Pipeline: Скилл для комплексной обработки данных

--- name: pipeline description: > End-to-end processing pipeline that chains the full 6R sequence: seed, reduce, reflect, reweave, verify. Runs per-item or in batch mode. Three depth levels control thoroughness vs. speed. The orchestrator for all processing skills. Triggers on: "pipeline", "process", "full pipeline", "6R" --- # /pipeline > End-to-end processing — seed through verify in one command. ## Purpose Chain the full processing sequence in a single invocation: seed (initialize) → reduce (extract) → reflect (connect) → reweave (update old content) → verify (quality gate). Handles orchestration, error recovery, and progress tracking. Use this when you want the full treatment rather than running individual steps. ## Usage ```bash # Process a single file through the full pipeline /pipeline path/to/article.md # Process with explicit depth /pipeline path/to/transcript.md --depth deep # Process multiple files in batch /pipeline path/to/inbox/*.md --batch # Quick single-pass processing (no subagents) /pipeline path/to/notes.md --depth quick # Process from URL /pipeline --url "https://example.com/article" --depth standard # Resume a failed pipeline run /pipeline --resume run-2026-03-20-143000 # Skip specific stages /pipeline path/to/file.md --skip reweave,verify ``` ## Arguments | Flag | Type | Default | Description | |------|------|---------|-------------| | `<input>` | positional | required | File path, glob, or `--url` for web content | | `--url` | string | — | URL to fetch and process | | `--depth` | enum | `standard` | Processing thoroughness (see Depth Levels below) | | `--batch` | flag | false | Process multiple inputs as separate pipeline runs | | `--skip` | string[] | — | Comma-separated stages to skip: `seed`, `reduce`, `reflect`, `reweave`, `verify` | | `--stop-after` | string | — | Stop after this stage (e.g., `reduce` to just extract) | | `--resume` | string | — | Resume a previous pipeline run by ID | | `--parallel` | int | 1 | Number of parallel items in batch mode | | `--output-dir` | path | `./processed/` | Directory for pipeline output | | `--dry-run` | flag | false | Show what would happen without executing | ## Depth Levels | Depth | Seed | Reduce | Reflect | Reweave | Verify | Time | Token Cost | |-------|------|--------|---------|---------|--------|------|------------| | `quick` | inline | key points only | skip | skip | structural only | ~30s | ~2K | | `standard` | full metadata | claims + context | direct connections | dry-run suggestions | full checks | ~3min | ~15K | | `deep` | full + provenance | claims + evidence + counterpoints | 2-hop traversal + synthesis | apply with approval | full + fidelity | ~10min | ~50K | ## Workflow 1. **Initialize** — Create pipeline run directory with unique ID. Log start time, input, and parameters. 2. **Seed** (`/seed`) — Ingest source material. Create inbox item with provenance metadata. Output: seeded item with metadata. 3. **Reduce** (`/reduce`) — Extract atomic claims from seeded content. Output: structured claims file. 4. **Reflect** (`/reflect`) — Find connections between new claims and existing knowledge. Output: connection map. 5. **Reweave** (`/reweave`) — Backward pass over existing content. Flag items that need updating. At `deep` depth, apply with human approval. Output: update report. 6. **Verify** (`/verify`) — Quality gate on all new and modified content. Output: verification report. 7. **Report** — Emit pipeline summary: items processed, claims extracted, connections found, updates suggested, quality score. 8. **Cleanup** — Archive pipeline run metadata for resume capability. ## Output ```markdown ## Pipeline Report — run-2026-03-20-143000 **Input:** ai-masters/signals/2026-03-20-ed-strategy-call.md **Depth:** standard **Duration:** 2m 34s **Token cost:** ~12,400 tokens ### Stage Results | Stage | Status | Output | |-------|--------|--------| | Seed | done | 1 item, 2,340 words, source: voice transcript | | Reduce | done | 14 claims extracted (9 facts, 3 decisions, 2 action items) | | Reflect | done | 7 connections found, 2 synthesis opportunities | | Reweave | done | 3 items flagged for update (1 critical, 2 stale) | | Verify | done | 14/14 claims pass structural, 1 fidelity warning | ### Key Outputs - Claims file: `processed/run-2026-03-20-143000/claims.md` - Connection map: `processed/run-2026-03-20-143000/connections.md` - Reweave report: `processed/run-2026-03-20-143000/reweave.md` - Verification: `processed/run-2026-03-20-143000/verify.md` ### Action Items Extracted 1. Roberto: Send revised pricing doc to Ed by Friday 2. Ed: Confirm enterprise pilot with Acme Corp 3. Roberto: Update money-revenue/context.md with new pricing ``` ## Dependencies - `/seed` — Stage 1 - `/reduce` — Stage 2 - `/reflect` — Stage 3 - `/reweave` — Stage 4 - `/verify` — Stage 5 - File system write access for output directory - Subagent spawning capability (for `deep` depth)