Replatform Cloudflare submission execution onto Agents SDK Tasks

[astrobot-houston]

Originally proposed by @mattzcarey in #700. Their description is reproduced verbatim below.

All community-submitted pull requests are automatically converted to issues (bugs) & discussions (feature requests, enhancements) where they can be triaged and prioritized. Once prioritized, a PR implementation is created automatically.

Draft. Depends on the Agents SDK release that carries cloudflare/agents#2274 (step.signal, run({ deadline }), onError(error, run)). The lockfile resolves agents@0.23.0, which does not yet have those; against it the Task handlers throw on step.signal. Bump the range once the release lands and this is ready for review.

What

Flue's Cloudflare target already extends the Agents SDK Agent class; this replatforms its execution layer onto the SDK's Tasks capability and deletes the hand-rolled supervision: the zero-delay drain plus 30s heartbeat on schedule(), the in-isolate supervisor chain, detached runFiber attempt fibers with onFiberRecovered, deadline enforcement over live fibers with a 60s force-settle grace, and the __flueWakeAgentSubmissions schedule target.

Two Task definitions are declared on the generated class via taskDefinitions (resolved by the SDK on every wake, so replay is correct by construction):

• flue:drive@v1 — one bounded reconcile pass: materialize unready rows, finalize pending settlements, recover interrupted running rows, claim runnable heads, start an attempt run per claim, complete. One run id per object; every former "arm the drain" boundary (admission, abort, onStart, attempt settlement, Task failures) ensures it and joins a live one for free.
• flue:attempt@v1 — one submission processed to settlement in the handler body. The flue abort controller is linked to step.signal; the submission's durability timeout is the run's deadline. The SDK owns the claim backstop that survives eviction, replays the run on a fresh isolate after an interruption (the body reconciles from stream evidence first, then processes the replacement attempt), and settles over a signal-deaf attempt at the deadline, fencing its later writes; the next drive settles the submission through the existing timeout classification.

Unchanged: the submission ledger (now the inbox), the canonical stream, admission and dispatch paths, processSubmission, reconcileInterruptedSubmission, the extension mechanism, and the alarm() context wrapper for extension-authored schedule callbacks.

Behavior change worth knowing: abort of an attempt that ignores its signal now settles at the submission timeout rather than after a fixed 60s grace. Stale __flueWakeAgentSubmissions schedule rows from earlier deploys log callback not found once and complete.

@flue/runtime still never imports agents: the Tasks surface is typed structurally, as before.

Verification

pnpm check:types and biome lint are green (knip's unused-export findings are pre-existing on main). Against a local build of the agents PR branch, examples/cloudflare under vite dev with the Workers AI binding:

• single prompt settles completed;
• two prompts back to back: the second joins the running response at a turn boundary, both settle completed;
• POST /abort mid-response settles aborted;
• killing the dev server mid-response and restarting: the SDK replays the attempt run (attempt 2), flue claims a replacement attempt, the partial is closed out, the response completes and settles, and the non-retained Task rows are gone from cf_agents_task_runs.

---

Original implementation from #700 by @mattzcarey

[astrobot-houston]

Possibly related contribution from @mattzcarey — opened #700 which appears to be addressing this.

Their summary: This PR replaces the hand-rolled supervision machinery in Flue's Cloudflare target (drain scheduling, heartbeats, fiber management, deadline enforcement) with the Agents SDK's Tasks capability. Two task definitions handle reconciliation and per-submission attempt processing, leveraging the SDK's built-in claim backstop, replay, and signal/deadline management. It also changes abort behavior so signal-ignoring attempts now settle at the submission timeout instead of a fixed 60-second grace period.

Implementation: withastro/flue@feat/replatform-coordinator-on-agents-tasks (diff)

This comment was posted automatically when #700 was redirected. The implementation is preserved on the branch above so it can inform the work here.