AskUserQuestion in plugin subagents — BLOCKED upstream (anthropics/claude-code#12890, #34592, #18721)

[ZaxShen]

Definitive status — blocked upstream

Verified against anthropics/claude-code issues (2026-04-24 search):

• #12890 — "BUG: AskUserQuestion tool not available to subagents regardless of configuration". Closed as not-planned. Worked in CC 2.0.55, broke in 2.0.56, Anthropic will NOT be restoring it.
• #34592 — "AskUserQuestion unavailable in all sub-agent contexts despite docs stating foreground pass-through".
• #18721 — "DOCS: Missing warning and workflow guidance for AskUserQuestion limitation in Subagents".
• #20275 — "DOCS: AskUserQuestion tool lacks documentation; conflicts with Agent SDK docs".

Root cause: CC filters AskUserQuestion at the system level when spawning subagents, regardless of tools: frontmatter. Configuration is ignored.

No workaround: none documented. Doesn't matter if the subagent is foreground or background, spawned via Task or otherwise — tool is stripped.

What this means for TMB

All four plugin agents (bro, architect, swe, pr-reviewer) are subagents spawned via the Task tool from the main CC conversation. None of them can use AskUserQuestion, ever.

The onboarding skills have been reverted to text-based multi-choice (Reply with 1/2/3). This is permanent until upstream reverses the not-planned close.

Possible workarounds (none pursued)

1. Main-Claude-mediated UI — emit a structured signal from bro, have main Claude intercept and render AskUserQuestion. Requires main-Claude cooperation via a plugin hook; CC doesn't currently expose the necessary seam.
2. Native terminal menu via Bash — select prompts or similar. Works but not rendered in the CC chat UI; appears as opaque terminal output.
3. Text-based multi-choice with clean labels — what we're doing. Slightly clunkier, universally works.

Resolution

• Onboarding skills reverted to text-based multi-choice (2026-04-24, commit 67f43b7).
• AskUserQuestion removed from bro's tools: allowlist.
• Upstream issues referenced here for the record.
• Monitor anthropics/claude-code for any policy change. If restored, re-enable in first-run-onboarding and tmb-reonboard skills.

Recommendation

Close this once the revert is merged. Re-file a new issue only if Anthropic reopens #12890 and restores the tool.

Related

• PR 📝 dogfood setup: SCENARIOS.md + bun workspaces + doc fixes (closes #40, #42) #41 (this PR) — contains the revert.
• Onboarding UX: use AskUserQuestion for multi-choice; remove press-enter-for-default pattern #48 — the original onboarding UX issue; now fully resolved via text-based multi-choice + clean labels (GitHub Flow/Git Flow naming clarity).

[ZaxShen]

Closing — no action item on our side. Anthropic closed anthropics/claude-code#12890 as not-planned; AskUserQuestion filtered from subagent tool registries at the system level, ignoring tools: frontmatter. Our plugin agents are subagents (spawned via Task), so this is permanent until upstream reverses. Current state: onboarding uses text-based 1/2/3 with clean labels. Re-open only if Anthropic restores the tool.

[ZaxShen]

Reopening — I was wrong to close this. User demonstrated AskUserQuestion rendering live in an interactive plugin-subagent session (screenshot attached to PR #41 discussion). My test that said otherwise was claude -p (headless), which strips interactive tools regardless of subagent status — not a subagent limitation. The upstream anthropics/claude-code#12890 may also be out-of-date for current CC builds.

Revert-revert shipped in PR #41. Keeping this issue open as a tracking note for: (a) the -p mode gotcha (AskUserQuestion doesn't expose in headless), (b) any future regression where CC strips it again at the subagent level. Will close when we have more confidence the current behavior is stable across CC versions.