harness-3-retro

The Retro stage of the harness loop. The place the loop turns observation into encoded improvement. Two modes:

Encode, don't document. A wiki paragraph that says "remember to do X" is worth nothing; an automated step that does X for you is worth everything. The whole point of retro is that friction observed during work becomes executable knowledge — a justfile recipe, a skill edit, a staged diff — not a prose note that rots. The [e]ncode action stages the change as a diff; that diff IS the encoding. Prefer an automated command over a wiki paragraph, a seed script over a manual setup step, a pre-flight check over a "remember to…". Executable knowledge > prose.

Input

$ARGUMENTS
# Modes:
# --drain      Session-end: read the per-agent buffer, present the soft prompt, route entries to .retro.md
# --harvest    Long-horizon: scan + cluster + prioritize .retro.md files; print the curated view
# --harvest --json   Machine-readable render of the harvest view (for `just compound-value`, CI hooks)
# Plus --harvest runtime filters: --plan <slug> / --agent <slug> / --since <date> / --kind <kind>
# Plus --harvest --prune --older-than <Nd> [--apply]   Reversible stale-retro pruning (dry-run by default)