| name | reflect |
| description | Orchestrates Reflexion-based learning: dispatches to /reflect-on-outcome (hypothesis ABC chains, execution patterns, batch micro-hypotheses), /reflect-on-self (pattern synthesis, Level 2 self-model, calibration), or /reflect-maintain (memory curation, active forgetting, aspiration grooming) based on --mode. Use whenever the aspirations loop hits a reflection cadence, a hypothesis resolves, the user asks to "reflect on recent work", or after major outcomes that warrant pattern extraction. |
| user-invocable | false |
| triggers | ["/reflect"] |
| parameters | [{"name":"mode","description":"Reflection mode: --on-hypothesis, --on-execution, --extract-patterns, --calibration-check, --full-cycle, --curate-memory, --curate-aspirations, --batch-micro","required":false}] |
| execution_history | {"total_invocations":0,"outcome_tracking":{"successful":0,"unsuccessful":0,"success_rate":0},"last_invocation":null,"known_pitfalls":[],"reconsolidation_trigger":"After 10 invocations with declining success rate, trigger skill review"} |
| conventions | ["pipeline","reasoning-guardrails","pattern-signatures","handoff-working-memory"] |
| minimum_mode | assistant |
| revision_id | skill-bootstrap-reflect-8f539f |
| previous_revision_id | null |
/reflect — Reflexion-Based Self-Learning Engine
Generates structured reflections from hypothesis outcomes, extracts reusable strategies, tracks violations of expectation, and synthesizes hierarchical insights. This is the core self-learning mechanism — it turns raw outcomes into institutional knowledge.
Based on: Reflexion (Shinn 2023), ABC Method, Generative Agents (Park 2023), VoE metacognitive framework.
Quick Links
| Sub-skill | Mode | Purpose |
|---|
| /reflect-on-outcome | --on-hypothesis, --on-execution, --batch-micro | Outcome reflection: hypothesis ABC chains, execution patterns, batch micro |
| /reflect-on-self | --extract-patterns, --calibration-check | Self-model: pattern synthesis, strategy extraction, calibration |
| /reflect-maintain | --curate-memory, --curate-aspirations | Maintenance: memory curation, aspiration grooming |
| /reflect-tree-update | (shared protocol) | Propagate tree changes upward |
Related skills: /replay (hippocampal replay), /aspirations-spark (Phase 6.5 immediate learning)
Parameters
--on-hypothesis <hypothesis-id> — Reflect on a single resolved hypothesis (session/short/long horizon)--on-execution — Reflect on a goal execution outcome (pattern signatures, contradiction detection, investigation goals)--batch-micro — Batch-reflect on micro-hypotheses from working memory (session-end)--extract-patterns — Mine all resolved hypotheses for reusable strategies--calibration-check — Analyze confidence calibration across all hypotheses--full-cycle — Run all reflection modes in sequence (includes --batch-micro)--curate-memory — Retire stale/low-utilization strategies, guardrails, reasoning bank entries, and pattern signatures--curate-aspirations — Groom stuck goals whose evidence has converged (backlog grooming)--level N — Reflection depth (0=episode, 1=pattern, 2=strategic). Default: auto-detect
Step 0: Load Conventions
Step 0: Load Conventions — Bash: load-conventions.sh with each name from the conventions: front matter. Read only the paths returned (files not yet in context). If output is empty, all conventions already loaded — proceed to next step.
Step 0.1: Stamp reflection cadence (G5 of Phase 1.5)
Step 0.1: Stamp reflection cadence — Bash: reflection-cadence-stamp.sh <mode> (mode is the --mode arg, e.g. "on-hypothesis", "on-execution", "full-cycle", "batch-micro", etc.). Records this reflection in wm.last_reflection_at AND appends to world/reflection-history.jsonl. Signal #13 of the Self/Program evolution metric vector (world/conventions/self-program-evolution.md) reads the history file to compute "reflection_cadence — fraction of expected reflection windows that contained ≥1 fire over the last 50 goals." Fail-silent — never blocks /reflect. Runs BEFORE Step 0.5 context retrieval so cadence accounting captures the reflection regardless of downstream success.
Step 0.5: Load Context for Reflection
Before loading the hypothesis, load background knowledge for informed analysis.
Bash: retrieve.sh --category {hypothesis.category} --depth medium
Returns JSON with tree_nodes, reasoning_bank, guardrails, pattern_signatures,
experiences, beliefs, experiential_index. All retrieval counters already incremented.
Use retrieved context to:
- Compare ABC chain against known patterns (did we use the right strategy?)
- Check if any guardrail should have fired (failure prevention analysis)
- Assess whether beliefs need updating based on this outcome
- Identify if a pattern signature matched (or should have matched)
Step 0 runs ONCE per /reflect invocation. Context is available to the invoked sub-skill. For --full-cycle mode with multiple hypotheses, cache context per category — don't re-retrieve for same category.
Step 0.3: Load Meta-Reflection Strategy
# Step 0.3: Load Meta-Reflection Strategy
Bash: meta-read.sh reflection-strategy.yaml
# The agent's learned reflection preferences:
# - depth_allocation: episode/pattern/strategic weight distribution
# (overrides developmental stage defaults when non-default)
# - trigger_overrides: conditions that modify reflection behavior
# - skip_conditions: conditions where reflection can be safely skipped
# - category_depth_overrides: per-category depth preferences
# - reflection_effectiveness_by_type: MR-Search quality tracking (Priority 2)
# - adaptive_depth: MR-Search adaptive reflection scaling (Priority 6)
# These are advisory — structural rules (horizon gating) still apply.
# MR-Search reflection quality-driven depth allocation (Priority 2):
# Use reflection_effectiveness_by_type to allocate more depth to reflection
# types that historically produce downstream improvement.
IF reflection_effectiveness_by_type exists AND has data:
# Only adjust depth for types with sufficient data (total >= 3).
# total == 0 means "no data" not "ineffective" — use default allocation.
# Currently only spark reflections are tracked (Phase 6.5 tags artifacts).
# Hypothesis and execution types will show total=0 until those sub-skills
# also tag their artifacts with source_reflection_id.
types_with_data = {type: data for type, data in reflection_effectiveness_by_type if data.total >= 3}
IF types_with_data is non-empty:
Apply effectiveness rates as advisory weight on depth_allocation for those types only
# Types without sufficient data: keep default depth_allocation unchanged
# MR-Search adaptive reflection depth (Priority 6):
# Scale reflection effort dynamically based on task properties.
# Only applies when goal context is available (--on-execution, --on-hypothesis from Phase 8.75).
# Skipped in --full-cycle mode where reflect iterates over hypotheses without goal context.
IF adaptive_depth exists AND goal context is available:
depth_multiplier = 1.0
IF adaptive_depth.scale_on_surprise AND surprise_level > 7:
depth_multiplier = min(depth_multiplier * 1.5, adaptive_depth.max_depth_multiplier)
IF adaptive_depth.scale_on_chain_length AND goal has episode_history with length > 1:
depth_multiplier = min(depth_multiplier * 1.25, adaptive_depth.max_depth_multiplier)
IF adaptive_depth.scale_on_importance AND goal.priority == "HIGH":
depth_multiplier = min(depth_multiplier * 1.25, adaptive_depth.max_depth_multiplier)
# Apply multiplier as advisory guidance to sub-skill invocations
Step 0.35: Adversarial Review Check (rb-356 origin)
Originated from the PreCompact-gate release-scope bug: tests passed 15/15 while
the release predicate was a strict subset of the acquire predicate. The bug
was only caught by adversarial self-review with verification questions. This
step automates the trigger so the protocol fires without the user having to
ask for it.
# Step 0.35: Adversarial Review trigger check
# Only applies when goal context is available (--on-execution, --on-hypothesis)
IF adversarial_review exists AND goal context is available:
goal_category = goal.category OR ""
goal_topic = goal.description OR goal.title OR ""
combined = lowercase(goal_category + " " + goal_topic)
files_touched = length(goal.files_changed OR [])
matched_substring = first(s for s in adversarial_review.triggers.category_substrings
if s in combined)
IF matched_substring is not null
AND files_touched >= adversarial_review.triggers.min_files_touched:
# Surface the protocol questions so the sub-skill incorporates them
# into its ABC chain / pattern synthesis / encoded lessons.
Emit: "ADVERSARIAL_REVIEW_TRIGGERED (matched: {matched_substring}, files: {files_touched})"
Emit: adversarial_review.protocol # the 4 questions
# Bump counters — tracks whether the pattern pays for itself over time.
Bash: meta-set.sh reflection-strategy.yaml adversarial_review.last_triggered "$(date +%Y-%m-%dT%H:%M:%S)"
Bash: meta-set.sh reflection-strategy.yaml adversarial_review.times_fired $(($OLD_VALUE + 1))
# If the sub-skill surfaces a bug through the protocol, the sub-skill
# bumps adversarial_review.times_caught_bug via the same mechanism.
# These questions are advisory guidance for the reflection sub-skill — they do
# not block the reflection or change routing. They join retrieval_context so
# the sub-skill naturally addresses them during pattern extraction.
Mode Routing
--on-hypothesis <hypothesis-id> — Single Hypothesis Reflection
Reflect on one resolved hypothesis with full ABC chain analysis, pattern extraction,
belief updates, and knowledge reconciliation. Handles horizon gating (micro → error,
session → lightweight path, short/long → full pipeline).
invoke /reflect-on-outcome Mode: Hypothesis with: hypothesis-id, retrieval_context from Step 0
Sub-skill handles Steps 0.5-9: horizon gate, load, ABC chain, differentiated extraction,
contrastive extraction, experience archival, encoding score, textual reflection,
domain-specific, violation, source tracking, journal, accuracy, pattern signatures,
entities, beliefs, contradictions, process-outcome, context gaps, strategy tracking,
experiential index, spark check, knowledge reconciliation, tree growth, snapshot invalidation.
--on-execution — Execution Outcome Reflection
Reflect on a goal execution outcome that was notable (mistake, surprise, recurring
pattern). Handles pattern signatures and contradiction detection that Phase 6.5
(immediate learning) does not cover. Creates investigation goals for findings
that need follow-up. Lightweight — no ABC chains, no horizon gating.
invoke /reflect-on-outcome Mode: Execution with: goal, result, outcome_class, retrieval_context from Step 0
# Sub-skill handles Steps 0.5-5: notability gate, pattern signatures, contradiction
# detection, investigation goal creation, experience archival, journal entry.
--batch-micro — Batch Micro-Hypothesis Reflection
Batch-process micro-hypotheses from working memory (session-end).
Computes batch stats, promotes surprises, updates aggregate pipeline stats.
invoke /reflect-on-outcome Mode: Batch Micro with: retrieval_context from Step 0
Sub-skill handles Steps 1-7: load micros, batch stats, surprise promotion,
aggregate stats, journal, actionable work check, return batch result.
--extract-patterns — Pattern Extraction
Mine all resolved hypotheses for reusable strategies and Level 2 strategic self-model.
invoke /reflect-on-self Mode: Extract Patterns with: retrieval_context from Step 0
Sub-skill handles Steps 1-5: load resolved, Level 1 pattern synthesis,
strategy extraction, Level 2 strategic self-model, update knowledge base.
--calibration-check — Confidence Calibration
Analyze confidence calibration across all hypotheses.
invoke /reflect-on-self Mode: Calibration with: retrieval_context from Step 0
Sub-skill handles Steps 1-4: bin by confidence, calculate accuracy,
self-consistency check, update calibration data.
--curate-memory — Memory Curation
Retire stale/low-utilization strategies, guardrails, reasoning bank entries,
and pattern signatures. Includes active forgetting reference formulas.
invoke /reflect-maintain Mode: Curate Memory with: retrieval_context from Step 0, scope (if provided)
Sub-skill handles Steps 1-4: gather candidates, evaluate (agent judgment),
execute retirements, journal. Plus active forgetting: decay model,
retrieval strengthening, interference detection, reconsolidation.
--curate-aspirations — Aspiration Grooming
Detect stuck goals whose evidence has converged. Cross-reference pending/blocked
goals against experience archive and knowledge tree. Complete, skip, or re-scope
goals that can be resolved from existing data.
invoke /reflect-maintain Mode: Curate Aspirations with: retrieval_context from Step 0, scope (if provided)
Sub-skill handles Steps 1-4: gather candidates, evidence cross-reference,
execute decisions, journal.
--full-cycle — Full Reflection Cycle
Run all reflection modes in sequence. This is the comprehensive learning pass.
Phase A: Outcome Reflection
- Bash: wm-read.sh micro_hypotheses --json → if non-empty, invoke /reflect-on-outcome Mode: Batch Micro
1.5. invoke /reflect-on-outcome Mode: Execution for goals completed this session with notable outcomes
(only if not already reflected via --on-hypothesis pathway — check goal IDs)
1.75. invoke /reflect-maintain Mode: Curate Aspirations (groom stuck goals before reflecting on hypotheses)
- Bash: pipeline-read.sh --unreflected → get unreflected resolved hypotheses
- For each unreflected hypothesis:
invoke /reflect-on-outcome Mode: Hypothesis with: hypothesis-id
Phase B: Self-Model Reflection
- invoke /reflect-on-self Mode: Extract Patterns
- invoke /reflect-on-self Mode: Calibration
Phase C: Maintenance
5.5. invoke /reflect-maintain Mode: Curate Memory (light sweep scoped to categories touched this session)
5.55. Weakness Analysis (AutoContext-inspired):
# Aggregates signals from pattern signatures, experience archive, and
# backpressure rollbacks into a coherent weakness report. HIGH-severity
# weaknesses auto-create investigation goals. (The guardrail times_active
# source is RETIRED — g-115-2141, re-applied by g-115-2470; see the
# retirement block below.)
# Only runs during --full-cycle.
Read agents/<agent>/weakness-report.yaml (create with {last_analyzed: null, analysis_count: 0, weaknesses: [], signal_baseline: {captured_at: null, guardrail_times_active: {}, rollback_count: 0, pattern_totals: {}}} if missing)
# WINDOWED SIGNAL SOURCES (g-115-2002 / sig-27 CONFIRMED): several raw signals
# below are MONOTONIC CUMULATIVE counters that never stop firing once they
# cross an absolute threshold — they measure lifetime accumulation, not recent
# activity, so they are non-discriminating. utilization.times_active is the
# worst: Phase 0.5a runs `guardrail-check.sh --context any --phase
# pre-selection`, matches_context("any") is unconditionally True, and
# matches_phase filters on STATIC rule-text keywords (PHASE_PRE_SELECTION_KEYWORDS)
# — NOT the runtime condition. So every guardrail whose text matches those
# keywords increments ~once per iteration whether or not it genuinely fired,
# making times_active iteration-correlated (top ~3500). backpressure
# rollback_history is lifetime-cumulative the same way. THE FIX: snapshot each
# signal's value at the END of every pass into weakness_report.signal_baseline;
# this pass thresholds on the DELTA since that baseline — and for the guardrail
# counter, on the delta ABOVE the iteration-correlated BACKGROUND (median delta),
# so a guardrail signals only when it fired MORE than the per-iteration
# keyword-match rate. Windowing keeps agent-judgment synthesis unchanged below;
# only the signal INPUTS change. (The guardrail half of this fix is now
# HISTORY — even windowed, the counter carries no fire information; the
# source is retired outright in the g-115-2141 block below. The windowing
# doctrine stands for pattern totals + any future counter-derived source.)
baseline = weakness_report.signal_baseline (default {captured_at: null, guardrail_times_active: {}, rollback_count: 0, pattern_totals: {}})
first_pass = (baseline.captured_at is null) # no history yet → cumulative-counter sources emit nothing; baseline is written at end-of-pass
# Gather signals — 3 CONSUMED sources: pattern_signatures (windowed,
# script-side), experience, backpressure. Signature windowing (g-115-1905):
# lifetime counters stop discriminating on a mature store (127/722
# guardrails matched the old lifetime `times_active >= 3`; guard-054 sat
# at 3520 because the precheck guardrail-check mass-matches ~58 rules
# every iteration). The baseline lives in weakness-report.yaml
# `signal_baseline:` (advisory-ratchet shape, same pattern as
# meta/audit-baselines.yaml) and is advanced by the script. The guardrail
# source is computed by the script but NOT consumed — retired, block below.
signals = []
# 1. Pattern signatures (windowed accuracy) — computed by
# core/scripts/weakness-signals.py (guard-399: mechanical arithmetic is
# script baseline; LLM synthesis on top). A sig flags when window
# accuracy < 0.70 over >=3 window outcomes. First run seeds the baseline
# and emits no signals. (The script also computes guardrail windowed
# deltas — those are baseline-bookkeeping only, NOT consumed as signals;
# see the retirement block below.)
# rb-245 pre-read gates: verify schema before the script consumes the
# fields. If a gate fails, SKIP this sub-phase (other phases continue).
# Do NOT --override: fix field paths in the script instead.
Bash: source core/scripts/_paths.sh && bash core/scripts/audit-schema-gate.sh \
--jsonl-path "$WORLD_DIR/pattern-signatures.jsonl" \
--field-names "outcome_stats.accuracy,outcome_stats.total"
Bash: source core/scripts/_paths.sh && bash core/scripts/audit-schema-gate.sh \
--jsonl-path "$WORLD_DIR/guardrails.jsonl" \
--field-names "utilization.times_active"
# g-115-2002 windowing is SCRIPT-SIDE (weakness-signals.py, g-115-1905):
# the two implementations met at the 2026-07-11 g-115-2022 merge — upstream
# carried equivalent LLM-side pseudocode (signature total-advance gate;
# guardrail delta-minus-median-background). The script supersedes it:
# signatures get true WINDOW accuracy (delta_confirmed/delta_total >= 3
# outcomes, < 0.70 signals); guardrails get windowed deltas with ambient
# discrimination (>= ambient_mult x median of nonzero deltas — the
# iteration-correlated keyword-match background sig-27 warned about);
# baseline seeding + storage owned by the script (signal_baseline section).
# Script-enforced beats LLM-discretionary (Phase 3.7 doctrine) — do NOT
# re-inline the windowing math here.
Bash: py -3 core/scripts/weakness-signals.py --agent $MIND_AGENT
Parse JSON: seeded, guardrail_signals[], signature_signals[], window_start
IF seeded == true:
Log: "▸ Weakness signals: baseline seeded — windowed signals available from the next analysis"
FOR EACH sig in signature_signals:
signals.append({source: "pattern_signature", id: sig.id, detail: sig}) # window_accuracy < 0.70, window_total >= 3
# Guardrails that fired frequently — RETIRED as a weakness signal source
# (g-115-2141, 2026-07-14; re-applied 2026-07-17 by g-115-2470 after the
# ac3730ea31d7 sync merge silently dropped the retirement). The script's
# guardrail_signals[] output is deliberately NOT consumed here.
#
# WHY RETIRED (rationale updated for the script-side impl this lineage runs):
# (a) guard-841 spirit: a weakness signal must reflect GENUINE activity.
# utilization.times_active increments on keyword-scan matches
# (Phase 0.5a guardrail-check.sh: matches_context("any") is always
# True; matches_phase filters on STATIC rule-text keywords) — so even
# a WINDOWED delta measures keyword-match rate. Windowing satisfies
# the freshness letter of guard-841, but the counter never contained
# fire information to begin with.
# (b) times_active cannot distinguish a real fire from a keyword-scan
# bump — no arithmetic downstream of the counter can recover a
# distinction the counter never recorded.
# (c) residual false-signal paths survive the script's nonzero-median x2
# ambient filter (which DOES defeat the uniform-cohort shape that
# killed the inline variant — pass 19: 61 guards all delta=31 →
# nonzero-median 31 → threshold 62 → zero signals): the <6-movers
# path (AMBIENT_MIN_COHORT) thresholds at bare min_delta, and a
# guard whose rule text matches ~2x more scan keywords than the
# cohort clears 2x-median EVERY pass — a chronic FP carrying zero
# fire content. Empirical yield: 19 passes → zero real weaknesses
# (origin lineage, inline variant); 6 passes → zero (this box,
# script variant).
#
# RE-ADD CONDITION: only when guardrail-check.py logs REAL fires
# (action_hint ran AND surfaced an issue) as a field DISTINCT from
# keyword-scan matches. A real_fires_since_baseline delta can then replace
# this block; a keyword-bump-derived signal never can (guard-841).
# (weakness-signals.py still computes + baselines guardrail deltas for
# schema stability and a clean future re-add; CONSUMPTION alone is
# retired. The rb-245 schema gate on guardrails.jsonl above stays — it
# protects the script's computation, not this retired consumption.)
# 3. Experience records with negative relative_advantage clustered by approach
# (already windowed — recent 20; unchanged by g-115-1905)
Bash: experience-read.sh --recent 20
negative_experiences = filter WHERE relative_advantage < -0.1
IF len(negative_experiences) >= 3:
# Cluster by category
clusters = group_by(negative_experiences, "category")
FOR EACH cluster WHERE len(cluster.items) >= 2:
signals.append({source: "experience_cluster", category: cluster.key, count: len(cluster.items)})
# 4. Backpressure rollback patterns — windowed to the last 14 days
# (lifetime history counted months-old rollbacks as current weakness
# signal; entries carry timestamps — filter, no baseline needed)
Bash: meta-backpressure.sh status
# (g-115-2002 upstream variant used a baseline.rollback_count slice; the
# timestamp filter below is self-contained — entries carry timestamps, no
# baseline dependency — and was kept at the g-115-2022 merge.)
FOR EACH rollback in result.rollback_history WHERE rollback timestamp within last 14d:
signals.append({source: "backpressure_rollback", id: rollback.meta_change_id, detail: rollback})
# Synthesize weaknesses from signals
IF len(signals) >= 2:
# Detect weakness types
# regression: declining performance in a category over time
# stagnation: category with many goals but no capability improvement
# dead_end: same approach keeps failing (feeds into dead end registry)
# systematic_bias: agent consistently over/under-estimates
FOR EACH detected weakness:
existing = find in weakness_report.weaknesses WHERE description matches
IF existing:
existing.last_confirmed = now
existing.times_confirmed += 1
ELSE:
new_weakness = {
id: "wk-{next_num}",
type: detected_type,
description: synthesized_description,
evidence: {
pattern_signatures: [relevant sig IDs],
guardrail_triggers: [relevant guard IDs],
experience_ids: [relevant exp IDs],
meta_log_entries: count_of_relevant
},
severity: HIGH if regression/dead_end else MEDIUM,
first_detected: now,
last_confirmed: now,
times_confirmed: 1,
status: "active",
remediation: {proposed: null, applied: null, goal_id: null}
}
weakness_report.weaknesses.append(new_weakness)
# Create investigation goals for HIGH-severity active weaknesses
FOR EACH weakness WHERE severity == "HIGH" AND status == "active" AND remediation.goal_id is null:
# Check dedup against existing goals
Bash: load-aspirations-compact.sh → IF path returned: Read it
IF no existing "Investigate: {weakness.description}" goal:
goal_json = {
title: "Investigate: {weakness.description (60 chars)}",
status: "pending", priority: "MEDIUM",
skill: null, participants: ["agent"],
description: "Weakness detected by aggregated failure analysis.\nType: {weakness.type}\nEvidence: {weakness.evidence}\nDiscovered by: Step 5.55 Weakness Analysis",
origin_signal: "investigate:weakness-{weakness.type}"
}
# Route to most relevant aspiration
target_asp = aspiration matching weakness category, or most recent active
echo '<goal_json>' | bash core/scripts/aspirations-add-goal.sh {target_asp}
weakness.remediation.goal_id = created_goal_id
Output: "▸ WEAKNESS ANALYSIS: Created investigation goal for {weakness.description}"
weakness_report.last_analyzed = now
weakness_report.analysis_count += 1
# Snapshot every windowed signal's CURRENT value so the NEXT pass diffs
# against it (g-115-2002). This write is what turns the cumulative counters
# above into windowed deltas — without it the delta math has no baseline and
# every source degrades to first_pass (emits nothing). Snapshot ALL active
# guardrails/patterns (not just the ones that signalled) so a guardrail that
# crosses threshold NEXT window is measured from its value NOW. Runs on EVERY
# pass — including passes with < 2 signals that skip synthesis — so the window
# advances every cycle, never stalls at a stale baseline.
weakness_report.signal_baseline = {
captured_at: now,
guardrail_times_active: {g.id: g.utilization.times_active FOR EACH active guard},
rollback_count: len(result.rollback_history),
pattern_totals: {sig.id: sig.outcome_stats.total FOR EACH active sig}
}
Edit agents/<agent>/weakness-report.yaml with updated content
Output: "▸ Weakness analysis: {len(signals)} signals, {new_weakness_count} new weakness(es), {goal_count} investigation goal(s)"
5.7. Meta-Reflection ROI Tracking:
For each reflection mode invoked in this cycle, track:
- Did it produce a reasoning bank entry, guardrail, or pattern signature?
- Did it add an encoding queue item?
- Did it change a belief or knowledge node?
Compute: reflection_roi = artifacts_produced / modes_invoked
Append via meta-yaml.py append to reflection-strategy.yaml roi_history:
{date: today, modes_invoked: N, artifacts_produced: N, roi: N, session: N}
5.8. Reflection Quality Consolidation (MR-Search Priority 2):
Update reflection_effectiveness_by_type from reflection_quality_log:
For each entry in reflection_quality_log:
Derive type from reflection_id prefix (ref-{goal_id} → look up goal's spark/execution context)
Count by type: entries where helpful == true are "effective"
For each reflection type (execution, hypothesis, spark):
total = count entries of this type
effective = count entries of this type where helpful == true
rate = effective / total (or 0.0 if total == 0)
Bash: meta-set.sh reflection-strategy.yaml reflection_effectiveness_by_type '<updated_json>'
This closes the meta-learning loop: reflection quality → depth allocation → better reflections
6. invoke /replay --sharp-wave --selective (if violations detected)
7. Tree Health Lint (wiki integrity check):
# Periodically verify the knowledge tree's structural and content health.
# Inspired by Karpathy's wiki "health checks": find inconsistencies,
# flag stale data, discover missing cross-references.
Bash: tree-read.sh --stats
# Staleness check: heavily-used nodes that haven't been updated recently
FOR EACH node where retrieval_count > 10 AND last_updated older than 5 sessions:
echo '{"node_key": "<key>", "reason": "stale-high-retrieval", "retrieval_count": <N>, "sessions_since_update": <M>, "priority": "MEDIUM"}' | wm-append.sh knowledge_debt
Log: "▸ Tree lint: {node.key} flagged stale (retrieved {retrieval_count}x, last updated {sessions_ago} sessions ago)"
# Cross-reference discovery: nodes that share entities but aren't linked
Bash: world-cat.sh knowledge/tree/_tree.yaml # entity_index
IF entity_index is non-empty:
FOR EACH entity appearing in 2+ nodes:
Check if those nodes have cross-references to each other in their .md files
IF no cross-reference exists:
Add "See also: [{other_node}]({other_node.file})" to both nodes
Log: "▸ Tree lint: cross-reference added between {node_a.key} and {node_b.key} (shared entity: {entity})"
# Width check: interior nodes exceeding K_max children
Read core/config/tree.yaml for K_max
FOR EACH interior node where child_count > K_max * 2:
Log: "▸ Tree lint: {node.key} has {child_count} children (K_max={K_max}) — consider reorganization"
bash core/scripts/tree-update.sh --set <node.key> growth_state ready_to_decompose
Report: "Tree lint: {stale_count} stale nodes flagged, {xref_count} cross-references added, {wide_count} wide nodes flagged"
Integration Points
- Called by
/aspirations: After every goal completion (spark check); --batch-micro at session-end consolidation
- Called by
/aspirations-state-update: Step 8.75 after productive goal execution with notable outcomes
- Called by
/review-hypotheses --learn: For each resolved hypothesis with reflected: false (horizon gate routes session→lightweight, short/long→full)
- Calls
/replay: During full-cycle mode (Step 2.5) for hippocampal replay
- Calls
/research-topic: When a knowledge gap is identified
- Calls
/aspirations add: When a new aspiration emerges from patterns
- Calls
/reflect-tree-update: Shared tree update protocol used by reflect-on-outcome (Hypothesis mode) and reflect-on-self (Patterns mode)
- Calls
/tree maintain: When new categories detected or article counts cross thresholds (Step 8.5)
- Updates discovery filters: Adds new trap types to discovery lessons-learned
- Updates evaluation calibration: Adjusts evaluation weights based on calibration data
- Updates pattern signatures (via
pattern-signatures-add.sh, pattern-signatures-record-outcome.sh): New signatures, accuracy updates, separation markers
- Updates working memory (via
wm-append.sh): Encoding queue items from Step 2.5
- Updates
agents/<agent>/developmental-stage.yaml: Schema operations (assimilation/accommodation)
- Updates
meta/skill-gaps.yaml: Capability gap detection (Spark Q6) and skill underperformance (Spark Q7)
Active Forgetting — Decay Formula Reference
retention_score = base_decay^days_since_last_access × (1 + retrieval_count × retrieval_boost)
Default parameters: base_decay=0.95, retrieval_boost=0.15.
See /reflect-maintain (Memory Curation mode) for full active forgetting procedures (interference detection,
reconsolidation windows, protection rules).
Return Protocol
See .claude/rules/return-protocol.md — last action must be a tool call, not text.
