Harvest

---
description: Apply the FORGE HARVEST workflow (Extract → Analyze → Restructure → Verify → Extend → Synthesize → Transform) for documentation and knowledge extraction
argument-hint: <what to document or understand>
---

You are operating under the FORGE framework's **HARVEST workflow** for UNDERSTANDING problems — documentation, knowledge extraction, codebase comprehension.

## Setup (read these before starting)

1. Workflow definition: `workflows/harvest.md`
2. Master spec sections for protocol context: `FORGE_MASTER.md` — sections 5 (Confidence), 6 (Assumptions), 7 (Meta-Prompting), 8 (Verification), 12 (Operating Principles)

## Phase 0 — Prior learnings retrieval

If `learnings/_index.jsonl` exists, search for prior HARVEST entries on similar domains/codebases per `core/omega_retrieval.md`. Surface top 3 before Phase 1.

## Phases 1-7

Apply HARVEST. Hard rules:
- Every claim must be traceable to source — no speculation as fact
- Confidence Protocol on every uncertain finding; mark gaps as `UNKNOWN` or `NEEDS VERIFICATION`
- Iterate until quality reaches 95%+ (completeness, accuracy, cross-references, consistency)
- Time-budget the phases per `workflows/harvest.md` (Extract 25%, Analyze 20%, Restructure 15%, Verify 15%, Extend 10%, Synthesize 10%, Transform 5%)

## Closing emissions

After delivery:

**1. Telemetry** — Append one JSON line to `telemetry/sessions.jsonl` per `telemetry/schema.json`. Self-report honestly:
- `task_id` (UUID4), `timestamp` (ISO 8601 UTC), `workflow: "harvest"`
- `classification`: category (UNDERSTANDING), complexity, domain, confidence
- `gates_passed` / `gates_failed` against the gate names in `schema.json`
- `assumptions_count`, `iterations` (refinement passes; HARVEST iterates until 95%+ quality), `outcome`
- `prompt_hash`: SHA-256 of $ARGUMENTS, first 16 hex chars
- `raw_prompt`: $ARGUMENTS only if `FORGE_TELEMETRY_RAW=1`, else `null`
- `notes`: brief commentary on which extraction sources were richest, which patterns generalized

Skip emission entirely if `FORGE_TELEMETRY_DISABLED=1`.

**2. Learnings** — Append a learning entry capturing which extraction sources were richest and which patterns generalized.

---

## User task

$ARGUMENTS