Intent Gate

A TypeSafe Jev powered plugin for OpenCode that makes the agent confirm intent before diving into underspecified requests.

Agents love to silently start digging. When a request is vague ("just fix it somehow"), they pick an interpretation and run with it — investigating, editing, sometimes wrecking things. opencode-intent-gate inserts a judgment step before the turn starts: if the request looks underspecified, it injects a system directive telling the agent to ask 1-3 clarifying questions first instead of calling tools.

How it works

user message
  → session hook "context" (right before model dispatch)
      → judge the latest user message with Jev (4 noul questions, 1 request)
          is_work_request / ambiguous / missing_user_info / scope_unclear
      → plain-code thresholds decide: gate or pass
  → if gated: push a directive into the system prompt
      "ask before acting; do not start tool calls this turn"

The decision is made from calibrated probabilities, not vibes: Jev returns noul scores and the thresholds are ordinary code constants you can tune. Verdicts are cached per message, so each user message costs one request of roughly 500 tokens (~0.5s, ~$0.00002).

Measured behavior (Japanese requests)

message	verdict
いい感じに直しといて	gate (ambiguous 0.81 / scope 0.97)
app/main.go のビルドエラーを直して	pass
これってどういう意味？	pass (not a work request)
予想ロジックを改善して。あとで見るから、いい感じに。	gate
tmp/... を読んで要約して	pass
掃除は今して	pass (specific follow-up)

Requirements

• OpenCode V2 beta (tested on opencode2 0.0.0-beta-19271)
• A TypeSafe API key: sign in at https://console.typesafe.ai and issue one (a trial credit is included)

Beta note: this plugin judges from the context hook (which runs before every model dispatch) instead of the prompt hook, so it works across OpenCode V2 betas. Tested on beta-19271.

Compaction requests and other runtime-generated messages are never gated. The hook returns before judging when the agent is compaction, when the latest user message is a compaction prompt (You MUST summarize the conversation above... or Update the existing checkpoint...), when it is an id-less synthetic prompt about summarizing, or when it is a runtime notice (interruption/server-restart resume, max-steps, user-executed tool reports, plan-mode reminders, subagent reports). Skips are recorded in the log as event: "skip" with a reason.

Install

opencode plugin add opencode-intent-gate

Or run from a local checkout:

git clone https://github.com/hoshinodis/opencode-intent-gate ~/app/opencode-intent-gate
ln -s ~/app/opencode-intent-gate ~/.config/opencode/plugins/opencode-intent-gate

Provide the API key to the process that runs the OpenCode service:

export TYPESAFE_API_KEY=...

or drop the key into ~/.config/opencode/typesafe/api_key.

Options

Options can be passed where the plugin is registered; defaults shown below.

option	default	meaning
enabled	true	also disabled with TYPESAFE_INTENT_GATE=off
model	jev-latest	pin e.g. jev-1.13.0 for reproducibility
isWorkThreshold	0.5	minimum is_work_request score to consider gating
dimensionThreshold	0.75	any ambiguity dimension at or above this gates
timeoutMs	2500	request timeout; on timeout the gate is skipped (fail-open)
minChars	2	skip messages shorter than this
maxChars	4000	clip the text sent to Jev to this many characters (head + tail), so attached-file expansions cannot blow past the request limit
apiKeyEnv / apiKeyFile	TYPESAFE_API_KEY / ~/.config/opencode/typesafe/api_key	key lookup order
logFile	~/.config/opencode/intent-gate/decisions.jsonl	JSONL decision log

Safety

• fail-open: API errors and timeouts never block the prompt; after 3 consecutive failures the gate pauses for 5 minutes
• acknowledgements (ok, はい, thanks, ...) and slash commands are skipped
• the gate is a system directive, not a hard block. Pair it with tool permissions if you want enforcement.

Development

npm install
npm run typecheck

License

MIT. Not affiliated with TypeSafe or OpenCode.

---

日本語

曖昧な依頼でコーディングエージェントが勝手に走り出すのを防ぐ、TypeSafe Jev 製の OpenCode プラグイン。モデル送信直前の context フックで最新のユーザーメッセージを4つの noul 質問（作業依頼か / 曖昧か / ユーザーしか知らない情報が必要か / スコープ不明か）で判定し、「確認すべき」と判定されたときだけ「ツールを使う前に質問しろ」という指示を system プロンプトに注入します。判定は確率スコアとコード側の閾値で決まり、失敗時は素通し（fail-open）。