kenoxa/run-discuss

Propose-and-refine interview producing discuss_artifact. Standalone or composed by workflow orchestrators (frame and design phases).

Accepts optional scope (frame|design), goal, and codebase_signals from parent skill. Standalone: infer scope from context, default frame.

Read lens reference matching scope at session start: frame → discuss-lens-frame.md, design → discuss-lens-design.md.

Protocol

Read discuss-protocol.md at session start.

| # | Phase | Type | |---|-------|------| | 1 | Seed | mainthread | | 2 | Propose | mainthread (interactive) | | 3 | Probe | G — dispatch @scout (breadth) or @researcher (depth) on demand | | 4 | Declare | mainthread |

1. Seed

Initialize known/unknown from codebase_signals and goal. Classify unknowns by type: scope, parameter, constraint.

Challenge initial framing: if user presents a diagnosis or solution as the problem ("the cache is broken", "we need to refactor X"), treat it as a hypothesis — add to unknown with type: scope, not known. Probe the underlying symptom before building on the assumption.

2. Propose

For each blocking unknown: propose interpretation, ask user to confirm/redirect/reject. Batch 2-3 per exchange. Never ask blank questions — every question includes a proposed answer.

Preemptive probe gate. Before posing any question to the user, ask: can the answer be derived from the source artifact, codebase Read/Grep, or a @scout/@researcher dispatch? If yes, dispatch Phase 3 Probe FIRST and present the derived answer. Only escalate to the user when the answer genuinely requires their judgment, domain knowledge, or a preference between materially different deliverables.

Calibrated depth: simple (2-3 questions), complex (5-8).

3. Probe

Two trigger paths:

• Preemptive (preferred): Propose's preemptive-probe gate fires — dispatch BEFORE asking the user. Feed result into known table; if probe resolves the unknown, do not surface a question for that item.
• Reactive: User response surfaces codebase question → dispatch on demand.

Dispatch @scout (breadth) or @researcher (depth). Skip user turn. Zero-dispatch when no codebase question is open — log with justification.

4. Declare

Convergence: no blocking unknowns remain. Before emitting artifact: verify Phase Trace has 4 rows (Seed, Propose, Probe, Declare).

Emit discuss_artifact with known table, open items (informational only), confidence.

Stall: 3 exchanges with no state change → surface gaps, suggest next step.

Output

Write discuss_artifact to .scratch/<session>/discuss-artifact.md. Schema in references/discuss-protocol.md.

Anti-Patterns

• Asking questions answerable from codebase_signals (use run-explore probes instead)
• Open-ended questions when enough signal exists to propose an interpretation
• Turn-count termination instead of state-delta convergence
• Accepting user's initial diagnosis as fact without probing the underlying symptom
• Scope drift: WHAT questions in design lens, HOW questions in frame lens