Agent Open Questions Protocol

The single source of truth for how a subagent surfaces multi-choice questions to the user. Subagents do not call AskUserQuestion themselves — the tool's user-visibility from inside a subagent is undefined. Instead, a subagent emits a structured envelope in its final assistant message, the orchestrator parses it, calls AskUserQuestion, and resumes the subagent via SendMessage(to: <agentId>, message: ...) with the user's selections.

Rule: do not call AskUserQuestion from inside a subagent

A subagent MUST NOT call AskUserQuestion. Even if the tool is wired in the agent's tools: frontmatter, the subagent has no reliable way to surface the prompt to the user. Emitting prose ("I would ask the user but the tool is unavailable...") is also forbidden — it replaces the contract with a rationalization. Use the envelope below instead.

The envelope shape

The subagent's final assistant message ends with a fenced JSON block whose top-level object carries an envelope array. Each item mirrors the shape AskUserQuestion itself accepts: question, header, and options (each option carries label and description).

Example envelope:

{
  "openQuestions": [
    {
      "question": "How should rate limiting be enforced for unauthenticated requests?",
      "header": "Rate limit",
      "options": [
        { "label": "Token bucket per IP (Recommended)", "description": "Simple, no shared state. Trade-off: NAT'd users share a bucket." },
        { "label": "Sliding window per IP",             "description": "More accurate burst handling. Trade-off: needs Redis." },
        { "label": "No limit on unauthenticated path",  "description": "Smallest change. Trade-off: leaves DoS surface open." }
      ]
    }
  ]
}

After emitting the envelope, STOP. Do not write any artifact on the same turn — wait for the orchestrator to resume you via SendMessage with the user's selections.

Resume contract

After the user answers, the orchestrator resumes the same subagent via the Task tool's SendMessage(to: <agentId>, message: ...) mechanism. The message is the user's selections verbatim (plus any free-text follow-up answer the orchestrator collected — see "Free-text escape hatch" below). The subagent receives the selections as a new user turn with its full prior transcript intact; it is not a fresh dispatch.

Size caps

• ≤ 4 questions per envelope. AskUserQuestion accepts 1–4

questions per call. If you have more than 4 sharp questions, split or defer the lower-priority ones to a ## Open questions (deferred) section in your artifact.

• `description` ≤ ~200 characters. Each option's description

names the approach and its trade-off in roughly one sentence.

Zero-question case

If the subagent has nothing to ask, omit the envelope entirely and proceed to write its artifact. Do not emit an empty envelope with an empty array — the orchestrator treats the absence of an envelope as "nothing to render."

Free-text escape hatch

AskUserQuestion returns only the chosen label — there is no free-text field. When a subagent genuinely needs free-text input (for example: collecting absolute repo paths after a Single/Multi-repo choice), state in the chosen option's description that the orchestrator will follow up with a plain-text question after the user selects that option. The orchestrator asks the plain-text follow-up, collects the answer, and concatenates it into the resume SendMessage. The canonical worked example is the questioner's multi-repo flow.

Revision-dispatch exception

On a revision dispatch (the human gate rejected an earlier draft and the subagent is re-run with the user's verbatim feedback), the envelope is optional. Emit it only if the feedback raises new ambiguities. If everything the feedback asks for is unambiguous, write the revised artifact directly.

Malformed envelope path (two attempts)

If the orchestrator's parse step fails (malformed JSON, missing label on an option, etc.):

1. Attempt 1: the orchestrator SendMessages the subagent with the

exact parse error verbatim and asks for a corrected envelope.

1. Attempt 2 fails: the orchestrator writes the raw tool result to

docs/plans/<id>/dispatch-failure.md with frontmatter phase: <current-phase>, status: parse-failed, marks the phase halted in TodoWrite, and surfaces a clear message to the user naming the artifact path.

Missing description on an option is filled with the empty string and rendered. Missing label is a parse failure (the user has nothing to click).

Envelope + summary collision

A subagent that emits both an envelope and its summary JSON on the same turn is harmless: Decision 5 (first-block-wins) selects the envelope; the summary is ignored on that turn. The subagent re-emits the summary after the resume, when the artifact is written.

Parallel-dispatch limitation

The only parallel dispatch in the pipeline today is RESEARCH (file-finder + researcher), neither of which uses this protocol. If a future phase dispatches two agents in parallel and both emit envelopes, the orchestrator does not have a defined ordering rule yet. Avoid emitting envelopes from parallel agents until the protocol is extended.

Nested sub-agents: never emit envelopes below depth 2

This protocol works exactly ONE level deep. The orchestrator parses the final message of its DIRECT child only. A nested sub-agent — a helper spawned by a pipeline agent that holds the Agent tool (Claude Code >= 2.1.172) — MUST NOT emit an openQuestions envelope: it would never reach the user, and the helper would stall awaiting a resume that cannot come. Parents must not delegate question-asking downward; questions discovered by nested helpers surface in the parent's own envelope or artifact. See skills/nested-agents/SKILL.md for the full guardrails.

Orchestrator parse rule (Decision 5: first-block-wins)

The orchestrator scans the subagent's final tool result for fenced

object contains an `openQuestions` array is the envelope. Other fenced
JSON blocks (such as a subagent's summary return at the very end of an
artifact-complete message) are ignored when an envelope is present.
Subagents MUST NOT include that key in their summary returns.

## Reference implementations

- `agents/design-author.md` — DESIGN phase open questions.
- `agents/questioner.md` — multi-repo disambiguation, including the
  free-text escape hatch for path collection.