Promote Learning to standalone pack, add MARR standards

[virtualian]

Context

During pack reinstall (2026-03-19), the Utilities pack was replaced with upstream source. Local customizations were deliberate and need to be reapplied from the backup.

Backup location

~/backups/pai/utilities-skill-20260319-095100/Utilities/

Changes to reapply

1. Learning sub-skill (removed by upstream)

• Local had Learning/ as a dedicated sub-skill under Utilities
• Upstream absorbed it into PAIUpgrade with inline MineReflections.md and AlgorithmUpgrade.md
• Decision needed: Keep as separate Learning/ sub-skill or accept upstream consolidation into PAIUpgrade?

2. SKILL.md routing table

• Local routing table included Learning/SKILL.md as a separate route
• Upstream merged Learning triggers into the PAIUpgrade row
• If Learning is restored, re-add the routing row

3. PAIUpgrade/SKILL.md

• Local description referenced "learning pipeline"
• Local Upgrade workflow delegated Thread 3 to Learning skill
• Upstream rewrote Thread 3 as inline reflection mining

4. PAIUpgrade/State/github-trending.json

• Runtime state file that existed locally but isn't in pack source
• Will be regenerated on next upgrade run — no action needed

How to reapply

# Diff backup against current install
diff -rq ~/backups/pai/utilities-skill-20260319-095100/Utilities ~/.claude/skills/Utilities --exclude=node_modules --exclude=bun.lock

# Restore specific files from backup
cp -r ~/backups/pai/utilities-skill-20260319-095100/Utilities/Learning ~/.claude/skills/Utilities/Learning
# Then update SKILL.md routing table and PAIUpgrade references as needed

Labels

maintenance, packs

[virtualian]

Findings from restoring Learning sub-skill

PR #94 restores the Learning sub-skill and is ready to merge. During validation (running all 4 Learning workflows), we discovered an architectural gap in the learning pipeline:

The pipeline disconnect

The Learning skill has 4 workflows that form a pipeline, but they don't actually pass data to each other:

MineReflections ──┐
                  ├─→ Synthesize ──→ (gap) ──→ AlgorithmUpgrade
MineRatings ──────┘

Synthesize runs MineReflections + MineRatings in parallel, cross-references them, and produces a combined Internal Signals report with a prioritized upgrade list. This is the richest analysis.

AlgorithmUpgrade is supposed to take upgrade candidates and map them to specific Algorithm spec sections with concrete diffs. But it re-mines algorithm-reflections.jsonl from scratch — it doesn't consume Synthesize's output. This means:

1. Lost enrichment — The cross-referenced priority list (reflections × ratings) that Synthesize produces is thrown away
2. No ratings signal — AlgorithmUpgrade only reads reflections, never ratings. The behavioral STOP/DO MORE patterns don't inform the proposed changes
3. Redundant work — Re-mines the same JSONL file that Synthesize already processed

What should happen

AlgorithmUpgrade should accept Synthesize's output as input (when available), not re-derive it. Two options:

Option A: File-based handoff — Synthesize writes its report to a known location (e.g., MEMORY/LEARNING/last-synthesis.md). AlgorithmUpgrade checks for it and uses it if fresh (< 24h), otherwise falls back to direct mining.

Option B: Workflow chaining — AlgorithmUpgrade's Step 1 becomes "Run Synthesize if no recent output exists, then proceed with section mapping." This makes AlgorithmUpgrade the top-level entry point that includes synthesis.

Recommendation

Option A (file-based handoff) is simpler and more composable — each workflow stays independent but can benefit from prior work. It also means the PAIUpgrade Upgrade workflow's Thread 3 (which calls Synthesize) automatically creates the handoff file for a subsequent AlgorithmUpgrade run.

Additional context

The Synthesize run on 52 reflections + 780 ratings produced 7 upgrade candidates with 3 cross-referenced CRITICAL signals. The top finding: "read before acting" appeared in both reflections (5 occurrences) and ratings (5 occurrences, avg 2.6/10) — the strongest combined signal in the dataset.

[virtualian]

Discovery: Two downstream consumers have no learning signal path

While validating the pipeline handoff fix, mapped all destinations that should consume Synthesize output. Found two gaps beyond AlgorithmUpgrade:

Current state of learning signal flow

Destination	What it receives	Updated by	Status
Algorithm v3.7.0.md	Structural process fixes (phase gates, pre-read steps)	AlgorithmUpgrade workflow	✅ Fixed (synthesis handoff added in commit 28cba3b)
Dynamic context (session start)	Recent signals, failure patterns, performance score	learning-readback.ts + hooks	✅ Automated
AISTEERINGRULES.md	Behavioral rules ("never do X", "always do Y")	Manual only	❌ No automated path
Feedback memories	Persistent behavioral corrections (DO MORE / STOP)	Manual only	❌ No automated path

Why this matters

AISTEERINGRULES.md is force-loaded at every session start. The STOP patterns from Synthesize (wrong target avg 2.6/10, answering wrong question avg 1.75/10, not following instructions avg 1.67/10) are exactly the kind of behavioral rules that belong here — but today those rules were all written by hand.

Feedback memories persist across conversations. The DO MORE patterns (comprehensive research avg 8.77, smooth execution avg 9.67) should become feedback memories so future sessions know what Ian values — but today they only exist in raw ratings.jsonl.

Options to explore

1. Extend AlgorithmUpgrade — Add a "Behavioral Rules" section that proposes new AISTEERINGRULES entries alongside Algorithm spec diffs
2. New workflow: BehavioralUpgrade — Dedicated workflow that takes Synthesize output and proposes steering rules + feedback memories
3. Extend Synthesize itself — Add Steps 6-7 that generate rule/memory proposals directly

This is getting into multi-workflow architecture territory — parking for tomorrow.

[virtualian]

Here's a spec for the new work flow:

2026-03-22-learning-pipeline-spec.md