fix(observability): preserve AI span hierarchy across WebSocket turns

[mattzcarey]

Summary

• use the existing AI SDK namespace wrapper for both v6 and v7 so invoke_agent is active before model and tool work starts
• keep chat, execute_tool, and requested/approved/denied tool_approval spans in the expected hierarchy
• preserve provider/tool AsyncLocalStorage state instead of replaying a captured snapshot
• close WebSocket-owned Agent and AI spans when async work hands off beyond the native invocation
• keep RPC, alarm, and other non-WebSocket spans open through normal completion

Root cause

Think used the AI SDK v7 lifecycle adapter directly. The SDK captures its asynchronous operation context before lifecycle callbacks run, so opening invoke_agent from onStart is too late: later chat and execute_tool spans inherit the pre-operation context.

Delayed hibernatable WebSocket work has a second constraint. A later model step can retain the original workerd tracing context after that native invocation ends. An open span created in that stale context can then be force-closed at the earlier invocation timestamp, producing a negative duration or span_not_ended warning.

Fix

Think now uses wrapAISDK for both supported majors. AI SDK v7 still exposes the structural model middleware used by the existing wrapper, so no duplicate v7 operation wrapper or tracing snapshot is needed. The shared wrapper now reads v7's renamed telemetry and runtimeContext fields and handles v7's top-level toolApproval policy.

Invocation-bounded spans use one finishOnAsyncHandoff policy. WebSocket Agent spans and streaming operation/model/tool spans close before pending async work escapes the invocation, while their ended context remains active so later children retain the correct parent ID.

Regression coverage

The normal Agents test suite covers:

• real AI SDK v7 multi-step model/tool hierarchy
• provider and tool auth-store replacement
• needsApproval and top-level v7 approval policies
• delayed model → tool → model WebSocket turns for both v6 and v7
• the tracer's async-handoff lifetime contract

No CI-only test job, demo project, generated fixture, documentation rewrite, or lockfile change is included.

Verification

• observability: 7 files / 112 tests
• Agents: 124 files / 2,412 tests
• Think: 52 files / 1,000 tests
• AIChat: 50 files / 737 tests
• agent-think: 11 files / 39 tests
• actual AI SDK v6: focused regressions 19 tests; Think typecheck; Think workers 46 files / 920 tests
• monorepo build: 25 projects
• pnpm run check: exports, formatting, lint, and all 117 TypeScript projects

---

[changeset-bot]

🦋 Changeset detected

Latest commit: 3ff1e76

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

This PR includes changesets to release 3 packages

Name	Type
agents	Patch
@cloudflare/think	Patch
@cloudflare/agent-think	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

[devin-ai-integration]

✅ Devin Review: No Issues Found

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

[pkg-pr-new]

Open in StackBlitz

agents

npm i https://pkg.pr.new/agents@1982

@cloudflare/ai-chat

npm i https://pkg.pr.new/@cloudflare/ai-chat@1982

@cloudflare/codemode

npm i https://pkg.pr.new/@cloudflare/codemode@1982

create-think

npm i https://pkg.pr.new/create-think@1982

hono-agents

npm i https://pkg.pr.new/hono-agents@1982

@cloudflare/shell

npm i https://pkg.pr.new/@cloudflare/shell@1982

@cloudflare/think

npm i https://pkg.pr.new/@cloudflare/think@1982

@cloudflare/voice

npm i https://pkg.pr.new/@cloudflare/voice@1982

@cloudflare/worker-bundler

npm i https://pkg.pr.new/@cloudflare/worker-bundler@1982

commit: 3ff1e76