When an agent’s memory gets compressed, it should still know the job.

When a coding session grows too long for the model to hold, the tool compresses older context into a summary so work can continue. That is compaction, and it is necessary. The trouble is that a coding project is operational state: which worker is already running, which audit passed, which file is authoritative, and what the next valid action is.

I noticed my agents coming back from compaction knowing the project but not where they were inside it: rereading files, repeating finished audits, and asking what to do next. In one two-day build, the agent’s context compacted 38 times. The field report tells that story in the agent’s own voice.

1. Before compaction

   A pre-compaction hook writes a session-specific receipt: bounded lifecycle details and a hash of the current checkpoint, pointing at the durable notebook for this task.

2. During compaction

   The compact prompt asks for a handoff, not a diary: objective, completed, running, and pending work, binding decisions, evidence paths, and the next action.

3. After compaction

   On resume, a short recovery message tells the agent to reread the named checkpoint, inspect actual saved state, and never duplicate a running worker.

This is a summary, not a spec. The repository’s code and README are the authority on the exact hooks, fields, and install steps.

A notebook that takes the whole context window to read would recreate the problem it is meant to solve, so it stays small on purpose.

Local continuity: verified in one real task Credit savings: not measured

1. Verified locally

   Real lifecycle chain in one long task

   On one local Codex Desktop installation, repeated automatic compactions were followed by continued work from durable state, with PreCompact → PostCompact → SessionStart receipts preserved.

2. Tested

   Focused safety tests

   Session isolation, withholding secret-looking checkpoint data, not copying transcript or hidden-reasoning text, allowlisted fields, and no global continuation loop.

3. Kept in

   One unmatched event

   One compaction has no matching resume receipt. It may be a lifecycle edge case. It stays in the record until someone investigates it.

4. Not measured

   Does it save credits or time?

   Unknown. That needs an A/B trial on equivalent long-running tasks.

5. Not shown

   Does it behave the same everywhere?

   Universal behavior across installations, versions, or other agent tools has not been shown.

The field report’s event counts are preserved from the original report and were not independently revalidated for this website.

• A bad plan, failing code, or an ordinary reasoning mistake.
• A provider outage, a rate limit, or a tool waiting for permission.
• A checkpoint that was never updated, or external state that is simply wrong.