Harness Tools

The harness's operations are tools on the MCP server harness (Claude Code
shows them as mcp__harness__<tool>, other harnesses by server and tool
name). Each runs the same code as a script in <root>/scripts/ and returns
what that script prints, except that the results that can be large come back
compact by default (the scan as a summary, recent fires as lines, JSON on one
line), with an argument for the full form. Agents call the tools, in
autonomous and interactive runs alike; a person at a terminal runs the
scripts. None places, changes, or cancels an order.

When to Use

• Any time a skill or role needs bars in a file, a strategy scan, indicator
  numbers, the multi-timeframe read, the news blackouts, the prop attempt's
  state, the journal's lessons, or a day's reconcile.
• When a step says to record something for the order gate (record: true).
• To know which script a person would run for the same result.

How It Works

• bars and daily are bars files by absolute path (no ~): the runner's bars
  file (the prompt names it, under the bar directory) or a bars_fetch
  file under <FTH_HOME>/bars-scratch (use the path bars_fetch returns:
  each harness home, a replay's run home included, has its own, and
  /tmp/fth is no longer read). Nothing else is read: another path, a symlink, or
  a file that isn't .json or .csv is refused (Parquet and Excel files
  are the user's, for backtests).
• symbol is a contract root in capitals (MNQ, MES).
• No tool takes a time: the harness clock decides (a replay broker's time
  in a replay). No tool takes an output path or an environment.
• An unknown argument is refused, with the arguments the tool takes.
• Give numbers as JSON numbers ("timeframe": 3) and flags as true or
  false. A number or flag sent as a string ("3", "true") is read as
  the number or flag the tool takes (surrounding spaces ignored), then
  checked the same way (a timeframe of "90" is still refused); "1.5"
  where a whole number is taken, or any other string, is not.
• A refusal or failure comes back as an error result: [<script>] <why>,
  as the script prints it. Read it, don't retry the same call blindly.
• A reply over 256 KB is refused with how to narrow it (fewer bars, a lower
  count, no json or detail). Read a result from the reply itself: some
  harnesses move a long reply to a file you may not be allowed to read, so
  ask for the compact form (or name for one strategy), not the full one.
• In an autonomous run record is ignored, not refused: the call succeeds
  with its normal result plus a line [<script>] note: record ignored - the autonomous runner records .... The result is correct; nothing was
  recorded because the runner records every bar's read and signals itself.
  Don't retry or doubt the read over it; just leave record off. Outside an
  autonomous run, record is refused whenever the autonomous runner is
  running, for the same reason.

Only the user's commands stay scripts: starting, stopping, or recording a day
of a prop attempt (combine.js start | stop | record-day), backtests,
and replays. In a trading run (autonomous or interactive) ask the
user to run those. The one exception is authoring: the strategy-researcher
role, which the user starts interactively for a new strategy, may run a
backtest itself (the strategy-authoring skill); no trading role does.

Examples

bars_fetch {"symbol": "MNQ", "timeframe": 15, "count": 160}
→ wrote 160 15-minute bars for MNQ (tick 0.25, $0.5/tick) to /home/u/.futures-trading-harness/bars-scratch/MNQ-15m.json; last bar 2026-10-07T13:45:00.000Z, closed 2026-10-07T14:00:00.000Z

market_snapshot {"bars": "/home/u/.futures-trading-harness/bars-scratch/MNQ-15m.json"}
strategies_scan {"bars": "/home/u/.futures-trading-harness/bars/MNQ-3m.json", "symbol": "MNQ"}
→ Scan of MNQ on the last closed bar (opened 2026-10-07T13:57:00.000Z): 16 strategies, 1 candidate. Regime trend-up. Trend 4h up, 1h up, 15m up, prevailing 4h up.
  Candidates:
    orb long | active, 3m, rules, mtf trend | entryRef 21503.25, stop 5.25 pts = 21 ticks ($10.50/MNQ contract), target 1.5R = 31 ticks ($15.50/MNQ contract), minRR 1.5, exit: target 1.5R | with: none; against: none
  Exact order numbers (stop and target prices and ticks, size, $ risk, place_order arguments) for a strategy and side you choose: order_plan.
  Every strategy (name | status | timeframe | direction | candidate | why not):
    bos | active | 3m | - | no | no rule set fired (long: close crosses_above swing_high; short: close crosses_below swing_low)
    ...
strategies_scan {"bars": "/home/u/.futures-trading-harness/bars/MNQ-3m.json", "symbol": "MNQ", "name": "orb"}
→ orb's full result as JSON
order_plan {"bars": "/home/u/.futures-trading-harness/bars/MNQ-3m.json", "symbol": "MNQ", "strategy": "vwap_reclaim", "side": "long", "accountId": 7}
→ vwap_reclaim long: 1 MNQ at market (entry ref 21503.25) | stop 21498.00, below the entry: 5.25 pts = 21 ticks ($10.50/MNQ contract), stopLossBracket.ticks 21 | target 21513.75 (2R): 42 ticks, takeProfitBracket.ticks 42 | risk $10.50 = 1 x 21 ticks x $0.50/tick (before fees) | size 1: the minimum; no account budget sizes a rules strategy (position-sizing: size 1 until the setup has 30+ reviewed trades)
  place_order arguments: {"accountId":7,"contractId":"MNQ","side":"buy","type":"market","size":1,"stopLossBracket":{"ticks":21,"type":"stop"},"takeProfitBracket":{"ticks":42,"type":"limit"},"rationale":"setup:vwap_reclaim long at market from 21503.25 (the signal bar's close), stop 21498.00 (21 ticks), target 21513.75 (42 ticks), 1 MNQ risks $10.50"}
mtf_read {"bars": "/home/u/.futures-trading-harness/bars-scratch/MNQ-3m.json", "daily": "/home/u/.futures-trading-harness/bars-scratch/MNQ-1d.json"}
blackouts_add {"start": "2026-10-07T12:25:00Z", "end": "2026-10-07T12:40:00Z", "reason": "CPI 08:30 ET"}
combine_status {}

A person does the same at a terminal: node <root>/scripts/bars.js --symbol MNQ --timeframe 15 --count 160.