hotato AGENTS.md

AGENTS.md

You are an agent evaluating or wiring up hotato, the local-first AI engineering platform, to trace, evaluate, test, and gate an AI agent (voice, chat, or LLM). This file is written to be executed top to bottom. Every command below is offline, reversible, and needs no account or key. Prefer the CLI's own contract (hotato describe --format json) over this prose when they disagree.

The core loop (run this first)

One path, first touch to a CI gate. Each step prints the exact next command, so you can chain the whole loop from stdout alone:

1  hotato start --demo                    see it catch a failure on two bundled calls
2  hotato investigate ./call.wav          score your own two-channel recording
3  hotato investigate label <ref>         commit the caught moment as a regression
4  hotato pr create --fixtures <bundle>   open the pull request that adds the CI gate
5  hotato contract verify contracts/      the gate re-runs the stored evidence

• Step 1 sweeps the two bundled demo calls, builds one failure contract, and runs one say-do conversation check. It exits 0 because setup finished; the gate command it points at exits 1.
• Step 2 opens a local dual-channel WAV (or pulls one live with --stack vapi --call-id ID), authenticates its capture origin, runs the input-health gate, and ranks the timing moments. It marks the top-ranked one most likely failure and prints one investigate label command for it. No intent is inferred; a suspected channel swap refuses the verdict path.
• Step 3 writes a content-addressed contract bundle to contracts/<id>.hotato/ from the candidate ref (.hotato/investigate-state.json#N) and the --expect yield|hold label you supply. The label is the human decision; hotato measures whether the timing matched it.
• Step 4 stages that bundle byte-identical under tests/hotato/contracts/ and opens the PR (dry run by default; --yes runs git and gh).
• Step 5 re-measures the stored evidence deterministically. This is the CI gate: exit 0 pass, exit 1 fail.