fix(sessions): truncate at incomplete tool calls instead of synthetic results by kiranjd · Pull Request #3565 · openclaw/openclaw

kiranjd · 2026-01-28T20:46:28Z

Summary

When session history contains an assistant message with tool calls but missing results (due to branching, interruption, or crash), the API rejects the request with errors like:

"an assistant message with 'tool_calls' must be followed by tool messages responding to each 'tool_call_id'"

This change makes agents recover automatically instead of getting stuck.

What users experience

Before (broken)

User messages agent
Agent is stuck, returns errors, won't respond
Only fix: manually delete entire conversation history
User loses everything

After (fixed)

User messages agent
Agent responds normally
Agent may have forgotten one thing from earlier (the broken part)
User asks again if needed, conversation continues

Technical details

Root cause

Session files use a tree structure with branching. When linearizing the tree for API calls:

Tool calls and their results can end up on different branches
Linearization picks one path, potentially leaving tool results orphaned
API rejects the malformed history

Previous approach (fragile)

Scan transcript, try to re-associate orphaned tool results
Insert synthetic error results for missing tool calls
Drop orphaned results

Problems:

Complex repair logic with many edge cases
Synthetic error results confuse the conversation
Still failed in some scenarios

New approach (simple)

Detect the first assistant message with incomplete tool calls
Truncate history before that message
Always produces valid history (just shorter)
Log truncation details for debugging

Code changes

session-transcript-repair.ts: New findFirstIncompleteToolCallIndex() function, truncation logic
google.ts: Log when truncation occurs for monitoring
Tests updated to verify new behavior

Alternatives considered

Approach	Why rejected
Insert synthetic error results	Confusing conversation state, fragile
Show error, ask user to fix	Users shouldn't need to understand session internals
Delete entire conversation	Too much data loss
Prevent all interruptions	Impossible (network issues, crashes, user edits)

Testing

Unit tests updated and passing
Lint passing
Build passing

🤖 Generated with Claude Code

Greptile Overview

Greptile Summary

This PR changes session transcript recovery to avoid provider rejections when an assistant tool-call turn lacks matching tool results. Instead of inserting synthetic error tool results, repairToolUseResultPairing now detects the first assistant message with missing tool results and truncates history before it, then reorders any in-span tool results and drops duplicates/orphans. sanitizeSessionHistory is updated to use the new repair report and log when truncation occurs, and tests were updated to cover the new truncation behavior.

Confidence Score: 3/5

Mostly safe to merge, but the new truncation detection may miss the branch/orphan scenario it is meant to fix.
Core change is localized and tests cover the new truncation behavior, but findFirstIncompleteToolCallIndex treats any matching toolResult anywhere in the transcript as satisfying a tool call, which can prevent truncation when results exist off-path (the described root cause). Logging change is low risk.
src/agents/session-transcript-repair.ts

Context used:

Context from dashboard - CLAUDE.md (source)
Context from dashboard - AGENTS.md (source)

… synthetic results When session history contains an assistant message with tool calls but missing results (e.g., due to branching, interruption, or crash), the API rejects the request. Previously, we tried to repair this by inserting synthetic error results, but this approach was fragile and could lead to confusing conversation states. New approach: truncate history before the first incomplete tool call sequence. This is simpler and always produces valid history: - Agent continues working (just with less context) - No confusing synthetic error results - User may need to repeat their last request The truncation is logged with details about which tool call IDs were missing, helping track and diagnose the root cause. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

greptile-apps

_{2 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

greptile-apps · 2026-02-03T03:54:56Z

Additional Comments (2)

src/agents/session-transcript-repair.ts
[P1] Truncation check can false-positive due to global toolResult scan

findFirstIncompleteToolCallIndex precomputes allResultIds by scanning all tool results anywhere in the transcript (indexAllToolResultIds). That means an assistant tool call will be considered “complete” if a matching toolResult exists later on a different branch/turn, even if it won’t be included after linearization (the exact failure mode described in the PR). In that scenario, this function won’t truncate and the provider will still reject the request.

This is especially likely when the missing results are actually “orphaned on another branch” rather than truly absent. Consider checking for tool results only within the linearized path/after the call turn (or the same span the reordering logic already scans) instead of globally.

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/agents/session-transcript-repair.ts
Line: 80:83

Comment:
[P1] Truncation check can false-positive due to global toolResult scan

`findFirstIncompleteToolCallIndex` precomputes `allResultIds` by scanning *all* tool results anywhere in the transcript (`indexAllToolResultIds`). That means an assistant tool call will be considered “complete” if a matching `toolResult` exists later on a different branch/turn, even if it won’t be included after linearization (the exact failure mode described in the PR). In that scenario, this function won’t truncate and the provider will still reject the request.

This is especially likely when the missing results are actually “orphaned on another branch” rather than truly absent. Consider checking for tool results only within the linearized path/after the call turn (or the same span the reordering logic already scans) instead of globally.


How can I resolve this? If you propose a fix, please make it concise.

src/agents/pi-embedded-runner/google.ts
[P2] Log may include untrusted tool call IDs and get very large

The truncation warning logs missingToolCallIds.join(", ") inline. If tool call IDs can be influenced by user/session contents (or just be numerous), this can produce noisy/oversized logs and make parsing harder. Consider logging them as structured data (array field) and/or truncating the list/count (e.g., first N ids + total) to keep logs bounded.

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/agents/pi-embedded-runner/google.ts
Line: 341:348

Comment:
[P2] Log may include untrusted tool call IDs and get very large

The truncation warning logs `missingToolCallIds.join(", ")` inline. If tool call IDs can be influenced by user/session contents (or just be numerous), this can produce noisy/oversized logs and make parsing harder. Consider logging them as structured data (array field) and/or truncating the list/count (e.g., first N ids + total) to keep logs bounded.


How can I resolve this? If you propose a fix, please make it concise.

openclaw-barnacle bot added the agents Agent runtime and tooling label Jan 28, 2026

This was referenced Feb 2, 2026

fix(session): strip malformed tool_use blocks to prevent session corruption #5557

Closed

fix(agents): skip extracting tool calls from errored assistant turns #1859

Closed

greptile-apps bot reviewed Feb 3, 2026

View reviewed changes

thewilloftheshadow force-pushed the main branch from bfc1ccb to f92900f Compare February 15, 2026 18:46

kiranjd closed this Feb 15, 2026

kiranjd deleted the fix/session-truncate-incomplete-tool-calls branch February 15, 2026 23:04

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

fix(sessions): truncate at incomplete tool calls instead of synthetic results#3565

fix(sessions): truncate at incomplete tool calls instead of synthetic results#3565
kiranjd wants to merge 1 commit intoopenclaw:mainfrom
kiranjd:fix/session-truncate-incomplete-tool-calls

kiranjd commented Jan 28, 2026 •

edited by greptile-apps bot

Loading

Uh oh!

greptile-apps bot left a comment

Uh oh!

greptile-apps bot commented Feb 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

kiranjd commented Jan 28, 2026 • edited by greptile-apps bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

What users experience

Before (broken)

After (fixed)

Technical details

Root cause

Previous approach (fragile)

New approach (simple)

Code changes

Alternatives considered

Testing

Greptile Overview

Greptile Summary

Confidence Score: 3/5

Uh oh!

greptile-apps bot left a comment

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot commented Feb 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

kiranjd commented Jan 28, 2026 •

edited by greptile-apps bot

Loading