agentjido / req_llm

49%

main: 50%

LAST BUILD ON BRANCH egomes/fix-claude-multi-turn

branch: egomes/fix-claude-multi-turn

CHANGE BRANCH

x

Reset

Sync Branches

• egomes/fix-claude-multi-turn
• add-vllm-provider
• bedrock-clean
• bug/api-return-types
• bug/codec-tool-calls
• bug/debug-stream-return
• bug/incorrect-model-spec-docs
• bug/openai-tool-calls
• bug/streaming-race-condition
• cerebras
• chore/2025-10-14-update-fixtures
• chore/object-fixtures
• chore/object-fixtures-resurrected
• chore/refine-fixtures
• chore/refresh-coverage-tests
• chore/update-models-2025-09-21
• copilot/fix-32
• devtools
• feat/context-json-serialization
• feat/google-upload-file
• feat/in-type-support
• feat/structured-output-openai-google
• feature/cerebras-provider
• feature/model-catalog
• feature/normalize-bedrock-inference-profiles
• feature/pre-release-fixes
• feature/refactor-llm-api-fixtures
• feature/refined-key-management
• feature/unique-model-provider-options
• feature/upgrade-ex-aws-auth
• feature/zai-fixtures
• fix-anthropic-streaming
• fix-duplicate-clause
• fix-google
• fix-groq-stream-error
• fix-mix-task-docs
• fix-openai-max-tokens-param
• fix/cost-calculation-in-usage
• fix/google-file-support
• fix/google-structured-output
• fix/issue-65-http-status-validation
• fix/issue-96-validation-error-fields
• fix/proxy-options
• fix/registry-get-provider-nil-module
• fix/tool_calls
• google-vision
• improve-metadata-provider-errors
• main
• patch-1
• put-max-tokens-model-options
• refactor/context-tools
• refactor/req-streaming
• refactor/xai-structured-objects
• remove-jido-keys
• zai

Build # 159d39f357ad33b30d5c6328587075e1bcd51591-PR-94

Build Type

Pull #94

github

Committed by lostbean

Commit Message

Fix Anthropic provider tool result encoding for multi-turn conversations

Resolves critical bugs preventing tool calling from working with
Anthropic models:

1. Transform :tool role to :user - Anthropic API only accepts "user" or
"assistant"
   roles, not "tool". Tool results must be sent in "user" messages.

2. Encode tool_result content blocks - Previously returned nil, causing
content to be
   filtered out. Now properly encodes with tool_use_id and serialized
output.

Changes:
- lib/req_llm/providers/anthropic/context.ex: Add role transformation
and tool_result encoding
- test/providers/anthropic_test.exs: Add multi-turn tool calling test
- test/req_llm/integration/tool_result_encoding_test.exs: Add
comprehensive provider-agnostic
  test suite for tool result encoding (7 tests covering edge cases)

All tests pass. Multi-turn tool calling now works correctly with
Anthropic models.

Pull Request Pull Request #94: fix(anthropic): Provider tool result encoding for multi-turn conversations

Run Details

5 of 5 new or added lines in 1 file covered. (100.0%)

2690 of 5502 relevant lines covered (48.89%)

177.66 hits per line