opencode-for-claude-code

A Claude Code plugin that lets Claude delegate well-scoped work to OpenCode Go models via the headless opencode run CLI. Claude acts as the conductor (handling planning, the hard 20%, and verification), while OpenCode Go models serve as the cheap executors for high-volume tasks. It is modeled directly on the cost-aware architecture of antigravity-for-claude-code.

What it is

This plugin enables hybrid agentic engineering. Keep judgement-heavy work on Claude (the frontier model) and route deterministic, bulk work (scaffolding, well-specified implementation, test generation, migrations, first-pass review) to OpenCode Go. It provides the oc-delegate bash wrapper to handle routing, timeouts, and structured error signals.

Requirements

• Claude Code: The conductor.
• opencode CLI: The executor tool.
• OpenCode Go Subscription: A $10/mo subscription providing curated open-source models with dollar-based limits (~$12/5h, ~$30/wk, ~$60/mo).

Install

# Add the marketplace
claude plugin marketplace add katsu1110/opencode-for-claude-code

# Install the plugin
claude plugin install opencode

Quickstart Examples

In Claude Code, you can trigger delegations directly or use slash commands:

• /opencode:setup - Run the doctor script to check your installation and models.
• /opencode:delegate Write unit tests for src/auth.py - Delegate a task.
• /opencode:review - Run a cross-model first-pass code review of your uncommitted changes.

Or converse with Claude: "Migrate the legacy endpoints in src/api/ to the new format. Delegate this to opencode to save tokens, then run the tests to verify."

Tier Routing

Tier	Default Model	Best For	Quota Note
flash	opencode-go/deepseek-v4-flash	Bulk reads, test generation, digests	Cheap (~158k req/mo)
code	opencode-go/kimi-k2.7-code	Implementation against specs	Standard
pro	opencode-go/glm-5.2	Hard reasoning, review, retries	Scarce (~4.3k req/mo)

Tiers can be remapped via plugin options.

oc-delegate Options

Usage:
  oc-delegate.sh [options] "the task prompt"
  echo "long prompt" | oc-delegate.sh [options] -      # read prompt from stdin

Options:
  -t, --tier <flash|code|pro>   Model tier (default: flash)
  -d, --dir <path>              Directory to run in (single dir — opencode limit)
      --timeout <dur>           Wall-clock timeout, e.g. 10m, 300s (default: 10m)
      --write                   Allow file writes / commands (DANGEROUS; run on a branch)
      --mode accept-edits|plan    Execution mode (accept-edits maps to --auto; plan touches nothing)
      --digest                  Append a digest-only output contract to the prompt
  -c, --continue                Resume the most recent opencode session
  -s, --session <id>            Resume a specific opencode session by id
  -m, --model <provider/model>  Exact model (any from `opencode models`)
      --variant <v>             Model variant / reasoning effort
      --print-command           Print the resolved opencode command and exit
  -h, --help                    Show this help

Exit Codes & OC_SIGNAL

• 0 ok | 1 usage | 2 run failed | 3 empty
• 10 quota | 11 auth | 12 timeout | 13 opencode missing | 14 model unavailable
• On classifiable failures, a machine-readable JSON line is printed to stderr: OC_SIGNAL {"status":"...","reason":"..."}

Cost Discipline Summary

• Break-even: Only delegate above the break-even point. Do not delegate tiny edits.
• Digest-not-dump: Use --digest for bulk reads to prevent large dumps from inflating Claude's context.
• Batch: Consolidate tasks.
• Lean Context: Never paste raw code back into the thread.

Safety

• --write automatically approves commands and file edits. Always run write tasks on a dedicated branch.
• Verify files actually changed.
• Never trust the executor's self-reported green status.

Configuration

Available in .claude-plugin/plugin.json:

• default_tier: Default tier (flash).
• timeout: Default timeout (10m).
• coding_policy: Inject the cost-aware routing policy at SessionStart (on/off).
• delegation_nudge: Add a delegation nudge on bulk-looking prompts (on/off).
• default_model: Exact model to override tier mapping.
• tier_flash, tier_code, tier_pro: Custom mappings for the three tiers.
• digest_warn_chars: Warning threshold for large outputs (default 8000).

FAQ

• Why a single directory? opencode run is limited to a single --dir argument. Always pass the repo root.
• Why did I hit a quota limit quickly? Quotas are shared with interactive TUI usage and are dollar-based. A runaway pro tier task can burn through the 5h window quickly. Use flash when possible.
• How do I remap models? Use the CLAUDE_PLUGIN_OPTION_TIER_* options in your plugin config to update models as OpenCode's lineup changes.