Remove default pruneMessages call from Think to preserve client-side tool results

[mattzcarey]

Summary

Fixes #1455.

Think hardcoded pruneMessages({ toolCalls: "before-last-2-messages" }) in the inference loop, which silently stripped client-side tool results (no execute, output supplied via addToolOutput) from any turn beyond the second. The user's choices live in those tool results, so the conversation effectively lost its memory.

This PR drops the pruneMessages call from Think._runInferenceLoop entirely. truncateOlderMessages still runs (size-trimming tool outputs >500 chars and text >10k chars in messages older than the last 4), so context cost stays bounded.

What if you want the old behavior?

Apply pruneMessages yourself in beforeTurn:

import { pruneMessages } from "ai";

beforeTurn(ctx: TurnContext) {
  return {
    messages: pruneMessages({
      messages: ctx.messages,
      toolCalls: "before-last-2-messages"
    })
  };
}

Or scope it per-tool with the array form so client-side tool results survive across turns:

beforeTurn(ctx: TurnContext) {
  return {
    messages: pruneMessages({
      messages: ctx.messages,
      toolCalls: [
        { type: "before-last-2-messages", tools: ["read_file", "search"] }
      ]
    })
  };
}

Both snippets are added to docs/think/lifecycle-hooks.md next to the existing beforeTurn examples.

Behavior change

This is a behavior change — the changeset is marked minor and explains the diff. Existing agents that benefit from the previous aggressive pruning will see longer model contexts; the size cap from truncateOlderMessages keeps the increase bounded.

Test plan

• npm run typecheck — all 80 projects pass
• npm run check:exports — all 8 packages valid
• npx sherif — no issues
• oxfmt --check . — clean
• oxlint examples/ packages/ guides/ openai-sdk/ site/ — 0 warnings, 0 errors
• New regression test in packages/think/src/tests/think-session.test.ts confirms client-side tool outputs (user-choice-0/1/2) all reach the model after follow-up turns
• Full think test suite passes (290/290)
• Manual verification by the issue reporter against their client-side-tool flow

[changeset-bot]

🦋 Changeset detected

Latest commit: 2aaf22b

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

This PR includes changesets to release 1 package

Name	Type
@cloudflare/think	Minor

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

[threepointone]

Feels like we need a better fix than making this configurable now when we'll likely remove it later?

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 2 additional findings.

[mattzcarey]

Feels like we need a better fix than making this configurable now when we'll likely remove it later?

id would prefer to remove it cause it bit me before. @threepointone I cant see it being a good default.

[pkg-pr-new]

Open in StackBlitz

agents

npm i https://pkg.pr.new/agents@1456

@cloudflare/ai-chat

npm i https://pkg.pr.new/@cloudflare/ai-chat@1456

@cloudflare/codemode

npm i https://pkg.pr.new/@cloudflare/codemode@1456

hono-agents

npm i https://pkg.pr.new/hono-agents@1456

@cloudflare/shell

npm i https://pkg.pr.new/@cloudflare/shell@1456

@cloudflare/think

npm i https://pkg.pr.new/@cloudflare/think@1456

@cloudflare/voice

npm i https://pkg.pr.new/@cloudflare/voice@1456

@cloudflare/worker-bundler

npm i https://pkg.pr.new/@cloudflare/worker-bundler@1456

commit: 2aaf22b

[mattzcarey]

Pivoted per your feedback — dropped the default pruneMessages call entirely instead of making it configurable. truncateOlderMessages still runs so context cost stays bounded, and the beforeTurn docs now include opt-in pruning snippets for users who want the old behavior. Force-push avoided; updated PR description and changeset accordingly.