Transport Configuration

Transport config lives at ~/.abtars/config/transport.json. It defines which AI providers are available and which models each agent uses.

Terminology

Term	Definition
Provider	A named service entry — identifies how to reach a model
Transport	Communication method: acp (CLI via JSON-RPC), tmux (CLI via pane scraping), api (HTTP endpoint)

Example

{
  "agents": {
    "professor": {
      "model": "claude-sonnet-4.6",
      "provider": "kiro",
      "fallbacks": [
        { "model": "kimi-k2.5:cloud", "provider": "ollama" },
        { "model": "gemini-2.5-flash", "provider": "openrouter" }
      ]
    },
    "dreamy": { "model": "minimax-m2.5:cloud", "provider": "ollama" },
    "browsie": { "model": "minimax-m2.5:cloud", "provider": "ollama" },
    "coding": { "model": "qwen3.5:cloud", "provider": "ollama" }
  },
  "providers": {
    "kiro": { "transport": "acp", "cli": "kiro-cli" },
    "gemini": { "transport": "acp", "cli": "gemini" },
    "ollama": { "transport": "api", "endpoint": "http://localhost:11434/v1" },
    "openrouter": { "transport": "api", "endpoint": "https://openrouter.ai/api/v1", "apiKeyEnv": "OPENROUTER_API_KEY" }
  },
  "maxTurns": 50,
  "healthPolicy": {
    "skipThreshold": 0.7,
    "leakPerMinute": 0.03,
    "rateLimitFill": 0.5,
    "transientProgressive": [0.1, 0.2, 0.4, 0.8],
    "transientCooldownAfter": 3,
    "transientMaxCooldownSec": 300
  }
}

Agents

Agent	Role
professor	Main conversation agent
dreamy	Sleep cycle / dream processing
browsie	Web browsing tasks
coding	Coding mode (switched via /coding)

Each agent has a model and provider. Co-agents are independent — they can use different providers than professor.

Fallback Chain

Professor supports an ordered fallbacks array. When the primary model fails:

1. Try fallback 1, 2, 3, ... N (in order)
2. Skip models with full health buckets (rate-limited, auth failed)
3. If all exhausted → error (use /model emergency for paid hailMary)

Fallbacks must use the same transport type and CLI binary as the primary. API fallbacks can span providers freely.

Add as many fallbacks as you want — just edit the array in transport.json.

Health Policy

The bridge tracks model health using a leaky-bucket algorithm. Errors fill the bucket, time drains it. Full bucket = model skipped.

All values in healthPolicy are optional (defaults shown above):

Field	Default	Description
skipThreshold	0.7	Bucket level (0-1) above which model is skipped
leakPerMinute	0.03	How fast a penalized model recovers
rateLimitFill	0.5	Penalty per rate-limit hit (429)
weakFill	0.05	Mild penalty for weak responses
transientProgressive	[0.1, 0.2, 0.4, 0.8]	Escalating penalty per consecutive error
transientCooldownAfter	3	Errors before hard cooldown
transientMaxCooldownSec	300	Maximum cooldown duration

Adding a Provider

Add an entry to providers and restart:

"my-provider": {
  "transport": "api",
  "endpoint": "https://api.example.com/v1",
  "apiKeyEnv": "MY_PROVIDER_KEY"
}

Then assign an agent to it or use /model change to switch interactively.

Model Switching

Action	What happens
/model restore	Undo last switch (swaps transport.json with .old backup)
/model default	Factory reset from transport.default.json
Model change, same transport	New session (conversation reset)
Provider change (different transport)	Full reset (transport rebuilt)