feat(opencode): allow skipping the agent turn after a command - #46579

Issue for this PR

Closes #28292 (the noReply half; custom dialogs are a separate surface)

Type of change

• New feature

What does this PR do?

Slash commands always run an agent LLM turn after execution, even when the command needs no LLM involvement. prompt() already short-circuits on noReply (returns the user message without running the loop), but the flag was unreachable from the command path: CommandInput had no noReply field, and the command.execute.before hook could only mutate parts.

This PR plumbs noReply through the command path:

• CommandInput gains an optional noReply field, so server API callers (POST /session/:id/command) can skip the turn
• The command.execute.before hook output gains a noReply flag (default false) that plugins can set, forwarded to prompt() alongside the client-provided value
• Semantics are OR: either party can suppress the turn; neither can force a turn the other explicitly declined

Real-world usage (from our own plugin and others in #28292):

• Config-management commands. A plugin slash command that lists/sets/deletes entries in a JSON file, validated against the provider list — pure text I/O, no LLM needed. Every invocation cost a full LLM turn. Worse: run right after an edit turn, the model received an effectively empty user message plus stale context and re-emitted the previous turn's tool calls — redoing edits the user had already made. The plugin-side workaround (splicing a synthetic anchor part instructing the model to reply with a one-line confirmation) stops the redo hazard but still burns an LLM call per command; only server-side noReply fixes the cost.
• Auto-review toggle (opencode-review, @Svtter in the thread): "toggle auto-review without LLM round-trip".
• Queue commands (@mirsella in the thread): a queue plugin handling /queue commands entirely locally.
• The thread also documents why the ignored: true parts workaround fails: the turn still fires, still costs tokens, and suppression is probabilistic.

How did you verify your code works?

• bun run typecheck clean in packages/opencode, packages/plugin, packages/sdk/js
• bun test test/session/prompt.test.ts: 61 pass, 0 fail, including three new tests:
  • client noReply: true → user message returned, zero LLM calls
  • plugin sets output.noReply = true in command.execute.before → same, plus plugin-pushed parts land in the persisted message
  • plugin sets output.noReply = false while client sent noReply: true → turn still skipped (a hook cannot override the client's suppression)
• SDK regen produces only the expected noReply additions, no unrelated drift

Screenshots / recordings

Not a UI change.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

Critical changes: the command.execute.before hook output type changes from { parts } to { parts, noReply } — additive; existing plugins that only touch parts are unaffected (structural typing; runtime always initializes noReply: false). Command.Event.Executed's messageID is the user message when noReply is set (vs the assistant message otherwise); noted in the SDK docs row. The only in-repo consumer reads only data.name and is unaffected.