DAXQO Trajectory Collector

Local GUI action recording, privacy filtering, replay and dataset export for any harness.

本地 GUI Agent 轨迹采集工具，可通过 Python API、命令行或 MCP 接入不同 harness。数据处理无需运行 DAXQO 或 Motor；Android 原生观察器另有明确的 bridge 依赖。

• Python 3.10+, standard-library-only runtime, MIT license.

• Durable JSONL + SQLite, conflict detection, idempotent ingestion and leakage-aware exports.

• Sensitive UI/input values and hidden reasoning filtered before persistence.

• No model provider, cloud upload or training service included.

Install

python -m pip install "git+https://github.com/axiomzero-qed/DAXQO-Trajectory-Collector.git@v0.1.0"

From a source checkout: python -m pip install .. No sibling checkout is needed. Motor is optional.

Related MCP server: apk-preview-mcp

Use from another harness

from pathlib import Path
from trajectory import Dataset, Recorder, snapshot

with Dataset(Path("./data")) as dataset:
    writer = Recorder(dataset,
        device_id="b683d45a-e298-40b3-90c2-3bcd8f59cb84",
        session_id="7468c2e4-dae4-4d3b-b959-b6e671f6f951",
        source="my_harness")
    receipt = writer.record_action(
        task_id="example-task", action_id="example-action",
        action={"type": "click", "target": {"resource_id": "details"}},
        before=snapshot("demo.app", [{"text": "Details"}]),
        after=snapshot("demo.app", [{"text": "Demo details"}]),
        status="succeeded",
        started_at="2026-10-24T10:00:00Z",
        finished_at="2026-10-24T10:00:01Z")

This records synthetic observations and performs no device action. In production, supply observations/outcomes from the existing executor and assign your own stable device UUID. Keep the device/session/action IDs, source, timestamps and contents unchanged when retrying; conflicting retries are rejected.

Reported success remains an unverified agent claim. Independent human/fixture verification is required for positive training examples by default.

CLI and MCP

daxqo-collect --data ./data stats
daxqo-collect --data ./data import ./actions.jsonl
daxqo-collect --data ./data replay --output ./dist/replay.html
daxqo-collect --data ./data export --output ./exports/scoring --split-by task
python -m trajectory.mcp_server --data ./data --device-id b683d45a-e298-40b3-90c2-3bcd8f59cb84

MCP tools: trajectory_record, trajectory_stats, trajectory_export. They record supplied actions; they never execute phone actions. See harness integration, host config and schema.

Optional Android observer

USB collection needs a compatible DAXQO Android bridge, an authorized ADB connection and the owner's existing phone-control permissions. No root is required. The core API/MCP path above needs no ADB, APK or DAXQO harness.

daxqo-collect --adb /path/to/adb --serial DEVICE_SERIAL devices
daxqo-collect --adb /path/to/adb --serial DEVICE_SERIAL --data ./data start --task "Browse the fixture" --allow-app io.daxqo.trajectoryfixture
daxqo-collect --adb /path/to/adb --serial DEVICE_SERIAL --data ./data watch

Coordinate one ACK receiver and one explicit Dataset root per device stream. Replay only inspects records; it never repeats historical actions. See Android build/integration. Native observer sources are canonical in this repository.

Development

python -m pip install ".[dev]"
python -m unittest discover -s tests
python -m build
python examples/mcp_smoke.py

The independent official MCP client smoke test uses synthetic inputs and an isolated working directory. Historical device results do not replace acceptance on a new device/harness.

Source releases exclude device data, credentials, generated reports, APKs and website sources. See dataset behavior.