feat(sessions): per-job provider routing — distribute across quota buckets

[brianjones-v4n]

Summary

Today SessionManager.spawnSession launches config.claudePath for every session — interactive and scheduled. This adds optional per-job provider routing so background scheduler jobs can be routed through a cheaper local provider (e.g., an Ollama wrapper running claude against DeepSeek) while interactive Telegram-driven sessions stay on native Anthropic Claude. Operators get to balance work deliberately across two quota buckets.

API surface

// types.ts
export interface ProviderConfig {
  claudePath: string;
  env?: Record<string, string>;
  modelTiers?: Partial<Record<ModelTier, string>>;
}

// SessionManagerConfig
providers?: Record<string, ProviderConfig>;

// JobDefinition
provider?: string;   // key into SessionManagerConfig.providers

.instar/config.json:

"sessions": {
  "claudePath": "/usr/local/bin/claude",
  "providers": {
    "anthropic": {
      "claudePath": "/usr/local/bin/claude",
      "modelTiers": { "haiku": "claude-haiku-4-5", "sonnet": "claude-sonnet-4-6", "opus": "claude-opus-4-7" }
    },
    "ollama": {
      "claudePath": "/usr/local/bin/ollama-launch-claude",
      "env": { "ANTHROPIC_BASE_URL": "http://localhost:11434" },
      "modelTiers": { "haiku": "qwen2.5:7b", "sonnet": "deepseek-v4-pro:cloud" }
    }
  }
}

.instar/jobs.json entries gain "provider": "ollama" to opt in. Omitting provider keeps the existing behavior verbatim.

Backward compatibility

• All new fields are optional.
• Jobs without provider produce the same spawn command as before (same binary, same env, same --model flag).
• Configs without a sessions.providers block work unchanged.
• Existing jobs.json files without provider work unchanged.

Test plan

• Static-analysis tests assert the new types and field signatures
• Resolution test: provider name → provider's claudePath
• Resolution test: no provider → global config.claudePath (unchanged)
• Env merge: provider env overrides base env
• Model tier translation: model: "sonnet" → provider's modelTiers.sonnet value
• Model tier passthrough: tier not in modelTiers falls through to bare tier name
• JobScheduler passes job.provider to SessionManager.spawnSession
• npx tsc --noEmit — no new errors (9 pre-existing unchanged)
• npm run build — pre-existing 9 tsc errors unchanged (build script exits non-zero on pre-existing errors that exist on main; tsc emits dist correctly with noEmitOnError unset)

Why tier semantics over provider-specific model names

Jobs stay portable. "model": "sonnet" in a job means "the medium-tier model on whatever provider runs you" — providers translate to their own identifier. Operators retune the translation table without touching any jobs.json.

🤖 Generated with Claude Code

[vercel]

@brianjones-v4n is attempting to deploy a commit to the sagemind Team on Vercel.

A member of the Team first needs to authorize it.