/atlas: intent not captured up front (no plan-mode-style confirm) + interactive clarification is Claude-AskUserQuestion-only (no Codex/Gemini/other harness)

[anombyte93]

AI work-order (via issue-from-feedback). Confirmed against the actual skill flow in skills/. The "Fix Approaches" are hints, not a spec — try the recommended path and let does the entrypoint actually capture intent + ask correctly on each harness be the source of truth. Acceptance Criteria are the real spec.

Problem

Invoking /atlas does not immediately establish/understand the user's intent (no plan-mode-style "here's what I understood, let me clarify" moment up front), and it does not drive the harness to interactively ask the user clarifying questions using the correct per-harness mechanism (Claude AskUserQuestion, and the Codex / Gemini / other equivalents). New users don't get a clear signal that their intent was understood, and on non-Claude harnesses the structured questioning isn't wired at all.

Current Behavior (verified in skills/)

• skills/atlas/SKILL.md is a thin alias (allowed-tools: [Skill]) that immediately invokes the go orchestrator and captures/echoes no intent itself.
• skills/go/SKILL.md routes purely by pipeline state: on a fresh run current_phase is null/SETUP, so the FIRST phase is /prd:setup (engine_preflight: backend/provider resolution, capability detection). Intent discovery (DISCOVER → superpowers:brainstorming) only runs AFTER setup advances the phase — so the first thing the user sees is environment/config work, not intent capture.
• The engine deliberately avoids Plan Mode: skills/handoff/SKILL.md states "Plan Mode is NOT used (spec section 13.5): AskUserQuestion is the sole [mechanism]". So there is no plan-mode-on-invoke by design.
• Interactive clarification lives in skills/discover/SKILL.md (delegates to superpowers:brainstorming, one adaptive question at a time) and skills/handoff/SKILL.md — both built around Claude's AskUserQuestion tool, with a prose fallback only when AskUserQuestion is hook-blocked. There is no harness-detecting abstraction that selects the correct interactive-question primitive for Codex / Gemini / other.

Expected Behavior

• On the Atlas entrypoint (/atlas on Claude, $atlas on Codex, "use atlas"/activate_skill on Gemini), the engine FIRST confirms it understood the user's intent and asks the most important clarifying question(s) — an immediate "intent captured + clarify" moment — before/alongside any setup/config output.
• The clarifying-question step uses the harness-correct interactive tool: AskUserQuestion (Claude Code), the Codex equivalent, the Gemini equivalent, and a clean structured-prose fallback for any other harness. The harness is detected, not assumed.

Files to Touch

• skills/atlas/SKILL.md — surface an immediate intent-confirmation + first clarifying question before delegating (or instruct go to front-load it).
• skills/go/SKILL.md — routing currently puts SETUP before any intent capture; front-load a lightweight intent-confirm (a fast intent question before/within SETUP).
• skills/discover/SKILL.md — owns the brainstorming/AskUserQuestion flow; route the question step through a harness adapter.
• skills/handoff/SKILL.md — also AskUserQuestion-centric; same adapter.
• (new) a shared "ask-the-user" harness adapter (Claude AskUserQuestion / Codex / Gemini / structured-prose), driven by the engine's existing capability detection (detect_capabilities).
• root SKILL.md (the standalone single-file variant) — mirror the same behavior.

Researched Fix Approaches

1. [Recommended] Front-load intent + a harness-aware "ask" adapter (confidence: 75%)

• Pattern: (a) On invoke, emit an immediate intent-confirmation + the single highest-value clarifying question BEFORE setup. (b) Add one small "ask the user a structured question" abstraction that detects the harness (reuse detect_capabilities) and dispatches: Claude → AskUserQuestion; Codex → its interactive mechanism; Gemini → its activate_skill/structured flow; other → explicit numbered-options prose. discover and handoff call the adapter instead of hardcoding AskUserQuestion.
• Why: one engine, fixes BOTH defects ("intent not understood up front" + "Claude-only questioning"); aligns with the existing capability-detection design.
• Risk: Codex/Gemini may NOT expose a structured AskUserQuestion equivalent — then the "correct tool" there is explicit-options prose (confirm per harness, see Trust). Mild tension with spec-13.5 ("no Plan Mode") — reconcile: an intent-confirm can be a single AskUserQuestion-style question, not the harness Plan Mode.

2. [Alternative] Use each harness's NATIVE plan mode on invoke (confidence: 55%)

• Pattern: /atlas → enter Claude Code Plan Mode (EnterPlanMode) / Codex/Gemini plan equivalents to capture intent.
• Why not recommended: directly contradicts the engine's deliberate "Plan Mode is NOT used" design (handoff §13.5) and risks re-opening the user-agency hole AskUserQuestion was chosen to close. Only take this path after deliberately revisiting §13.5.

3. [Fallback] Prose-only intent + questions on non-Claude harnesses (confidence: 60%)

• Pattern: Keep AskUserQuestion for Claude; on Codex/Gemini emit an explicit "Here's what I understood; choose 1/2/3" prose block.
• Why last resort: loses structured/machine-readable answers off Claude; acceptable interim until per-harness primitives are verified.

Reference

• The engine already has the building blocks: handoff/SKILL.md enumerates Codex/Gemini external modes and a prose fallback for hook-blocked AskUserQuestion; detect_capabilities already detects installed harnesses. The fix generalizes these into one interactive-ask adapter.
• Cross-harness primitives to verify: Claude Code = AskUserQuestion; Gemini CLI = skillz extension + activate_skill; Codex = $name skill triggers + free-text.

Acceptance Criteria

• Invoking the Atlas entrypoint immediately surfaces an intent-confirmation + first clarifying question (user sees their intent was understood) BEFORE setup/config output.
• On Claude Code the clarifying step uses AskUserQuestion.
• On Codex and Gemini the clarifying step uses that harness's correct interactive mechanism (verified), degrading to explicit-options prose only where no structured primitive exists.
• The harness is detected, not assumed Claude; an unknown harness gets the structured-prose fallback.
• No regression to the existing Claude /atlas → SETUP → DISCOVER → … flow once intent is confirmed.

Complexity: M–L (cross-harness)

Trust Level: HINT (not specification)

1. Verify each harness's ACTUAL interactive-question primitive before wiring (Claude AskUserQuestion exists; confirm whether Codex/Gemini expose a structured equivalent or only free-text — if only free-text, "correct tool" = explicit-options prose there).
2. Reconcile with spec §13.5 ("Plan Mode is NOT used") — the intent-confirm should be an AskUserQuestion-style question, not necessarily the harness's Plan Mode, unless §13.5 is deliberately revisited.
3. The Acceptance Criteria are the spec; the approaches above are starting points.

---

_{via issue-from-feedback. Grounded in skills/{atlas,go,discover,handoff}/SKILL.md rather than web research, since this is an engine-UX issue where the skill flow is the source of truth.}