Miniqmt CLI

miniqmt-cli: Drive miniQMT / xtquant from the Command Line

Access Policy — read first

All programmatic access to the trading daemon MUST go through miniqmt-cli. Direct HTTP calls to http://127.0.0.1:8765/... are no longer a supported integration surface — the URL space, payload shapes, and error semantics are internal implementation details and may change without notice.

Reasons:

• One source of truth. The CLI's command set is the only documented API. The HTTP routes underneath exist solely to let the CLI talk to the daemon over an SSH tunnel; their paths are unstable.
• Drift kills callers. When a parallel HTTP reference existed, a polling agent followed a stale /positions / /orders / /trades (unprefixed) path and silently 404'd for hours. The CLI shields callers from this class of bug.
• Safety pipeline. The CLI carries the masking, idempotency client_req_id generation, and --confirm-live handshake conventions. Bypassing it weakens those guarantees even though the daemon enforces most of them.

Callers from non-Python runtimes should wrap the CLI as a subprocess (subprocess.run(["miniqmt-cli", "account", "asset", "--account", "main", "--format", "json"], capture_output=True)) rather than constructing HTTP requests by hand. SSE consumers should use miniqmt-cli stream tick|kline|order ... and parse its stdout, not subscribe to /stream/* directly.

If you're an agent on a different host than the daemon, configure remote mode in ~/.miniqmt_cli/client.toml:

[client]
mode = "remote"
server_url = "http://127.0.0.1:8765"   # the daemon, via SSH tunnel

Then bring up the tunnel (see SSH Tunnel below). The CLI then works transparently — no proxy config needed.

Architecture

Mac CLI (Click) --> SSH tunnel --> Windows FastAPI daemon (port 8765) --> xtquant/miniQMT

• Client (Mac): ~/.miniqmt_cli/client.toml -- sends HTTP requests through an SSH tunnel
• Server (Windows): ~/.miniqmt_cli/server.toml -- FastAPI daemon wrapping xtquant, runs as a Scheduled Task
• Communication: all commands hit http://127.0.0.1:8765 via ssh -N -L 8765:127.0.0.1:8765 <host>

Prerequisites

Before running any command, verify:

1. SSH tunnel is up -- without it, all commands fail with "cannot reach daemon"
2. Daemon is running on Windows -- check with miniqmt-cli health
3. miniQMT client is open on Windows -- required for xtquant to connect to the broker

/health response shape

Two independent top-level blocks. There is no flat state enum — callers compose decisions from the structured fields. (The old composite strings like daemon_up_no_trader mixed two unrelated concerns into one name and have been removed.)

{
  "daemon": {
    "state": "up",              // up | degraded
    "xtquant_loaded": true,
    "xtquant_error": "..."      // present only when state == degraded
  },
  "accounts": {
    "main": {
      "trader": {
        "state": "alive",       // never_connected | alive | lost
        "last_connect_at": "2026-05-13T09:30:01Z",
        "last_disconnect_at": null
      },
      "risk_breaker": "ok",     // ok | tripped
      "baseline": "captured"    // captured | pending
    }
  }
}

Scope of each field:

• daemon.state is purely process-local: this Python process is up + xtquant module loaded. It says nothing about miniQMT or the broker.
• accounts.<name>.trader.state reflects what the daemon last heard on the xtquant SDK channel — alive after a successful login, lost after xtquant fires on_disconnected. The daemon cannot independently probe miniQMT's GUI or broker connectivity, so a lost flag is a positive signal (SDK told us the channel dropped), but a stale alive flag is not a guarantee the channel is currently up. The only definitive liveness probe is to run an account command — a real broker failure surfaces as trader.connect failed rc=... or trader.subscribe failed rc=....
• risk_breaker and baseline are computed from the daemon's risk manager.
• When the daemon was started with serve --dry-run, the daemon block adds "dry_run": true and xtquant_loaded is forced to false (xtquant is never touched in that mode). All account commands raise; this mode is for offline development only.

Reading rules of thumb:

miniqmt-cli health prints a human summary of this body. Exit code rules:

• Exit 1 when any of these are true: daemon unreachable (cannot reach daemon …); daemon.state != "up" (e.g. xtquant failed to load); any accounts.<name>.trader.state == "lost"; any accounts.<name>.risk_breaker == "tripped"; any accounts.<name>.baseline == "pending" while that account's trader is alive.
• Exit 0 otherwise. In particular, a fresh daemon where every account is never_connected exits 0 — that's the normal lazy-load state, not an error.

Global Options

miniqmt-cli [--format table|json|csv] [--config <path>] <command>

Default format comes from client.toml; override per-call with --format.

Commands Reference

Market Data

# List sectors (e.g. 沪深A股, 上证50)
miniqmt-cli sector list

# List instruments in a sector
miniqmt-cli instrument list --sector "沪深A股"
miniqmt-cli instrument list --limit 10

# Instrument detail
miniqmt-cli instrument info --code 000001.SZ

# Latest tick snapshot (supports multiple codes)
miniqmt-cli tick --code 000001.SZ --code 600519.SH

# Historical K-line (periods: 1d, 1m, 5m)
miniqmt-cli kline --code 000001.SZ --period 1d --start 20260101 --end 20260415

# Historical ticks (raw tick-by-tick data)
miniqmt-cli ticks --code 000001.SZ --start 20260415093000 --end 20260415100000

Real-time Streaming (SSE)

Streams run until Ctrl+C. Output is one event per line; use --format json for programmatic consumption (the table format is for ad-hoc inspection).

Known failure mode: xtquant snapshot cache wedge. The timetag field of xtdata.get_full_tick() can lock at a past value while lastPrice / bidVol / askVol continue updating normally. Observed 2026-05-18: four codes (600760.SH / 600011.SH / 000333.SZ / 002179.SZ) all locked at timetag=20260518 09:32:06 for 13 minutes, daemon process otherwise healthy, watchdog silent, all HTTP requests returning 200, trader=alive. The daemon is a pure pass-through to xtquant — grep timetag across the daemon source returns zero hits — so the wedge lives inside xtquant's process-internal snapshot cache, not in daemon code.

Empirically supported trigger: daemon uptime crossing a market-close boundary. Access-log analysis of the 5/18 incident window (09:20–09:59) showed zero stream tick activity from any client, stable ~150 req/min on /trade/* polling, zero non-200 responses, zero WARN/ERROR lines. The only standing variable was a 64-hour daemon uptime that had spanned the weekend close. Best fit: xtquant's internal timetag write path has weak day-rollover handling. An earlier hypothesis that short-lived stream tick subscribe/unsubscribe cycles trigger the wedge is not supported by the logs and should not be relied on.

Primary mitigation: nightly daemon restart on trading days. scripts/windows/bootstrap.ps1 registers a MiniqmtDaemonNightlyRestart Scheduled Task that runs scripts/windows/nightly-restart.ps1 at 08:50 Mon-Fri — one known-clean xtquant process per day eliminates the cross-close failure mode. Polling via /data/tick reads the same xtquant cache that stream tick populates, so polling does not recover a wedged cache — only a daemon restart does.

When to use stream tick vs tick polling. Persistent stream tick is the recommended consumer pattern (lower latency than respawning miniqmt-cli tick per evaluation, no subprocess startup cost). One subscription per consumer process for the whole trading day; do not subscribe/unsubscribe per evaluation. This is a performance/ergonomics recommendation — it does not by itself prevent the wedge above; pair it with the nightly restart.

Tick stream

# Interactive (Ctrl+C to stop)
miniqmt-cli stream tick --code 000001.SZ --code 600519.SH

# Programmatic consumption
miniqmt-cli --format json stream tick --code 000001.SZ --code 600519.SH

Event payload (JSON mode). First line is the subscription envelope; subsequent lines each wrap one tick event. Codes that do not tick in an interval emit nothing for that interval.

{"event": "subscribed", "codes": ["000001.SZ"], "seqs": [42]}
{"tick": {"code": "000001.SZ", "timetag": "20260518 09:32:06", "lastPrice": 10.48,
          "high": 10.52, "low": 10.41, "open": 10.45, "lastClose": 10.40,
          "volume": 12345, "amount": 128943.50,
          "bidPrice": [10.47, 10.46, 10.45, 10.44, 10.43],
          "askPrice": [10.48, 10.49, 10.50, 10.51, 10.52],
          "bidVol":   [100, 200, 300, 400, 500],
          "askVol":   [150, 250, 350, 450, 550]}}
{"tick": {...}, "dropped": 3}

• event: subscribed arrives exactly once at stream start — consumers must skip it before processing ticks.
• dropped: N appears alongside a tick when the daemon's per-stream queue overflowed since the last emit (cap = 256). The dropped events are the oldest; the emitted tick is the freshest. Log it as a backpressure warning, not an error — it means your consumer is slower than the broker pushes.
• Field set varies by xtquant SDK version. code, timetag, lastPrice, high, low, bidVol, askVol are present on standard A-share L1 feeds; anything else is best-effort.
• timetag is a "YYYYMMDD HH:MM:SS" string; parse before using as a sort key. If you see timetag static while other fields advance, you've hit the cache wedge — restart the daemon.

Consuming stream tick from a long-running subprocess. The boring details that bite:

1. bufsize=1 + env={..., "PYTHONUNBUFFERED": "1"} on Popen. Without both, Python's stdio block-buffers the daemon's per-line output into ~4 KB chunks; consumers see silent stretches where ticks have been emitted but are stuck in the pipe.
2. Drain stderr in a separate thread. miniqmt-cli rarely writes to stderr, but if the pipe buffer fills (tens of KB) the subprocess blocks on write(stderr) and the tick stream goes silent without any visible error.
3. No SSE keepalive from the daemon. During a quiet stretch the daemon emits no bytes — a wedged daemon and a quiet market look identical on the wire. Consumers MUST run their own stale watcher: if no tick has arrived for N seconds in trading hours (~30s for liquid watchlists), kill the subprocess and reconnect. Trigger on "no event from ANY watched code" rather than per-code so a single illiquid name doesn't false-positive.
4. Reconnect with exponential backoff (1s → 30s cap). SSH-tunnel blips and daemon restarts will close the SSE stream; resetting backoff after one successful tick keeps reconnects fast in steady state.
5. SIGTERM bypasses the CLI's finally block. miniqmt-cli stream tick only reaches its unsubscribe path on KeyboardInterrupt (SIGINT). For long-running consumers this is intentional — not unsubscribing avoids the cache wedge above. Use SIGINT only when you actually want a clean unsubscribe (testing/debug); for normal restart, SIGTERM is fine.
6. The tick callback must be near-instant. Block the read loop and the daemon's per-stream queue (cap 256) fills, you start seeing dropped: N, and the freshest tick is all you get. Push ticks to a worker queue immediately; evaluate strategies elsewhere.

Kline stream

miniqmt-cli stream kline --code 000001.SZ --period 1m

Each event wraps one bar: {"bar": {"code": ..., "time": ..., "open": ..., "high": ..., "low": ..., "close": ..., "volume": ...}}. The daemon coalesces by (code, bar_start_ts) and yields only the latest bar per key, so a slow consumer never sees duplicate stale bars.

Order stream

# Stream order lifecycle events (submitted / partially filled / filled / cancelled / rejected)
# Essential for agents: subscribe before placing an order, then consume fill/reject events.
miniqmt-cli stream order --account sim

# JSON format for programmatic parsing
miniqmt-cli --format json stream order --account sim

Event payload shape (JSON mode). The first line after subscription is an envelope, then each order state change emits an order_status dict:

{"event": "subscribed", "filter_account": "sim"}
{"type": "order_status", "account": "sim", "order_id": 12345, "code": "000001.SZ",
 "side": "buy", "status": "filled", "volume": 100, "filled_volume": 100,
 "avg_price": 10.48, "frozen": 0.0}

Also emitted on the same stream: {"type": "order_response", ...} (async submit ack) and {"type": "trade", ..., "trade_id": ..., "price": ..., "amount": ...} (per-fill detail).

Possible status values (from xtquant order_status): submitted, confirmed, partially_filled, filled, cancelled, rejected, expired, pending_cancel, unknown. Unknown codes are returned as unknown_<n> — agents should have a default branch.

Account & Portfolio

# List configured accounts (names + masked IDs)
miniqmt-cli account list

# Query account asset (cash, total, frozen)
miniqmt-cli account asset --account sim

# Query positions
miniqmt-cli account position --account sim

# Today's orders
miniqmt-cli account orders --account sim

# Today's trades (fills)
miniqmt-cli account trades --account sim

Trading (Three-Layer Safety)

Terminology — "live account" (实盘账户): Any account with requires_confirm_live = true in server.toml. This is a property, not an account name. Set to false for sim/paper accounts. In the examples below, sim is a paper account and real is a live account.

Where --confirm-live <last-4-digits-of-account_id> is required on live accounts (current behavior):

Command	--confirm-live required on live?	Enforced by
order buy / order sell (incl. --dry-run)	Yes	CLI (order.py) + daemon (/trade/order)
risk reset	Yes	CLI (risk.py) + daemon (/risk/reset)
order cancel	No — the CLI command has no --confirm-live flag and /trade/cancel does not re-check the live gate. Cancels are still gated by the daemon account whitelist.	—

Orders go through three independent safety checks:

1. CLI layer: --dry-run preview, interactive "yes" confirmation, --confirm-live digit match
2. Daemon layer: account whitelist, live-gate re-verification, idempotency dedup
3. Audit layer: every order logged to ~/.miniqmt_cli/orders.jsonl with pre/post phases

# Buy -- preview only (no order sent)
miniqmt-cli order buy --account sim --code 000001.SZ --volume 100 --price 10.50 --dry-run

# Buy -- with interactive confirmation
miniqmt-cli order buy --account sim --code 000001.SZ --volume 100 --price 10.50

# Buy -- skip confirmation (scripting)
miniqmt-cli order buy --account sim --code 000001.SZ --volume 100 --price 10.50 --yes

# Sell
miniqmt-cli order sell --account sim --code 000001.SZ --volume 100 --price 11.00 --yes

# Live (real-money) account requires last-4-digit verification
miniqmt-cli order buy --account real --code 000001.SZ --volume 100 --price 10.50 --confirm-live 1234

# Cancel an order
miniqmt-cli order cancel --account sim --order-id 12345 --yes

Order types: --type limit (default) or --type market.

Complete Trading Workflow (agent-facing)

Follow these steps in order. Each step has a purpose — skipping is how agents end up with "order placed, no idea what happened".

1. Pre-flight health — probe real connectivity, not just /health.

/health showing accounts.<name>.trader.state == "never_connected" is normal on a fresh daemon and is NOT a login failure (see the Prerequisites section). To confirm the account is actually usable, hit a trader-touching endpoint:

miniqmt-cli --format json account asset --account sim

Success (cash / total_asset numeric) proves: daemon up → xtquant loaded → trader connected → account subscribed. After this, accounts.<name>.trader.state flips to alive and baseline to captured.

If this fails with trader.connect failed rc=... or trader.subscribe failed rc=..., stop — investigate before sending any order.

2. Check risk state.

miniqmt-cli --format json risk status --account sim

If breaker_tripped: true, opening orders will be rejected at the daemon layer regardless of CLI confirmations. Closing trades and cancels still work. Resolve with risk reset (requires a --note) before opening.

3. Preview first (dry run).

miniqmt-cli order buy --account sim --code 000001.SZ --volume 100 --price 10.50 --dry-run

Dry-run hits /trade/preview (no order sent), shows masked account id, last price, and estimated cost. Exit code 3 (GuardExit: "dry-run: order not sent") is the success path. Agents should read the preview numbers to catch typos (volume unit is shares not lots; price is yuan per share).

4. Place the order, wait for a terminal status in one call.

miniqmt-cli --format json order buy --account sim --code 000001.SZ \
    --volume 100 --price 10.50 --yes --wait 30

--wait N blocks up to N seconds after the POST returns, subscribing to /stream/order and waiting for a terminal status (filled / cancelled / rejected) on the returned seq. The subscription happens after submit, so fast fills can theoretically race — in practice miniQMT fills take enough time that --wait catches them, but early lifecycle events (submitted, confirmed) may be missed. For zero-loss event capture, use the separate subscribe/place pattern below.

• The JSON response carries {"seq": <order_id>, "status": "ok"|"rejected", "client_req_id": ...}. Persist seq — you need it to cancel or to correlate stream events.
• client_req_id is a UUID the CLI auto-generates for idempotency. Re-submitting with the same id within the TTL window returns the original response with "idempotent_hit": true instead of placing a duplicate.
• Note on --format json output: the order command still prints human-readable preview lines (Account:, Code:, Side:, etc.) to stdout before the final JSON line. Parsers must read the last line, e.g. ... | tail -1 | jq -r '.seq'.

If you want to run the submit and the event loop separately (e.g. long-running agent) to avoid missing early events:

# Terminal A — subscribe BEFORE placing the order
miniqmt-cli --format json stream order --account sim > events.jsonl &

# Terminal B — place the order, capture seq from the JSON response (last line)
SEQ=$(miniqmt-cli --format json order buy --account sim --code 000001.SZ \
      --volume 100 --price 10.50 --yes | tail -1 | jq -r '.seq')

# Watch events.jsonl for order_status where order_id == $SEQ

Subscribing before placing is the only way to guarantee you see the full lifecycle (submitted → confirmed → filled). If the stream subscriber is registered after the daemon has already dispatched an event, that event is lost for this subscriber (the daemon fans out only to currently-registered subscribers; see session.py dispatch_order_event).

5. Verify the outcome.

miniqmt-cli account orders --account sim   # all of today's orders
miniqmt-cli account trades --account sim   # all of today's fills
miniqmt-cli account position --account sim # current positions

The broker is the source of truth, not the stream. If a stream event is missed (connection blip, subscriber queue overflow — drops are logged server-side), these queries reconcile.

Cancelling

Know which scenario you're in before cancelling:

Cancel takes the same form for sim and live accounts — the CLI has no --confirm-live flag on cancel (see the Terminology table above for the full matrix). Whitelist enforcement still happens at the daemon layer.

miniqmt-cli order cancel --account sim --order-id 12345 --yes
miniqmt-cli order cancel --account real --order-id 12345 --yes

# JSON output for scripting
miniqmt-cli --format json order cancel --account sim --order-id 12345 --yes

Cancel flow:

1. Get order-id from the place response (seq field) or from account orders.
2. POST /trade/cancel returns {"status": "ok", "seq": <cancel_seq>} synchronously — this is the submit ack, not confirmation the order is cancelled.
3. Watch /stream/order for order_status where order_id == <original order_id> and status == "cancelled". The actual cancel can take a moment, and the broker may reject if the order filled first.
4. Reconcile with account orders if you need certainty.

Cancels are also idempotent via client_req_id (CLI auto-generates). Re-running the same cancel command without changes creates a new client_req_id, so it will hit the broker again — use the HTTP API with a stable client_req_id if you need true idempotency across retries.

Error Handling

Risk rejections come back as HTTP 400 with body {"error": "risk_reject", "code": "<limit_name>", "message": "..."} — in CLI they surface as exit 1 with the message. Agents using HTTP directly should branch on code (e.g. max_daily_loss, max_position_pct, max_orders_per_minute).

Risk Control

The daemon enforces risk limits independently (v0.2.0+). When a limit trips, the breaker enters block-open-allow-close mode: new opening orders are rejected, but closing / cancel operations still work.

# Show risk status for one account (baseline, PnL, breaker state, pending orders)
miniqmt-cli risk status --account sim

# Show all accounts
miniqmt-cli risk status

# JSON for agents
miniqmt-cli --format json risk status --account sim

# Reset the breaker (operator action — requires a justification note)
miniqmt-cli risk reset --account sim --note "false positive: baseline re-captured"

# Live (real-money) account reset requires last-4-digit confirmation
miniqmt-cli risk reset --account real --note "manual unfreeze" --confirm-live 1234

risk status fields of interest:

server.toml [risk] defaults:

[risk]
enabled = true
max_daily_loss = 50000          # yuan
max_position_pct = 30           # % of total asset per single stock
max_orders_per_minute = 10
max_positions = 10

Per-account overrides live in [accounts.<name>.risk].

Daemon Management

# Check versions (local + remote)
miniqmt-cli version

# Health check
miniqmt-cli health

# Start daemon locally on Windows (normally managed by Scheduled Task)
miniqmt-cli serve [--host 127.0.0.1] [--port 8765] [--dry-run]

Configuration

# Client config
miniqmt-cli config client init          # Create template ~/.miniqmt_cli/client.toml
miniqmt-cli config client show          # Print resolved config
miniqmt-cli config client set-server-url http://127.0.0.1:8765

# Server config (run on Windows)
miniqmt-cli config server init          # Create template ~/.miniqmt_cli/server.toml
miniqmt-cli config server show          # Print config (account IDs masked)

Deployment (Mac to Windows)

The setup wizard walks through 9 steps: parameters, client.toml, SSH check, remote Python, Windows service registration, server.toml, code deploy, SSH tunnel, smoke test.

# Full wizard (idempotent, remembers progress)
miniqmt-cli setup

# Re-run a specific step (1-9)
miniqmt-cli setup --step 9

# Start fresh
miniqmt-cli setup --reset

For day-to-day code updates after initial setup:

# Deploy script (uses env from wizard state)
WIN_HOST=<user>@<windows-host> WIN_REPO=C:/apps/trading-skills bash scripts/deploy.sh

Restart the daemon on Windows:

ssh <user>@<windows-host> "schtasks /run /tn MiniqmtDaemon"

SSH Tunnel

The tunnel is the lifeline between Mac CLI and Windows daemon.

# Manual tunnel (foreground)
ssh -N -L 8765:127.0.0.1:8765 <user>@<windows-host>

# Persistent tunnel via autossh + launchd (setup wizard can generate the plist)
launchctl load ~/Library/LaunchAgents/com.miniqmt.tunnel.plist

Config File Reference

client.toml (~/.miniqmt_cli/client.toml)

[client]
mode = "auto"                         # auto | local | remote
server_url = "http://127.0.0.1:8765"  # required when mode = remote

[client.output]
format = "table"

Env overrides: MINIQMT_CLI_MODE, MINIQMT_CLI_SERVER_URL, MINIQMT_CLI_FORMAT

server.toml (~/.miniqmt_cli/server.toml, on Windows)

[server]
host = "127.0.0.1"
port = 8765
qmt_path = "C:/国金QMT交易端/userdata_mini"

[accounts.sim]
account_id = "1230001"
account_type = "STOCK"

[accounts.real]                       # real-money account; name is arbitrary
account_id = "1230002"
account_type = "STOCK"
requires_confirm_live = true          # marks this as a "live account" — extra confirmation required

[audit]
log_path = "~/.miniqmt_cli/orders.jsonl"

Env overrides: MINIQMT_CLI_SERVER_HOST, MINIQMT_CLI_SERVER_PORT, MINIQMT_CLI_SERVER_QMT_PATH

Troubleshooting

Exit Codes

Exit code 5 is the contract for state-changing operations (place/cancel) that didn't get a definitive response within the submit timeout. Scripts that branch on exit code should treat it as fundamentally different from exit 1: code 1 means "we know it failed", code 5 means "we don't know whether the broker accepted it". A naive retry on code 5 risks doubling the position.

Related Skills

• trading-analysis — Money flow + signals built on top of miniqmt-cli ticks