Cloudflare’s Pi 1.6 Upgrade Can Break Your Usual Rollback
Cloudflare’s November 20 Agents SDK v0.34.0 release adds Pi 1.6 support, including tools that can wait without keeping a Durable Object awake. Its upgrade instructions contain a release-planning constraint: once a conver
Cloudflare’s November 20 Agents SDK v0.34.0 release adds Pi 1.6 support, including tools that can wait without keeping a Durable Object awake. Its upgrade instructions contain a release-planning constraint: once a conversation runs on Pi 1.6, Pi 1.5 cannot run its tool calls or model requests. Deploying the previous Worker may therefore fail to restore the conversations that need help.
AI disclosure: This article was generated with AI assistance using Keith Azodeh’s public project documentation and the linked primary sources. It has not been reviewed by Keith before publication.
This concerns the Pi harness integration, not every Cloudflare Worker. Cloudflare’s Pi documentation calls Pi Durable experimental, ties each Agents SDK release to a supported Pi minor version, and advises fixing a bad deployment by rolling forward after conversations have used Pi 1.6.
A rollback plan that checks only whether the old application deploys is incomplete for this upgrade.
The deployment artifact is not the whole release
For a stateless handler, replacing the executable can remove much of a bad release’s behavior. A durable agent also leaves work behind: conversations, queued input, and tasks that another invocation must continue.
Pi’s versioned 1.6.0 changelog identifies new versions for tool and generation records. It says an older harness blocks those tasks. The compatibility boundary is in persisted work, not just the application bundle.
That changes the recovery question. “Can we deploy yesterday’s build?” and “Can yesterday’s build continue today’s work?” need separate answers.
A fresh conversation is a weak test of the second question. It does not contain the upgraded work that an older runtime would have to handle.
A staged platform needs a staged agent rollout
The Garden case study describes a platform built in stages, with migration from Planning Center still ongoing. Its documented workflows include service planning, team scheduling, follow-up ownership, communications, and staff work. Demonstration workflows use isolated data.
The case study distinguishes implementation checks from completed rollout and staff acceptance. It does not establish a Pi deployment or an AI staff assistant.
Consider a proposed assistant added to that kind of staff workspace. It might help a coordinator prepare a service plan, ask for a missing detail, and return later when the coordinator responds.
The product value is continuity. The coordinator expects to find the same unfinished work after a deployment. Starting a new chat may restore an input box while abandoning the task they were trying to finish.
For that proposed rollout, the acceptance unit should include an existing conversation and its pending work, not just the assistant’s welcome screen.
Rehearse the boundary with isolated conversations
Build a synthetic release exercise around the versions actually resolved by the application’s lockfile.
Start a conversation on the earlier version and retain its unfinished state in an isolated test environment. Upgrade that environment to the Pi 1.6-compatible build, continue the conversation, and exercise a tool call and model request.
Then evaluate recovery against the state written by the newer version. The point is to observe the documented incompatibility and prove the alternative recovery path, rather than to demand that an unsupported downgrade succeed.
Keep the original test environment separate if it is needed as a baseline. A newer and an older deployment should not be casually pointed at the same writable conversation store for comparison.
Also include a conversation created entirely after the upgrade. A migration test that covers only older conversations can miss the recovery needs of new work started during the release.
These are proposed checks, not tests performed against The Garden or Cloudflare.
Prepare a forward fix before widening the rollout
For this release, a practical proposed recovery candidate would retain the compatible harness dependencies while correcting or disabling the faulty application behavior.
That candidate needs a resumed-conversation check of its own. Pinning dependencies is useful, but the application’s tool registrations and behavior still need to match the work it is expected to continue.
If a new feature must be disabled, make the pending-work policy explicit. Should affected conversations wait for repair, finish through a supported path, or be handed to a staff member? The interface should say which outcome occurred.
Do not describe “start again” as recovery of the same task unless the application actually preserves the necessary context and work state.
This preparation has a cost: another recovery build, a test environment, and a release exercise. It is easier to justify when the assistant is responsible for work that staff will return to tomorrow.
Make the canary include persistence
A limited rollout should isolate which conversations can encounter the upgrade. A small percentage of HTTP traffic is not enough if different requests can send one durable conversation between incompatible builds.
Define the cohort around the application’s conversation ownership and storage routing. Verify that all operations for a test conversation reach a compatible deployment.
During the proposed canary, check unfinished conversations as well as completed answers. Have the operator reopen a pending task after deployment and explain how it will continue. Retain a record of which build and dependency versions handled the exercise.
The useful release evidence is specific: this upgraded conversation can resume under the recovery build, and its unfinished work is accounted for.
Pi 1.6’s waiting tools make longer-lived agent work more practical. The upgrade is also a reminder that durable work can outlive the code that created it. Plan recovery around both.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.