[Axellum]

Summary

• Sync miroir public avec moteur_agents#157
• Routes setup/vocal/ha/benchmarks + ihm-v2
• IPs LAN → placeholders

Test plan

• CI secrets scan + pytest
• Pas d'IP 192.168.0.* hors outil d'anonymisation

---

Note

High Risk
DreamCoder exécute des pipelines LLM sur un dépôt Git configurable avec commits et drainage multi-tâches ; les changements BudgetGuard et billing touchent la sélection de providers et les coûts en prod.

Overview
Synchronisation moteur ↔ IHM v2 autour du principe « pas de donnée inventée » : nouvelles routes de lecture (/api/setup/diagnostics, /api/vocal/*, santé et liste d’entités HA, benchmarks GET/POST avec tables SQLite), agrégation Comptes API via GET /api/billing (quotas, soldes live, saisie manuelle), et doc ihm-v2 (auth REST + ticket SSE, pas de mocks).

DreamCoder passe d’une tâche par cycle à une boucle de drainage bornée (dreamcoder_max_cycle_minutes), avec exécuteur vs gouverneur (DeepSeek / GLM selon budget et retries), rollback, rapports de cycle, et dépôt Git optionnel dreamcoder_repo_path (clone Deck + pull --ff-only).

BudgetGuard réordonne la cascade (rotation round-robin des API cloud gratuites/abonnements avant Ollama), ajoute CLI Gemini/Claude, DashScope, plafond total_daily_budget_usd, solde DeepSeek caché, cooldown après échec, et sync_source / get_scoped_spend_usd pour l’auditeur.

Home Assistant : TLS centralisé avec HA_TLS_SERVER_HOSTNAME (cert IP/DNS), sessions requests/aiohttp partagées (MCP, vocal, Tab5). Vocal : priorité entités « vivantes », consignes TTS sans JSON.

LLM : provider DashScope Coding Plan, clés Cerebras free vs payant séparées. .env.example réécrit (sécurité, ToolMaker, workspace root, placeholders ${HA_HOST} etc.). /version préfère .deploy_meta.json au git local sur le Deck.

^{Reviewed by Cursor Bugbot for commit 54ecc61. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 4 potential issues.

Bugbot Autofix prepared fixes for all 4 issues found in the latest run.

• ✅ Fixed: Shell placeholders in Python URLs
  • Replaced literal ${HA_HOST:-…} fallbacks with os.environ.get('HA_HOST', …) in HA health, MCP tools, vocal credentials, and Tab5 pusher defaults.
• ✅ Fixed: Ollama LAN URL not resolved
  • ollama_pc base_url now resolves OLLAMA_HOST via os.environ at registry load and again at provider creation in create_provider/_make_compat.
• ✅ Fixed: Provider cooldown on git errors
  • mark_provider_failed is gated behind provider_execution_failed so successful pipeline runs that later fail on git/IO no longer cool down the LLM provider.
• ✅ Fixed: Governor cap uses wrong scope
  • Governor budget check now uses BudgetGuard.get_total_spend_usd (all billing sources) against total_daily_budget_usd, matching get_available_provider.

Or push these changes by commenting:

@cursor push b2d07f8258

_{You can send follow-ups to the cloud agent here.}

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 54ecc61. Configure here.}

[cursor]

Shell placeholders in Python URLs

High Severity

When HASS_URL / HA_URL are unset, fallbacks use literal strings like http://${HA_HOST:-192.168.1.x}:8123. Python does not expand that shell syntax, so HA health, MCP tools, vocal reads, and Tab5 pushes target an invalid host instead of a LAN IP.

Additional Locations (3)

• core/mcp_tools/homeassistant.py#L40-L41
• core/vocal_tools.py#L107-L108
• core/tab5_pusher.py#L33-L34

 

^{Reviewed by Cursor Bugbot for commit 54ecc61. Configure here.}

[cursor]

Ollama LAN URL not resolved

Medium Severity

The ollama_pc provider base_url is set to http://${OLLAMA_HOST:-192.168.1.x}:11434/... as a literal string. HTTP clients use that hostname verbatim, so Deck-to-PC Ollama calls fail unless another layer rewrites the URL.

 

^{Reviewed by Cursor Bugbot for commit 54ecc61. Configure here.}

[cursor]

Provider cooldown on git errors

Medium Severity

DreamCoder’s shared except always calls mark_provider_failed(provider) after any pipeline-stage error. Git commit/diff failures or timeouts can mark a healthy LLM provider unavailable for 30 minutes even when the model call succeeded.

 

^{Reviewed by Cursor Bugbot for commit 54ecc61. Configure here.}

[cursor]

Governor cap uses wrong scope

Medium Severity

Governor degradation compares get_scoped_spend_usd("dreamcoder", …) to total_daily_budget_usd, described as a combined daily cap. That ignores non-DreamCoder spend recorded in billing_history, so paid planner/reviewer models may keep running after the real combined budget is exhausted.

 

^{Reviewed by Cursor Bugbot for commit 54ecc61. Configure here.}