Summary

When a tool result is never persisted to message storage (e.g., due to a crash, abort, or compaction gap), the appendMissingToolResults function in transport-message-transform.ts synthesizes a placeholder with isError: true during history replay. This causes the model to treat the tool call as failed, when the actual state is unknown/ambiguous — the result simply wasn't recorded.

Root Cause

appendMissingToolResults (currently in anthropic-vertex-stream-YyrYtvts.js) fires during message history reconstruction. It walks the stored message sequence and plugs any tool call that has no corresponding toolResult record. Current behavior:

result.push({
  role: "toolResult",
  toolCallId: toolCall.id,
  toolName: toolCall.name,
  content: [{ type: "text", text: "No result provided" }],
  isError: true,  // ← wrong: treats unknown as failure
  timestamp: Date.now()
});

The exec tool itself always emits a proper result ("(no output)" + exit code) on success. The gap occurs upstream — the result record was never written to storage. The model did receive the result in-context during the original turn, but on subsequent turns the replayed history shows a failure, causing the model to:

1. Conclude the previous tool call failed
2. Retry unnecessarily
3. Build follow-up reasoning on a false premise

Reproduction

1. Run any exec command in a turn that gets aborted or crashes mid-session-write
2. On the next turn, the history replay synthesizes isError: true for that tool call
3. Model treats the successful (but unrecorded) exec as a failure

Proposed Fix

Change the synthesized placeholder to signal ambiguity rather than failure:

result.push({
  role: "toolResult",
  toolCallId: toolCall.id,
  toolName: toolCall.name,
  content: [{ 
    type: "text", 
    text: "Tool result not recorded — status unknown. Verify state before retrying." 
  }],
  isError: false,  // ambiguous, not failed
  timestamp: Date.now()
});

This forces the model to verify state rather than assume failure, which is the correct behavior when the actual outcome is unknown.

Additional Recommendation

Add a health check to doctor --fix that detects tool calls with synthesized missing results and surfaces them as warnings, so users can identify sessions where result persistence failed.

Environment

• OpenClaw: 2026.4.14
• Affected file: dist/anthropic-vertex-stream-YyrYtvts.js line ~4532
• Trigger: message history replay when toolResult records missing from storage