packet-worker

Packet worker

Never spawn agents as coders. One traced 25-page packet cost 3.8M tokens as a tool-using subagent and ≈40K as one stateless structured-output call — 95×, same model, same work. Your job is the codebook, the packet boundaries, the adjudication of holds and the sampled QA. The coding itself is an API call.

Everything here is limbic.cerebellum.packet and limbic.hippocampus.resolve; read docs/packet.md and docs/resolve.md before writing a runner, and docs/refuse.md + docs/audit.md before writing an apply.

Order of work — do not reorder

1. Retrieve deterministically first. Build a candidate index over the local KB (resolve.build_index), then resolve.text_candidates per span. If a plain join already produces the records, stop: in one campaign 2,336 of 2,342 proposals came from the join and 384 model calls produced 6.
2. Probe 50 items before building anything. probe(packets, n=50, yield_fn=…, min_yield=…, execute=True). yield_fn counts actionable outputs, not items returned. 12.3k lines of sealed machinery, 26 prompt versions and 13 per-packet test files were built around a stream that produced 0 writes because nobody did this. Low yield means a deterministic join, not a better prompt.
3. Then write the runner: make_packet → lint_packet → run_packets.
4. Validate, union the passes, propose. Run the QA loop below. Apply through hippocampus.apply.apply_proposal, never by writing the record yourself.

Packet rules

• Fixed slot enums, never per-item enums. A per-item enum of that item's candidate IDs changes the JSON schema every call, and the schema sits ahead of the input in the provider's cache prefix: measured 0% cached input; a fixed c01..cNN enum on the same batch measured 58%. Use resolve.slot_enum(cards, n_slots) and resolve.unslot to map back. It also removes the last place the model handles an identifier of yours — asking a model to echo your hashes made valid answers fail.
• "None of these" is always in the enum, and it is a wanted answer. Scoring coverage without it produced 130 fabricated biographies.
• Run lint_packet(packets) and act on every line. It flags a varying schema, a shared prefix under the ~1,024-token provider cache minimum (below it, caching billed 92.5% of one campaign's input as cache writes and cost 9.7% more), derivable body fields (46% of one packet was an identical evidence_fields list) and packets past ~25 items, where models start dropping items silently.
• Freeze the prefix with the packet. input_sha256 covers the prefix hash; editing a shared mutable prefix file later silently invalidated a batch that had already been paid for.