fix(opencode): ensure tool-call/tool-result pairing after normalizeMessages by bvironn · Pull Request #22001 · anomalyco/opencode

bvironn · 2026-04-11T06:24:50Z

Issue for this PR

Type of change

Bug fix
New feature
Refactor / code improvement
Documentation

What does this PR do?

Adds ensureToolIntegrity() in the provider transform pipeline, running after normalizeMessages for Anthropic/Bedrock/Vertex-Anthropic. It detects orphaned tool-call parts (no matching tool-result) and injects synthetic error results to keep the session alive.

Why here? normalizeMessages can drop messages with empty content, breaking tool-call/tool-result pairs. By the time the request leaves ProviderTransform.message(), orphaned tool-calls cause a permanent 400 from Anthropic — the corrupted history replays from SQLite on every retry, making the session unrecoverable.

The function is a no-op when pairs are intact (single pass, Map+Set lookups, zero allocation on happy path). It catches corruption from 6 independent vectors in the pipeline — normalizeMessages filtering (V1), error-skip logic in toModelMessages (V2), lost step boundaries during retry (V3), tool-error race conditions (V4), filterCompacted cutting pairs (V5), and the AI SDK producing empty assistant messages (V6). Full analysis in #21326 comment.

How did you verify your code works?

130 existing transform tests pass (one updated for new pipeline ordering)
8 new unit tests: matched pairs (no-op), single/multiple orphans, partial orphans appended to existing tool messages, empty input, string content, full end-to-end through ProviderTransform.message()
bun typecheck clean
Manual reproduction: corrupted a session simulating V1+V5, confirmed the fix prevents the 400

Checklist

I have tested my changes locally
I have not included unrelated changes in this PR

github-actions · 2026-04-11T06:25:41Z

The following comment was made by an LLM, it may be inaccurate:

Great! I found several related PRs. Let me identify the most relevant ones:

Related PRs Found

PR fix(provider): skip empty-text filtering for assistant messages in normalizeMessages (#16748) #16750 - fix(provider): skip empty-text filtering for assistant messages in normalizeMessages (#16748)
- Related: Directly addresses empty-text filtering in normalizeMessages, which is the root cause issue being fixed in PR fix(opencode): ensure tool-call/tool-result pairing after normalizeMessages #22001
PR fix(session): fix root causes and reconstruction of tool_use/tool_result mismatch (#16749) #16751 - fix(session): fix root causes and reconstruction of tool_use/tool_result mismatch (#16749)
- Related: Addresses tool_use/tool_result mismatches, which is the exact problem that PR fix(opencode): ensure tool-call/tool-result pairing after normalizeMessages #22001 aims to prevent with the ensureToolIntegrity function
PR fix: blank assistant text - finish-reason regression after AI SDK v6 migration #20467 - fix: blank assistant text - finish-reason regression after AI SDK v6 migration
- Related: Deals with blank assistant text issues that can break message pairing logic
PR fix: prevent Kimi session poisoning from empty MCP tool-result text #14174 - fix: prevent Kimi session poisoning from empty MCP tool-result text
- Related: Addresses session corruption from empty tool-result text, similar problem domain
PR fix: handle dangling tool_use blocks for LiteLLM proxy compatibility #8497 - fix: handle dangling tool_use blocks for LiteLLM proxy compatibility
- Related: Handles orphaned tool_use blocks without matching results

These are not exact duplicates of PR #22001, but they address related problems in the tool-call/tool-result pairing system. PR #22001 appears to be a comprehensive fix that consolidates lessons from these previous issues.

github-actions · 2026-04-11T06:26:27Z

Thanks for updating your PR! It now meets our contributing guidelines. 👍

Copilot

Pull request overview

This PR adds a post-normalization safety check to prevent Anthropic-style tool-call history corruption by ensuring every tool-call has a corresponding tool-result, even when upstream transforms drop/alter messages.

Changes:

Added ProviderTransform.ensureToolIntegrity() to detect orphaned tool-calls and inject synthetic error tool-results.
Wired ensureToolIntegrity() into ProviderTransform.message() after normalizeMessages() for Anthropic/Bedrock SDKs.
Expanded/updated provider transform tests to cover multiple orphan scenarios and the full normalize→repair pipeline.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
packages/opencode/src/provider/transform.ts	Introduces `ensureToolIntegrity()` and applies it after `normalizeMessages()` for specific providers.
packages/opencode/test/provider/transform.test.ts	Updates an existing Anthropic normalization test expectation and adds a new test suite covering integrity repair cases.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-04-11T06:30:36Z

packages/opencode/src/provider/transform.ts

  export function message(msgs: ModelMessage[], model: Provider.Model, options: Record<string, unknown>) {
    msgs = unsupportedParts(msgs, model)
    msgs = normalizeMessages(msgs, model, options)
+    if (model.api.npm === "@ai-sdk/anthropic" || model.api.npm === "@ai-sdk/amazon-bedrock") {


The tool integrity repair is only enabled for models using @ai-sdk/anthropic and @ai-sdk/amazon-bedrock. This omits the Vertex Anthropic provider (@ai-sdk/google-vertex/anthropic / providerID google-vertex-anthropic), which is treated as Anthropic elsewhere in this module and likely has the same tool-call/tool-result pairing requirement. Consider including that provider in the condition so Vertex Anthropic sessions can’t be corrupted the same way.

Suggested change

if (model.api.npm === "@ai-sdk/anthropic" || model.api.npm === "@ai-sdk/amazon-bedrock") {

if (

model.api.npm === "@ai-sdk/anthropic" ||

model.api.npm === "@ai-sdk/amazon-bedrock" ||

model.providerID === "google-vertex-anthropic"

) {

Already covered — the condition at line 332 includes model.providerID === "google-vertex-anthropic" alongside the npm checks for @ai-sdk/anthropic and @ai-sdk/amazon-bedrock. This matches the same triple-check pattern used by applyCaching below (lines 336-347).

The npm package for Vertex Anthropic is @ai-sdk/google-vertex/anthropic (a subpath export), so checking model.api.npm alone would miss it — that's why the providerID fallback is there.

…ssages

bvironn · 2026-04-11T07:23:40Z

CI Note: E2E timeouts are unrelated to this change

Both e2e (linux) and e2e (windows) failed with runner timeouts (31m / 33m exceeding the 30-minute limit). Unit tests pass on both platforms.

This PR only modifies the message transform pipeline (transform.ts) — it doesn't touch session lifecycle, TUI rendering, or any code path exercised by the Playwright E2E suite. The function itself is a no-op when tool-call/tool-result pairs are already matched (which they are in normal operation).

Happy to re-trigger if needed, but this appears to be a runner infrastructure issue rather than a regression.

bvironn · 2026-04-11T17:36:10Z

Here's a corrupted session: https://gist.github.com/bvironn/b513c1eafae54291890355ddc7fd1941

Claude Opus 4.6, no interruption — the corruption happens on its own during normal tool execution. The agent made 5 tool calls across 2 steps (mem_search, glob, read), all completed successfully, but 2 tool_results were missing from the ModelMessage[] by the time the next request was sent.

messages.3: `tool_use` ids were found without `tool_result` blocks immediately after:
toolu_01APxrADs7VozN8uWzw9WwHr, toolu_01N8mDEzG8DSTs7UPHFtmgCT

Copilot AI review requested due to automatic review settings April 11, 2026 06:24

github-actions bot added the needs:compliance This means the issue will auto-close after 2 hours. label Apr 11, 2026

Copilot started reviewing on behalf of bvironn April 11, 2026 06:25 View session

github-actions bot removed the needs:compliance This means the issue will auto-close after 2 hours. label Apr 11, 2026

Copilot AI reviewed Apr 11, 2026

View reviewed changes

fix(opencode): ensure tool-call/tool-result pairing after normalizeMe…

8521eb6

…ssages

bvironn force-pushed the fix/ensure-tool-result-integrity branch from a7b2dae to 8521eb6 Compare April 11, 2026 06:33

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix(opencode): ensure tool-call/tool-result pairing after normalizeMessages#22001

fix(opencode): ensure tool-call/tool-result pairing after normalizeMessages#22001
bvironn wants to merge 1 commit intoanomalyco:devfrom
bvironn:fix/ensure-tool-result-integrity

bvironn commented Apr 11, 2026 •

edited

Loading

Uh oh!

github-actions bot commented Apr 11, 2026

Uh oh!

github-actions bot commented Apr 11, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Apr 11, 2026

Uh oh!

bvironn Apr 11, 2026

Uh oh!

bvironn commented Apr 11, 2026

Uh oh!

bvironn commented Apr 11, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

bvironn commented Apr 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Issue for this PR

Type of change

What does this PR do?

How did you verify your code works?

Checklist

Uh oh!

github-actions bot commented Apr 11, 2026

Related PRs Found

Uh oh!

github-actions bot commented Apr 11, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Apr 11, 2026

Choose a reason for hiding this comment

Uh oh!

bvironn Apr 11, 2026

Choose a reason for hiding this comment

Uh oh!

bvironn commented Apr 11, 2026

CI Note: E2E timeouts are unrelated to this change

Uh oh!

bvironn commented Apr 11, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

bvironn commented Apr 11, 2026 •

edited

Loading