4.3 KiB

Raw Blame History

RUNBOOK — Bifrost VK OpenCode Fix + Veille Fallback

Date : 2026-06-23
Contexte : Veille bolbol.tn — "Resume indisponible" systematique

Symptomes

00:00 et 19:00 : could not auto resolve a provider
07:00 : GoUsageLimitError — Weekly usage limit reached

Diagnostic

Erreur 1 — could not auto resolve

Node n8n Enrichir via Bifrost (workflow 9378593adccf42dc) envoie x-bf-vk: bfk-e5832bfe (Nabil-Key), model: deepseek-v4-flash. La Nabil-Key VK n'avait pas OpenCode dans governance_virtual_key_provider_configs. Bifrost cherchait deepseek-v4-flash dans openrouter/google/nvidia_nim -> introuvable.

Erreur 2 — GoUsageLimitError

Job cron hermes-nyora veille-hebdo-nexum-kanban-a1b2c3d4 avait model/provider null. Orchestrateur -> provider par defaut = OpenCode Go direct -> quota hebdo epuise lundi. NOTE: Les 6 workers kanban utilisent le profil veilleur (openrouter) -> OK. Seul l'orchestrateur est affecte.

Fixes appliques

Fix 1 — Bifrost DB: OpenCode ajoute a Nabil-Key VK

conn = sqlite3.connect('/mnt/docker/bifrost/data/config.db')
cur = conn.cursor()
vk = 'dc17fe0c-f7bf-4752-acb8-798156b2ec9e'
models = '["deepseek-v4-flash","deepseek-v4-pro","glm-5.1","kimi-k2.6"]'
cur.execute(
    "INSERT INTO governance_virtual_key_provider_configs "
    "(virtual_key_id, provider, weight, allowed_models, blacklisted_models, allow_all_keys) "
    "VALUES (?, 'opencode', 1.0, ?, '[]', 1)", (vk, models))
conn.commit()

Fix 2 — Bifrost DB: OpenCode key persistee dans config_keys

kid = str(uuid.uuid4())
now = datetime.datetime.utcnow().isoformat() + 'Z'
cur.execute(
    "INSERT INTO config_keys (name,provider_id,provider,key_id,value,models_json,"
    "blacklisted_models_json,weight,enabled,encryption_status,created_at,updated_at) "
    "VALUES ('opencode-1',36,'opencode',?,?,?,'[]',1.0,1,'plain_text',?,?)",
    (kid, BIFROST_OPENCODE_KEY, models, now, now))
conn.commit()

Fix 3 — Bifrost restart (requis apres modif DB)

Via Portainer API http://172.17.0.1:9000 (endpoint_id=2, container 3136d565d6dd). Auth: POST /api/auth -> jwt. Actions: stop + start sur /api/endpoints/2/docker/containers/{id}/{action} ATTENTION: container ID change apres recreate -> verifier avant.

Fix 4 — jobs.json hermes-nyora: orchestrateur job 2

Fichier: /mnt/docker/hermes-platform/hermes-nyora/data/cron/jobs.json Job veille-hebdo-nexum-kanban-a1b2c3d4 -> model: deepseek/deepseek-v4-flash, provider: openrouter

Fix 5 — n8n workflow 9378593adccf42dc: fallback gemini-2.5-flash

Node Enrichir via Bifrost modifie: si GoUsageLimitError -> retry avec gemini-2.5-flash. Via n8n API: PUT /api/v1/workflows/9378593adccf42dc avec {name, nodes, connections, settings}.

Architecture finale (VK Nabil-Key bfk-e5832bfe)

Provider	Modeles autorises
openrouter	Gemma-4-31B/26B free, Nemotron, GPT-OSS free, Hermes-3, Qwen3-coder
google	gemma-4-31b/26b, gemini-2.5-flash, gemini-2.5-flash-lite
nvidia_nim	Llama-3.3-70B, Nemotron-70B, Llama-405B, Qwen2.5-Coder
opencode	deepseek-v4-flash, deepseek-v4-pro, glm-5.1, kimi-k2.6 (NOUVEAU)

Pieges a eviter

Bifrost ne hot-reload PAS les providers -> restart obligatoire apres modif DB
config.json charge providers au boot en memoire; config_keys DB = source persistante
auth Bifrost = x-bf-vk header, PAS Authorization: Bearer -> 401
Portainer container ID change apres recreate -> verifier avant restart
N8N PUT /workflows: accepte uniquement {name, nodes, connections, settings} -> 400 sinon

MCP NAS — Regles de scripting fiables

NE PAS faire	Faire a la place
Heredoc << 'EOF'	Base64 encode -> transfer -> decode
printf 'long code python'	python3 >400c
Commands dans session apres plusieurs err	Nouvelle session ou session differente
Pipes complexes inline dans printf	Script intermediaire en /tmp/

4.3 KiB Raw Blame History