band-once 0.1.0

Exactly-once lifecycle shell for Band agents, with a runnable proof and a public conformance check.

band-once

Exactly-once delivery for Band agents, as a tiny, reusable library.

Band serves an agent's inbox through a lifecycle cursor: GET /agent/chats/{id}/messages/next returns the oldest not-yet-processed message and re-serves the same message on every call until you mark it processed or failed. That cursor is correct, but it means a crash anywhere in the middle of handling a message (before you post, after you post but before you ack, or with a lost ack on the way back) leads to the same message being delivered again. Handle it naively and you double-post; guard it wrong and you lose work. There is no reference implementation of getting this right, so most Band agents either fake the lifecycle or get exactly-once subtly wrong.

band-once is that reference implementation, lifted out of a production system and stripped of every application concern. You supply one handle() function; the shell owns the drain, the processing/processed/failed lifecycle, the post, the read-then-act dedup guard, and bounded retry on transient HTTP. It ships with a runnable proof and a public conformance check that an LLM cannot self-certify.

Install

pip install band-once          # once published to PyPI

Locally, from this directory:

pip install -e .

The shell: write only handle()

from band_once import BandAgentShell, strip_mention_markers

def handle(message: dict, context: list[dict]) -> dict | None:
    body = strip_mention_markers(message["content"])
    # ... do your framework-specific work (LLM call, graph, tool) ...
    return {
        "content": f"echo: {body}",
        "mentions": [peer_uuid],          # Band requires at least one mention
        "dedup_key": f"echo:{message['id']}",  # the natural key of this unit of work
    }

shell = BandAgentShell(api_key=KEY, agent_name="my_agent",
                       dedup_namespace="echo", max_attempts=3)
shell.whoami()
shell.join(chat_id)
shell.run(handle, idle_breaks=5)   # drain -> processing -> handle -> post -> processed

dedup_key is the natural key of the unit of work (for example draft:nis2:inc-8842:round-1). Before re-posting, the shell reads the room and drops the post if the key is already present. A crash and re-delivery therefore re-runs the work but never double-posts. See examples/echo_agent.py for a complete agent.

The proof: a receipt, not a vibe

python -m band_once.proof

drives a large, seeded space of kill + duplicate schedules (kill before post, kill after post, lost-ack) through the real BandAgentShell dedup guard and the real IdempotencyLedger, then prints:

exactly-once held across 10000 schedules: 0 double-posts, 0 lost messages

The master seed and the schedule-space size are printed in the output, so anyone re-running the exact command gets the exact same number. No API keys, no network. Exit code is 0 only when exactly-once held on every schedule.

The conformance check: an external prover

Ofer's thesis for Band is that your coding agent cannot review its own work. The same applies to exactly-once: an agent author should not mark their own homework. verify_exactly_once is the independent grader.

from band_once import verify_exactly_once, clean_echo_agent

result = verify_exactly_once(clean_echo_agent)
assert result.ok           # passes the reference correct agent
print(result.schedules_checked)

You hand it an agent_factory that builds a fresh agent over a shared IdempotencyLedger; it drives seeded kill + duplicate schedules through that agent and asserts exactly-once on every one. On a violation it returns a typed ConformanceResult naming the first schedule that broke the invariant (its seed and a one-clause reason), the same way first_broken_index names the exact link a hash chain breaks at:

def buggy_factory():
    # records the work BEFORE checking the ledger, so a redelivery double-posts
    def agent(ledger, dedup_key, attempt, ts):
        ...
    return agent

result = verify_exactly_once(buggy_factory)
assert not result.ok
print(result.first_violating_seed, result.reason)
# 20260616 key 'work:k3:job-1:round-1' was ACCEPTED 2 times ...: a double-post

The grader never trusts the agent's claim about itself: it re-derives the invariant from the deliveries and the agent's own dispositions.

What is in here

File	What
band_once/shell.py	BandAgentShell: the lifecycle shell. Imports only stdlib + requests + the sibling retry helper.
band_once/ledger.py	IdempotencyLedger: one ACCEPTED per natural key, every redelivery dropped.
band_once/retry.py	Bounded full-jitter exponential backoff on transient HTTP only.
band_once/fake_band.py	An in-process Band double for the proof and tests (delivered / processing / processed, the three kill positions).
band_once/proof.py	The kill-storm receipt. python -m band_once.proof.
band_once/verify.py	verify_exactly_once, the external prover, and clean_echo_agent, the reference correct agent.
examples/echo_agent.py	A complete agent built on the shell.
BAND_API_FIELD_GUIDE.md	The verified Band Agent API surface this shell encodes.

Verified Band API facts this shell encodes

• Base https://app.band.ai/api/v1, auth header X-API-Key: <agent key> (not Authorization: Bearer).
• mentions is required on every message; the mentioned agent must already be a participant; you cannot mention yourself.
• /messages/next is per-chat and re-serves the same message until it is marked processed/failed; it is a cursor, not a destructive pop.
• The lifecycle endpoints are per-chat under the message; processed/failed must follow processing; failed carries {"error": <string>}.
• Exactly-once is owned by the poster via a read-then-act dedup guard, never by relying on re-delivery.

See BAND_API_FIELD_GUIDE.md for the full table.

License

MIT. See LICENSE.