# Task: Create anthropic_messages Tool Config
## Context
Create the base Anthropic Messages API tool configuration. This is a pure YAML config that chains to `http_client` primitive.
## Dependencies
- Must complete: Phase 8.1 (streaming support)
## Files to Create
- `.ai/tools/llm/anthropic_messages.yaml`
## Implementation Steps
1. Create `.ai/tools/llm/` directory if it doesn't exist
2. Create `anthropic_messages.yaml` with tool metadata
3. Configure HTTP request (method, url, headers, auth)
4. Configure body template with parameters
5. Configure streaming (SSE transport, return sink)
6. Add retry configuration
## Code Snippet
From doc lines 512-578:
```yaml
# .ai/tools/llm/anthropic_messages.yaml
tool_id: anthropic_messages
tool_type: http
version: "1.0.0"
description: "Call Anthropic Messages API"
executor_id: http_client
config:
method: POST
url: "https://api.anthropic.com/v1/messages"
auth:
type: bearer
token: "${ANTHROPIC_API_KEY}"
headers:
Content-Type: application/json
anthropic-version: "2023-06-01"
# Body is templated with params
body:
model: "{model}"
max_tokens: "{max_tokens}"
stream: "{stream}"
system: "{system_prompt}"
messages: "{messages}"
tools: "{tools}"
# Streaming config (used when stream=true)
stream:
transport: sse
destinations:
- type: return # Buffer for caller
parameters:
- name: model
type: string
required: true
default: "claude-sonnet-4-20250514"
- name: max_tokens
type: integer
required: true
default: 4096
- name: stream
type: boolean
default: false
- name: system_prompt
type: string
required: false
- name: messages
type: array
required: true
description: "Array of {role, content} message objects"
- name: tools
type: array
required: false
description: "Tool definitions for function calling"
# A.3: Retry configuration is data-driven in tool YAML
retry:
max_attempts: 3
backoff_ms: [250, 1000, 3000] # Exponential backoff
retryable_errors:
- STREAM_INCOMPLETE
- CONNECTION_RESET
- TIMEOUT
```
## Success Criteria
- [ ] YAML file exists with correct structure
- [ ] All required fields are present
- [ ] Body templating uses `{param}` syntax
- [ ] Streaming config is correct
- [ ] Retry configuration is included
- [ ] YAML validates without errors
## Verification Command
```bash
# Validate YAML syntax
python -c "
import yaml
with open('.ai/tools/llm/anthropic_messages.yaml') as f:
config = yaml.safe_load(f)
assert config['tool_id'] == 'anthropic_messages'
assert config['executor_id'] == 'http_client'
print('YAML validation passed!')
"
```
## Notes
- This is pure config - no Python code needed
- Body templating uses `{param}` placeholders
- Streaming is configured but only used when `stream=true`
- Retry config is data-driven (not hardcoded)
- Auth token comes from environment variable
