fix: fail on truncated model tool calls

[Geonwoo Jeong (geonwoo-jeong)]

Fixes #458

Summary

• Fail the run when a model reaches its output limit before completing a tool call.
• Correlate the invalid_tool_call and length signals by namespace, node, and model run ID.
• Preserve existing behavior for valid tool calls, non-length finishes, and unrelated concurrent or subgraph runs.

Reproduction and root cause

During a long Bedrock agent run, stock @langchain/aws@1.4.2 emitted:

1. content-block-finish with content.type = "invalid_tool_call"
2. message-finish with reason = "length"

The affected route reported output_tokens = 4096; this PR does not assume that value is a universal Bedrock default.

OpenWiki currently ignores both non-text message events. The incomplete tool call is therefore not executed, but the outer stream can finish without surfacing why the requested file change was missing.

Exact behavior

Stream signals	Result
Same message: invalid_tool_call then length	Throw an actionable error
invalid_tool_call then another finish reason	Preserve existing behavior
length without an invalid tool call	Preserve existing behavior
Signals from different runs, namespaces, or nodes	Do not combine them

The error is raised inside the existing stream try block, so temporary-plan cleanup and interrupted-run metadata handling remain unchanged. Tool arguments and generated file contents are not included in the error.

Scope

This is a provider-neutral guard at the LangChain v3 protocol boundary. It does not change model defaults, expose a maxTokens setting, patch dependencies, or alter prompts and workflows.

It is also independent of langchain-ai/langchainjs#11231, which addresses replay of unstamped v1 tool-call history rather than truncation of the current streamed response.

Validation

• Regression tests confirmed RED without the guard and GREEN with it, including namespace/node isolation.
• An integration regression test exercises the real createDeepAgent(...).streamEvents(..., { version: "v3" }) path and verifies run_id correlation plus message-finish.reason = "length".
• pnpm install --frozen-lockfile
• pnpm test — 58 files, 697 tests
• pnpm run build
• pnpm run typecheck
• pnpm run lint:check
• pnpm run format:check
• Manual stock-@langchain/aws@1.4.2 stream reproduction confirmed the invalid_tool_call → length sequence.