qfai-grill

qfai-grill - the session a user asks for

[DRIFT-PROTOCOL:REQUIRED]

Run a grilling session on whatever the argument names. This skill holds no
method of its own: it directs the run to qfai-grilling, which is the one
implementation. It holds a user session (.agents/rules/grilling.md § Two kinds
of session): the user asked for the interview, so every decision is put to them.

disable-model-invocation: true is the mechanism, and it is stated rather
than implied. A host offers a skill to the model from its description, so
omitting the description keeps it out of reach — on a host that tolerates a
skill without one. Others require it and register nothing, which loses the front
door the skill exists to be. Declaring the opt-out keeps the skill loadable
everywhere and reached only when the user names it, which is the whole
difference between the two: a session the user asked for, and a session a skill
started.

What this is for

Anything. A feature not yet designed, a decision that keeps being deferred, a
constraint nobody has written down. It needs no repository, no spec and no
prior stage — a session is a conversation, and the argument is its subject.

Non-goals

• Not a place to state the method. The method is qfai-grilling. Two
  statements of it drift, and the one a user reads is the one that gets
  forgotten.
• Not a writer. This skill produces no file. Nothing downstream reads a
  session it ran, so nothing is owed an artifact.
• Not model-invoked. The agent never fires this on its own. An agent that
  wants the method invokes the primitive.

Preconditions

Condition	Effect
qfai-grilling is installed	Read it and follow it
qfai-grilling is absent	Stop, and name what is missing
The argument is empty	Ask for the subject; a session needs one
A no-question mode is active	Run without asking; leftovers stay open

Without the primitive, stop. Report that
.qfai/assistant/skill/qfai-grilling/SKILL.md is not present and that
npx qfai init installs it. Do not interview from memory of how a session goes:
an improvised interview is the failure this split exists to prevent — it looks
like the method and answers to nothing, so a session that skipped the frontier
reads exactly like one that honoured it.

The run

1. Read .qfai/assistant/skill/qfai-grilling/SKILL.md.
2. Take the argument as the subject at the root of the tree.
3. Follow that skill, round by round, until it ends.

Nothing is added here. A step this skill performs that the primitive does not
describe is a second method, which is what the split exists to stop.

User Questions (AskUserQuestion Protocol)

Follow .qfai/assistant/rule/shared-skill-operating-baseline.md#user-questions-askuserquestion-protocol.
Do not open a second path for a question.

Inputs Priority

When unsure, read inputs in this order:

• P1: .qfai/assistant/rule/*
• P2: .qfai/assistant/skill/qfai-grilling/SKILL.md — the method
• P3: when the subject belongs to a QFAI project, its policy files under <paths.specsDir>/01_policy/ and its technology contract <paths.contractsDir>/tech.md
• P4: whatever the subject names, as facts to read rather than decisions to
  re-open

Sub-agent Delegation (MANDATORY)

Follow .qfai/assistant/rule/shared-skill-delegation-baseline.md. The
primitive's delegation rules apply unchanged: one thing is delegated, reading a
fact the environment holds, and the questions are never delegated.

Orchestrator Protocol

• The orchestrator reads the primitive, puts each round and reads the answers.
• It must not decide a frontier question on the user's behalf, and must not
  self-approve the session's end condition.

Capability Probe

1. Read the primitive at session start. That read is the capability check.
2. Attempt the first fact lookup a decision waits on.
3. If either fails, classify per the baseline taxonomy before doing anything
   else.

Delegation Failure (Hard Stop)

• The primitive unreadable: stop and report it, as above.
• A lookup unavailable: read what can be read directly, as the baseline's
  unavailable response sets out. Report the class, report every fact that
  stayed unread, and hold the decisions downstream of it open.
• A lookup saturated: use the baseline's bounded retry branch.
• Do not simulate roles. Do not answer a dispatched lookup from recollection —
  the frontier would then treat a guess as settled.

Work Orders Summary

A session that dispatched any lookup must report a ## Work Orders Summary
table to the user. Use the shared schema from
.qfai/assistant/rule/shared-skill-delegation-baseline.md, including
the Agent instance column. It is reported, not written: this skill produces no
file.

Step	Role (sub-agent)	Agent instance	Task title	Input (refs)	Output (refs)	Status (PASS/REVISE/PENDING)
1	Reviewer	<instance id>	Read the facts a round waits on	Decision and what it depends on	The fact, and where it was read	PASS/REVISE

Reviewer Gate

There is no invoking stage, so the gate is the user reading the record. That is
why this skill writes no files: a record nobody reads is not a gate, and a file
nobody asked for is not a record.

This entry point is therefore exempt from the baseline's reviewer gate. That
gate rules on an artifact and this skill writes none, so no PASS is requested,
none is awaited, and a run never blocks for one. The remit row for /qfai-grill
records the same thing from the reviewer's side: there is no artifact to review.

What the user is shown at the end:

• the ending the session reached, named: confirmed, user-closed,
  no-question or stopped — four of the five the rule names, because
  adopted ends only a delegated session. It is the one thing the rest cannot be read off:
  a confirmed session and a closed session that left nothing open produce the
  same three lists, and only the first says the user agreed the understanding
  was shared;
• every decision they answered, and what they answered;
• every node left open, labelled as an assumption or as unasked — decisions and
  facts only the user holds alike, because a fact no lookup reached is exactly
  the node that disappears when the report names decisions only;
• every fact taken as settled, with where it was read.

The exemption covers the verdict and nothing else. A lookup dispatched during the
session is a delegation like any other, so:

• Reviewer independence is defined normatively in
  .qfai/assistant/rule/shared-skill-delegation-baseline.md#definition-independent-reviewer-normative,
  and an agent that answered a lookup in this session is disqualified from
  answering it again as a check on itself.
• The Drift Protocol applies to the run, as it does to any run. Where a
  session touches tests, test-layers.md is what their placement is read
  against, and a ratio there is a signal rather than a gate.
• A lookup that could not be run at all is reported as unread, with the decisions
  downstream of it held open. That is this skill's PENDING, and it never reads
  as settled.

Default Autopilot Policy

Every decision this skill meets falls in one of three named buckets. They are
the primitive's, because the decisions are.

• auto-decide:
  • output formatting
  • equivalent-option pick
• ask-user:
  • every decision on the frontier
  • the confirmation that closes the session
• hard-required:
  • grilling subject (the design to interrogate; the argument supplies it, and
    an empty argument is asked for rather than guessed)

The asking is what this skill performs, so its ask-user entries are its own operations
rather than an entry added to the prototype.

Related

• The method this skill runs: .qfai/assistant/skill/qfai-grilling/SKILL.md
• Question form and the fallback without the tool:
  .qfai/assistant/rule/shared-skill-operating-baseline.md#user-questions-askuserquestion-protocol
• The clarification budget and its exemptions:
  .qfai/assistant/rule/constitution.md#article-vi--clarification-budget-avoid-endless-qa

project_memory:

• This skill states no method. The method is qfai-grilling, read at session start; a step performed here that the primitive does not describe is a second method.
• Without the primitive the run stops and names it. An improvised interview reads exactly like the method and answers to nothing.
• No file is written. There is no invoking stage, so the record is what the user is shown.
• The agent never fires this skill on its own. disable-model-invocation: true is what keeps a host from offering it, and it is declared rather than left to the absence of a description — a host that requires one would otherwise register nothing.