Chapter one uses a ninety-second AI music video to ask how a long-running job survives processes. Chapter two switches to “How a City Wakes at Night,” an animated explainer, because OpenMontage's animated-explainer manifest exposes research, proposal, script, scene plan, assets, compose, and publish. The examples do not share one business pipeline, but they face the same waits, failures, revisions, external calls, and human acceptance. That production pressure contains at least two different systems problems.

The first asks how the job continues to exist: after a browser disconnect, Worker restart, or provider timeout, does the job retain the same ID, and can the system tell what happened? The second asks how the agent advances production: how does it know the current stage, approved decisions, required approval pauses, available tools, and the checks that still follow a successful render?

Reading contract. After both chapters, you should be able to say who preserves long-running identity, makes creative decisions, calls external tools, and approves delivery. You should also be able to explain why checkpoint files help a new session continue but do not start a new execution process.

1. Each chapter solves a different kind of interruption

Temporal and OpenMontage are not competing products, and OpenMontage does not use Temporal internally. Reading them together compares two meanings of “continue”: Temporal preserves execution identity and history so available Workers can continue processing tasks; OpenMontage lets a new agent continue production by reading operating instructions, manifests, checkpoints, and a tool registry. 

User request: produce an AI video

Job recovery
  Which job is this?
  What is confirmed to have happened?
  Who can continue after a process disappears?
  -> Chapter 01 · Temporal

Production progress
  Which stage is active?
  Which creative decisions were approved?
  Which providers and renderers are available?
  What makes the final file deliverable?
  -> Chapter 02 · OpenMontage

2. Reading route

01 · Temporal: When an AI Video Job Must Outlive Its Process Follow Workflow ID, Event History, Task Queue, Workers, Workflows, and Activities to see which records survive process loss and who continues. 02 · OpenMontage: When a Coding Agent Runs the Video Workflow Follow AGENT_GUIDE, pipeline manifests, stage directors, checkpoints, ToolRegistry, render locks, and final review to see how each stage produces an artifact and pauses for approval.

3. How the two kinds of recovery differ

ResponsibilityTemporal chapterOpenMontage chapter
Durable identityA Workflow Execution and stable Workflow ID.A project directory and project.json marker.
Historical factsServer-side append-only Event History used for replay.Stage checkpoints, decision_log.json, and files under history/.
Next stepA Workflow decides deterministically from history.The agent follows the manifest and stage-director instructions.
External actionAn Activity executes through a Worker.ToolRegistry discovers and routes tools by capability.
Waiting for approvalSignals or Updates plus application code.The manifest marks required approval; writing completed requires human_approved=True.
Recovery promiseA compatible Worker can rebuild control flow once it polls the appropriate Task Queue.A new agent session reads project files and resumes the first unfinished stage in manifest order.

The final row is the essential distinction. OpenMontage turns expensive creative work into inspectable, resumable project state. Temporal turns control flow itself into an execution identity recoverable across processes. A production system may need both, but shared words such as “checkpoint” and “resume” do not make them the same layer.

If production stops halfway through, Temporal's PollWorkflowTaskQueue is polled by an application Worker. Recovery requires a compatible Worker to be running; the service does not start that process for the application. OpenMontage's get_next_stage() instead finds the first unfinished stage in pipeline order: later files do not let an agent skip an earlier awaiting_human stage.

The checkpoint writer checks approval at write time. Reading a checkpoint or finding the next stage does not recheck approval, but advancing a later stage to awaiting_human or completed does recheck predecessor completion and required approval fields. Those booleans do not authenticate the approver. The caller must still verify the approval source and reconcile existing external outputs before resuming, to avoid repeated generation or charges.

Sources