to-opencode

Beta. These skills are public on purpose — try them and tell me what breaks.

Hand work to OpenCode from Claude Code or Codex: run prompts headless, inspect the live model catalog, resume sessions, and quality-gate the output.

Install

1. Install the OpenCode CLI with the official installer:

   curl -fsSL https://opencode.ai/install | bash

   Review and run this command manually; the skills never execute remote installers automatically. If you prefer a package manager, use npm install -g opencode-ai or follow the official installation guide.

2. Clone this repository and link the skills:

   Replace <repository-url> with the repository's Git URL.

   git clone <repository-url> skill-opencode
   cd skill-opencode
   ln -s "$PWD/skills/"* ~/.claude/skills/
   ln -s "$PWD/skills/"* ~/.codex/skills/

3. Optional: subscribe to OpenCode Go through this referral link, then authenticate when needed:

   opencode auth login -p opencode-go

   OpenCode currently lists Go at $5 for the first month, then $10/month; check the current terms before subscribing.

Prerequisites

• An installed opencode CLI, either on PATH or exposed through an absolute OPENCODE_BIN.
• An existing local OpenCode login for paid providers.

If the CLI is missing during a delegated run, stop and ask before installing it. Do not silently execute a remote install script.

Executable discovery

Resolve the executable in this order: an absolute executable OPENCODE_BIN, command -v opencode, then the official installer locations $OPENCODE_INSTALL_DIR/opencode, $XDG_BIN_DIR/opencode, $HOME/bin/opencode, and $HOME/.opencode/bin/opencode. Verify the resolved file with "$OPENCODE_BIN" --version.

Do not search the whole filesystem or execute an unverified file merely named opencode. If no candidate is accessible inside the current environment, report that boundary and ask the user to expose the absolute path; a host-only binary cannot be executed from an isolated container where it is not mounted.

Local runtime

Reuse the opencode binary, user home, environment, credentials, configuration, and plugins already available to the host process. A detached worktree changes the project directory only; it must not replace HOME, XDG_*, or copy credentials into the worktree.

Do not ask the user to log in preemptively. First run "$OPENCODE_BIN" models opencode-go in the inherited environment. If the requested model appears, continue immediately. Only after an authentication failure should the skill inspect "$OPENCODE_BIN" auth list and tell the user to run opencode auth login -p opencode-go. If the provider is authenticated but the model is missing, refresh the catalog once and report model unavailability instead of starting a login loop.

Model selection

The local catalog is the source of truth:

"$OPENCODE_BIN" models --refresh
"$OPENCODE_BIN" models <provider>

Use an exact provider/model ID that appears in that catalog:

• opencode-go/deepseek-v4-flash — OpenCode Go; favorite and default.
• opencode-go/deepseek-v4-pro — OpenCode Go; harder work.
• opencode/deepseek-v4-flash-free — no-cost option only when explicitly selected.

If an actual authentication failure shows that OpenCode Go has no usable local credentials, stop, tell the user to run opencode auth login -p opencode-go, and retry after login. Never fall back to a free model silently.

Free OpenCode models may retain submitted data for model improvement. Do not send confidential or proprietary material to a free model without explicit user approval.

OpenCode Go's model list and privacy terms can change. Check the live catalog and the current OpenCode Go documentation instead of relying on a cached list.

Execution safety

OpenCode must use the build agent when delegated work needs to execute commands or edit files.

• Isolated worktree (default for edits): run build in a detached Git worktree, verify its diff, then promote only authorized changes.
• Direct worktree: run in the active tree only when the user explicitly requests it.
• --fork isolates session history, not files.
• --auto auto-approves permissions; it is not a sandbox and requires explicit approval.

A detached worktree protects the active checkout from accidental edits, but it is not an OS sandbox: OpenCode can still reach network, credentials, and other host resources allowed by its process environment.

Project context

--dir selects OpenCode's project. In isolated mode it must point to the detached worktree, where project rules load normally:

AGENTS.md (project root or nearest parent) > CLAUDE.md (project)
~/.config/opencode/AGENTS.md > ~/.claude/CLAUDE.md

The instructions setting in opencode.json can load additional standards. Put ticket or architecture paths that are not covered by those rules in the prompt explicitly. If uncommitted context is needed, copy only the task-relevant files into the isolated worktree and never copy ignored secrets.

Skills

Skill	When
to-opencode	Router — selects the right fragment
opencode-invoke	Fresh headless build run
opencode-models	Live catalog and exact model IDs
opencode-resume	Continue or fork a session
opencode-qa	Ledger, claims, diff, and verification gate
opencode-visual	Translate captures for models without vision

Each skill is self-contained; the router only points to one fragment.