Agent tool prompt parameter re-injected as next user-turn input (no user typing)

Description

Claude Code main agent receives a user-turn input whose verbatim content is identical to the prompt parameter of the immediately preceding Agent (sub-agent) tool call. The user confirmed they did not type anything. This is reproducible (user reports it has happened before).

Reproduction

1. Main agent calls the Agent tool with a short structured prompt, e.g.:

``
task_id: .governance/projects/Koni/weekly/2026-W19/WO-20260511-003-koni-tencent-three-bugs-hotfix-redeploy.md
``

1. Sub-agent completes and returns its result (agentId a9475babc8fa3284f).
2. Main agent reports back to the user and waits for the next user turn.
3. Next user turn arrives — but the content is the exact verbatim string from step 1 (task_id: .governance/...same-path...).
4. User explicitly denies typing anything.

Environment

• Claude Code (CLI) main agent
• Model: claude-opus-4-7
• Custom sub-agent (koni-devops slug, defined in .claude/agents/)
• Project has no UserPromptSubmit hook configured (only SessionStart and PreToolUse Write|Edit)

Investigated and ruled out

• Project-side hooks: no UserPromptSubmit hook; SessionStart writes system context only; PreToolUse runs after user turn, cannot inject one
• Sub-agent return content: sub's return body does not contain a task_id: prefixed line (only the path appears in unrelated documentation context)
• SendMessage mis-route: sub→main reverse direction is not supported

Suspected related issues

• #51304 — \ScheduleWakeup\ re-enters slash-command prompt on completed task. Different surface but same underlying mechanism: a tool's prompt parameter being re-injected into conversation context as user input
• #30730 — sub-agent dispatch injects hardcoded post-prompt instructions, evidencing a context re-injection pipeline for \Agent\ tool prompts

Hypotheses

1. Framework-side: \Agent\ tool prompt history dedupe failure — main agent's conversation history reconstruction misclassifies the prior tool-call \prompt\ field as a standalone user message
2. Client-side: IDE / CLI history navigation key (VS Code \Ctrl+↑\, JetBrains \Cmd+↑\, terminal \↑\) accidentally surfaces the prior \Agent\ tool call's prompt as a draft user message in the input box, sent on Enter

Documentation gap

\code.claude.com/docs/en/sub-agents.md\ does not document the lifecycle of the \Agent\ tool's \prompt\ parameter on the main-agent side after the sub returns (caching, dedupe, merging behavior unspecified).

Impact

• Main agent may interpret the phantom user input as a real instruction and take action (e.g., re-dispatch a task that's already complete, mutate state, push commits)
• Defensive workaround pushed to user: every main-agent turn must check whether a structured user-input string matches a recent tool-call prompt verbatim and challenge the user before acting

Ask

• Confirm whether this is a known framework behavior or a client-side issue
• Document the \Agent\ tool \prompt\ parameter lifecycle on the main side
• If framework-side: consider stamping injected/replayed turns with a sentinel marker visible to the model