# agent-doc
Interactive document sessions with AI agents.
## Conventions
- Use `clap` derive for CLI argument parsing
- Use `serde` derive for all data types
- Use `serde_yaml` for frontmatter parsing
- Use `similar` crate for diffing (pure Rust, no shell `diff` dependency)
- Use `serde_json` for agent response parsing
- Use `std::process::Command` for git operations (not `git2`)
- Use `toml + serde` for config file parsing
- No async — sequential per-run
- Use `anyhow` for application errors
- **NEVER swallow errors** — no `let _ =` on fallible operations. Always log at minimum a warning to stderr. Silent failures make bugs invisible and waste debugging cycles.
- **All deterministic behavior in the binary** — document manipulation (compact, diff, merge, patch, write), snapshot management, git operations, and component parsing must live in Rust. The SKILL.md skill is the non-deterministic orchestrator (reads diff, generates response, decides what to write). Never implement deterministic document logic in the skill or ad-hoc scripts.
## Binary vs Agent Responsibility
| Responsibility | Owner | Why |
|---------------|-------|-----|
| Component parsing, patch application, mode resolution | **Binary** (Rust) | Deterministic, testable, consistent across agents |
| CRDT merge, snapshot management, atomic writes | **Binary** (Rust) | Concurrency safety requires flock + atomic rename |
| Diff computation, comment stripping, truncation detection | **Binary** (Rust) | Reproducible baseline comparison |
| Git operations (commit, history, clean) | **Binary** (Rust) | Direct `std::process::Command` calls |
| Tmux routing, session registry, pane management | **Binary** (Rust) | Process-level coordination |
| Pre-response snapshots, undo, extract, transfer | **Binary** (Rust) | File-level atomicity |
| Reading diff, interpreting user intent | **Skill** (SKILL.md) | Requires LLM reasoning |
| Generating response content | **Skill** (SKILL.md) | Non-deterministic |
| Deciding what to write to which component | **Skill** (SKILL.md) | Context-dependent |
| Streaming checkpoints, progress tracking | **Skill** (SKILL.md) | Response-generation timing |
| Pending item management (parse, populate, process) | **Skill** (SKILL.md) | Semantic understanding of prompts |
**Rule of thumb:** If the operation can be unit-tested with fixed inputs → binary. If it requires understanding natural language → skill.
- **Inline component attributes:** `<!-- agent:name patch=append -->` — patch mode is configurable on the tag itself. `mode=` is accepted as a backward-compatible alias; `patch=` takes precedence if both are present. Precedence: inline attr > `components.toml` > built-in defaults.
- **`agent_doc_format: inline`** is the canonical name for the old "append" format (`append` accepted as backward-compat alias). Template mode uses components; inline mode uses User/Assistant blocks.
## Module Layout
Use this layout when adding modules. Add new subcommands in their own file, wired through `main.rs`.
```
src/
main.rs # CLI entry point (clap derive)
submit.rs # Core loop: diff, send, merge-safe write, snapshot, git
init.rs # Scaffold session document
reset.rs # Clear session + snapshot
diff.rs # Preview diff (dry run) + comment stripping
clean.rs # Squash git history
component.rs # Component parser (<!-- agent:name --> markers) + name validation
patch.rs # Replace/append/prepend component content, config + shell hooks
watch.rs # Watch daemon: auto-submit on file change with debounce + loop prevention (reactive mode for stream docs)
frontmatter.rs # YAML frontmatter parse/write
snapshot.rs # Snapshot path/read/write
git.rs # Commit, branch, squash (includes `commit` subcommand)
config.rs # Global config (~/.config/agent-doc/config.toml)
sessions.rs # Session registry (sessions.json) + Tmux struct
route.rs # Route /agent-doc commands to correct tmux pane (pub auto_start for sync.rs)
start.rs # Start Claude session inside tmux pane
claim.rs # Claim document for current tmux pane
focus.rs # Focus tmux pane for a session document
layout.rs # Arrange tmux panes to mirror editor split layout
outline.rs # Markdown section structure + token counts
prompt.rs # Detect permission prompts from Claude Code sessions (strip_ansi is pub(crate))
skill.rs # Manage bundled SKILL.md (install/check)
resync.rs # Validate sessions.json, remove dead panes, detect wrong-session/wrong-process panes (--fix)
history.rs # Exchange version history from git + restore
upgrade.rs # Self-update via crates.io / GitHub Releases
plugin.rs # Editor plugin install/update/list via GitHub Releases
crdt.rs # CRDT foundation (yrs-based conflict-free merge)
merge.rs # 3-way merge + CRDT merge path
stream.rs # Stream command: real-time CRDT write-back loop
agent/
mod.rs # Agent trait
claude.rs # Claude backend (Agent + StreamingAgent)
streaming.rs # StreamingAgent trait + stream-json parser
terminal.rs # Launch external terminal with tmux session
parallel.rs # Parallel fan-out with git worktrees
worktree.rs # Git worktree management for parallel sessions
audit_docs.rs # Audit instruction files (via instruction-files crate)
editors/
jetbrains/ # IntelliJ plugin (Kotlin/Gradle)
vscode/ # VS Code extension (TypeScript)
```
## Release Process
1. Bump version in `Cargo.toml` + `pyproject.toml` (keep in sync)
2. `make check` (clippy + test)
3. Branch → PR → squash merge to main
4. Tag: `git tag v<version> && git push origin v<version>`
5. `cargo publish` (crates.io)
6. `maturin publish` (PyPI)
7. `gh release create v<version> --generate-notes` with prebuilt binary (GitHub Release)
8. Install binary: `cargo install --path .`
## Agent Backend Contract
Each agent backend implements: take a prompt string, return (response_text, session_id).
The prompt includes the diff and full document. The agent backend handles CLI
invocation, JSON parsing, and session flags.
### StreamingAgent Contract
Streaming backends implement `StreamingAgent::send_streaming()` → `Iterator<StreamChunk>`.
Used by `agent-doc stream` for real-time write-back. Currently only `claude` supports streaming
(via `--output-format stream-json`). Each `StreamChunk` has cumulative text, optional
`thinking` content, `is_final` flag, and optional `session_id` on the final chunk.
## Stream Mode
Stream mode (`agent_doc_format: template` + `agent_doc_write: crdt`) enables real-time agent output with CRDT-based conflict-free merge. Legacy: `agent_doc_mode: stream` is still supported as a deprecated alias.
**Usage:** `agent-doc stream <FILE> [--interval 200] [--agent claude] [--model opus] [--no-git]`
**How it works:**
1. Validates document uses CRDT write strategy (`resolved.is_crdt()`), reads `StreamConfig` from frontmatter
2. Computes diff, builds prompt requesting patch-block format
3. Spawns streaming agent (`claude -p --output-format stream-json`)
4. Timer thread (default 200ms) periodically flushes accumulated text to document:
`flock → read file → apply template patch (replace mode) → atomic write → unlock`
5. On completion: saves CRDT state + snapshot, updates resume ID, optional git commit
**Frontmatter:**
```yaml
agent_doc_format: template
agent_doc_write: crdt
agent_doc_stream:
interval: 200 # write-back interval (ms), default 200
strip_ansi: true # strip ANSI codes from output
target: exchange # target component name
thinking: false # include chain-of-thought (default: false)
thinking_target: log # route thinking to separate component (optional)
```
### Chain of Thought
Stream mode can capture the agent's chain-of-thought (thinking blocks) from Claude's
`stream-json` output. Controlled by `thinking` and `thinking_target` in `StreamConfig`:
| Config | Behavior |
|--------|----------|
| `thinking: false` (default) | Thinking blocks silently skipped |
| `thinking: true` (no `thinking_target`) | Thinking interleaved in target component as `<details><summary>Thinking</summary>...</details>` |
| `thinking: true` + `thinking_target: log` | Thinking routed to separate `<!-- agent:log -->` component; response goes to target |
**Parser:** `extract_assistant_content()` in `streaming.rs` extracts both `"type": "text"` and
`"type": "thinking"` content blocks. Thinking is buffered separately (`thinking_buffer`) and
flushed on the same timer interval as response text.
Use this frontmatter structure to enable thinking with a separate log component:
```markdown
---
agent_doc_format: template
agent_doc_write: crdt
agent_doc_stream:
target: exchange
thinking: true
thinking_target: log
---
<!-- agent:exchange -->
User prompt here.
<!-- /agent:exchange -->
<!-- agent:log -->
<!-- /agent:log -->
```
### Flush Behavior
Stream flushes use the normal mode resolution chain (inline attr > `components.toml` > built-in default). `run_stream()` no longer hardcodes replace mode for exchange — mode resolution applies normally.
Implementation: `flush_to_document()` uses `template::apply_patches()`.
### IPC-First Writes (v0.17.5)
All write paths (`run`, `stream`, `write`) try IPC to the IDE plugin when `.agent-doc/patches/` exists (plugin installed) and `--force-disk` is not set. IPC writes a JSON patch file to `.agent-doc/patches/<hash>.json`; the IDE plugin applies it via Document API (preserving cursor, undo stack, no "externally modified" dialog) and deletes the file as ACK. On IPC timeout (2s), exits with code 75 (`EX_TEMPFAIL`) instead of falling back to disk write. Use `agent-doc write --force-disk` to bypass IPC and write directly to disk.
- `try_ipc()` — component-level patches for template/stream documents
- `try_ipc_full_content()` — full document replacement for inline-mode documents
- Both are safe to call unconditionally; they return `false` immediately if `.agent-doc/patches/` does not exist
**Key files:** `crdt.rs` (CRDT foundation), `merge.rs` (CRDT merge path), `stream.rs` (command),
`agent/streaming.rs` (StreamingAgent trait + chunk parser), `agent/claude.rs` (streaming impl)
**Reactive file-watching:** CRDT-mode documents (`resolved.is_crdt()`) get reactive file-watching (zero debounce) from the watch daemon. The `WatchEntry` has a `reactive: bool` field set by `discover_entries()` for CRDT docs. Reactive paths are tracked in a `HashSet<PathBuf>` and use `Duration::ZERO` for the debounce check, enabling instant re-submit on file change.
**One session per document:** Each `agent-doc stream` spawns its own Claude CLI process.
Multiple documents stream in parallel via separate tmux panes.
**CRDT state storage:** `.agent-doc/crdt/<hash>.yrs` — persisted after each stream for
subsequent merges. Compacted via `agent-doc compact` to GC tombstones.
## Domain Ontology
agent-doc extends the existence kernel vocabulary (defined in `~/.claude/philosophy/src/`) with domain-specific terms. See the full ontology table in [README.md](README.md#domain-ontology).