opencode.md

OpenCode integration

Version 0.3.2 bundles 19 engineering/shared skills. Writing Method bundles its own 18 writing/shared skills and a distinct writing-method.js loader. Both use the same adapter from Method Core. Earlier release evidence under docs/verification/ describes the combined historical packages and is not fresh discovery evidence for this split. A local plugin config hook adds the skill directory and OpenCode adapter to effective configuration. It leaves existing model, permissions, provider and MCP settings untouched.

Install from a retained checkout

Requires Python 3.11+ and OpenCode; native validation used OpenCode 1.18.30 on macOS. Use version 0.1.1 or newer; the earlier v0.1.0 release predates this integration. Clone a retained checkout and install from it:

git clone --branch v0.3.2 --depth 1 https://github.com/phense/engineering-method.git
cd engineering-method

python3 scripts/install-opencode --check
python3 scripts/install-opencode
opencode debug skill > /private/tmp/engineering-method-skills.json

The installer creates only ~/.config/opencode/plugins/engineering-method.js and missing parent directories. It respects XDG_CONFIG_HOME; pass --config-dir /absolute/config/directory for a custom OpenCode configuration. If you run OpenCode with OPENCODE_CONFIG_DIR, pass that same directory explicitly to the installer. Symlinked destination parents and foreign or modified loaders are refused. On macOS use canonical /private/tmp, not its /tmp symlink, for disposable installations. --check describes an action and writes nothing; a successful check does not mean installation already exists.

The loader imports the retained source with a file URL. Keep the checkout or extracted package at that path. Updates become effective after the OpenCode helper restarts. Moving a source requires uninstalling its exact loader with the old --source-root and installing again with the new path. Do not overwrite an unrecognized loader. Relative resources continue to resolve within the original skill tree; scripts operate on the user's target project.

Use from T3 Code

Install on the machine running OpenCode (the Mac in a remote T3 environment). Finish active work, refresh provider status and start a new thread. A cached helper may need to be idle for at least 30 seconds or restarted before the new plugin is loaded. T3 and standalone CLI sessions can expose different tool and permission sets; verify the new thread can load a named skill.

For a small check, ask the model to load verification-before-completion with the skill tool, execute a harmless verification command and report the observed result. Skill discovery alone is not evidence of actual invocation. A conflicting same-name skill may win discovery: inspect its returned source location. OpenCode 1.18.30 can truncate large piped JSON on exit, so write debug skill output directly to a file before parsing it.

Remove

From the original source, or with its original absolute --source-root:

python3 scripts/install-opencode --uninstall --check
python3 scripts/install-opencode --uninstall

Only the byte-for-byte matching generated loader is removed. Other plugins, configuration and project artifacts remain. Restart the helper or start a fresh host session to remove the effective instructions.

Scope of verification

See EM-008 evidence. Native discovery and a bounded DeepSeek skill invocation cover this adapter, not every lifecycle or multi-agent workflow. The shared Claude/Codex skills were not modified.

Model choice remains the user's selection. The integration installs no automatic Flash/Pro quality router or LiteLLM. Existing read-only agentic-rag MCP access is separate; session recall, transcript ingestion and compaction hooks are not implemented here. Do not infer those capabilities from skill installation.

API references

Checked 2026-09-11 against the official skills, plugins and configuration documentation, plus the v1.18.30 plugin hook interface.