fix(core): keep deferred provider calls only on the last surviving assistant turn

[roaminro]

Fixes #15668 (similar class of bug to #14148).

When a provider drops a tool-result chunk (Vertex/Gemini code_execution) or a run aborts mid-stream after a tool-call (Anthropic web_search), the unresolved provider-executed call gets persisted in input-available state. sanitizeV5UIMessages was unconditionally keeping every input-available + providerExecuted part to protect Anthropic's legitimate N→N+1 deferral — which also meant orphans were replayed to the model forever, returning empty text on Gemini and violating the tool-call/tool-result invariant on Anthropic.

The fix tightens the keep-rule: input-available provider-executed parts are only preserved on the last surviving assistant message after sanitization, and only when no later user message has closed the turn. Anywhere else, they're treated as orphans and dropped.

Conceptually:

const lastSurvivingAssistantIdx = findLastAssistantAfterSanitization();
const assistantTurnStillOpen = idx === lastSurvivingAssistantIdx && idx > lastUserIdx;

if (p.state === 'input-available' && p.providerExecuted && assistantTurnStillOpen) {
  return true;
}

Regression tests cover the Gemini code_execution orphan (#15668), the Anthropic web_search orphan (#14148 class), a multi-assistant history where a stale provider call sits on an earlier assistant turn behind a later assistant message, and the case where a trailing assistant sanitizes away so the previous surviving assistant still keeps the legitimate deferred provider call. The existing deferred provider-executed coverage still passes unchanged.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
mastra-docs-1.x	Skipped		Apr 23, 2026 7:52pm

[changeset-bot]

🦋 Changeset detected

Latest commit: d44668f

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 21 packages

Name	Type
@mastra/core	Patch
mastracode	Patch
@mastra/mcp-docs-server	Patch
@internal/playground	Patch
@mastra/client-js	Patch
@mastra/opencode	Patch
@mastra/longmemeval	Patch
mastra	Patch
@mastra/deployer-cloud	Patch
@mastra/deployer-vercel	Patch
@mastra/playground-ui	Patch
@mastra/react	Patch
@mastra/server	Patch
@mastra/deployer	Patch
create-mastra	Patch
@mastra/express	Patch
@mastra/fastify	Patch
@mastra/hono	Patch
@mastra/koa	Patch
@mastra/deployer-cloudflare	Patch
@mastra/deployer-netlify	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[coderabbitai]

Walkthrough

Sanitization of V5 UI messages now prevents orphaned provider-executed tool parts from persisting across turns: input-available provider-executed tool parts are retained only on the most recent assistant message when no subsequent user turn exists; earlier orphaned tool parts are dropped.

Changes

Cohort / File(s)	Summary
Changeset note \.changeset/puny-rats-talk.md	Adds a patch Changeset documenting the bug fix for orphaned provider-executed tool calls causing threads to become unusable.
Tests packages/core/src/agent/message-list/conversion/output-converter-provider-executed.test.ts	Adds regression tests for #15668: builds multi-turn histories with provider-executed tool orphans (Gemini code_execution, Anthropic web_search, stale deferred calls) and asserts converted messages contain no unpaired tool-call parts; includes a preservation case for a deferred tool on the last assistant.
Core sanitization logic packages/core/src/agent/message-list/conversion/output-converter.ts	Updates sanitizeV5UIMessages: precomputes last user and surviving assistant indices, introduces getSafeParts, and restricts retention of provider-executed input-available parts to the most-recent assistant without a following user, dropping earlier orphaned parts.

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~45 minutes

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	The code changes directly address the objectives from #15668: preventing persistence and replay of orphaned provider-executed tool calls by sanitizing orphaned parts on closed assistant turns.
Out of Scope Changes check	✅ Passed	All changes are scoped to fixing orphaned provider-executed tool calls in message conversion and sanitization, with comprehensive test coverage for related issues.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly summarizes the main bug fix: constraining provider-executed tool calls to only the last surviving assistant turn to prevent orphaned calls from being replayed.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ 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 roamin/github-issue-analysis-request

---

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

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@packages/core/src/agent/message-list/conversion/output-converter.ts`:
- Around line 111-136: Compute the "open assistant turn" against the
post-sanitization view instead of the raw history: add a first pass that
iterates messages from the end to find lastSurvivingAssistantIdx (skip assistant
messages that would sanitize to zero parts using the same logic you apply later,
e.g., check m.parts and any sanitizer conditions) and use that index in place of
lastAssistantIdx when computing assistantTurnStillOpen inside the msgs.map;
update the variable name (or set lastAssistantIdx = lastSurvivingAssistantIdx)
so deferred providerExecuted/input-available parts on an earlier surviving
assistant are preserved when a trailing assistant would have been dropped.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 2f6be80c-3d5e-4575-8a7f-4828c13a352f

📥 Commits

Reviewing files that changed from the base of the PR and between 3497c1d and b4857f5.

📒 Files selected for processing (3)

• .changeset/puny-rats-talk.md
• packages/core/src/agent/message-list/conversion/output-converter-provider-executed.test.ts
• packages/core/src/agent/message-list/conversion/output-converter.ts