discuss-case

name: discuss-case description: This skill should be used when the user asks to "discuss the case", "brainstorm about the investigation", "I'm stuck", "what do you think", or when the investigation loop detects a deadlock, hypothesis ambiguity, or needs user input to break a tie. Facilitates structured case discussion.

Discuss Case — Structured Investigation Dialogue

Conduct a case discussion with the user at critical investigation junctures. Present the situation clearly, surface key uncertainties, and integrate user input as high-weight evidence.

MCP-first user guidance

When the user provides guidance, write it through detective_apply_user_guidance:

• Facts → guidance_type="fact"
• Theories → guidance_type="theory"
• Boundaries or disallowed directions → guidance_type="constraint"
• New uncertainties → guidance_type="question"

After applying guidance, call detective_graph_overview before continuing. User guidance has priority over automated scheduling.

When This Triggers

Automatic triggers (from investigate loop):

• Multiple hypotheses with confidence spread < 0.1 (ambiguity)
• All candidate actions scoring below 0.3 (deadlock)
• Previously strong hypothesis just eliminated (direction change)
• Budget > 80% consumed (approaching limit)
• Phase regression detected (new evidence contradicted main theory)

Manual trigger: user explicitly calls /detective:discuss-case

Discussion Structure

Every case discussion follows this four-part format:

Part 1: Case Summary (Brief)

One paragraph summarizing the current state in plain language. No jargon, no fragment IDs — speak like a detective briefing a colleague:

"We're investigating [problem]. So far we've established [key facts]. Our leading theory is [hypothesis] but [complication]."

Part 2: Key Uncertainty (The Fork)

Identify the 1-2 most critical points of uncertainty that are blocking progress:

"The main question right now is: [specific question]. This matters because [why it's blocking]."

If there's a tie between hypotheses, present them as alternatives:

"Two theories are equally plausible right now:

1. [Hypothesis A] — supported by [evidence], confidence [X]
2. [Hypothesis B] — supported by [evidence], confidence [Y]"

Part 3: Detective's Inclination

State what the system would do next if acting autonomously, and why:

"My instinct says [direction] because [reasoning]. But I'm [X]% confident in this — [what could be wrong]."

Part 4: Ask the User

Pose a specific, actionable question. Prefer structured options when possible:

• "Do you have any domain knowledge that could distinguish between these two theories?"
• "Should I prioritize [A] or [B] given your context?"
• "Is there a source of information I haven't considered?"
• "Should I abandon [direction] and try something new?"

Processing User Response

For v2/v2.1 cases, user input enters the graph only through detective_apply_user_guidance:

User-authority fragments have elevated status:

• source: user_authority fragments start at higher maturity
• Constraints from user are treated as hard constraints (anchor-level)
• User hypotheses get a confidence boost (+0.2 over default)

After Discussion

1. Apply any user input with detective_apply_user_guidance if it was not already applied.
2. Call detective_graph_overview to review the updated MCP-backed graph state.
3. Use Detective MCP tools for any follow-up state changes, exports, or scheduling updates.
4. Report what changed:

"Based on your input, I've [updated X, eliminated Y, added Z]. The investigation now points toward [direction]. Continuing with /detective:investigate."

5. Return control to the investigate loop (or pause if user prefers)

Tone

Speak as a detective colleague — professional, direct, no unnecessary formality. Acknowledge uncertainty honestly. Never pretend confidence that doesn't exist in the data.

Legacy v1 fallback resources

These scripts are deprecated fallback references only for legacy v1 flat JSON case files when the Detective MCP server is unavailable. For v2/v2.1 cases, user guidance and any state changes must go through Detective MCP tools.

Scripts

• scripts/board.py — Legacy v1 board inspection and deprecated user-fragment helpers
• scripts/scoring.py — Legacy v1 scoring and deprecated propagation helpers