Table of Contents generated with DocToc
Capability: capability:platform
Harness: OpenCode
OpenCode uses Magpie's skill sources, deterministic action guard, spec-loop runner, and clean-environment wrapper. Nothing about a Magpie skill or a guard rule is harness-aware: the same files that serve Claude Code serve OpenCode, through a different registration mechanism.
| Magpie requirement | OpenCode implementation |
|---|---|
| Skill discovery | OpenCode is part of the cluster that converged on .agents/skills/, which is the canonical target the universal row in agents.md already writes. It also reads native Agent Skills from .opencode/skills/ (project) or ~/.opencode/skills/ (personal). |
| Deterministic guard | A plugin on tool.execute.before forwards each bash command to agent-guard.py --opencode and throws the deny reason, which aborts the tool call. |
| Spec-loop | The opencode profile passes the prompt positionally to opencode run, with --auto, --model, and --format json. |
| Credential isolation | opencode-iso launches the CLI through the clean-environment wrapper. |
OpenCode aborts a tool call whose tool.execute.before handler throws, which is the OpenCode equivalent of a Claude Code PreToolUse deny. The plugin in tools/agent-guard/opencode/plugin.js forwards each bash command to the shared dispatch() core and throws with the reason the core returns.
Register it by dropping the plugin into OpenCode's plugin directory:
mkdir -p .opencode/plugin ln -s "<framework>/tools/agent-guard/opencode/plugin.js" .opencode/plugin/agent-guard.js
The plugin locates the engine at .claude/hooks/agent-guard.py under the worktree by default, so a repository already wired for Claude Code needs no second copy of the script, and honours MAGPIE_AGENT_GUARD=/abs/path/agent-guard.py to point elsewhere.
Because both harnesses call the same dispatch(), the bundled and skill-contributed guards, the MAGPIE_* overrides, and the deny reasons are byte-for-byte identical across the two. Full wiring notes are in the agent-guard OpenCode recipe.
The opencode profile in tools/spec-loop/lib.sh runs opencode run --auto --model … "<prompt>" from the repository root and asks for stream JSON with --format json. Setting SPEC_LOOP_AGENT=opencode is normally the whole configuration:
SPEC_LOOP_AGENT=opencode SPEC_LOOP_MODEL=anthropic/claude-sonnet-4-5 \ tools/spec-loop/loop.sh <spec>
OpenCode is the one harness whose reasoning-effort band is not an identity mapping: the runner translates the neutral low / medium / high to --variant minimal / medium / max. See the spec-loop guide.
source /path/to/magpie/tools/agent-isolation/agent-iso.sh opencode-iso
opencode-iso applies the same env -i credential strip as claude-iso, and adds no settings grant: OpenCode takes its filesystem isolation from the OS-level sandbox rather than from a per-launch allowlist. Network restrictions and tool approval are separate controls, described in the secure-agent setup.
uv run --directory tools/agent-guard --project . python -m pytest bash tools/spec-loop/tests/test_runner_fixtures.sh
The guard tests cover the OpenCode path end-to-end — the plugin's command routing, the throw-on-deny contract, and that a decision made by the shared core is preserved unchanged. The runner fixtures verify the positional prompt, the --auto flag, and the effort-band translation, including the low → minimal mapping that is unique to this harness.
To verify in a live session, start OpenCode from the checkout root and ask it to run git status --short; it should pass the guard. For a harmless denial probe, ask it to run git commit --dry-run --no-verify -m hook-probe — the expected result is an agent-guard[no-verify] denial before Git starts. The --dry-run flag prevents the probe from creating a commit even if the plugin is not registered.