docs(backlog): epic consolidation & maturité ulk (v7) + 5 cards#194
Merged
Conversation
Epic capturing the 'harness > modèle' lesson into operational work: ulk grows faster than it consolidates (171 agent files, divergent counts, manual numbering collisions). - milestone-consolidation-maturite: epic + governance rule (un entre = un sort) - FEAT-agent-usage-stats: ulk agents stats → mesurer l'usage réel, archiver la traîne morte - FEAT-agent-id-automation: auto-ID + golden test anti-collision (harvey/thor #73) - FEAT-token-benchmark-harness: benchmark reproductible versionné (prouver les gains) - FEAT-ulk-init-onboarding: ulk init interactif (abaisser la barrière d'entrée) - SPEC-declarative-workflows: orchestration déterministe (à cadrer, anti sur-ingénierie) https://claude.ai/code/session_01JrFhrsWDBZVZBRMN6MiXG4
… grooming, agent verify) Second batch from the 'tu vois autre chose ?' deep-dive: - FEAT-dogfooding-self-optimization (🔴): CLAUDE.md ≈4005 tokens loaded every session; run caveman-compress + /context-audit on ulk itself, in CI. Ultimate credibility test + best marketing angle. - TASK-coordinate-detachment-maturity (🟠): unify consolidation epic with the in-flight GitHub/cloud detachment thread (5 OPS cards 2026-05-31) under one 'maturité v7' banner; decide CI infra before wiring new jobs. - FEAT-backlog-grooming-routine (🟠): 21 todo cards accumulating (oldest 30 apr); no WIP limit — paradoxical for a project shipping a Kanban + checkpoint agent. - SPEC-agent-behavioral-verify (🟡): agents have structural golden tests only; wire verify (65) for semantic meta-conformance (after consolidation #1). Epic milestone updated 5→9 cards. https://claude.ai/code/session_01JrFhrsWDBZVZBRMN6MiXG4
…5→3736 tok) Passe 1 of the dogfooding chantier (#6): remove pure duplication from the always-loaded CLAUDE.md, content already present in pointed-to files. - Skills (opt-in): design/gstack paragraphs → pointer to install-reference.md - Effort + Dynamic Workflows: trim verbose prose, keep existing pointers - Multi-LLM: model alias table → pointer to multi-provider-protocol.md -1076 chars / -269 tokens / -7 lines. vault:* block untouched, export unaffected (export.go does not read CLAUDE.md), 7 pointers preserved. Passes 2 (aggressive cuts) and 3 (CI) noted in card pending human arbitration / detachment coord. https://claude.ai/code/session_01JrFhrsWDBZVZBRMN6MiXG4
…DIGEST.md Dogfooding passe 2 (#6): move the auto-generated vault block out of the always-loaded CLAUDE.md into docs/_memory/VAULT-DIGEST.md, leaving a pointer. CLAUDE.md: 4005 → 3398 tokens (-15% total vs baseline, -607 tok). Retargeted the lovecraft memory loop (prompt-driven, no script): - _shared/vault-memory.md: flux, subcommand table, Distribute-4/5/6, template links - _shared/memory-protocol.md: flux, section 3, idempotence, gandalf check, sandbox - _shared/auto-dream-protocol.md: projection target, write sandbox, obsolescence refs - session/34-gandalf.md: Phase 5 vault health check (digest + pointer) - orchestrators/47-lovecraft.md: description, mode table, command table, promotion rule - regenerated registry + compiled gandalf/lovecraft commands Tests: 685/689 pass — the 4 failures (georges golden, managed-agents yaml) are pre-existing, confirmed on a clean tree, unrelated to this change. Tradeoff: critical lessons no longer auto-surface in raw sessions — they surface via godspeed (00) phase 1.5 / bruce, accepted to save ~400 tok on every session. https://claude.ai/code/session_01JrFhrsWDBZVZBRMN6MiXG4
Follow-up to the Knowledge Vault externalization: blog articles described the vault as living 'in CLAUDE.md'. Update the factual references (distribute target tables + pro/cons mechanism lines, FR+EN) to docs/_memory/VAULT-DIGEST.md (pointed to from CLAUDE.md). High-level narrative mentions left as-is. https://claude.ai/code/session_01JrFhrsWDBZVZBRMN6MiXG4
izo
force-pushed
the
claude/agentic-ai-kitchen-metaphor-4svND
branch
from
e195471 to
42e1bf7
Compare
izo
added a commit
that referenced
this pull request
Jun 8, 2026
…gistry, fix georges golden Porte de pré-release Black Emperor (mode release, verdict GO) : - Supprime les clés `durability:` dupliquées dans 8 frontmatter d'agents (astride, sargeras, ed209, bruce, lovecraft, 2b3, gandalf, rodin) — doublons introduits par l'édition manuelle du commit 7e77729, pas par apply-durability.cjs (idempotent). Impact runtime nul (last-wins), hygiène YAML. - Régénère registry.json (drift post-merge : dates + hallmark/pm-skills/desc lovecraft). - Réaligne le golden de test georges sur l'agent réécrit (claude-code-video-toolkit) — fixture obsolète depuis 233d505, antérieure aux PR #194-198. Gates : Go 130 pass · @ulk/core ✓ · @ulk/status-board ✓ · Biome ✓ · Node 689/689. Sécurité ed209 9/10 (0 blocker) · Qualité vision 8/10 (0 bug fonctionnel). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Contexte
Issue de la discussion « harness > modèle » (diagramme Theory Ventures / Tunguz). ulk est une preuve par l'implémentation du harness à 7 couches — sa force (documenter le harness en Markdown plutôt que le coder dans un SDK) est devenue son risque : le harness grossit plus vite qu'il ne se consolide.
Diagnostic factuel du repo (2026-06-01)
CLAUDE.mddit « 95 agents » · registry « prochain #76 » · 171 fichiers.mdharvey #73/thor #73→ commit « renumber to resolve #73 collision »Ce que cette PR ajoute
Uniquement de la documentation Faru (6 cards backlog) — aucun code touché.
milestone-consolidation-maturite— epic + règle d'or de gouvernance : un agent ne peut entrer que si un autre sort, ou si l'usage réel prouve un manqueFEAT-agent-usage-stats—ulk agents stats: mesurer l'usage réel viaaccountability.jsonl, réconcilier le comptage, identifier la traîne morte (171 → ~40)FEAT-agent-id-automation— auto-ID + golden test anti-collision (CI bloque les doublons de numéro)FEAT-token-benchmark-harness— benchmark reproductible versionné pour prouver les gains tokensFEAT-ulk-init-onboarding—ulk initinteractif (3 questions → profil) pour abaisser la barrière d'entréeSPEC-declarative-workflows— spec d'orchestration déterministe, cadrée pour éviter la sur-ingénierie (defer si pas de besoin prouvé)Séquencement
#1 et #2 (les deux 🔴) sont indépendants et partent en parallèle. #1 débloque la décision de consolidation. #4 dépend de #1. #5 est exploratoire (go/no-go/defer après spec).
https://claude.ai/code/session_01JrFhrsWDBZVZBRMN6MiXG4
Generated by Claude Code