The head conversation is durable: reuse the saved session when it is still alive.

SKU × warehouse × day. ~250 × 3 × 90 ≈ 67k rows.

Idempotent: one agent + environment per participant, reused across runs. IDs and a hash of the spec are cached in .agent_cache.json so re-runs

Compute the metric. String results are printed verbatim; numbers go through `format`.

A single normal request: start + end log lines.

* Map a value within [min, max] to a red→yellow→green RGB color. `good: "low"` * flips the spectrum so smaller is greener.

Jump to a specific step. Pauses (so it mounts, verifies, and holds).

Report only on every Nth turn, for use inside agent loops.

Send Claude one turn at a time until the bot has `task` in its inventory. Counts only non-chat actions toward max_turns. Aborts if the same a

A checkout request after the bad deploy: N+1 queries, slow, maybe 500.

Check that target SKU appears in the first N body rows of the first markdown table.

Ask Claude to rewrite its strategy notes after a task. Plain text reply, no tools — fast and cheap. On API failure, leave notes unchanged.

Full first-turn render: run context + state + current task.

Compact subsequent-turn render: just current state + task line. No RunContext header — that's already in the conversation history. The goal i