delegate-autoquant
GitHub将定量研究任务委派给AutoQuant智能体,涵盖因子分析、组合回测及稳健性验证。需明确决策目标、数据源与约束,通过CLI发起持久化项目并获取证据报告,避免直接交易或猜测结果。
Trigger Scenarios
Install
npx skills add TraderAlice/OpenAlice --skill delegate-autoquant -g -y
SKILL.md
Frontmatter
{
"name": "delegate-autoquant",
"description": "Delegate reproducible quantitative research from a general Chat Workspace to the durable AutoQuant desk and consume its evidence-bound handoff. Use when a request needs a quantitative Project, historical data intake, factor or portfolio research, backtesting, governed experiments, robustness evidence, or an AutoQuant follow-up. Do not use for ordinary market commentary, broker state, order execution, or a question Chat can answer directly."
}
Delegate quantitative research to AutoQuant
Treat AutoQuant as a specialist coworker, not a function call. OpenAlice owns the conversation and routing; the AutoQuant Agent owns Project selection, research method, dependencies, experiments, Git history, and evidence.
Bound the assignment
Before dispatch, preserve the caller-owned parts of the question:
- the decision the research should support;
- assets or universe, direction, horizon, and cadence when the caller knows them;
- available data and its provenance;
- benchmark, costs, constraints, risk limits, and requested deliverable;
- material unknowns that must be clarified rather than invented.
Do not translate the request into AutoQuant CLI steps, choose a Project or Study on its behalf, promise a positive result, or grant trading authority. A negative result, unsupported route, or request for clarification can be a correct handoff.
Recruit the default desk
Use the alice-workspace collaboration surface:
alice-workspace conversation ask --harness autoquant --await --prompt '
Research question: <question>
Decision this supports: <decision>
Caller-owned scope and constraints: <assets, direction, horizon, cadence, benchmark, costs, limits>
Available evidence or data: <paths and provenance, or say what is missing>
Expected handoff: answer in plain language; name the Project and every applicable immutable Run, Experiment, Report, or Dossier id; give the absolute Project root and end with `Primary deliverable directory: <absolute path>` pointing to the Dossier bundle, Report bundle, or Project directory that contains the named evidence; distinguish validation, visible-test, and external-holdout evidence; state assumptions, unsupported claims, and that trading authority is none.
Ask before proceeding if a missing caller-owned fact would materially change the study. Do not manufacture a Report or Dossier solely for transport.
'
Add --await when the current turn needs the answer. For longer delegation,
omit it, retain the returned taskId, and use conversation await, read, or
collect as described by the alice-workspace skill. There is no unsolicited
Agent-to-Agent completion notification bus. AutoQuant assignments deliberately
have no default runtime deadline; add --timeout-ms only when the caller has
chosen a real hard stop for that particular study.
If AutoQuant is not initialized, report that boundary and direct the user to initialize the Quant desk. Do not silently perform the research in Chat or guess another Workspace.
Understand the returned state
The universal result is the Agent's ordinary assistantText handoff, not one
mandatory file type:
| Research route | Applicable durable evidence |
|---|---|
| Fixed or descriptive Study | Usually an immutable Run and Explorer; no Session Report is required |
| Editable single lane | Runs/Experiments and often one immutable lane Report |
| Multi-lane research program | Current lane Reports and, when applicable, one Project Dossier |
| Blocked, rejected, or unsupported study | Plain-language boundary plus whatever exact evidence ids exist |
A Report is an immutable Markdown/JSON deliverable for one research lane. A Dossier synthesizes compatible current lane Reports into a Project-level deliverable. Their absence is not failure when the route does not support or need them. AutoQuant does not automatically publish either artifact to the OpenAlice Inbox.
Consume and continue
-
Read the direct handoff first. Require the Project identity and absolute root, the primary deliverable's absolute directory, applicable evidence ids, assumptions, limitations, and no-trading boundary.
-
If the absolute deliverable directory is missing, continue the returned Session immediately instead of searching the desk or recruiting another worker:
alice-workspace conversation ask --resume-id <resumeId> --await \ --prompt 'Please provide the absolute directory path of the primary Report, Dossier, or Project evidence from your completed assignment. Do not rerun the research.' -
Use native Read/Search/Git capabilities directly on the returned directory and only the named artifacts. Use
alice-workspace peer path --id <workspaceId>only as an addressing fallback when the reported absolute path is unavailable. -
If the answer needs substantive clarification, continue that same
resumeId; do not recruit a new AutoQuant Session and discard its context. -
Translate the evidence for the user, separating AutoQuant's findings from your judgment. Any live account or execution decision returns to
alice-utaand its approval flow.
A normal attended Chat reply already reaches the user. Ask AutoQuant to commit and push an exact file to Inbox only when the user explicitly wants a durable human-facing delivery or asynchronous notification; do not use Inbox as the ordinary peer reply channel.
Version History
- 0f89c70 Current 2026-08-20 08:03


