Skip to content

harvest

Source: skills/harvest/SKILL.md


Nightly usage (2026-04-09): /dream start now runs harvest as part of its bounded compounding loop. Use /harvest for manual sweeps, CI runs, or when Dream is disabled. Dream holds .agents/overnight/run.lock while running — manual ao harvest will refuse until the lock releases.

Sweep all .agents/ directories across the workspace, extract learnings, patterns, and research, deduplicate cross-rig, and promote high-value items to the global knowledge hub (~/.agents/learnings/).

Naming gotcha. $harvest promotes into ~/.agents/learnings/, not ~/.agents/. Users often say "harvest all to ~/.agents" and mean the promotion hub. If you really want every raw artifact (not just the promotion set) mirrored verbatim, you want rsync, not $harvest.

Which skill do I need?

See docs/skills-decision-tree.md for the full "which skill next?" decision table covering harvest, compile, dream, knowledge-activation, and quickstart.

What This Skill Does

The knowledge flywheel captures learnings per-rig, but they stay siloed. Harvest closes the loop by walking all rigs, extracting artifacts, deduplicating by content hash, and promoting high-confidence items to the global hub where every rig can access them via ao inject.

When to use: Before an evolve cycle, after a burst of development across multiple rigs, or weekly as part of knowledge governance.

Output: .agents/harvest/latest.json (catalog) + promoted files in ~/.agents/learnings/

Execution Steps

Step 1: Preview Scope (Dry Run)

Bash
ao harvest --dry-run --quiet

Read .agents/harvest/latest.json and report: - Rigs discovered - Total artifacts extracted - Unique vs duplicate count - Promotion candidates (artifacts >= min confidence)

Step 2: Confirm Execution

Skip if --auto is set. Otherwise, show the dry-run summary and ask:

Text Only
Harvest will promote N artifacts from M rigs to ~/.agents/learnings/.
Proceed? [Approve / Adjust threshold / Abort]

Step 3: Execute Harvest

Bash
ao harvest --roots ~/gt/ --promote-to ~/.agents/learnings --min-confidence 0.5

Step 4: Post-Harvest Cleanup

Run dedup on the promotion target to clean up any remaining duplicates:

Bash
ao dedup --merge ~/.agents/learnings/ 2>/dev/null || true

Step 5: Report Results

Report to user: - Rigs scanned - Artifacts extracted and unique count - Duplicates found (with top duplicate groups) - Artifacts promoted (with provenance) - Top discoveries (highest-confidence cross-rig patterns)

Flags

Flag Default Description
--auto off Skip confirmation gate
--roots ~/gt/ Override root directories to scan
--min-confidence 0.5 Minimum confidence for promotion
--include learnings,patterns,research Artifact types to extract

Quick Start

Bash
/harvest                          # Full sweep with confirmation
/harvest --auto                   # Hands-free sweep
/harvest --min-confidence 0.7     # Only promote high-confidence items
/harvest --roots ~/gt/,~/projects/ # Scan additional directories

Governance

See references/governance.md for ongoing governance model: size budgets, sweep frequency, staleness thresholds, and cross-rig synthesis triggers.

See Also

Reference Documents