What bee held back

bee steering
bee steering --since 7d --json

bee spends context on every turn. Until this existed it recorded not one delivery, so the question “which of my rules never actually arrive” was not hard to answer — it was unanswerable, because the data did not exist.

Why the withheld half

What bee delivers is already in the agent’s own transcript. What bee held back is recorded nowhere by anything:

┌──────────┬──────────┬──────────────┬──────────┬──────────┐
│ what     │ id       │ held back by │ at least │ surfaces │
├──────────┼──────────┼──────────────┼──────────┼──────────┤
│ rule     │ bb29871f │ rule-cap     │ 14       │ prompt   │
│ rule     │ 7197a5e8 │ rule-cap     │ 14       │ prompt   │
└──────────┴──────────┴──────────────┴──────────┴──────────┘

A rule added last month can be starved by a cap on every turn since, while bee hook rule list keeps showing it as enabled. That is what this finds.

The block itself now names them too, so it is visible in the moment:

… 3 more rule(s) over the per-turn budget: bb29871f, 7197a5e8, 7a83d14f

Counts are a lower bound

Said on every render, not in a footnote. The ledger is capped and drops oldest, and a concurrent drain can lose a write, so every number is at least this often, never exactly. A sampled instrument that implies completeness is worse than no instrument, because it will be believed.

What it cannot tell you

It records what bee sent. It cannot show whether the model read it. Those are different claims and only the first one is bee’s to make.

It cannot break the hook

Every failure is swallowed. An unwritable home, a full disk, a locked file — each costs one ledger line and nothing else. Verified with the bee directory made read-only: the hook exited 0, emitted its full block, and wrote nothing to stderr.

It is also bounded without anyone running a background worker, which is the state most machines are in. The file has a ceiling enforced by the writer, dropping oldest, checked with one stat per write.

Why one stat per write, and not a counter

The first version trimmed every 250 writes using an in-process counter — and every hook is a fresh process, so the counter reached one and the process exited. Seeded to 6,000 lines against a 5,000-line cap, one emission left 6,001. The self-bounding property the design claimed did not exist at all.

The caps it reports on

surface	rules	memories
session start	12	8
every prompt	12	5
on a file write	20	3

The per-write caps are higher because those items were selected by matching that exact path, so they are far likelier to apply than the standing set.

See what each costs:

bee budget
bee budget --path src/Catalog/Data/ProductDataSource.cs