seokan-jeong/team-shinchan:fierce-option-panel

EXECUTE IMMEDIATELY

fierce-option-panel is a main-loop Workflow that hardens the quality of interview recommendation options generated by Misae's DESIGN_NEXT_QUESTION. It runs N diverse generators (structure-free — no A/B/C schema), filters them with a SelfCheckGPT majority-vote consensus (an option survives only if backed by ≥ ceil(N/2+1) generators — no any-pass promotion of a single generator's hallucination, HR-2), scores the survivors with a SteerConf cautious-confidence judge rubric, and returns a deterministic top-K.

Default-ON: explicit exception to the fierce-* opt-in convention

Every other fierce-* skill (fierce-debate, fierce-compete, fierce-ralph, fierce-review) is opt-in — the user invokes it explicitly because it is expensive. fierce-option-panel is the single intentional exception: it is ON by default under the quality-over-cost principle. Option quality at the interview stage compounds across the entire downstream workflow, so the extra per-turn cost is worth it. Documented identically in docs/fierce-option-panel.md and agents/misae.md (FR-10.2).

Step 0: Validate + opt-out

• This skill calls the Workflow tool (main-loop only). Never delegate to a subagent — workflow() throws inside a Task child (R-5).
• Opt-out / escape hatch (FR-6.3): read .shinchan-config.yaml. If interview.fierce_option_panel: false, do NOT run the panel — run the basic B-path (structure-free gen → verbalized sampling → missing-alternative critic → DINCO calibration, see agents/misae.md) and record current.interview.option_source: basic.
• Read interview.fierce_panel_generators (default 3), interview.fierce_panel_k_max (default 6), and interview.fierce_panel_token_budget_per_turn (default 60000). If the estimated panel cost for this turn exceeds the budget (HR-3), skip the panel, run the basic B-path, and record option_source: basic_fallback.

Step 1: Resolve question + personas (main loop — the script can't read files)

• Question + context: the interview question to design options for; pass the relevant files so generators read the right code.
• Personas (DRY): resolve generator + judge personas from the agent files:

  node ${CLAUDE_PLUGIN_ROOT}/src/workflow-personas.js misae actionkamen
  

Step 2: Run the fierce-option-panel Workflow

Workflow({
  scriptPath: "${CLAUDE_PLUGIN_ROOT}/skills/fierce-option-panel/fierce-option-panel.workflow.js",
  args: {
    question: "<the interview question>",
    generators: 3,                          // interview.fierce_panel_generators
    kMax: 6,                                // interview.fierce_panel_k_max
    files: ["<path>", "..."],              // context the generators should read
    generatorPersona: "<workflow-personas.js misae>",
    judgePersona: "<workflow-personas.js actionkamen>"
  }
})

Generate (N generators each emit one structure-free option + evidence + weight) → Consensus (SelfCheckGPT majority-vote, ≥ ceil(N/2+1); diversity floor bypass if < 2 survivors, R-3) → Judge (SteerConf cautious-confidence rubric, deterministic top-K). Returns { question, generators, k_max, consensus_threshold, options, scores, winner, rationale, dissent, option_source }.

Step 3: Graceful degradation (NFR-3)

If the Workflow returns an error (any generator threw, judge returned no verdict, budget overrun), the result carries option_source: 'basic_fallback'. Run the basic B-path to produce the options for this turn instead. No interview turn is ever blocked by panel failure.

Step 4: Record + format

• Record current.interview.option_source (fierce_panel | basic | basic_fallback) in WORKFLOW_STATE.yaml (FR-6.4) so retrospectives can audit which path each turn used.
• Apply A/B/C labels to the returned options only now (separate formatting step, FR-1).
• Raw confidence (FR-4 / HR-1): never write raw/uncalibrated self-confidence to WORKFLOW_STATE, logs, or debug output — only DINCO-normalized values.

Limitations / transferability gap (NFR-5)

The ECE/AUROC calibration metrics (in src/option-metrics.js) transfer from the factual QA literature, where "correct" is objectively defined. Here, "correct" is a proxy: the user's eventual option selection is treated as ground truth. This transfer is unvalidated for the design-option domain. The gating bars (ECE < 0.10, AUROC >= 0.70, Distinct-2 >= 0.55, self-BLEU <= 0.40) are pragmatic targets under the proxy, not universally validated thresholds. The same caveat appears in the option-metrics.js module docstring.