Rebuild provider-executed Anthropic blocks for a restored session

[shibayan]

What this changes

A session restored from persistence can replay a turn that used the code-execution beta: when rawRepresentation is gone — serialization deliberately strips it — the Anthropic conversion rebuilds server_tool_use and the code-execution result blocks from the typed contents, so calls stay paired with their results instead of dropping out and drawing a 400. Closes #126.

Parity

• Reference checked: the block shapes follow the @anthropic-ai/sdk request param schemas (v0.122: BetaServerToolUseBlockParam, BetaCodeExecutionToolResultBlockParam, BetaBashCodeExecution…, BetaTextEditorCodeExecution…), and the receive-side mapping being inverted is this package's own (from-anthropic.ts), which itself matches Python's _parse_tool_result_from_mcp-era handling. Python does not rebuild these blocks — its transcripts hit the same loss — so the reconstruction rules are this repository's own, pinned by round-trip equivalence: for every fixture (code success and error, bash success and timeout, text-editor view/str_replace/create/error), receive → serialize → restore → send is asserted byte-identical to the in-memory raw replay.
• Wire format affected: yes
• Public API affected: no
• Breaking change: no

The server_tool_use family is inferred from the kind of content answering the call, since the typed code_interpreter_tool_call folds all three code-execution tools into one kind; bash files are gathered from the hosted_file siblings the receive side places before the shell result; return_code and file_type are not in the typed model and default to 0 / 'text'. Error outputs parsed from code-execution and text-editor failures now also carry the wire's code under ErrorContent.errorCode (additive; both layers proven individually necessary by negative control). A transcript still holding rawRepresentation replays the exact bytes — a raw block with an unmodelled extra field is pinned to survive — and rawRepresentation is still never serialized. A function_result answering a local call keeps the local tool_result mapping after a restore. Restoring web_search/web_fetch turns stays out of scope, per the issue.

Checklist

• pnpm check passes (lint, typecheck, build, test)
• Behaviour changes are covered by a test that fails without the change
• Public API changes are reflected in the package README and CHANGELOG.md

🤖 Generated with Claude Code

[Copilot]

🟡 Changes recommended

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

This PR fixes Anthropic “provider-executed” (code execution / bash / text editor) turn replay after restoring a persisted session by rebuilding the missing Messages API blocks when rawRepresentation has been stripped during serialization.

Changes:

• Reconstruct server_tool_use + provider tool-result blocks from typed Content when rawRepresentation is absent (restored sessions).
• Preserve error codes through persistence by additionally storing Anthropic tool error codes in ErrorContent.errorCode.
• Add replay/round-trip tests to assert restored replay matches in-memory replay for provider-executed turns, and document the fix in the changelog.

File summaries

File	Description
packages/anthropic/src/to-anthropic.ts	Adds restored-session reconstruction logic for provider-executed tool call/result blocks.
packages/anthropic/src/from-anthropic.ts	Captures tool error codes into ErrorContent.errorCode to enable faithful rebuild after persistence.
packages/anthropic/src/code-execution.test.ts	Updates expectations for new errorCode field on error outputs.
packages/anthropic/src/code-execution-replay.test.ts	New fixtures asserting restored replay is byte-identical to in-memory replay for provider-executed turns.
CHANGELOG.md	Documents restored-session replay behavior fix for the Anthropic package.

Review details

Suppressed comments (2)

packages/anthropic/src/to-anthropic.ts:407

• These restored-session cases should not emit blocks with missing/empty ids (callId is optional on hosted tool contents). As written, callId ?? '' can produce id: '' / tool_use_id: '', which is not pairable and can cause Messages API rejection.

    case 'code_interpreter_tool_call': {
      // Reached only without a raw block — a restored transcript — since the raw path above
      // replays the exact bytes otherwise. The same holds for the result cases below.
      blocks.push(rebuiltServerToolUse(content, ctx.families.get(content.callId ?? '') ?? 'code_execution'));
      return;

packages/anthropic/src/to-anthropic.ts:421

• For bash replays, malformed file entries are preserved by the receive side as unknown content items (unknownType bash_code_execution_output) placed immediately before the shell_tool_result. The rebuild currently only gathers hosted_file siblings, so restored replays drop those entries from the rebuilt payload and may incorrectly emit them as standalone unknown blocks.

    case 'shell_tool_result': {
      // The files a bash run produced sit as `hosted_file` siblings *before* the shell result —
      // that is where the receive side put them — so the rebuilt payload gathers the contiguous
      // run immediately preceding this content.
      const files: Content[] = [];

• Files reviewed: 5/5 changed files
• Comments generated: 3
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

🟡 Changes recommended

The restored-session reconstruction path can mis-rebuild code_execution_tool_result_error into a success-shaped code_execution_result when ErrorContent.errorCode is present but not in the known enum, which changes the payload variant/semantics instead of degrading to a valid error payload.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

• Files reviewed: 5/5 changed files
• Comments generated: 1
• Review effort level: Lite

[Copilot]

🟡 Changes recommended

codeExecutionFamilies currently correlates missing callId values via '', which can mis-infer tool families for keyless hosted-tool contents and produce mismatched rebuilt server_tool_use.name in degenerate transcripts.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Review details

• Files reviewed: 5/5 changed files
• Comments generated: 1
• Review effort level: Lite

[Copilot]

🟢 Approval recommended

The rebuild behavior is well-scoped, preserves raw replay when present, and is backed by comprehensive round-trip equivalence tests covering key provider-executed variants and edge cases.

Review details

• Files reviewed: 5/5 changed files
• Comments generated: 0 new
• Review effort level: Lite

[Copilot]

🟢 Approval recommended

The reconstruction behavior is well-scoped, preserves exact raw replay when available, and is backed by byte-equivalence replay tests covering the restored-session path.

Review details

• Files reviewed: 5/5 changed files
• Comments generated: 0 new
• Review effort level: Lite