Honey Loop

A loop is the one place token waste compounds: cost = per_tick × ticks. The honey levers still apply inside each tick (less code, less prose), but a loop adds two leaks the single-shot levers don't cover — re-paying for context every wake-up, and re-doing work that didn't change. Cut those.

The five levers

1. Pace to the cache, skip the dead zone. The prompt cache TTL is 5 min.

- < 270s → context stays warm; right for actively polling external state. - ≥ 1200s → one cache miss amortized over a long idle wait; the default for idle ticks (use 1200–1800s). - ~300s is the worst choice — pays the cache miss without amortizing it. Never round to "5 minutes." Drop to 270s or commit to 1200s+.

1. Event-driven over polling. Harness-tracked work (background Bash, Agent,

Workflow) re-invokes you automatically on completion. Don't schedule a short tick to poll it — that re-reads full context for nothing. Set a long fallback heartbeat (1200s+) and let the completion notification drive you. Poll only external state the harness can't see (CI, a deploy, a remote queue), at a cadence matched to how fast that state actually changes.

1. No-change short-circuit. Re-generating output is the bulk of per-tick cost.

Check cheaply first (a hash, a timestamp, a git rev-parse, a status field). If nothing changed: emit one line and reschedule. Don't redo the task, don't re-read the world, don't write a report.

1. Carry a compact state handle, don't re-derive. Keep what the loop has done

so far in a small file (.honey-loop/state.json or a <<honey:HASH>> stash via eso stash), and read that each tick — not the full history re-discovered from scratch. Lever 3b: the cheapest input token is the one that never re-enters context.

1. Define done; then stop. An unbounded loop is unbounded cost. State the exit

condition up front (target reached, N empty ticks, deadline). When met, end the loop — omit the wake-up reschedule. "Loop forever" is a bug, not a feature.

Per-tick output: ultra by default

A loop tick is an agent-to-self / agent-to-log message, not a deliverable. Default to ultra: status line only, near-zero prose. One line is enough — tick 4/∞ · no change · next 1800s. Step up to a full report only on the tick where something actually needs the user (state changed, a decision, an error).

Don't cut

• The stop/exit check — dropping it is how a loop runs all night.
• Error surfacing — a tick that hit an error reports it; silent failure in a

loop hides for hours.

• The cheap pre-check in lever 3 — skipping it to "just redo it" defeats the point.