[backnotprop]

Found in the 0.28.5 smoke, OpenCode 2 with the plannotator tool turned on.

Bug

A background subagent (subagent with background: true) opens a review with the tool. Its "Plannotator session ready: " notice was posted into the ROOT session, which is idle at that point. The pending steer row then got promoted on its own as a model turn the next time the root woke up. The model answered the URL line, and that exchange stayed in every later request. The same thing happens for a child session created through the API (parentID). Decisions were always delivered correctly. Foreground subagents were not affected because the root is mid-turn while it waits on them.

Cause

runPlannotatorTool hands the launch the root session (which is correct for the decision), and runNativeCommand posted the notice into that same session. With alwaysQueue, the queued decision promotes one row at a time, so the steer notice ahead of it ran as its own turn. Because it was then the last message after an assistant reply, dropSessionUrlNotices kept it on purpose.

Fix

• A tool launch now posts its notice into the CALLING session (the subagent's own session, or the main session for a direct call), and only while the tool call is still open. runPlannotatorTool closes that window the moment it answers. The caller is therefore always mid-turn when the notice arrives, so the notice is promoted inside that turn and dropped from the model's context, exactly like a main-session tool call today.
• If the server comes up after the tool already answered "starting", no notice is posted. That URL is still in list and in the log.
• createV2BridgeClient gains notice: { sessionID, open }. Co-promotion is now only attempted for a notice in the decision's own session.
• The decision still goes to the root, queued. Slash commands are unchanged: same session, same steer co-promotion (opencode v2: Plan annotations get queued after "Plannotator session ready:" message #1515).
• The AGENTS.md OpenCode 2 tool section is updated.

Tests

• plannotator-tool.test.ts, through the real launch path with a stub CLI:
  • a subagent's notice lands in its own session, and the decision goes to the root with queue
  • a main-session call still posts into its own session
  • a server that is ready only after the "starting" answer posts no notice
  • The subagent test and the late-ready test fail with the fix removed.
• native-commands.test.ts: a notice in another session never steers the decision, and a closed window posts nothing.
• bun test in apps/opencode-plugin: 342 pass. bun run typecheck passes. The touched opencode files also typecheck under strict tsc. build:review + build:hook + build:opencode succeed.

Live check (OpenCode 2.0.22, fake LLM, sandboxed HOME/XDG/data dir, local registry)

Run	Plugin	Notice row	Model turn answering the notice
child session of an idle root	0.28.5 build	root	yes (ECHO: Plannotator session ready: …)
child session of an idle root	this branch	child	none
background subagent	this branch	subagent	none; decision delivered to root
main-session tool call	this branch	main (unchanged)	none

In my harness the background-subagent run did not reproduce the leak on the 0.28.5 build. The root had already finished its turn before the notice arrived, so the notice was co-promoted with the subagent's completion. The smoke's leak needs that race. The child-session case reproduces it every time, and it is the same code path.