edge-candidate-agent（边缘候选 Agent）

Edge Candidate Agent

Overview

Convert daily market observations into reproducible research tickets and Phase I-compatible candidate specs. Prioritize signal quality and interface compatibility over aggressive strategy proliferation. This skill can run end-to-end standalone, but in the split workflow it primarily serves the final export/validation stage.

When to Use

• Convert market observations, anomalies, or hypotheses into structured research tickets.
• Run daily auto-detection to discover new edge candidates from EOD OHLCV and optional hints.
• Export validated tickets as strategy.yaml + metadata.json for trade-strategy-pipeline Phase I.
• Run preflight compatibility checks for edge-finder-candidate/v1 before pipeline execution.

Prerequisites

• Python 3.9+ with PyYAML installed.
• Access to the target trade-strategy-pipeline repository for schema/stage validation.
• uv available when running pipeline-managed validation via --pipeline-root.

Output

• strategies/<candidate_id>/strategy.yaml: Phase I-compatible strategy spec.
• strategies/<candidate_id>/metadata.json: provenance metadata including interface version and ticket context.
• Validation status from scripts/validate_candidate.py (pass/fail + reasons).
• Daily detection artifacts:

- daily_report.md - market_summary.json - anomalies.json - watchlist.csv - tickets/exportable/*.yaml - tickets/research_only/*.yaml

Position in Split Workflow

Recommended split workflow:

1. skills/edge-hint-extractor: observations/news -> hints.yaml
2. skills/edge-concept-synthesizer: tickets/hints -> edge_concepts.yaml
3. skills/edge-strategy-designer: concepts -> strategy_drafts + exportable ticket YAML
4. skills/edge-candidate-agent (this skill): export + validate for pipeline handoff

Workflow

1. Run auto-detection from EOD OHLCV:

- skills/edge-candidate-agent/scripts/auto_detect_candidates.py - Optional: --hints for human ideation input - Optional: --llm-ideas-cmd for external LLM ideation loop

1. Load the contract and mapping references:

- references/pipeline_if_v1.md - references/signal_mapping.md - references/research_ticket_schema.md - references/ideation_loop.md

1. Build or update a research ticket using references/research_ticket_schema.md.
2. Export candidate artifacts with skills/edge-candidate-agent/scripts/export_candidate.py.
3. Validate interface and Phase I constraints with skills/edge-candidate-agent/scripts/validate_candidate.py.
4. Hand off candidate directory to trade-strategy-pipeline and run dry-run first.

Quick Commands

Daily auto-detection (with optional export/validation):

python3 skills/edge-candidate-agent/scripts/auto_detect_candidates.py \
  --ohlcv /path/to/ohlcv.parquet \
  --output-dir reports/edge_candidate_auto \
  --top-n 10 \
  --hints path/to/hints.yaml \
  --export-strategies-dir /path/to/trade-strategy-pipeline/strategies \
  --pipeline-root /path/to/trade-strategy-pipeline

Create a candidate directory from a ticket:

python3 skills/edge-candidate-agent/scripts/export_candidate.py \
  --ticket path/to/ticket.yaml \
  --strategies-dir /path/to/trade-strategy-pipeline/strategies

Validate interface contract only:

python3 skills/edge-candidate-agent/scripts/validate_candidate.py \
  --strategy /path/to/trade-strategy-pipeline/strategies/my_candidate_v1/strategy.yaml

Validate both interface contract and pipeline schema/stage rules:

python3 skills/edge-candidate-agent/scripts/validate_candidate.py \
  --strategy /path/to/trade-strategy-pipeline/strategies/my_candidate_v1/strategy.yaml \
  --pipeline-root /path/to/trade-strategy-pipeline \
  --stage phase1

Export Rules

• Keep validation.method: full_sample.
• Keep validation.oos_ratio omitted or null.
• Export only supported entry families for v1:

- pivot_breakout with vcp_detection - gap_up_continuation with gap_up_detection

• Mark unsupported hypothesis families as research-only in ticket notes, not as export candidates.

Guardrails

• Reject candidates that violate schema bounds (risk, exits, empty conditions).
• Reject candidate when folder name and id mismatch.
• Require deterministic metadata with interface_version: edge-finder-candidate/v1.
• Use --dry-run in pipeline before full execution.

Resources

skills/edge-candidate-agent/scripts/export_candidate.py

Generate strategies/<candidate_id>/strategy.yaml and metadata.json from a research ticket YAML.

skills/edge-candidate-agent/scripts/validate_candidate.py

Run interface checks and optional StrategySpec/validate_spec checks against trade-strategy-pipeline.

skills/edge-candidate-agent/scripts/auto_detect_candidates.py

Auto-detect edge ideas from EOD OHLCV, generate exportable/research tickets, and optionally export/validate automatically.

references/pipeline_if_v1.md

Condensed integration contract for edge-finder-candidate/v1.

references/signal_mapping.md

Map hypothesis families to currently exportable signal families.

references/research_ticket_schema.md

Ticket schema used by export_candidate.py.

references/ideation_loop.md

Hint schema and external LLM ideation command contract.