fix(opencode-plugin): persist agent/model context across injected turns

[mmkzer0]

Summary

Follow-up to #23.

Workers launched with hcom opencode --agent <name> could run turn 1 on the intended agent/model, then drift on follow-up hcom-injected turns to OpenCode's default agent/model path. This surfaced during multi-agent orchestration, where frequent hook deliveries made fallback behavior visible in real usage.

Root cause

deliverPendingToIdle() injected through promptAsync with only body.parts, omitting body.agent and body.model.

When those fields were absent, OpenCode resolved agent context through its default path (defaultAgent()), which could diverge from the launched worker agent/model on subsequent injected turns.

Fix

• Track currentAgent and currentModel in plugin state and refresh on each chat.message
• Scope the currentAgent/currentModel writes behind an isBoundSession check so foreign-session chat.message events can't cross-contaminate the bound session's injection context
• Pass agent and model on every injected promptAsync call
• Preserve fix(opencode-plugin): close TOCTOU race in deliverPendingToIdle #23 delivery hardening in this branch (promptAsync failure handling and deliveryInFlight cleanup on session reset)
• Keep the original bootstrap payload available across transforms (OpenCode transform mutations are prompt-local and not persisted to stored session history)
• Add reconcileInFlight guard to prevent overlapping interval reconciles
• Remove dead helper code and normalize --detail shell args with String(x ?? "")
• Add an explicit runtime-contract comment near the promptAsync as any cast

Known limitation: Tab-switch without submit

currentAgent/currentModel only update when the user submits a turn (chat.message fires). If the user Tab-switches the TUI agent but doesn't submit, an hcom message arriving in that window still runs on the previously-submitted agent. The plugin has no way to observe TUI-only state — the agent selection is client-side until submit. Fully closing this window requires an upstream OpenCode change (exposing the TUI's current agent selection to plugins).

CLI parse limitation (known)

parseCliArgValue and parseCliModelArg are currently null in PTY-launched OpenCode because the plugin process sees OpenCode binary argv, not outer hcom opencode --agent/--model argv.

This PR documents that behavior. Effective turn-to-turn tracking still works through the chat.message hook. Long-term fix is to seed launch agent/model from the opencode-start payload.

Validation

• cargo test -q
• cargo build -q
• Real OpenCode usage after patch: follow-up injected turns remain on the intended agent/model, with no observed default fallback drift

[Copilot]

Pull request overview

Fixes OpenCode plugin turn injection so hcom-injected follow-up turns preserve the intended agent/model context instead of drifting to OpenCode defaults.

Changes:

• Track currentAgent / currentModel in plugin state and include them in client.session.promptAsync injections.
• Add a reconcileInFlight guard to prevent overlapping reconcile ticks, and clean up/reset state on session reset.
• Improve robustness/observability: normalize --detail args, add logging, clear bootstrap text after first injection, and document the promptAsync typing gap.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

currentAgent/currentModel are global and are updated on every chat.message regardless of which session the hook event belongs to. If OpenCode emits chat.message for multiple sessions in the same process, this can cross-contaminate agent/model context and cause deliverPendingToIdle(sessionA) to inject using sessionB’s agent/model. Consider scoping these fields per-session (e.g., Map<sessionID, {agent, model}>) or at minimum only updating them when input.sessionID === sessionId (and clearing them when sessionId changes / on bind).

Suggested change

	if (input.agent) currentAgent = input.agent
	const resolvedModel = normalizePromptModel(input.model)
	if (resolvedModel) currentModel = resolvedModel
	const isBoundSession =
	!input.sessionID || !sessionId || input.sessionID === sessionId
	if (isBoundSession) {
	if (input.agent) currentAgent = input.agent
	const resolvedModel = normalizePromptModel(input.model)
	if (resolvedModel) currentModel = resolvedModel
	} else {
	log("DEBUG", "plugin.chat_message_ignored_foreign_session", instanceName, {
	session_id: input.sessionID,
	bound_session_id: sessionId,
	})
	}

[Copilot]

deliverPendingToIdle(sid) injects agent/model using the global currentAgent/currentModel, which are not keyed to sid. If more than one session can be active, this can send an injection under the wrong agent/model for the target session. Use agent/model values associated with the specific sid being delivered to (e.g., look up from a per-session map) or validate that sid === sessionId before injecting.