dev_agentandClaude Opus 5.5
3b36f07cb8
Add internal/ai/llm/anthropic, an llm.Provider over plain net/http:
- POST {base}/v1/messages and GET {base}/v1/models with x-api-key and
anthropic-version 2023-06-01. An empty key, a non-positive
Config.MaxTokens and a base URL that is not absolute http(s), or that
carries userinfo, a query or a fragment, are construction errors.
- System text goes in the top-level system field. A Request with a
Schema becomes one tool whose input_schema is the schema, forced by
tool_choice; Response.JSON is that tool's input. Otherwise the text
blocks are Response.Text. Usage and stop_reason are parsed.
- ListModels follows has_more/after_id pages; Verify fetches one page
with limit=1.
- 401/403 map to ErrAuth, 429 and 529 to ErrRateLimited with
Retry-After, 408 and deadlines to ErrTimeout, other statuses to
ErrProvider with a redacted excerpt.
- Errors carry the OpenAI adapter's redaction chain, so no Error(),
Unwrap layer, errors.As target or %#v output holds the key.
Contract tests replay sanitized fixtures adapted from Anthropic's public
Messages API docs over httptest, covering tool-forced JSON, text, each
error mapping and key redaction with the keys x-api-key, anthropic,
HTTP, rate, a one-letter key and a URL-escaped key.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Anthropic adapter fixtures
These are the Messages API response bodies the contract tests in
anthropic_test.go replay. They are not live captures. Product AI is
tested with fakes, and live vendor calls are opt-in and never use
credentials, so no request was sent to Anthropic to make them. Each fixture
is a sanitized adaptation of the response shapes in Anthropic's public API
documentation:
- Messages: https://docs.anthropic.com/en/api/messages
- Tool use, including
tool_choice{type: tool}: https://docs.anthropic.com/en/docs/build-with-claude/tool-use - List models: https://docs.anthropic.com/en/api/models-list
- Errors: https://docs.anthropic.com/en/api/errors
Sanitizing:
- IDs, request IDs, timestamps and token counts are illustrative values in the documented format, not values from a real account.
- Message content and tool input are replaced with MealQuest dinner text.
- No real credential appears anywhere.
{{KEY}}is a placeholder the tests replace with a fake key, to prove the adapter redacts a key echoed back in a vendor body.
Below, not in the public example marks fields or values that were added for a test case and do not come from the documented example.
| Fixture | Shape | Not in the public example |
|---|---|---|
messages_tool_use.json |
Message with one tool_use block, stop_reason: tool_use, usage with cache fields. |
The tool name and input. |
messages_text.json |
Message with text blocks, stop_reason: end_turn. |
A second text block, to test joining. |
messages_max_tokens.json |
Message with a text block, stop_reason: max_tokens. |
— |
models_page1.json |
Models list page: data, has_more: true, first_id, last_id. |
— |
models_page2.json |
Last models list page, has_more: false. |
The second model has no display_name, to test the fallback to id. |
error_authentication.json |
Error, authentication_error (401). |
The {{KEY}} text; added to test redaction. |
error_permission.json |
Error, permission_error (403). |
— |
error_rate_limit.json |
Error, rate_limit_error (429). |
The {{KEY}} text; added to test redaction. |
error_overloaded.json |
Error, overloaded_error (529). |
— |
error_not_found.json |
Error, not_found_error (404). |
— |
error_invalid_request.json |
Error, invalid_request_error (400). |
— |
error_api.json |
Error, api_error (500). |
The x-api-key: {{KEY}} text; added to test redaction. |