edge-pipeline-orchestrator

---
name: edge-pipeline-orchestrator
description: Use when coordinating the full edge research pipeline end-to-end, from candidate detection through strategy design, review, revision, and export.
category: Strategy R&D
version: 1.0.0
tools: []
---

# Edge Pipeline Orchestrator

Coordinate all edge-research stages into a single automated pipeline run.

## When to Use

- Run the full edge pipeline from tickets (or raw OHLCV) through to exported strategies.
- Resume a partially completed pipeline from the drafts stage.
- Review and revise existing strategy drafts through a feedback loop.
- Dry-run the pipeline to preview results without exporting.

## Methodology

1. Establish the pipeline configuration (source tickets or OHLCV, output location, run mode).
2. If starting from raw OHLCV, run the auto-detect stage first to generate tickets.
3. Run the hints stage to extract edge hints from the market summary and anomalies.
4. Run the concepts stage to synthesize abstract edge concepts from tickets and hints.
5. Run the drafts stage to design strategy drafts from concepts.
6. Run the review-revision feedback loop: review all drafts (maximum 2 iterations); accumulate PASS and REJECT verdicts; any REVISE verdict triggers a revision pass and re-review; anything still REVISE after the iteration cap is downgraded to a research_probe rather than looping forever.
7. Export only drafts that are PASS, flagged `export_ready_v1`, and belong to an exportable entry family.
8. Record a full execution trace in a run manifest for auditability.

Optional controls: an as-of date to forward to the hints stage for date filtering; a strict-export mode where any warn-level finding on an otherwise export-eligible draft downgrades it to REVISE instead of PASS; a max-synthetic-ratio cap (floor of 3x) on synthetic tickets relative to real ticket count; an overlap threshold (default 0.75) for concept dedup, or disabling dedup entirely.

Resuming from the drafts stage, or running review-only, skips the hints/concepts stages — hint-promotion flags are ignored in those modes since there are no hints to promote at that point.

## Output

All artifacts land under the run's output directory: a run manifest, tickets, hints, concepts, drafts, exportable tickets, per-iteration review results, and the final exported strategies (`strategy.yaml` + `metadata.json` per candidate).

## Agent-Augmented Workflow

The hint-generation step can be done by the calling agent directly instead of an external LLM call: run auto-detect to produce the market summary and anomalies, reason over them to draft hint objects (title, observation, symbols, regime bias, mechanism tag, preferred entry family, hypothesis type), save them, then run the orchestrator with hint-promotion enabled so those hints flow into concept synthesis alongside rule-based ones.