feat: rewind-on-interrupt — clean conversation state on abort/resume

[hannesrudolph]

Summary

When a user interrupts a task (abort or close VS Code) and later resumes it, the conversation history could contain corrupted state: orphaned tool_use blocks without matching tool_results, synthetic "interrupted" messages that confused the LLM, and attempt_completion calls that never received a response. This PR introduces a rewind mechanism that truncates conversation state to the last valid checkpoint on both abort and resume, ensuring the LLM always sees a clean, API-compliant conversation.

Changes

Core Implementation

• rewindToLastAskPoint() in Task.ts — truncates clineMessages to the last qualifying ask point, then sanitizes trailing API history (removes orphaned assistant tool_use blocks, incomplete user tool_result messages)
• Two-phase completion flow in AttemptCompletionTool.ts — completion_result ask followed by resume_completed_task ask holds execution context alive, ensuring attempt_completion always gets a tool_result
• getPendingAttemptCompletionToolUseId() in Task.ts — detects pending attempt_completion tool_use for routing follow-up text as tool_result on resume
• Neutral fallback message in validateToolResultIds.ts — "No result available for this tool call." replaces the old "Tool execution was interrupted" synthetic message

Code Quality

• Extracted normalizeContentBlocks() to src/shared/messages.ts — replaces 5 inline Array.isArray(content) patterns
• Type predicates replace 3 post-hoc as casts on Array.filter()/Array.find() results
• satisfies checks replace 2 as Anthropic.ToolResultBlockParam casts on conforming object literals
• Eliminated ! non-null assertions — direct truthiness checks on pendingAttemptCompletionToolUseId let TypeScript narrow naturally

Tests (65 total)

• 12 unit tests for rewindToLastAskPoint() covering truncation, ask-type filtering, API history cleanup, error handling, idempotency
• 7 integration tests including completion→restart→resume with/without follow-up, double rewind idempotency, multi-tool_use edge case
• 15 completion flow tests for AttemptCompletionTool covering accept, feedback, and abort paths
• 5 utility tests for normalizeContentBlocks()
• Persistence assertions verify overwriteClineMessages/overwriteApiConversationHistory receive correct truncated data
• Shared test fixtures in rewind-test-setup.ts reduce ~110 lines of duplicated mock setup

---

Important

Introduces a rewind mechanism to clean conversation state on task abort/resume, ensuring consistent conversation history by truncating to the last valid checkpoint.

• Behavior:
  • rewindToLastAskPoint() in Task.ts truncates clineMessages to the last valid ask point, sanitizing API history by removing orphaned tool_use blocks and incomplete tool_result messages.
  • Two-phase completion flow in AttemptCompletionTool.ts ensures attempt_completion always gets a tool_result by asking completion_result followed by resume_completed_task.
  • getPendingAttemptCompletionToolUseId() in Task.ts detects pending attempt_completion tool_use for routing follow-up text as tool_result on resume.
  • Neutral fallback message in validateToolResultIds.ts replaces synthetic "interrupted" messages with "No result available for this tool call."
• Code Quality:
  • Extracted normalizeContentBlocks() to messages.ts to replace inline Array.isArray(content) patterns.
  • Type predicates replace post-hoc as casts on Array.filter()/Array.find() results.
  • satisfies checks replace as Anthropic.ToolResultBlockParam casts on conforming object literals.
  • Eliminated ! non-null assertions by using direct truthiness checks.
• Tests (65 total):
  • 12 unit tests for rewindToLastAskPoint() covering truncation, ask-type filtering, API history cleanup, error handling, idempotency.
  • 7 integration tests including completion→restart→resume scenarios, double rewind idempotency, multi-tool_use edge case.
  • 15 completion flow tests for AttemptCompletionTool covering accept, feedback, and abort paths.
  • 5 utility tests for normalizeContentBlocks().
  • Persistence assertions verify overwriteClineMessages/overwriteApiConversationHistory receive correct truncated data.
  • Shared test fixtures in rewind-test-setup.ts reduce ~110 lines of duplicated mock setup.

^{This description was created by for 1f04bab. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Oroocle   See task

All previously flagged issues have been addressed. No new issues found in the latest changes.

• getPendingAttemptCompletionToolUseId should guard against multi-tool_use assistant messages where other tool_uses also lack tool_results, to avoid bypassing the rewind when the task wasn't fully completed

Previous reviews

• db8f472: Review #1

_{Mention @roomote in a comment to request specific changes to this pull request or fix all unresolved issues.}

[roomote]

If the last assistant message contains attempt_completion alongside other tool_use blocks (e.g., read_file + attempt_completion) and the process crashes before any tool_result is saved, this method returns the attempt_completion ID, which causes resumeTaskFromHistory to bypass the rewind and treat the task as completed. The other tool_uses never executed, so validateAndFixToolResultIds patches in synthetic "No result available" fallbacks, producing misleading context for the model. This is unlikely in practice (models rarely combine attempt_completion with other tools), but the guard is straightforward:

Suggested change

	const toolUse = content.find(
	(block): block is Anthropic.Messages.ToolUseBlockParam =>
	block.type === "tool_use" && block.name === "attempt_completion",
	)
	return toolUse?.id
	const toolUse = content.find(
	(block): block is Anthropic.Messages.ToolUseBlockParam =>
	block.type === "tool_use" && block.name === "attempt_completion",
	)
	// Only treat as pending completion if there are no other unanswered
	// tool_use blocks — otherwise the rewind path is more appropriate.
	if (toolUse) {
	const hasOtherToolUse = content.some(
	(block) => block.type === "tool_use" && block !== toolUse,
	)
	if (hasOtherToolUse) return undefined
	}
	return toolUse?.id

_{Fix it with Roo Code or mention @roomote and request a fix.}