Grilling

Grilling

Intent and applicability

Expose the questions on which a plan's conclusions depend. The default is a question-only interview in which the user answers. In a selected speculative mode, a respondent supplies researched candidate answers and the orchestrator closes with an advisory recommendation. The griller remains question-only in both cases; respondent answers and orchestrator synthesis are separate roles.

Inputs and local bindings

Resolve the topic/artifact, existing decisions, selected mode, and any requested depth, time/cost limits, backend/model or stop boundary from current context. Inspect recoverable local facts instead of asking the user to remember them. Bind genuinely variable artifact paths, host tools, transport, persistence and backend capabilities from local evidence. Record missing/conflicting bindings only when they affect the selected work; do not invent commands or reinterpret a missing capability as permission to change the requested profile.

A requested depth is a useful-exploration budget in ordinary/linear native modes; stop when further questions cannot change a decision. Gradient depth is a hard ceiling with separate executed-node/resource constraints. The explicit linear runtime instead bounds completed exchanges and each direct invocation's timeout; it does not provide a nested-call or token budget.

Procedure: select the mode before loading mechanics

factory changes the evidence wave inside a linear interview; gradient changes the topology of the whole inquiry. Mode names never imply debug output. Do not load lattice mechanics for an ordinary interview or linear run.

The native named modes preserve direct respondent-to-specialist delegation. Only an explicitly selected mediated compatibility profile may change that transport. Host capability alone does not select it. For an explicit raw graph/protocol request, load review-boundaries.md after preparing the human review. For private machine transport or any offered resume/replay, load rendering-and-replay.md whether or not the user asks to see internals.

Non-goals

An ordinary interview does not select agent-authored answers, external backend calls, a lattice, debug diagrams or implementation. Those can be selected by an explicit user request or authorized parent scope. A comparison does not select a software execution handoff. Supporting investigation and bounded recovery remain available within the agreed objective and permissions.

Must not

• Turn the griller into an advocate, critic, judge, answerer or recommender. Each substantive griller block must be a genuine question. A short verified fact may support a question; research gives the griller no decision authority.
• Present candidate answers, scheduled branches, model agreement or silence as user ratification. Preserve exact approved nodes/effects; invalidate only dependency descendants when an upstream premise changes.
• Put graph identifiers, ledgers, control deltas or diagnostics in an ordinary human turn. Render plain prose with turn-local question numbers. An explicit debug view follows the complete human review; it never substitutes for it.
• Claim an automated validator, private-content guarantee, hard nested token cap, atomic graph replay or resume merely because a skill describes it. Bind actual implementation and verification evidence. The linear runtime's checkpoints do not implement the native graph protocol or lattice scheduler.
• Fabricate findings, missing provider output, completion, sources or authority. Retain incomplete turns and unresolved evidence with their resolution paths.

Interaction and authority

Preserve the selected interview: ask dependency-ready questions and wait for the user's answers in ordinary mode. In speculative modes, reuse settled scope, backend/data authority, budgets and decisions, then research and conduct bounded turns autonomously. Ask at material unresolved choices about objective, external data/backend, execution topology, budget guarantees or ratification; prepare the concrete options and evidence first. Do not ask for routine method choices or reconfirm unchanged approval through an alias or composed skill.

The user ratifies decisions unless they have explicitly delegated that exact authority. A respondent cannot fill a user-judgment gap by guessing. Stop only dependent work at a real unresolved boundary and continue useful authorized preparation. Research, spawning and model agreement confer no implementation, external-write or installation authority. Existing authority for the same settled implementation scope remains valid; a grilling recommendation adds none.

Completion and evidence

An ordinary run ends at the user's stop, no useful ready questions, or explicit resolution paths for remaining uncertainties. Return questions during the interview and a small plain-language review at a review boundary.

A native speculative close includes a bounded internal debate whenever materially different positions remain; debate explicitly selects that phase. Skip it only when there is no material split, explaining why. Preserve minority findings. Then give the orchestrator's recommendation, implications, relevant trajectory and evidence-supported effort range. Include the full autonomous execution handoff when implementation planning is selected, not for every comparison. These remain proposals until the required authority is supplied.

Report observed execution, actual bounds/usage when available, unresolved capabilities, provisional claims and what requires user judgment. A file's presence is not evidence that its runtime is installed, and manual review is not automated proof. Resumability, when selected, requires verified state reconstruction and dependency/ratification checks even when all state stays private. Keep user prose readable without requiring any protocol knowledge.