Over a long run

Handover — what happened

bee handover --session      # what THIS run did
bee handover --since 12h    # or a window
bee handover --json

Decisions first, then fixes, then the rest — the order someone picking the work up needs it in.

--session reads the id the harness already exports — CLAUDE_CODE_SESSION_ID, CODEX_SESSION_ID — so “this run” is an exact boundary rather than a guess at one. BEE_SESSION overrides it for anything bee has not been taught.

It refuses rather than guessing

With no session id in the environment, --session fails instead of quietly falling back to a time window. A window catches a colleague’s parallel work and misses anything older than the guess — it would answer a different question while looking like it answered this one.

Budget — what the steering costs

bee budget
bee budget --json

         session start (once)
┌───────────┬───────┬───────┬─────────┐
│ component │ items │ chars │ ~tokens │
├───────────┼───────┼───────┼─────────┤
│ memory    │ 1     │ 108   │ 27      │
│ chrome    │       │ 48    │ 12      │
│ total     │       │ 156   │ 39      │
└───────────┴───────┴───────┴─────────┘

Broken down by component, for session start and for a sample prompt. This is how you find out that a rule set has quietly become expensive.

Compaction — what is costing recall

bee compact                 # report only
bee compact --archive       # act on it
bee compact --archived      # what is currently out of recall
bee compact --restore <id>  # bring one back

Over time a store accumulates memories that make it harder to find things: near-duplicates, superseded decisions, one-off notes that will never be asked for again. bee compact reports them; nothing moves unless you ask.

There is no delete

--archive copies to a separate collection before removing from recall, records why, and --restore brings it back. The question “should compaction delete, summarise or archive?” was answered by removing the first option: a memory you cannot get back is a memory you cannot have been wrong about discarding.

The background worker

bee worker              # stay up
bee worker --once       # run each job once and report
bee worker --once --json

Work that scales with the size of the store does not belong on the way to a prompt. The worker builds the relevance index and checks embedding health.

{"ok":true,"jobs":[{"job":"embedding-health"},
 {"job":"search-index","message":"built the relevance index — search now reaches
  the whole store, not just the recent end of it"}],"failures":[]}

Jobs are resolved from dependency injection, so adding one is a registration.