cost or network dependency. Primary E2E test strategy.
- **Live API tests:** Small suite behind feature flag / env var. Verifies real API integration. Run manually before releases, not in CI.
### Snapshot Testing
- `insta` crate for TUI visual regression testing from Phase 2 onward
- Capture rendered `TestBackend` buffers as string snapshots
- Catches layout, mode indicator, and token display regressions
### Benchmarking — SWE-bench
- **Target:** SWE-bench Verified (500 curated problems) as primary benchmark
- **Secondary:** SWE-bench Pro for testing planning mode on longer-horizon tasks
- **Approach:** Headless mode (core without TUI) produces unified diff patches, evaluated via SWE-bench Docker harness
- **Baseline:** mini-swe-agent (~100 lines Python, >74% on Verified) as calibration — if we score significantly below with same model, the issue is scaffolding
- **Cadence:** Milestone checks, not continuous CI (too expensive/slow)
- **Requirements:** x86_64, 120GB+ storage, 16GB RAM, 8 CPU cores
### Test Sequencing
- Phase 1: Unit tests for SSE parser, event types, message serialization
- Phase 2: Snapshot tests for TUI with `insta`
- Phase 4: Recorded session replay infrastructure (core loop complex enough to warrant it)
- Phase 6-7: Headless mode + first SWE-bench Verified run
## Configuration (Deferred)
- Single-user, hardcoded defaults for now
- Designed for later: global config, per-project `.agent.toml`, configurable keybindings
## Deferred Features
- Conversation branching (tree structure in log, linear UX for now)
- Direct sub-agent interaction
- MCP adapter
- Full markdown/syntax-highlighted rendering
- Session log viewer
- Per-project configuration
- Structured plan editor in TUI (use `$EDITOR` for now)

47
IDEAS.md Normal file
View file

@ -0,0 +1,47 @@
# IDEAS
Notes based on ideas I've had.
## Token Usage Visualization
- Per-turn token breakdown (input/output/cache) inline in conversation
- Cumulative session totals in status bar
- Estimated Cost of Usage
## Planning Mode
- Activity mode distinction in TUI (Plan vs Execute), visible in status bar
- Planner agent: has tool access (reads, search) but no write/exec permissions
- Plan output as structured markdown (steps, files, acceptance criteria)
- `:edit-plan` command to open plan in `$EDITOR` before execution
- Explicit plan approval gate before transitioning to execution
- Executor agent spawned with plan file + project context (not planning conversation)
- Plan-step progress tracking (complete/in-progress/failed) visible in TUI
- **Done when:** Can plan a task, review/edit the plan, then execute it as a separate activity
## Sub-Agents
- `spawn_agent` tool, independent `ConversationTree` per sub-agent
- Agent tree sidebar in TUI, navigable in Normal mode
- Sub-agents follow same approval policy with hard deny on unpermitted actions
- Plan executor refactored as a sub-agent specialization
- **Done when:** Agent delegates to sub-agent, user can inspect it, result flows back
## Context Window Management
- Token counting for outgoing payloads
- Compaction strategy: summarize older turns, preserve full history in session log
- Stable message prefix for prompt caching
- **Done when:** Conversations run indefinitely without hitting context limits
## Automated Anomaly Notation
- Similar to Jon's SESSION.md: https://github.com/jonhoo/configs/blob/master/agentic/AGENTS.md
- Allows the agents to note an anomaly or bad design decision.
## Defered TODO list
- Allow the user to notate things that should be fixed after the agent has iterated on its full loop.
- Potentially add a way to iterate through the todo list at the end.
## Session Logging
- JSONL `SessionWriter` with `Event` structure
- Parent IDs, timestamps, token usage per event
- Predictable file location with session IDs
- **Done when:** Session files are coherent, parseable, with token counts per turn

88
PLAN.md Normal file
View file

@ -0,0 +1,88 @@
# Implementation Plan
## Phase 1: Minimal Conversation Loop
**Done when:** Multi-turn streaming conversation with Claude works in terminal
### 1.1 Project Scaffolding
- `Cargo.toml` with initial dependencies:
- `ratatui`, `crossterm` — TUI
- `tokio` (full features) — async runtime
- `serde`, `serde_json` — serialization
- `thiserror` — error types
- `tracing`, `tracing-subscriber` — structured logging
- `reqwest` (with `stream` feature) — HTTP client for SSE
- `futures` — stream combinators
- Establish `src/{app,tui,core,provider}/mod.rs` stubs
- `cargo build` passes; `cargo clippy -- -D warnings` passes on empty stubs
### 1.2 Shared Types (`src/core/types.rs`)
- `StreamEvent` enum: `TextDelta(String)`, `InputTokens(u32)`, `OutputTokens(u32)`, `Done`, `Error(String)`
- `UserAction` enum (TUI → core channel): `SendMessage(String)`, `Quit`
- `UIEvent` enum (core → TUI channel): `StreamDelta(String)`, `TurnComplete`, `Error(String)`
- `ConversationMessage` struct: `role: Role`, `content: String`
- All types derive `Debug`; all public types have doc comments
### 1.3 Provider: `ModelProvider` Trait + Claude SSE (`src/provider/`)
- `ModelProvider` trait: `async fn stream(&self, messages: &[ConversationMessage]) -> impl Stream<Item = StreamEvent>`
- `ClaudeProvider` struct: API key from env, `reqwest` HTTP client
- Serialize messages to Anthropic Messages API JSON format
- Parse SSE byte stream → `StreamEvent` (handle `content_block_delta`, `message_delta` for tokens, `message_stop`)
- Unit tests: SSE parsing from hardcoded byte fixtures in `#[cfg(test)]`
### 1.4 Core: Conversation State + Orchestrator Loop (`src/core/`)
- `ConversationHistory`: `Vec<ConversationMessage>` with `push` and `messages()` (flat list, no tree yet)
- `Orchestrator` struct holding history, provider, channel senders/receivers
- Orchestrator loop:
1. Await `UserAction` from TUI channel
2. On `SendMessage`: append user message, call `provider.stream()`
3. Forward each `StreamEvent` as `UIEvent` to TUI
4. Accumulate deltas into assistant message; append to history on `Done`
5. On `Quit`: break loop
### 1.5 TUI: Layout + Input + Streaming Display (`src/tui/`)
- `AppState` struct: `messages: Vec<(Role, String)>`, `input: String`, `scroll: u16`
- Ratatui layout: full-height `Paragraph` output area (scrollable) + single-line `Paragraph` input
- Insert mode only — printable chars append to `input`, Enter sends `UserAction::SendMessage`, Backspace deletes
- On `UIEvent::StreamDelta`: append to last assistant message in `messages`, re-render
- On `UIEvent::TurnComplete`: finalize assistant message
- Crossterm raw mode enter/exit; restore terminal on panic or clean exit
### 1.6 App Wiring + Entry Point (`src/app/`, `src/main.rs`)
- `main.rs`: parse `--project-dir <path>` CLI arg
- Initialize `tracing_subscriber` (log to file, not stdout — avoids TUI interference)
- Create `tokio::sync::mpsc` channel pair for `UserAction` and `UIEvent`
- Spawn `Orchestrator::run()` as a tokio task
- Run TUI event loop on main thread (Ratatui requires main thread for crossterm)
- On `UserAction::Quit` or Ctrl-C: signal orchestrator shutdown, restore terminal, exit cleanly
### 1.7 Phase 1 Unit Tests
- Provider: SSE byte fixture → correct `StreamEvent` sequence
- Provider: `ConversationMessage` vec → correct Anthropic API JSON shape
- Core: `ConversationHistory` push/read roundtrip
- Core: Orchestrator state transitions against mock `StreamEvent` sequence (no real API)
## Phase 2: Vim Modes and Navigation
- Normal, Insert, Command modes with visual indicator
- `j`/`k` scroll in Normal mode
- `:quit`, `:clear` commands
- **Done when:** Fluid mode switching and scrolling feels vim-native
## Phase 3: Tool Execution
- `Tool` trait, `ToolRegistry`, core tools (`read_file`, `write_file`, `shell_exec`)
- Tool definitions in API requests, parse tool-use responses
- Approval gate: core → TUI pending event → user approve/deny → result back
- Working directory confinement + path validation (no Landlock yet)
- **Done when:** Claude can read, modify files, and run commands with user approval
## Phase 4: Sandboxing
- Landlock: read-only system, read-write project dir, network blocked
- Tools execute through `Sandbox`, never directly
- `:net on/off` toggle, state in status bar
- Graceful degradation on older kernels
- **Done when:** Writes outside project dir fail; network toggle works