[rekram1-node]

Problem

When a background subagent is running and the user adds an instruction ("also target the release branch"), agents send it with session_prompt instead of the subagent tool. session_prompt delivers the text, but the parent never gets the subagent's answer back.

The subagent tool already handles this: passing a running child's sessionID steers it mid-task and the parent still gets one completion notification. Nothing in the tool's text said so. The sessionID field only talked about continuing a "previous" conversation, and the background result only said what not to do.

Change

Two lines in packages/core/src/tool/plugin/subagent.ts. The tool description stays as is, so behavior for finished subagents doesn't change.

 background result
   "Work on non-overlapping tasks, or briefly tell the user what you launched and end your response.",
+  "To send it follow-ups or change its instructions while it runs, call subagent again with this sessionID.",

 sessionID
-  "Continue a specific previous subagent conversation by passing its sessionID. Calls without a sessionID start a new conversation.",
+  "Send a follow-up to an existing subagent, running or finished. Omit to start a new one.",

Also adds a test: three follow-ups to a running background child (with and without background: true) all return right away, and the parent gets exactly one completion notification.

Eval

I replayed fixed conversations against each model's API, 10 to 20 runs per cell. Failures when asked to redirect a running background subagent (two phrasings):

model	before	after
Opus 5.5	13/20	0/20
Gemini 4 Argon	7/20	0/20
Sonnet 5.5	5/20	0/20
Haiku 5.5	12/20	0/20
Sonnet 4.5	20/20	0/20
DeepSeek	20/20	0/20
MiniMax M2.5	20/20	0/40
MiniMax M2	20/20	2/40
Haiku 4.5	20/20	13/40

Checked for over-reuse too: when the user asks for something unrelated, every model still starts a fresh subagent, both while the old one runs and after it finishes. Related follow-ups to a finished explore agent are reused as often as before or more.

Out of scope

• No change to session_prompt.
• Haiku 4.5 still picks session_prompt about a third of the time. The wordings that fixed it fully either forbade other tools or made other models worse.

Verification

• bun test test/tool-subagent.test.ts in packages/core: 12 pass
• bun typecheck in packages/core: clean

Visual

<svg xmlns="http://www.w3.org/2000/svg" width="240" height="72" viewBox="0 0 240 72" fill="none" stroke="currentColor" stroke-width="1.5" font-family="system-ui, sans-serif" font-size="10">
  <style>
    .msg { animation: send 2.4s ease-in-out infinite; }
    .reply { animation: back 2.4s ease-in-out infinite; }
    @keyframes send { 0%, 10% { transform: translateX(0); opacity: 0 } 20% { opacity: 1 } 45% { transform: translateX(104px); opacity: 1 } 55%, 100% { transform: translateX(104px); opacity: 0 } }
    @keyframes back { 0%, 55% { transform: translateX(0); opacity: 0 } 65% { opacity: 1 } 90% { transform: translateX(-104px); opacity: 1 } 100% { transform: translateX(-104px); opacity: 0 } }
    @media (prefers-reduced-motion: reduce) { .msg, .reply { animation: none; opacity: 1; } .msg { transform: translateX(52px); } .reply { transform: translateX(-52px); } }
  </style>
  <rect x="8" y="20" width="64" height="32" rx="6" opacity="0.8"/>
  <text x="40" y="40" text-anchor="middle" fill="currentColor" stroke="none">parent</text>
  <rect x="168" y="20" width="64" height="32" rx="6" opacity="0.8"/>
  <text x="200" y="40" text-anchor="middle" fill="currentColor" stroke="none">subagent</text>
  <circle cx="190" cy="14" r="3" fill="currentColor" stroke="none" opacity="0.5"/>
  <line x1="76" y1="30" x2="164" y2="30" opacity="0.25"/>
  <line x1="76" y1="44" x2="164" y2="44" opacity="0.25"/>
  <rect class="msg" x="60" y="26" width="16" height="8" rx="2" fill="currentColor" stroke="none" opacity="0.7"/>
  <rect class="reply" x="164" y="40" width="16" height="8" rx="2" fill="currentColor" stroke="none" opacity="0.4"/>
</svg>

---

Added by the opencode bot (AI-generated).

Related: #38966

[opencodejr]

The change looks sound. The new text matches what the tool already does: a follow-up with sessionID goes through the normal prompt path, which redirects a running child, and the call returns straight away with status running. I couldn't test the effect on model behavior because it needs real models; the simulated model can't show tool choice. tool-subagent.test.ts passes (12/12) and typecheck:tsc is clean. The typecheck script fails with "Script not found check" on dev as well, so that failure isn't from this PR.

Public API: unchanged. The tool's input fields are the same; only the wording the model sees changes.

The new test can be dropped (see the inline comment).

Related: #38966

[opencodejr]

I ran this test on dev without the prompt change and it passes there too. It checks behavior that already exists and doesn't touch the new text, so for a two-line wording change it can be dropped to keep the diff minimal.

[rekram1-node]

Keeping it. It passes on dev on purpose: the new text tells models a follow-up to a running background subagent returns right away and still notifies once, so this pins that behavior. If it changes, the prompt would be lying.