edge-pipeline-orchestrator（边缘管道编排器）

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 OHLCV) to exported strategies
• Resume a partially completed pipeline from the drafts stage
• Review and revise existing strategy drafts with feedback loop
• Dry-run the pipeline to preview results without exporting

Workflow

1. Load pipeline configuration from CLI arguments
2. Run auto_detect stage if --from-ohlcv is provided (generates tickets from raw OHLCV data)
3. Run hints stage to extract edge hints from market summary and anomalies
4. Run concepts stage to synthesize abstract edge concepts from tickets and hints
5. Run drafts stage to design strategy drafts from concepts
6. Run review-revision feedback loop:

- Review all drafts (max 2 iterations) - PASS verdicts accumulated; REJECT verdicts accumulated - REVISE verdicts trigger apply_revisions and re-review - Remaining REVISE after max iterations downgraded to research_probe

1. Export eligible drafts (PASS + export_ready_v1 + exportable entry_family)
2. Write pipeline_run_manifest.json with full execution trace

CLI Usage

# Full pipeline from tickets
python3 scripts/orchestrate_edge_pipeline.py \
  --tickets-dir path/to/tickets/ \
  --output-dir reports/edge_pipeline/

# Full pipeline from OHLCV
python3 scripts/orchestrate_edge_pipeline.py \
  --from-ohlcv path/to/ohlcv.csv \
  --output-dir reports/edge_pipeline/

# Resume from drafts stage
python3 scripts/orchestrate_edge_pipeline.py \
  --resume-from drafts \
  --drafts-dir path/to/drafts/ \
  --output-dir reports/edge_pipeline/

# Review-only mode
python3 scripts/orchestrate_edge_pipeline.py \
  --review-only \
  --drafts-dir path/to/drafts/ \
  --output-dir reports/edge_pipeline/

# Dry run (no export)
python3 scripts/orchestrate_edge_pipeline.py \
  --tickets-dir path/to/tickets/ \
  --output-dir reports/edge_pipeline/ \
  --dry-run

Output

All artifacts are written to --output-dir:

output-dir/
├── pipeline_run_manifest.json
├── tickets/          (from auto_detect)
├── hints/hints.yaml  (from hints)
├── concepts/edge_concepts.yaml
├── drafts/*.yaml
├── exportable_tickets/*.yaml
├── reviews_iter_0/*.yaml
├── reviews_iter_1/*.yaml  (if needed)
└── strategies/<candidate_id>/
    ├── strategy.yaml
    └── metadata.json

Claude Code LLM-Augmented Workflow

Run the LLM-augmented pipeline entirely within Claude Code:

1. Run auto_detect to produce market_summary.json + anomalies.json
2. Claude Code analyzes data and generates edge hints
3. Save hints to a YAML file:

- title: Sector rotation into industrials
  observation: Tech underperforming while industrials show relative strength
  symbols: [CAT, DE, GE]
  regime_bias: Neutral
  mechanism_tag: flow
  preferred_entry_family: pivot_breakout
  hypothesis_type: sector_x_stock

1. Run orchestrator with --llm-ideas-file and --promote-hints:

python3 scripts/orchestrate_edge_pipeline.py \
  --tickets-dir path/to/tickets/ \
  --llm-ideas-file llm_hints.yaml \
  --promote-hints \
  --as-of 2026-02-28 \
  --max-synthetic-ratio 1.5 \
  --strict-export \
  --output-dir reports/edge_pipeline/

Optional Flags

• --as-of YYYY-MM-DD — forwarded to hints stage for date filtering
• --strict-export — export-eligible drafts with any warn finding get REVISE instead of PASS
• --max-synthetic-ratio N — cap synthetic tickets to N × real ticket count (floor: 3)
• --overlap-threshold F — condition overlap threshold for concept deduplication (default: 0.75)
• --no-dedup — disable concept deduplication

Note: --llm-ideas-file and --promote-hints are effective only during full pipeline runs. --resume-from drafts and --review-only skip hints/concepts stages, so these flags are ignored.

Resources

• references/pipeline_flow.md — Pipeline stages, data contracts, and architecture
• references/revision_loop_rules.md — Review-revision feedback loop rules and heuristics