fix anthropic PTC

[j4ys0n]

No description provided.

[j4ys0n]

Automated review 🤖

Summary of Changes
This PR fixes Anthropic programmatic tool calling (PTC) behavior by adjusting how code_execution_tool_result blocks are handled during message replay and result mapping. Key changes include: (1) introducing CodeExecutionResultInfo as a shared type for code execution diagnostics, (2) filtering out code_execution_tool_result blocks from assistant message replay when resuming pending programmatic tool calls, and (3) surfacing codeExecutionResults in both streaming and non-streaming GenerateResult. Version bumped to 1.11.1, and a new debug example (examples/anthropic-ptc-debug.ts) is added to validate PTC flows.

Key Changes & Positives

• Introduces CodeExecutionResultInfo type in src/types/common.types.ts to unify code execution diagnostics across providers 🟢
• Adds filterAnthropicReplayContentBlocksForProgrammaticToolResultResume to exclude code_execution_tool_result blocks during PTC resume, preventing duplicate execution attempts in src/core/mapping/anthropic.mapper.ts (lines 182–197) 🟢
• Surfaces codeExecutionResults in GenerateResult (non-streaming and streaming) with proper aggregation in mapFromProviderResponse and mapProviderStream (e.g., src/core/mapping/anthropic.mapper.ts, lines 681, 837–840) 🟢

Potential Issues & Recommendations

1. Issue / Risk: The isProgrammaticToolResultResume method checks only the immediate next message (messages[assistantIndex + 1]) for role 'tool', which may miss cases where tool messages are interleaved or delayed.
   Impact: Could incorrectly include/exclude code_execution_tool_result blocks, leading to malformed replay or missed tool results.
   Recommendation: Add unit tests covering multi-tool-call sequences and verify behavior when tool messages appear after multiple assistant turns.
   Status: 🟡 Needs review

2. Issue / Risk: In mapCodeExecutionResultData, the fallback case returns returnCode: 1 and errorCode without populating stdout/stderr, but the original implementation defaulted to empty strings for all fields—this may alter downstream error handling expectations.
   Impact: Consumers expecting empty stdout/stderr on error may need updates.
   Recommendation: Document error-case semantics explicitly in CodeExecutionResultInfo JSDoc and ensure tests cover all three content types (code_execution_result, encrypted_code_execution_result, error fallback).
   Status: 🟡 Needs review

Language/Framework Checks

• CodeExecutionResultInfo correctly imported into result.types.ts and stream.types.ts (lines 2, 3 in each)
• Zod schema usage preserved in examples/anthropic-ptc-debug.ts (z.object({})) for tool parameters
• TypeScript strictness maintained: CodeExecutionResultInfo fields are properly typed as optional (encryptedStdout?, errorCode?, contentFileIds?)

Security & Privacy

• No sensitive data exposure introduced; debug example uses mock data and writes artifacts to configurable outputRoot (examples/anthropic-ptc-debug.ts, lines 56–59)
• cloneJson utility used consistently for deep copying messages and artifacts, avoiding accidental mutation

Build/CI & Ops

• New example script added to package.json scripts (example:anthropic-ptc-debug)—ensure CI runs it or equivalent smoke test
• Version bump to 1.11.1 requires release notes highlighting PTC fix and codeExecutionResults field

Tests

• Unit tests added for code_execution_tool_result exclusion during replay (tests/unit/core/mapping/anthropic.mapper.spec.ts, lines 268–342)
• Tests added for non-streaming and streaming codeExecutionResults surfaced in results (lines 803–820, 1234–1275)
• Coverage need: Add integration test verifying end-to-end PTC flow with code execution (e.g., getCoreConfig → listProviders → summary) using the debug harness pattern

[Copilot]

Pull request overview

This PR addresses Anthropic programmatic tool calling (PTC) behavior around code execution blocks, ensuring replay/resume flows don’t resend code_execution_tool_result blocks while also surfacing code execution diagnostics consistently in both streaming and non-streaming results.

Changes:

• Filter code_execution_tool_result blocks out of assistant replay content when resuming programmatic tool results.
• Surface code execution diagnostics via a shared CodeExecutionResultInfo type on both GenerateResult and stream chunks.
• Add a debug example harness for Anthropic PTC and bump package version to 1.11.1.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
tests/unit/core/mapping/anthropic.mapper.spec.ts	Adds coverage for filtering replay blocks and for surfacing code execution diagnostics in streaming/non-streaming mappings.
src/types/stream.types.ts	Reuses the shared CodeExecutionResultInfo shape for code_execution_result stream chunks.
src/types/result.types.ts	Adds codeExecutionResults?: CodeExecutionResultInfo[] to GenerateResult.
src/types/common.types.ts	Introduces CodeExecutionResultInfo as the provider-agnostic diagnostics payload.
src/core/mapping/anthropic.mapper.ts	Implements replay filtering logic and collects code execution diagnostics into results/chunks.
package.json	Bumps version and adds an example script entry.
examples/anthropic-ptc-debug.ts	Adds a runnable debug harness for investigating Anthropic PTC streaming behavior.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review. Take the survey.