Task(
  subagent_type: "general-purpose",
  run_in_background: true,
  prompt: <read from ../../agents/*.md file>
)

User: "tokio latest version"

Claude:
1. Read ../../agents/crate-researcher.md
2. Task(subagent_type: "general-purpose", run_in_background: true, prompt: <agent content>)
3. Wait for agent
4. Summarize results

User: "What's new in Rust 1.85?"

Claude:
1. Read ../../agents/rust-changelog.md
2. Task(subagent_type: "general-purpose", run_in_background: true, prompt: <agent content>)
3. Wait for agent
4. Summarize features

1. actionbook: mcp__actionbook__search_actions("lib.rs crate info")
2. Get action details: mcp__actionbook__get_action_by_id(<action_id>)
3. agent-browser CLI (or WebFetch fallback):
   - open "https://lib.rs/crates/{crate_name}"
   - get text using selector from actionbook
   - close
4. Parse and format output

## {Crate Name}

**Version:** {latest}
**Description:** {description}

**Features:**
- `feature1`: description

**Links:**
- [docs.rs](https://docs.rs/{crate}) | [crates.io](https://crates.io/crates/{crate}) | [repo]({repo_url})

1. actionbook: mcp__actionbook__search_actions("releases.rs rust changelog")
2. Get action details for selectors
3. agent-browser CLI (or WebFetch fallback):
   - open "https://releases.rs/docs/1.{version}.0/"
   - get text using selector from actionbook
   - close
4. Parse and format output

## Rust 1.{version}

**Release Date:** {date}

### Language Features
- Feature 1: description
- Feature 2: description

### Library Changes
- std::module: new API

### Stabilized APIs
- `api_name`: description

1. Construct URL: "https://doc.rust-lang.org/std/{path}/"
   - Traits: std/{module}/trait.{Name}.html
   - Structs: std/{module}/struct.{Name}.html
   - Modules: std/{module}/index.html
2. agent-browser CLI (or WebFetch fallback):
   - open <url>
   - get text "main .docblock"
   - close
3. Parse and format output

## std::{path}::{Name}

**Signature:**
```rust
{signature}

{example_code}

### Third-Party Crate Docs (tokio, serde, etc.)

**Output Format:**
```markdown
## {crate}::{path}

**Signature:**
```rust
{signature}

{example_code}

### Clippy Lints

**Output Format:**
```markdown
## Clippy Lint: {lint_name}

**Level:** {warn|deny|allow}
**Category:** {category}

**Description:**
{what_it_checks}

**Example (Bad):**
```rust
{bad_code}

{good_code}

---

## Tool Chain Priority

Both modes use the same tool chain order:

1. **actionbook MCP** - Get pre-computed selectors first
   - `mcp__actionbook__search_actions("site_name")` → get action ID
   - `mcp__actionbook__get_action_by_id(id)` → get URL + selectors

2. **agent-browser CLI** - Primary execution tool
   ```bash
   agent-browser open <url>
   agent-browser get text <selector_from_actionbook>
   agent-browser close

actionbook → agent-browser → WebFetch (only if agent-browser unavailable)