Claude Code Extension Points
Agents (.claude/agents/)
Agents are auto-discovered by Claude Code. Each has YAML frontmatter with model, tools, memory, and skills configuration. The roster below lists 9 representative agents — the full roster (21 agents) with descriptions, mapping Epic → specialist, et mapping side-track est dans subagents-reference.md (la source de vérité).
| Agent | Model | Purpose |
|---|---|---|
| notebook-iterative-builder | inherit | Orchestrate build/improve/fix cycles |
| notebook-executor | sonnet | Execute notebooks via MCP |
| notebook-validator | sonnet | Validate all quality aspects |
| notebook-enricher | sonnet | Add pedagogical content |
| notebook-cleaner | sonnet | Fix markdown structure |
| notebook-designer | inherit | Create new notebooks |
| notebook-cell-iterator | sonnet | Fix specific cells iteratively |
| readme-updater | haiku | Update README files |
| readme-hierarchy-auditor | haiku | Audit README hierarchy |
Skills (.claude/skills/)
| Skill | Type | Description |
|---|---|---|
| notebook-helpers | Reference (auto) | Script reference for notebook manipulation |
| mcp-jupyter | Reference (auto) | MCP Jupyter tools and patterns |
| notebook-patterns | Reference (auto) | Enrichment patterns (GameTheory model) |
| verify-notebooks | User (/command) |
Verify and test notebooks |
| enrich-notebooks | User (/command) |
Enrich with pedagogical content |
| cleanup-notebooks | User (/command) |
Clean markdown structure |
| build-notebook | User (/command) |
Create/improve/fix notebooks |
| execute-notebook | User (/command) |
Execute via MCP |
| validate-genai | User (/command) |
Validate GenAI stack |
Slash commands utiles
/verify-notebooks [target] [--quick] [--fix] # Verify and test notebooks
/enrich-notebooks [target] [--execute] [--strict] # Add pedagogical content
/cleanup-notebooks [target] [--dry-run] # Clean markdown structure
/build-notebook <action> <path> [--quality=90] # Create/improve/fix notebooks
/execute-notebook <path> [--batch] [--save] # Execute via MCP
/validate-genai [target] [--local] # Validate GenAI stack
Rules (.claude/rules/)
Auto-chargées à chaque session, sauf celles portant un frontmatter paths: (voir CLAUDE.md §Règles modulaires) — leur contenu est déjà en contexte pour un agent en cours de travail. Cet inventaire sert à qui lit le dépôt sans session active, ou cherche quelle règle porte quel sujet.
| Règle | Ce qu’elle porte |
|---|---|
git-workflow |
branches, messages de commit, périmètre force-push, scan de branche orpheline |
pr-review-discipline |
critères CHANGES_REQUESTED §A-H (composites, Lean, ML multi-seed, notebooks, docs, audit, QC, SOTA) |
anti-regression |
patterns red-flag, protocole avant suppression de preuve/implémentation |
notebook-conventions |
structure, exécution kernel, C.1-C.5 |
exercise-example-labeling |
classification par contenu, fin du flip-flop exercice/exemple |
code-style |
PEP 8, .NET 9, pas d’emojis, i18n Lean FR/EN siblings |
genai-config |
services, env, scripts, architecture GenAI |
wsl-kernels |
kernels WSL (GameTheory/**, Lean/**) |
student-pr-reviews |
anti-fuite soutenance, review bienveillante |
public-contribution-triage |
accueil, triage et escalade des issues/PRs publiques |
lean-merge-discipline |
gates de merge propres au Lean |
secrets-hygiene |
content-based, Stop & Repair, canal RooSync privé + quorum (fusion de secrets-roosync-policy, #12051) |
audit-reassessment · audit-cross-source-distillation |
protocole 4 étapes, sortie = dashboard/issue jamais un fichier |
verify-before-claiming |
G.1, firsthand avant tout claim |
coordinator-discipline |
R1-R7 ai-01 (merge actif, no languishing, lanes indépendantes, jamais sanctionner l’idle, steer qui atteint, adjoint, registre re-parcouru des arbitrages user) |
proactive-coordination |
plancher 1 PR/wakeup, pool global, picker, never-idle, L721/L740/L898 |
user-blocker-signaling |
anti-dilution des bloqueurs user |
harness-hygiene |
3 tiers harnais / docs / dashboard |
bibliography-hygiene |
publications utilisées archivées sur le GDrive canonique, jamais dans Git |
catalog-pr-hygiene |
le catalogue appartient à l’automatisation |
model-delegation |
modèle explicite obligatoire, routage vision |
three-exercises-per-notebook |
richesse pédagogique |
sota-not-workaround |
5 verdicts, checklist 6 axes INTRINSIC, Prong B |
readme-french-first |
français d’abord, anglais préservé en .en.md |
variation-protocol |
tag Grain:, gates G-VAR-1/2/3, merge-gate |
lane-claim-protocol |
claim = commentaire d’issue, paths:, anti-collision cross-lane |
cell-interpretation-ordering |
ancrage sémantique des cellules d’interprétation |
Model Selection Strategy
Modèle explicite obligatoire sur tout Agent() — un sous-agent sans model hérite du tier parent (opus) et annule l’économie de la délégation. Règle complète : .claude/rules/model-delegation.md.
haiku: tâches simples (comptage, extraction format-imposé, grep/scan, vérification mécanique)sonnet: tâches intermédiaires (audit, recensement, diagnostic, rédaction, enrichissement, review structurelle)opus: uniquement sur justification écrite d’une ligne dans le prompt (décision architecturale cross-fichier, synthèse multi-sources contradictoires, régression profonde)
Proactive Behaviors
- After completing notebook work, update agent memory with lessons learned
- After enrichment, verify cell placement with git diff
- Before executing GenAI notebooks, validate the stack with
/validate-genai - When encountering repeated errors, record the pattern in memory for future reference
- When working with notebooks, use the helper scripts (not ad-hoc Python)