The original title is quite long and technical. Let me extract the key facts:

Short summary

A practical case study of fixing a vault-auto-ingest job that timed out repeatedly by splitting step2 into independent sub-steps (Claude log digestion and Codex log digestion), each with its own marker file for idempotent retries. The key principle: draw marker boundaries at units of work that can be independently re-run. This reduced 40-minute combined timeout windows to 25-minute independent slots.

• •Split a monolithic job step into independent sub-steps with separate marker files for idempotent retries
• •Draw marker boundaries at units of work that can be independently re-run
• •Composite markers confirm all sub-steps completed before setting the overall done flag

Generated with AI, which can make mistakes.