Summary

• sanitize both sides of Anthropic tool pairs before dispatch:
  • drop or convert orphaned tool_result blocks when their matching tool_use is no longer in the immediately preceding assistant message
  • drop or convert orphaned assistant tool_use blocks when the next user message no longer contains the corresponding tool_result
• truncate oversized native Anthropic requests before the final sanitization pass, so truncation cannot leave invalid tool pairs in the provider payload
• normalize Anthropic streaming errors so deferred stream failures still match the Anthropic SSE error envelope

Why

The earlier v0.2 sanitizer handles orphaned tool_result blocks, but current main can still send invalid compacted histories to Anthropic:

• assistant tool_use left without a following tool_result
• mixed user messages containing one valid tool_result plus another orphaned tool_result
• context truncation after sanitization, which can create new orphaned blocks

I reproduced this against current main with local services:

• a normal control /claude/v1/messages request returns 200 on both main and this branch
• compacted history with orphaned assistant tool_use: main returns 400 (tool_use ids were found without tool_result blocks immediately after); this branch returns 200
• mixed valid + orphaned tool_result: main returns 400 (unexpected tool_use_id found in tool_result blocks); this branch returns 200

Notes

This supersedes #33 for the current main branch. I rebased the change onto current main and removed the noisy/manual SDK investigation test from the older PR so this PR only contains deterministic unit/regression coverage.

Validation

• uv run ruff check ccproxy/llms/formatters/openai_to_anthropic/requests.py ccproxy/llms/utils/__init__.py ccproxy/llms/utils/context_truncation.py ccproxy/llms/utils/token_estimation.py ccproxy/plugins/claude_api/adapter.py ccproxy/plugins/claude_sdk/adapter.py ccproxy/streaming/deferred.py tests/test_tool_result_sanitization.py tests/plugins/claude_api/unit/test_native_anthropic_sanitization.py tests/unit/llms/test_context_window_management.py tests/unit/llms/test_truncation_creates_orphans.py tests/unit/streaming/test_deferred_stream_errors.py
• uv run ruff format --check ccproxy/llms/formatters/openai_to_anthropic/requests.py ccproxy/llms/utils/__init__.py ccproxy/llms/utils/context_truncation.py ccproxy/llms/utils/token_estimation.py ccproxy/plugins/claude_api/adapter.py ccproxy/plugins/claude_sdk/adapter.py ccproxy/streaming/deferred.py tests/test_tool_result_sanitization.py tests/plugins/claude_api/unit/test_native_anthropic_sanitization.py tests/unit/llms/test_context_window_management.py tests/unit/llms/test_truncation_creates_orphans.py tests/unit/streaming/test_deferred_stream_errors.py
• uv run mypy ccproxy/llms/formatters/openai_to_anthropic/requests.py ccproxy/llms/utils/__init__.py ccproxy/llms/utils/context_truncation.py ccproxy/llms/utils/token_estimation.py ccproxy/plugins/claude_api/adapter.py ccproxy/plugins/claude_sdk/adapter.py ccproxy/streaming/deferred.py
• uv run python scripts/check_import_boundaries.py
• uv run pytest tests/test_tool_result_sanitization.py tests/plugins/claude_api/unit/test_native_anthropic_sanitization.py tests/unit/llms/test_context_window_management.py tests/unit/llms/test_truncation_creates_orphans.py tests/unit/streaming/test_deferred_stream_errors.py (69 passed)
• uv run pytest --ignore=tests/plugins/copilot/integration -k 'not test_openai_chat_stream_to_anthropic_sample' (1249 passed, 6 skipped, 60 deselected, 1 warning)