fix(agent): recover truncated tool calls

[yyhhyyyyyy]

Summary

• Atomically reject DeepChat-owned tool calls when provider output is truncated by max_tokens
• Generate matching error tool results without executing potentially incomplete arguments
• Retry the provider once after settling the rejected batch, then terminate safely if truncation repeats
• Track DeepChat-owned and provider-owned tool calls to prevent local execution of provider-managed tools
• Preserve tool-call protocol integrity across persistence, context rebuilding, and Tape facts

Implementation

• Add optional per-call execution ownership with backward-compatible DeepChat ownership by default
• Map AI SDK providerExecuted calls and ACP tool calls to provider ownership
• Extend tool batch settlement with explicit execute and truncated-rejection dispositions
• Collect completed and pending DeepChat-owned calls in source order for atomic rejection
• Mark rejected and structurally incomplete blocks with distinct max_tokens reasons
• Keep rejected calls outside tool execution limits, metadata accounting, permissions, hooks, and no-progress detection
• Scope block updates to the current provider round when call IDs are reused
• Preserve existing normal tool_use, abort, pending-input, output fitting, and provider-round-limit behavior

Safety

Truncated tool arguments are never executed, even when some calls appear complete. The rejection path does not invoke permission prechecks, auto-approval reviewers, tool executors, question tools, or skill activation.

Automatic recovery is limited to one additional provider request per run.

Summary by CodeRabbit

• New Features

  • Added recovery for tool calls interrupted by output-length limits.
  • Preserves incomplete tool-call details and records clear skipped or incomplete statuses.
  • Distinguishes tool calls already handled by the provider from those requiring local execution.
  • Maintains tool-call ordering and context across recovery attempts.

• Bug Fixes

  • Prevents truncated tool calls from remaining unresolved or being executed with incomplete arguments.
  • Improves tool-result persistence and status reporting after interrupted responses.

• Documentation

  • Added a specification describing truncated tool-call recovery behavior and acceptance criteria.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 96326198-ff24-426d-8b2f-3cd8f3305d1a

📥 Commits

Reviewing files that changed from the base of the PR and between 38e8d66 and 71bc743.

📒 Files selected for processing (22)

• docs/issues/truncated-tool-call-recovery/spec.md
• src/main/agent/acp/runtime/acpContentMapper.ts
• src/main/agent/deepchat/loop/deepChatLoopEngine.ts
• src/main/agent/deepchat/runtime/accumulator.ts
• src/main/agent/deepchat/runtime/dispatch.ts
• src/main/agent/deepchat/runtime/imageGenerationBlocks.ts
• src/main/agent/deepchat/runtime/process.ts
• src/main/agent/deepchat/runtime/types.ts
• src/main/provider/aiSdk/streamAdapter.ts
• src/shared/chat.d.ts
• src/shared/types/agent-interface.d.ts
• src/shared/types/core/chat.ts
• src/shared/types/core/llm-events.ts
• test/main/agent/acp/runtime/acpContentMapper.test.ts
• test/main/agent/deepchat/loop/deepChatLoopEngine.test.ts
• test/main/agent/deepchat/runtime/accumulator.test.ts
• test/main/agent/deepchat/runtime/contextBuilder.test.ts
• test/main/agent/deepchat/runtime/dispatch.test.ts
• test/main/agent/deepchat/runtime/process.test.ts
• test/main/evals/nativeAgent/nativeAgentBehavior.eval.test.ts
• test/main/evals/nativeAgent/scenarios.ts
• test/main/provider/aiSdkStreamAdapter.test.ts

---

📝 Walkthrough

Walkthrough

Adds execution-owner metadata for tool calls, refactors tool execution into settlement, and introduces bounded recovery for max_tokens-truncated DeepChat-owned calls with ordered error results and persistence coverage.

Changes

Truncated tool-call recovery

Layer / File(s)	Summary
Execution ownership contracts src/shared/types/core/llm-events.ts, src/main/provider/aiSdk/streamAdapter.ts, src/main/agent/acp/runtime/acpContentMapper.ts, src/main/agent/deepchat/runtime/accumulator.ts, src/shared/...	Tool-call start events identify provider-owned calls; accumulated state excludes provider-owned calls from local execution; assistant metadata includes skipped and incomplete reasons.
Tool-batch settlement src/main/agent/deepchat/runtime/dispatch.ts, src/main/agent/deepchat/runtime/imageGenerationBlocks.ts	settleToolBatch handles execute or reject dispositions, commits staged results centrally, and scopes block updates to the current batch.
Recovery loop control src/main/agent/deepchat/runtime/process.ts, src/main/agent/deepchat/loop/deepChatLoopEngine.ts	max_tokens rounds derive ordered DeepChat-owned truncated calls, reject them without execution, allow one recovery round, mark other unresolved blocks incomplete, and terminate after the recovery limit.
Recovery specification and validation docs/issues/truncated-tool-call-recovery/spec.md, test/main/agent/deepchat/..., test/main/provider/..., test/main/evals/nativeAgent/*	Specification, unit tests, context reconstruction tests, provider-ownership tests, ordering coverage, and a native-agent recovery scenario validate the new behavior.

Estimated code review effort: 4 (Complex) | ~60 minutes

Sequence Diagram(s)

sequenceDiagram
  participant Provider
  participant processStream
  participant settleToolBatch
  participant ContextBuilder
  Provider->>processStream: Emit truncated tool-call round
  processStream->>processStream: Select DeepChat-owned calls
  processStream->>settleToolBatch: Reject output_truncated batch
  settleToolBatch->>ContextBuilder: Commit ordered error results
  processStream->>Provider: Request one recovery round

Loading

Possibly related PRs

• ThinkInAIXYZ/deepchat#1959: Adds the native-agent evaluation harness and scenario infrastructure used by the new recovery scenario.

Suggested reviewers: zerob13, zhangmo8

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 8.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly matches the main change: recovering truncated tool calls in the agent.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/truncated-tool-call-recovery

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}