Interrogate

Interrogate is a bounded adversarial review. It does not implement fixes, does not auto-apply findings, and never uses Roblox Studio MCP playtest controls. Read ../david-mode/references/agent-runtime.md, references/interrogate-panel.md, and ../../models.json before dispatching reviewers.

Every reviewer uses one worker tier selected by the active client's runtime policy. Use narrow/Haiku for a local, evidence-limited review; use complex/Sonnet-high for cross-module, debugging, or security-sensitive review that is still independently bounded. Codex uses Luna-max for either tier. This is a single-model reviewer panel within the selected tier, not a multi-model panel. The parent checks the actual code and failure paths; agreement among workers is supporting evidence, not a quality guarantee.

Step 1, Determine scope and risk

Identify the exact diff or files to review:

• a pointed diff or file set from the user's request;
• git diff <base>...HEAD for a branch changeset;
• the relevant recent files when the user references work just completed.

Read the smallest surrounding contracts needed to trace behavior: Roblox repository instructions, typed Luau interfaces, remotes, persistence schemas, authored instance rules, and local verification scripts. Do not pull unrelated history or raw logs into every reviewer prompt.

Classify the review using the panel contract and select its budget from models.json.riskPolicy. Use fewer reviewers when the diff is smaller than the class suggests. Add a reviewer only when it can see the complete evidence and the parent can wait for it; queue reviewers beyond the runtime concurrency limit.

Step 2, State the intent