fix(llm): recover from truncated tool-call arguments instead of failing the stream

[Robin1987China]

Created by a bot. Do not close or edit manually.

Issue for this PR

Fixes#36766

Note: #37220 (KlyneChrysler) targets dev, this PR targets the correct v2 branch for the 2.0-labeled issue. Both implement recovery at complementary layers — this PR at the LLM protocol/tool-stream layer, #37220 at the session runner layer.

Type of change

• Bug fix
• New feature
• Documentation update
• Refactoring
• Performance improvement
• Other (please describe): ______________________

What does this PR do?

When a provider (e.g., OpenAI native Responses) streams tool arguments with truncated JSON, the current parseToolInput() behavior produces an LLMError that terminates the entire execution stream. This PR replaces that fatal failure with a recoverable pattern: a placeholder tool-call event (with input: {}) followed by an error tool-result event. The model receives the error as a recoverable tool failure and can retry, instead of losing the entire execution.

This mirrors the existing AI SDK path, which already redirects malformed tool calls to the experimental_repairToolCall / "invalid" tool mechanism.

Changes

• packages/llm/src/protocols/utils/tool-stream.ts:
  • Added malformedRecovery() helper: produces a placeholder tool-call + error tool-result with diagnostic metadata (providerMetadata.malformedInput containing byte count and parse error).
  • Added catchMalformed() helper: uses Effect.catchIf to intercept LLMError with InvalidProviderOutput reason.
  • Modified finish(), finishWithInput(), and finishAll() to wrap toolCall() in catchMalformed(), so truncated inputs produce recovery events instead of stream failures.

How did you verify your code works?

• bun typecheck passes in packages/llm (no new errors)
• Unit tests (blocked by pre-existing effect v4 beta Schema.Defect runtime crash in test runner)
• End-to-end test with a provider that returns truncated tool arguments

Checklist

• I am mergable via the "Merge" button
• I have signed the CLA

[github-actions]

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Potential Duplicate PRs Found

Related PR #37220: fix(core): recover from truncated tool arguments with bounded retries

• fix(core): recover from truncated tool arguments with bounded retries #37220
• This appears to be directly related—both address recovery from truncated tool arguments, likely working on complementary layers (LLM vs core).

Historical PR #24289: fix: Repair truncated JSON tool inputs in LLM session

• fix: Repair truncated JSON tool inputs in LLM session #24289
• Older PR addressing similar JSON repair for truncated tool inputs; may provide context on prior approaches.

Related PR #26167: fix(session): retry empty stream truncations and discard partial parts

• fix(session): retry empty stream truncations and discard partial parts #26167
• Addresses stream truncation handling in sessions; may overlap with error recovery patterns.

The most likely duplicate is #37220, as both target truncated tool argument recovery. You should verify whether they're addressing the same issue or complementary parts of the fix.

[Robin1987China]

Closing this PR: the v2 branch has since rewritten packages/llm into packages/ai (this change targets the old package), the branch is hundreds of commits behind v2, and #37220 already covers truncated tool argument recovery with bounded retries on the new architecture. Thanks for considering it.

[Robin1987China]

Closing as superseded — see comment above.