opencode-vise

An opencode plugin that gates an agent's changes against a frozen behavioural baseline — using vise as a tamper-proof judge.

Every other safety plugin blocks bad actions (destructive commands, secret leaks) or applies judgement (LLM review). vise judges outcomes: it freezes a project's observable behaviour — command outputs, exit codes, metrics — into a committed vise.lock, then re-runs it against the agent's diff and returns a fail-closed verdict. Same behaviour → green. Anything drifted → refused, with a one-word instruction the agent obeys.

This plugin is a thin adapter. The vise binary stays the judge; the plugin runs it when the agent finishes a turn and feeds the typed verdict back into the session.

What it does

• Gates on session.idle. When the agent finishes a turn, the plugin runs vise gate --json in the worktree and surfaces the verdict.
  • Green (exit 0): a quiet success toast.
  • Not green: the typed verdict — exit code + next.action (revert, fix_probe, fix_invocation, human, …) + detail — is posted back into the session so the agent knows exactly what to do.
• Exposes two tools the agent can call on demand: vise_gate (judge now) and vise_status (read-only readiness).
• Injects vise's contract into the system prompt, so the agent reads the exit codes and the closed next.action vocabulary natively.

Requirements

• The vise binary on PATH (go install github.com/NakliTechie/vise/cmd/vise@latest).
• A committed baseline in the project: vise init && vise record, then commit vise.lock. Without a baseline the plugin stays inert.

Install

Add it to your opencode.json:

{
  "plugin": ["opencode-vise"]
}

With options:

{
  "plugin": [["opencode-vise", { "mode": "advisory", "bin": "vise", "gateOnIdle": true }]]
}

| Option | Default | Meaning | |---|---|---| | mode | "advisory" | advisory surfaces the verdict without forcing a turn; enforce posts it as a live prompt the agent must answer. | | bin | "vise" | Path or name of the vise binary. | | gateOnIdle | true | Gate automatically when the agent goes idle. Set false to gate only via the vise_gate tool. |

How it maps to vise's exit codes

| exit | verdict | next.action | agent does | |---|---|---|---| | 0 | green | proceed | continue | | 1 | behaviour changed | revert | undo the drift, or ask an operator to re-baseline | | 2 | harness error | fix_probe / fix_invocation / human | fix the probe, the command line, or stop | | 3 | indeterminate | human | stop and report | | 5 | metric | revert / human | a frozen metric regressed |

Run vise verify for the exact expected/got diff behind a red verdict.

Notes / limits

• Per-turn gating adds latency — the probes re-run each idle. Gate on the turn boundary (session.idle), not every edit. Keep probes fast, or set gateOnIdle: false and let the agent call vise_gate at checkpoints.
• A stable red state is surfaced once per distinct (exit, next.action) per session — no idle→prompt→idle loop.
• enforce mode nudges once per distinct verdict; it is not a hard commit block. A future version may deny via the permission.ask hook.

License

MIT