Skip to content

Anthropic provider implementation

Runtime classes

  • anthropic_plugins.provider.AnthropicProvider
  • anthropic_plugins.tools_extension.AnthropicToolsExtension
  • anthropic_plugins.reasoning_extension.AnthropicReasoningExtension
  • anthropic_plugins.reasoning_controls_extension.AnthropicReasoningControlsExtension
  • anthropic_plugins.usage_extension.AnthropicUsageExtension
  • anthropic_plugins.prompt_caching_extension.AnthropicPromptCachingExtension

All five extension classes plus the provider are registered by agent_plugin.json.

Provider boundary

AnthropicProvider owns:

  • Anthropic SDK client construction;
  • API-key consumption;
  • model discovery;
  • ordinary and streaming transport;
  • generic ordered content-block accumulation;
  • plain text conversion;
  • native history;
  • request cancellation;
  • provider-owned SDK retry configuration;
  • safe generic request failures.

The provider's content-block reducer preserves text, thinking, redacted-thinking, tool-use, and unknown block data. Preserving native data does not make the provider the owner of the corresponding public feature.

Extension boundaries

AnthropicToolsExtension owns tool schema conversion, call/result projection, partial input JSON, and Anthropic tool interop.

AnthropicReasoningExtension owns streamed reasoning, redacted thinking, signatures, reasoning metadata, and native-history reconstruction. It contributes no request config or UI.

AnthropicReasoningControlsExtension normalizes official model capability data, supplies known Claude regex fallbacks, contributes model-specific UI, and maps normalized settings into thinking and output_config. RequestOptionsFeature remains the final raw request override.

AnthropicUsageExtension owns public token metadata, formatting, turn/session accumulation, stable turn IDs and terminal/tool-call state, and the standard cumulative fields consumed by performance extensions.

AnthropicPromptCachingExtension owns automatic top-level cache configuration, enabled/TTL UI, and capability gating. Cache usage remains in AnthropicUsageExtension.

Extensions use shared provider state but keep feature state under namespaced keys:

anthropic_runtime
anthropic_tools
anthropic_usage

Streaming sequence

  1. The provider initializes a minimal Messages payload.
  2. Tools add schemas and the generic request-options feature adds direct request fields.
  3. The SDK yields events normalized to plain dictionaries.
  4. The provider updates a lossless ordered assistant block accumulator.
  5. Extensions project tool calls, reasoning partials, and usage.
  6. The provider appends one completed native assistant message.
  7. Extensions finalize public metadata.
  8. Core persists the full provider-native history.

Verification layout

  • test_provider_unit.py: transport/config/model/conversion/reducer contracts.
  • test_extensions_unit.py: isolated extension ownership and mapping.
  • test_stack_integration.py: deterministic Core stack and persistence.
  • test_ollama_integration.py: real local Messages transport and application loading, model discovery, effort passthrough, usage, and Duration/TPS.
  • test_ollama_reasoning_effort_integration.py: slow local Granite low/high reasoning-size effectiveness.
  • test_anthropic_api_integration.py: minimal official API verification.
  • test_anthropic_stage3_api_integration.py: official capability descriptors, prompt caching, adaptive interleaving, authentication errors, and cancellation.

The deterministic tests define exact behavior. Ollama tests prove real compatible transport. Hosted tests cover only official API behavior that the local endpoint cannot establish.