looptail

Record every AI decision as a signed, append-only trail. Every loop leaves a tail.

pip install looptail

import looptail

looptail.init(app="support-agent")

@looptail.trail
def handle_ticket(ticket):
    ...

# report an outcome signal when you learn how it went
looptail.outcome(looptail.last_event_id(), csat=5)

Every call becomes a hash-chained, Ed25519-signed loop event appended to .looptail/support-agent.jsonl. Nothing is edited in place; any modification, deletion, or reorder of past records is detectable:

npx @looptail/cli verify --app support-agent

✔ 2 loop events
✔ chain intact · signatures valid

Or from Python: looptail.verify_trail().

The whole loop, CLI-first

looptail-evals run --rubric rubrics/refund-policy.json --app support-agent
looptail-issues cluster --app support-agent
looptail-improve propose --app support-agent --issue refund-policy/v1/criterion-0 \
    --prompt-file prompts/support.txt
looptail-improve replay --app support-agent --proposal prop_... \
    --cases regression.jsonl --runner myapp.replay:run --rubric rubrics/refund-policy.json
looptail-improve approve --app support-agent --proposal prop_... --apply

Evaluators score recorded events against your rubrics with an LLM judge (pip install "looptail[anthropic]" or [openai]); failing verdicts cluster into tracked issues; the improve loop drafts a patch, replays your regression set, and gates approval on a passing replay. Every step lands in the trail as a signed event. Guide: looptail.ai/docs/improve.

How it works

• A signing key is created on first use at ~/.looptail/signing-key (override with LOOPTAIL_SIGNING_KEY, hex seed).
• Events follow the open trail format — spec/trail-format.md — and are verifiable by any implementation, including the JS CLI.
• The local trail is the source of truth. With an API key (looptail.init(app=..., api_key=...) or LOOPTAIL_API_KEY), events also sync to the hosted Tail (private beta); sync is best-effort and never blocks or crashes your app.

Read the trust model in the spec before relying on local trails as evidence: self-signed trails are tamper-evident; hosted anchoring (v0.2) makes them stronger.

Docs: looptail.ai/docs · License: Apache-2.0