eliasforge/fable-agent-orchestration/long-run-continuity

Long Run Continuity

A large engagement will outlive the working context it started in. Assume that context can be lost at any point, and make the run survive it.

Checkpoint

Persist to durable storage — a task tracker, a committed notes file, an issue — not just working memory:

• the queue: what is done, in flight, and not yet started;
• every open pull request and its evidence state;
• open findings, residuals, and security escalations, recorded the moment they surface;
• pending decisions and their blockers.

Rules

• Capture a finding when you find it, not at the end. The end may not arrive in the same context.
• The durable write is the guarantee. An in-context note that a boundary erases is not a checkpoint.
• Write state a fresh reader can act on with no prior context: exact paths, PR numbers, the precise next step.

When

Continuously through a long run, and always before a step that could exhaust or reset the working context.