Proxies and custom providers

None of the statusline’s numbers come from Anthropic or from an OpenCode plan. They are computed locally, which means a proxy changes what is available — and the bay is built to be honest about it.

Tokens always work

They come from the provider’s response. LiteLLM and anything else OpenAI-compatible returns usage, so tokens, the cache split and the composition bar populate with no configuration at all.

Cost and the context window need declaring

OpenCode computes cost as tokens × per-model price, and the context percentage needs a window size. For a custom provider both come from your own config:

~/.config/opencode/opencode.json

{
  "provider": {
    "litellm": {
      "npm": "@ai-sdk/openai-compatible",
      "options": { "baseURL": "https://llm.corp/v1" },
      "models": {
        "claude-opus-5": {
          "cost": { "input": 5, "output": 25, "cache_read": 0.5 },
          "limit": { "context": 200000, "output": 64000 }
        }
      }
    }
  }
}

Declare them and cost and context work normally. Leave them out and both stay silent rather than reporting $0.00 and 0%.

That silence is the point. An unpriced model is not a free one, and a percentage with no denominator is not zero. A confident wrong number in a statusline is worse than an absent segment, because you stop checking a number you have learned to trust.

Your proxy knows better than we do

LiteLLM reports the real cost of a response and tracks team and key budgets. A locally multiplied estimate cannot match that. Read the real figure with a command segment:

{
  "statusline": {
    "commands": {
      "budget": { "run": "curl -sf $LLM_PROXY/spend | jq -r .remaining", "intervalMs": 30000 }
    },
    "segments": [{ "type": "command", "name": "budget", "prefix": "budget " }]
  }
}

A failing request leaves the last good value on screen, and the call is off the draw path, so a slow proxy never makes the interface stutter.

Subagents

The cost segment reports the session’s own running total — the same number OpenCode prints in its sidebar. That is deliberate: two figures on one screen that disagree are worse than either, and an earlier version summed the messages it could see and read $0.30 against the host’s $0.56.

Whether a subagent’s spend is part of that total is OpenCode’s behaviour, not this bay’s. We show what the host shows, so the two always agree — and if the host starts counting child sessions, so will we, with no change here.

If you delegate heavily and want to be sure, the check takes a minute: run a turn that spawns a subagent, then compare the cost segment against the sum you would expect. If it is short, the host is not rolling child sessions up, and a command segment against your proxy’s own spend endpoint is the number to trust — it counts every request whoever made it.

What is never available

Anthropic plan quotas — the 5h and 7d windows a Claude Code statusline can show. They describe a subscription, not a model, and there is nothing to read them from behind a gateway. If your proxy exposes an equivalent, a command segment is the way to show it.