opencode-auto-fallback

OpenCode plugin that automatically detects model errors and switches to a fallback model chain — with intelligent retry backoff for transient failures.

Features

• Two-tier classification: immediate fallback for auth errors, exponential backoff retry for rate limits and transient failures — detected via structured SDK error types, not text matching
• Fallback chain: ordered list of fallback models per agent, with variant/reasoning/temperature support
• Per-model timed cooldown: failed models are skipped until cooldown expires
• Default‑retry safety net: any unrecognized error is treated as retryable
• Zero config startup: auto-generates fallback.json with sensible defaults on first run
• Toast notifications: terminal toasts when fallback is triggered
• Large context support: automatically switches to a larger model when context fills up

Installation

1. Register in opencode config

Add "opencode-auto-fallback" to the plugin array in ~/.config/opencode/opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["opencode-auto-fallback"],
}

2. Configuration (optional)

On first run, a default config is auto-created at ~/.config/opencode/fallback.json. You can customize it:

{
  "$schema": "https://raw.githubusercontent.com/HyeokjaeLee/opencode-auto-fallback/main/docs/fallback.schema.json",
  "enabled": true,
  "defaultFallback": ["anthropic/claude-opus-4-7"],
  "agentFallbacks": {
    "reviewer": [
      "zai-coding-plan/glm-5.1",
      {
        "model": "openai/gpt-5.5",
        "temperature": 0.5,
        "reasoningEffort": "medium",
      },
    ],
  },
  "cooldownMs": 60000,
  "maxRetries": 2,
  "logging": false,
}

Using with oh-my-openagent

When an agent is provided by oh-my-openagent, the runtime agent name can differ from the key in oh-my-openagent.json. Configure fallback with the name that appears in OpenCode session logs, normalized by removing whitespace and ignoring case.

For example, if logs show agent=Sisyphus - Ultraworker, configure it as Sisyphus - Ultraworker (whitespace is stripped, case is ignored):

{
  "agentFallbacks": {
    "Sisyphus - Ultraworker": [
      {
        "model": "opencode-go/deepseek-v4-pro",
        "variant": "high",
      },
    ],
    "hephaestus - deepagent": ["zai-coding-plan/glm-5.1"],
  },
  "largeContextFallback": {
    "agents": ["Sisyphus - Ultraworker", "hephaestus - deepagent"],
    "model": "opencode-go/deepseek-v4-pro",
  },
}

Agent matching is exact after normalization: Sisyphus - Ultraworker and sisyphus - ultraworker match the same entry, but sisyphus does not automatically match Sisyphus - Ultraworker.

Auto Updates

The plugin checks for updates on every startup and installs them automatically — no manual intervention needed.

opencode starts → check npm registry → newer version? → bun/npm update → done

If the auto-update fails for any reason, a toast notification appears with the manual update command.

Large Context Fallback

When an agent's context window fills up mid-task, automatically switch to a larger model to finish the work without interruption. After the task completes and the large model compacts, switch back to the original model with the compacted context.

{
  "largeContextFallback": {
    "agents": ["sisyphus", "explore"],
    "model": "openai/gpt-5.5",
    // Optional: define model context window sizes to guard against
    // switching to a model with similar context capacity
    "contextWindows": {
      "openai/gpt-4o": 128000,
      "openai/gpt-5.5": 2097152,
    },
    // Optional: minimum ratio difference required to trigger fallback
    "minContextRatio": 0.1,
  },
}

When contextWindows is configured, the plugin compares the current model's context window with the large model's. If the large model's window is within (1 + minContextRatio) × current (i.e., not sufficiently larger), the fallback is skipped and normal compaction proceeds instead.

Flow:

original model working → context full (auto compact) → switch to large model
    → large model finishes task → idle → large model compacts
    → switch back to original model (compacted context)

Note: Manual /compact commands do not trigger large context fallback — only automatic compaction (when context fills up from an assistant response) activates it.

Fallback Model Entry

Each entry in a fallback chain can be a simple string or an object:

// Simple
"openai/gpt-5.5"

// With options
{
  "model": "openai/gpt-5.5",
  "variant": "high",
  "temperature": 0.5,
  "reasoningEffort": "medium",
  "maxTokens": 8192
}

How It Works

Error Classification

The plugin detects errors through session.error events (structured statusCode and isRetryable flags) and session.status events (message-based pattern matching for rate limits and transient errors).

Retry Flow

With the default maxRetries: 2:

1st failure → abort → wait 2s   → re-prompt with SAME model
2nd failure → abort → wait 4s   → re-prompt with SAME model
3rd failure → FALLBACK CHAIN: try next model in ordered list

Immediate fallback errors (quota, auth) skip retries entirely and go straight to the fallback chain.

Fallback Chain

The plugin tries each model in the chain sequentially. Models in cooldown are automatically skipped. If all models are exhausted, the error is logged and a critical toast is shown.

Compatibility with Other Fallback Plugins

If another plugin with model fallback logic is installed alongside this one, place opencode-auto-fallback first in the plugin array. The first plugin in the list processes the model response first — by placing this plugin first, it intercepts the error before other fallback plugins see it.

// ✅ opencode-auto-fallback handles errors first
"plugin": ["opencode-auto-fallback", "other-fallback-plugin"]

// ❌ Other plugin may interfere
"plugin": ["other-fallback-plugin", "opencode-auto-fallback"]

Development

# Install dependencies
bun install

# Type check
tsc --noEmit

# Run tests (73 tests)
bun vitest run

# Bump version (CI auto-publishes)
npm version patch --no-git-tag-version
git push