feat(agents): resume dropped /run streams in createAI

[mattzcarey]

createAI (agents/models/ai-sdk and agents/models/pi-ai) now resumes a streamed answer from AI Gateway's /run when the connection drops mid-stream. It reconnects through the binding and replays from the last complete event, so the consumer's stream keeps going with nothing lost or repeated. Stacked on #2576.

A streamed /run answer names its run in cf-aig-run-id, and the Run Durable Object keeps the run's events. The binding replays them from an event index, through a route aig-entry serves for it:

env.AI.fetch(`https://workers-binding.ai/ai-gateway/gateways/${id}/run/${runId}/resume?from=${n}`)

• The core transport's run() wraps a streamed /run body that carries a run id. Every caller gets resume with nothing to configure: AI SDK catalog models and dynamic routes, pi-ai catalog models, and agents/models/opencode, which wraps the AI SDK provider. Workers AI streams and provider-native requests have no run id and are returned as they are.
• core/resume.ts passes on only complete SSE events and holds back a trailing partial one. from counts complete events, so after a drop the replay starts on the boundary: nothing duplicated, nothing cut.
• It stops with a CloudflareAIError after five drops, or on a 404 once the gateway no longer holds the run. An abort or a cancel never reconnects, and an abort reaches the consumer as the error it is.

Checked live on the agents account's default gateway. To get a real drop, a test Worker wrapped the real AI binding and cut the first /run stream partway through an event, as a dropped connection would. createAI then resumed against the real gateway buffer. Each answer came back complete: one to forty in order, nothing repeated or missing.

Provider and model	Cut after	Resumed from
AI SDK anthropic/claude-haiku-4.5	600 / 2,500 bytes	event 1 / 15
AI SDK openai/gpt-5-mini	600 / 2,500 / 5,000 bytes	event 1 / 8 / 16
pi-ai anthropic/claude-haiku-4.5	600 / 2,500 bytes	event 1 / 15
AI SDK dynamic/<route> (served by openai/gpt-5-mini)	1,500 bytes	event 4
pi-ai dynamic/<route>	1,500 bytes	event 4

The resume route also returns 404 {"error":"Request not found"} for an unknown run. pnpm run check and the model tests pass.

[changeset-bot]

🦋 Changeset detected

Latest commit: c3216f8

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

This PR includes changesets to release 2 packages

Name	Type
agents	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

[agent-think]

🟡 agents import sizes: 1 entry point grew

	Entry point	Exports	Largest gzip change	Size now
🟡	agents/models/ai-sdk	1 resized	+815 B (+3.92%)	21.1 KiB

Changed exports (1)

	Import	Gzip change	Size now
🟡	agents/models/ai-sdk#createAI	+815 B (+3.92%)	21.1 KiB

How this works

Each runtime export is bundled on its own, minified, and gzipped. Changes smaller than 100 B, or smaller than 1% and 1 KiB, are ignored. Growth over 10% or 5 KiB is marked 🔴. This report is informational and does not fail CI. The workflow artifact contains every measurement.

_{Compared fda892f8 → c3216f85 · workflow run · reported by agent-think[bot]}

[pkg-pr-new]

Open in StackBlitz

agents

npm i https://pkg.pr.new/agents@2589

@cloudflare/ai-chat

npm i https://pkg.pr.new/@cloudflare/ai-chat@2589

@cloudflare/codemode

npm i https://pkg.pr.new/@cloudflare/codemode@2589

hono-agents

npm i https://pkg.pr.new/hono-agents@2589

@cloudflare/shell

npm i https://pkg.pr.new/@cloudflare/shell@2589

@cloudflare/think

npm i https://pkg.pr.new/@cloudflare/think@2589

@cloudflare/voice

npm i https://pkg.pr.new/@cloudflare/voice@2589

@cloudflare/worker-bundler

npm i https://pkg.pr.new/@cloudflare/worker-bundler@2589

commit: c3216f8