Glossary

Active agent. The agent your messages and typed trades go to. Shown first in the status bar; agent <name> changes it.

Address check. An address the agent writes must have come from a tool this turn. One it wrote from memory is caught, and the agent is asked again.

Alias. A short word that stands for the start of a line, set in config.toml’s [aliases]: b5 = "buy $5 of".

Approvals. Per wallet: ask every time (every live trade is a card) or trade inside limits. Set in settings ▾ → approvals.

Bot token. What a code bot connects to the core with, in place of a key: made fresh each time the core starts it, ended when its process stops, revoked when you remove it. It reaches only that bot’s own trades and reads.

Bots folder. Where your code bots live, ~/candle-bots unless you picked another: each bot a folder of its own, beside the AGENTS.md, client and skills the Candle Harness writes there.

Build tab. Your own coding harness, signed in as you, in a terminal pane in your bots folder. It writes code bots, tests them on paper and proposes them; it isn’t given a key or a way to trade.

Candle Harness. This product: the terminal, its core and your agents. “The harness” alone is never the model’s tool: that’s an engine, or, in the Build tab, your coding harness.

Caps. Candle’s own limits on a key, held by Candle’s server whatever uses the key. See also figures and limits.

Card. A trade or other action the agent proposes, waiting for you. Priced fresh and checked against limits when it’s made. ⏎ approves, ctrl-n rejects, ctrl-o shows its details.

Chain. Up to five trades or candle commands on one line, split by ;, run one after another.

Chip. A small control drawn in a line or on a border, such as settings ▾ or ⧉. Every chip can be clicked.

Code bot. A program of yours that trades through the core with a bot token. Every trade it sends is held to the figures on its card (a trade’s, a day’s, open at most). The core runs it and restarts it after a crash. It’s a program you run: the core holds its trades, and doesn’t confine the program.

Coding harness. The AI coding tool the Build tab runs for you in a terminal pane, as you, to write bots: Claude Code, Codex, Grok, OpenCode, Pi or a command of yours. The same tools, run for your trading agent’s turns, are called engines.

Code runner. Off until you turn it on for an agent: then it runs a short script with python3, bun or node in its research folder, held there by the operating system, with no network.

Core. The one process on your machine that holds the agents’ keys and state and trades. Every view and engine attaches to it.

Embedded wallet. Your Candle account’s own wallet. The agents you allow share it.

Engine. The AI coding tool your trading agent’s turns run on: Claude Code, Codex, Grok, OpenCode or Pi, your own install, signed in as you. Picked per agent with ctrl-g.

Exposure. What the active agent’s open positions are worth, in the status bar.

Factor. What unlocks your vault for an owner step: your passphrase, or a security key.

Figures. Dollar amounts you approve that bound what runs without asking: a wallet’s two figures (a trade’s and a day’s) for trading inside limits, and a code bot’s (a trade’s, a day’s, open at most). The Candle Harness holds them. An agent’s own limits are in config.toml; Candle’s caps are on the key.

Flatten. Sells every open position of a bot, rules or code, now, at market, as its exits sell.

Figure mark. A ? after a price, market cap, liquidity, volume or balance the agent stated that no tool showed it that turn. It means “check this”.

Funds out. Sending funds out of a wallet. Only you do it, in the prompt pane, signed by the vault.

Hacc. Candle’s market data service: the signal tape, wallet and token reads, follows and alert rules, and the strategy lab.

Held exits. A take profit, stop, max hold or trailing stop the core watches on a code bot’s position, selling it with no card even while the bot is paused, crashed or held by the kill switch.

Keep-awake. On a Mac, keeps it from sleeping while a trigger or bot is on. Set in the Triggers overlay.

Kill switch. ctrl-x twice. Pauses buys, stops the agent, withdraws buy and swap cards, gives each open sell a fresh card, holds triggers’ buys, swaps and wakes, and holds bots’ entries while their exits still sell.

Linked token. The token the panes follow: the token pane shows it and Research reads it. The agent moves it by showing a token, and you by picking one.

Limits. An agent’s own per-trade and per-day limits, in config.toml, held by the Candle Harness: over one, a buy or swap is refused. See also figures and caps.

Live. Trades move real funds. The opposite of paper.

Login service. candle-harness service install: the core started at login, so triggers and bots run with no terminal open.

Memory. Each agent’s own notes, kept across turns and engines: an index and files in its research folder. Pinned lines are kept word for word.

Mint. A token’s address: the one name for a token that can’t be copied. Shown short (JUPy…vCN4), with a ⧉ that copies it in full.

Owner step. A step only the account’s owner takes, such as making a key or changing a wallet’s access. It runs in the prompt pane, with the vault’s factor first when there’s a vault.

Palette. ctrl-k: every action, tab, pane, setting and Candle command, found by typing.

Paper. Trades recorded at the quoted price and never sent. The Candle Harness starts on paper.

Pattern. One of Hacc’s kinds of signal, such as confluence (several tracked wallets buying one token) or new_ticker. Triggers and rules bots enter on one.

Pin. Holds a token pane on its token while the link moves on.

Pop-out. A pane moved into a tmux or cmux split of its own, attached to the same core. It approves nothing.

Preset. How an agent works: Quant (the default analyst), Trader or Bare.

Profile. A sign-in kept by Candle’s CLI: the Candle account the Candle Harness acts as, set at setup and kept as [candle] profile in config.toml.

Prompt pane. A pane where owner steps, passphrases and secrets are handled, as you. Nothing in it reaches the core, the agent or the Journal.

Receipt. What a trade that ran without a card leaves where its card would have been.

Rules bot. A Hacc strategy definition the core runs: it buys on Hacc’s tape when its pattern and filters match, and sells each position by its exits, with no card for its sells. Saved by your press on its card. Pro and Max.

Research folder. Each agent’s own folder of files, research/<agent>/. Its file reads are held to it.

Signal. An event on Hacc’s tape, such as a wallet buying or a token waking up.

Step cap. The most tool calls one turn may make: 30, unless set.

Tape. Hacc’s running list of signals, newest first. The Signals tab shows Hacc’s curated tape; rules bots read the full one.

TEE wallet. A wallet of its own for one agent’s key, signed in Candle’s secure enclave. Pro or Max.

Trading inside limits. A wallet setting: buys and swaps inside two figures of its own (per trade, and per day) run without a card, and every sell runs without one, whatever its size, unless its slippage bound is looser than the route’s default. The kill switch still holds its buys and swaps; sells still go.

Trigger. “When this happens, do that”, run by the core: a trade, a notification, or waking an agent. Saved by your press on its card. Pro and Max.

Turn. One run of the agent, from your message to its reply.

Vault. Keys kept encrypted on this machine, unlocked by the factor. Optional; needed for TEE wallets and funds out.

View. A Candle Harness window attached to the core: the main one, or a popped-out pane.

Woken turn. A turn a trigger started with wake, marked ⚡ woke. It trades as yours would, and can’t start a transfer out.

Window (P&L). The span every P&L figure is measured over: 24 hours, since midnight, since a time, or since your last session.