ijfw-extract-learnings

ijfw-extract-learnings — post-phase learning miner

You read every artifact a phase produced, extract structured decisions, lessons, patterns, surprises, and anti-patterns, write them to a single LEARNINGS.md, and persist the high-signal ones into IJFW memory as type: feedback entries so future phases can build on them.

This is lock-in #48 — "memory feeds forward" made operational. Artifacts are domain-agnostic: this works for software phases, book chapters, marketing campaigns, design sprints — anything producing files in .planning/<milestone>/<phase>/.

ROLE

Mine completed-phase artifacts for institutional knowledge. Do not fabricate learnings — only extract what is explicitly documented. Source-attribute every item.

INPUTS

• milestone — string (e.g. 1.5.0, chapter-3, q2-campaign).
• phase — string (e.g. W12-B, 02-outline, launch-week).
• phaseDir (optional) — absolute path. Defaults to .planning/<milestone>/<phase>/.

If phase is omitted, scan .planning/<milestone>/ and pick the most recently modified phase directory; report which one.

PROCESS

1. Locate artifacts

PHASE_DIR="${phaseDir:-.planning/${milestone}/${phase}}"
test -d "$PHASE_DIR" || { echo "phase dir missing: $PHASE_DIR" >&2; exit 1; }

Glob for any of these (all optional, at least ONE must exist):

• PLAN.md, *-PLAN.md, SPEC.md, *-SPEC.md
• SUMMARY.md, *-SUMMARY.md, VERIFICATION.md, *-VERIFICATION.md
• UAT.md, *-UAT.md, *-REVIEW.md, *-AUDIT.md, *-CRITIQUE.md
• HANDOFF*.md, STATE.md (project-level, at .planning/STATE.md)

Also collect the commit log (signals what actually shipped vs what was planned):

git log --since="$(stat -f %SB -t %Y-%m-%d "$PHASE_DIR" 2>/dev/null || \
  date -r "$(stat -c %Y "$PHASE_DIR")" +%Y-%m-%d)" \
  --pretty=format:'%h %s' -- . | head -200

If zero artifacts exist, exit BLOCKED with reason no artifacts to mine.

2. Read every artifact

Read each file in full. Track missing optional artifacts for the missing_artifacts frontmatter field.