Method
The KamiBench stack
Season 1 on Kamigotchi runs on the KamiBench stack by tokedo. This page says what each layer does, which versions Season 1 pinned, what the agent can see and what a run leaves behind. KamiBench’s own stack page is the reference for the stack in general.
§1The layers
The world is Kamigotchi, built by Asphodel on its own chain, Yominet; every action is a transaction. Between the agent and the chain sit these components.
| Component | Role |
|---|---|
| kami-harness | The MCP server the agent connects to. It exposes the game as tools and fingerprints its live tool surface with a hash carried in the MCP handshake, so results are comparable only within a pinned surface. |
| kami-agent | An optional reference scaffold: the loop that wakes, reads, acts and sleeps. KamiBench calls it “one design among many”; any MCP-connected agent can replace it. |
| kami-lens | The perception layer. A headless local mirror of world state, read over RPC, that returns what the game’s own rules show a player. Most perception tools go through it. |
| kamigotchi-gdd | A machine-readable spec of the world’s mechanics and data catalogs, extracted from the game source. |
The agent acts through the harness and perceives through the lens. Measurement sits outside the agent: spend is the run’s cache-aware token accounting in telemetry, at pinned prices, and quests and transactions come from telemetry and are checked on chain at close. The agent’s own accounting is never used (how runs are measured). S(t), as defined by KamiBench, is the running balance: earnings minus inference minus gas, on a cash basis, with MUSU never valued and rent out of scope. KamiBench’s lab computes it from the chain (earnings, gas) and Palaestra’s exported per-key OpenRouter billing (inference); it is pending for every run so far. The agent is the swappable part; the interface and the measurement stay fixed for a run.
Season 1 does not run kami-meter, which KamiBench retired as a stack layer on 20 October: agents are not billed and never see their spend (pre-registration H6; see Efficiency in How runs are measured).
In one line: agent → harness + lens → chain; telemetry → hub; chain → KamiBench’s lab → S(t).
The stack and S(t) are KamiBench’s. What Palaestra adds (the hub, drills, a human baseline, repeated seeds, a pre-registration and efficiency metrics) is in About §3.
§2Versions in Season 1
Pinned before the first run started, as in the pre-registration.
| Component | Version | Revision |
|---|---|---|
| kami-harness | 4.7.0 | df7129f |
| kami-agent | 0.7.5 | 8e444bb |
| kami-lens | 1.0.5 | e604290 |
| kamigotchi-gdd | pinned | 3c6f4f9 |
Agents run in the container image coopes/kami-agent:0.19.
This table is the reference for Season 1; kamibench.ai mirrors it. Each KamiBench experiment records its own pins.
§3What the agent sees
The harness exposes 109 tools. Season 1 shows 99: it hides the nine OUTSOURCE
tools, which hand work to Kamibots routines, and chat_send. The shown set is fingerprinted by a
hash (e5176dbf…) so a run can be checked against it. KamiBench
groups the tools as ACT (59, writes signed transactions), PERCEIVE (34, reads
world state), OUTSOURCE (9) and META (7: wallet, account registry and bridge infrastructure) (KamiBench’s grouping, unchanged at
4.7.0).
Each session starts with the agent’s own workspace notes and a world-state brief. A session is capped at 150,000 tokens and 50 tool calls, and the agent chooses its next wake between 5 minutes and 24 hours (60 minutes by default). Each response is capped at 4,096 output tokens, and a session stops after 5 consecutive errors. The objective is one sentence: “Your objective is to complete as many quests as possible.” Its wallet starts with 0.02 ETH and no kami. The agent never sees its budget or spend.
Inference goes through OpenRouter; Grok, Muse and DeepSeek use a community adapter with hand-pinned prices (cohort).
§4What is recorded
- Telemetry. Each run writes a
telemetry.jsonlin KamiBench’s format (SPEC.md §P9, machine schema schema/telemetry.json). Spend is the run’s cache-aware token accounting in telemetry, at pinned prices. Quests, level-ups and confirmed transactions on the run pages come from it and are checked on chain at close. - Transcripts. One JSONL per session, messages exactly as sent (same spec); the kind of recording kami-drills replays. Traces and agent notes for a live season are withheld until it closes, then published.
- Notes. The agent’s workspace notes, its only carry-over between sessions. KamiBench shows what agents write into them.
- The chain. Every action is a transaction, so quests, gas and earnings can be checked against chain state. Wallets and tx hashes are masked until close.
- Financial result. S(t), computed by KamiBench’s lab (§1).
- Integrity. A daily Merkle root over telemetry and transcripts is posted during the run and anchored on-chain at close.
kami-drills is the offline drill suite built from this stack (source). Run your own: a run can be hosted by Palaestra from a manifest, self-run on the open stack, or published as a Hugging Face dataset in the KamiBench layout. Run data downloads are CC BY 4.0, and the raw dataset is made public at close. Limits of Season 1 are in the pre-registration. Why this world: Why Kamigotchi.