Learning Harvest

Scope Boundary

Operational learnings are NOT rules. They are what you learn while applying rules.

Standards (refiner output, atom defaults)	Operational Learnings (this document)
"Domain layer must not import from infrastructure"	"When adding a new aggregate, we keep forgetting to define the repository interface first — design interface before implementation"
"Functions should have single responsibility"	"Service classes that start small grow past 500 lines within 3 features — split by command type proactively at ~200 lines"
"Value objects must validate in constructor"	"Date range VOs without explicit inclusive/exclusive documentation cause boundary bugs every time — document semantics alongside validation"

The standard is the rule. The operational learning is what we discovered while applying the rule on this project.

If an entry reads like a rule that should always be followed, it belongs in a standards document (run the relevant refiner). If it reads like "here's what we keep learning the hard way" or "here's an approach that keeps working for us" — it belongs here.

Patterns that recur frequently may graduate to standards via a refiner. That promotion path is part of the Tighten behavior.

Config Resolution

1. Check .lattice/config.yaml for paths.operational_learnings.
2. If set and the file exists at that path → use it.
3. If set but no file exists there → tell the user which configured path is missing, then use the default .lattice/learnings/operational-learnings.md.
4. If not set → use the default .lattice/learnings/operational-learnings.md.

Backward compatibility: If default path not found, check these legacy paths in order:

• .lattice/learnings.md — flat file at root
• .lattice/learnings/review-insights.md — prior naming convention

If found, offer migration to canonical path and format. If user declines, read as flat input. STOP: do not write to it.

Document Structure

# Operational Learnings

Experiential patterns from practice. Complements standards (what should be) with experience (what we keep learning).

## Design Patterns
<!-- Decomposition, architecture choices, scope decisions that proved good or bad -->

## Implementation Craft
<!-- Coding approaches, library gotchas, design-to-reality gaps -->

## Quality Signals
<!-- Recurring quality issues that keep appearing despite rules -->

## Reliability
<!-- Bug root causes, failure modes, fragile areas, boundary condition gaps -->

## Structural Health
<!-- Architectural drift, debt accumulation, coupling issues, migration lessons -->

Entry format: - YYYY-MM-DD [context] Pattern — actionable takeaway

• context: type of session (e.g., "design", "implementation", "review", "bug fix", "refactoring"). Not a feature name — learnings are cross-cutting.
• Each entry ONE bullet, max 2 lines, scannable in under 10 seconds.

Load Behavior

Invoked at session start. Composing workflow passes a focus hint (relevant categories).

1. Resolve file path per Config Resolution.
2. If file not found — "No operational learnings yet." Continue. Non-blocking.
3. If found — surface relevant entries (3-5 most recent from matching categories) as brief context. Treat as soft guidance, not hard constraints.

Active monitoring: Once loaded, maintain a silent harvest queue throughout the session. When a decision or trade-off passes the cross-cutting test below, add it to the queue. STOP: do not prompt immediately.

Cross-cutting test — a candidate must pass BOTH before queuing:

1. It names a pattern or approach, not a feature-specific fact.
2. A developer on a completely different feature could apply it without knowing this feature's context.

STOP: if either fails, skip entirely — do not queue.

Before queuing, check against entries loaded at session start. If the same pattern already exists — skip.

When to surface: Surface the queue as a single batch when EITHER condition is true — not at every level or layer:

• Queue reaches 3 candidates, OR
• A major phase completes (all design levels done, a full implementation layer done)

STOP: do not surface at every individual level approval or component completion — that is over-prompting. Once surfaced, clear the queue. Anything remaining at session end goes to Harvest.

"I noted [N] potential harvest candidates — worth a quick review?"

Mid-session interrupt (rare exception): surface a single pattern immediately, outside the queue, only when it would be impossible to reconstruct by session end — a live debate that resolved unexpectedly, a library gotcha caught mid-implementation. If in doubt, queue instead.

Session-end Harvest is the primary mechanism.

Harvest Behavior

Invoked at session end. Composing workflow passes a session context (what kind of work happened).

Governing principle: STOP: the atom never writes autonomously. Session-end Harvest is the primary capture event — mid-session prompting is the exception.

Steps:

1. Drain the queue. Collect all candidates from active monitoring queue plus any new ones surfaced by reviewing session decisions and outcomes. Each candidate must have passed the cross-cutting test (active monitoring) or pass it now.

2. Propose as a batch. Present queued candidates together — not one per message:

   Harvest candidates from this session:

   1. [Category] — [pattern in one line]
   2. [Category] — [pattern in one line]

   Accept, edit, add your own, or skip entirely.

   Empty queue and nothing new found? Say so in one line. STOP: do not force output.

3. Filter — apply before writing confirmed entries. For each entry the user accepts:

   Filter	Fail if...
   Evidence	No concrete session event — just prior knowledge
   Cross-cutting	Specific to this feature's domain, won't recur
   Actionable	Requires this conversation's context to understand
   Recurrence	No structural reason it will happen again

   Filter fails on a confirmed entry? Tell the user which filter — offer to reword. STOP: do not silently drop.

4. User decides. Accept, edit, reject, add their own, or skip all. STOP: do NOT argue for rejected entries.

5. Write confirmed entries only. Dedup against existing entries (update with a recurrence note if the same pattern exists). Create the file and directory if needed.

6. Assess health. Count entries per category and total (already read for dedup in Step 5). Any category exceeds ~10 entries, or total exceeds ~35 → note in one line: "Operational learnings is growing dense — say 'tighten learnings' to run Tighten standalone." Pattern recurred 4+ times → note it as a promotion candidate the same way. STOP: do not run Tighten in this session — flag only, never act.

Tighten Behavior

Invoked standalone only — Harvest may flag that tightening is due, but never launches it.

1. Read full document.
2. Identify: consolidation opportunities (same pattern, different words), noise (one-off, never recurred), promotion candidates (recurred 4+ times — suggest refiner), stale entries (project has changed).
3. Present each candidate individually — consolidation, noise, promotion, and staleness are different judgment calls. Accept / edit / reject per candidate, not as one batch.
4. Apply only what user confirms.

Self-Validation Checklist

Before writing any entry, verify ALL. STOP: if any fails, do not write.

1. User confirmed — STOP: Explicit user approval for every entry. No exceptions.
2. Evidence grounded — STOP: Produced by a specific session event, not prior knowledge.
3. Experiential, not prescriptive — STOP: Reads like "what we learned" not "what the rule should be." If it's a rule, it belongs in standards via a refiner.
4. Cross-cutting — STOP: Applies beyond this feature. Feature-specific decisions belong in context anchor doc.
5. Actionable standalone — STOP: a developer on a different feature can act on this without this conversation's context. Confidence level is not a gate — the user decides if it is worth capturing.
6. Not redundant — STOP: Not already in standards, atom defaults, or existing learnings. At most, add recurrence note.
7. Concise — STOP: Scannable in 10 seconds. Two lines max.

All checks pass on an entry → write it.

Standalone Invocation

When invoked directly — not composed by a molecule — match the user's phrase to exactly one behavior. STOP: if ambiguous, ask — never guess.

User says	Run
"tighten learnings", "compress learnings", "clean up learnings", "/learning-harvest tighten learnings"	Tighten Behavior
"harvest learnings", "capture this pattern", "log this learning"	Harvest Behavior
"what have we learned", "load learnings", bare "operational learnings" with no verb	Load Behavior

STOP: if the phrase doesn't clearly map to one row, ask — "Load recent entries, harvest something new, or tighten the document?" — before running anything.