agentleFS
jsboigejsboige/CoursIA/.vibe/AGENTS.md

CoursIA

AGENTS.md17 starsChanged yesterdayMITView on GitHub

Workspace: CoursIA Machine: myia-po-2025 Lane canonique: myia-po-2025:Microsoft VS Code Role: Worker autonome a cadence 4 h, livrables en relais humain (jamais de push/PR autonome) Commande: .vibe/commands/continue.md Config locale: .vibe/config.toml (override global si besoin) Outils disponibles (noms complets avec server prefix) : - roo-state-manager_roosync_dashboard - roo-state-manager_roosync_messages - roo-state-manager_roosync_search - roo-state-manager_codebase_search - roo-state-manager_conversation_browser - Et 10 autres... Permissions: permission = "always" dans config.toml Phase 1 : Contexte 1. roosync_messages(action:"inbox", status:"unread") 2. roosync_dashboard(action:"read", type:"workspace", section:"all") 3. Filtrer par lane : myia-po-2025:Microsoft VS Code…

The file

We split this AGENTS.md into the rules it gives an agent

2 rules. About 0 are common, 1 is distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Commit LOCAL uniquement (branche feature/<sujet>, message conventionnel) — JAMAIS de push

    line 65
  2. Distinctive:

    Commit local AVANT tout rapport

    line 68

≈ “Common” is an estimate: at least 1 in 5 of the 41,581 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
AGENTS.md

AGENTS.md - Mistral Vibe Workspace Configuration (CoursIA)

Instructions specifiques au workspace D:\dev\CoursIA

Workspace: CoursIA
Machine: myia-po-2025
Lane canonique: myia-po-2025:Microsoft VS Code Role: Worker autonome a cadence 4 h, livrables en relais humain (jamais de push/PR autonome) Commande: .vibe/commands/continue.md
Config locale: .vibe/config.toml (override global si besoin)


Configuration locale

Fichiers du harnais

D:\dev\CoursIA\.vibe\
├── config.toml          # Config MCP locale (override ~/.vibe/config.toml)
├── AGENTS.md           # Instructions workspace (ce fichier)
└── commands\
    └── continue.md      # Commande worker principale

MCP roo-state-manager (stdio local)

Outils disponibles (noms complets avec server prefix) :

  • roo-state-manager_roosync_dashboard
  • roo-state-manager_roosync_messages
  • roo-state-manager_roosync_search
  • roo-state-manager_codebase_search
  • roo-state-manager_conversation_browser
  • Et 10 autres...

Permissions: permission = "always" dans config.toml


Commande Worker : .vibe/commands/continue.md

Cadence : feeder 4 h, probe --wake 1 h

  • Le feeder externe, arme toutes les 4 h, envoie le payload de travail.
  • La schtask horaire reste un probe --wake : no-op tant qu'aucun payload n'est en attente.
  • Ce prompt ne s'exécute que lorsque le feeder/wake fournit un payload — aucun arbitrage de cadence in-prompt.

Workflow d'un payload

Phase 1 : Contexte

  1. roosync_messages(action:"inbox", status:"unread")
  2. roosync_dashboard(action:"read", type:"workspace", section:"all")
  3. Filtrer par lane : myia-po-2025:Microsoft VS Code
  4. git fetch origin main ; travail en worktree isole feature/<sujet> depuis origin/main

Phase 2 : Selection — le picker d'abord (P0 inclus)

  • Premier geste de selection : le picker calibre, apres lecture inbox/dashboard — jamais de scan manuel du pool avant lui : python scripts/pick_idle_grain.py --belt --lane "myia-po-2025:Microsoft VS Code" --json (verification des claims active ; --no-check-claims interdit)
  • Cle "repair" non nulle : P0 — le picker retourne le backlog rouge/review de la lane ; la reparation de la PR nommee (repair.grain) EST le grain du cycle, preparee en relais humain comme tout livrable Vibe.
  • Sinon : prendre les picks du tapis dans l'ordre (derniere visite la plus ancienne en tete), en ne sautant que les candidats hors profil Vibe : mono-sujet, CPU/local, verifiable textuellement. Exclus : GPU/vision, QuantConnect, Lean froid, notebooks a re-executer, travail relationnel/multi-fichiers, catalogue (enumeration des PRs de la lane via le tag Grain: du body, jamais --author @me : faux 0 silencieux sur cette machine multi-comptes).
  • Avant d'editer : check_lane_claim.py --lane "myia-po-2025:Microsoft VS Code" <N>, puis [CLAIMED] en commentaire d'issue (tag Grain: et clause paths: sur deux lignes distinctes).

Phase 3 : Execution

  • 1 sujet = 1 livrable, catalogue byte-identique a main
  • Outils SOTA, jamais workaround (regle F) ; env casse : diagnostiquer et relayer
  • Commit LOCAL uniquement (branche feature/<sujet>, message conventionnel) — JAMAIS de push

Phase 4 : Rapport + relais humain

  1. Commit local AVANT tout rapport
  2. [REVIEW-NEEDED] sur le dashboard : branche, commit, fichiers, resume du diff, preuves, tests, tag Grain: propose — l'humain relit puis pousse / ouvre la PR
  3. [DONE] myia-po-2025:Microsoft VS Code <resume> -- branche <nom> -- grade: <A/B/C>
  4. [ASK USER] separe si bloqueur

Fallback MCP

Lecture directe des fichiers GDrive

G:/Mon Drive/Synchronisation/RooSync/.shared-state/dashboards/workspace-CoursIA.md
G:/Mon Drive/Synchronisation/RooSync/.shared-state/dashboards/workspace-CoursIA-2.md
G:/Mon Drive/Synchronisation/RooSync/.shared-state/dashboards/workspace-roo-extensions.md

Commande git/gh

gh issue list --state open --limit 100 --json number,title,labels
# PRs ouvertes de la lane, via le tag Grain: du body (jamais --author @me)
gh pr list --state open --limit 100 --json number,title,body --jq '.[] | select(.body != null and (.body | contains("myia-po-2025:Microsoft VS Code"))) | "#\(.number) \(.title)"'

Scheduler

  • Feeder externe : arme toutes les 4 h, il envoie le payload de travail.
  • Schtask horaire : probe --wake, no-op tant qu'aucun payload n'est en attente.
  • Le prompt .vibe/commands/continue.md ne s'exécute que sur payload fourni par le feeder/wake — aucun arbitrage de cadence in-prompt.

Notes specifiques Vibe

  1. Pas de session persistante -> etablir contexte depuis dashboard
  2. CLI non-interactif -> checkpoints frequents dans dashboard
  3. mistral-medium-3.5 -> utiliser bash pour execution code
  4. Echeance a ~25 min : cloturer (commit local + [REVIEW-NEEDED] + [DONE]), jamais de travail non commite a l'echeance
  5. MCP stdio fonctionne -> utiliser noms complets: roo-state-manager_*
  6. Push et ouverture de PR ne sont JAMAIS autonomes : relais humain [REVIEW-NEEDED]

Exemples

# Lire dashboard
roosync_dashboard(action: "read", type: "workspace", section: "all")

# Poster un livrable en relais humain
roosync_dashboard(action: "append", type: "workspace", tags: ["ASK"],
  content: "[REVIEW-NEEDED] myia-po-2025:Microsoft VS Code -- issue #N : branche feature/<sujet>, commit <SHA>, diff/preuves/tests ci-dessous")

# Poster la fin de payload
roosync_dashboard(action: "append", type: "workspace", tags: ["DONE"], content: "[DONE] myia-po-2025:Microsoft VS Code : <resume> -- branche feature/<sujet> -- grade: A")

# Chercher
roosync_search(query: "fix main red", limit: 10)

Erreurs connues

ErreurSolution
Tool call failed (silent)Verifier config.toml, redemarrer
HTTP 502/429Attendre 5h, fallback GDrive
MCP not foundVerifier cwd dans config.toml

Changelog

DateActionPar
2026-08-20Creation + correction MCPjsboige + Mistral Vibe
2026-08-20Annonce dans 3 dashboardsMistral Vibe
2026-09-01Alignement #14114 : picker calibre en premier geste de selection (P0 rouge/review via mode repair), relais humain [REVIEW-NEEDED] (plus de push/PR autonome), cadence feeder 4 h / probe --wake 1 h, retrait de --author @mejsboige + Claude Code

"La mer, pas le burin" — Alexandre Grothendieck

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.