box/docs/MUSE-CHOICES-POLICY.md

---
title: "Muse-Choices Deny/Escalate & Agy Hold-All Exit Policy"
status: "signed-off"
coordinator: "operator-main"
scope: "muse-choices-policy"
accepted_at: "2026-10-07T05:55:03Z"
accepted_quote: "ACCEPT"
gate: "coordinator"
signoff_targets:
  - "choices-deny-escalate"
  - "agy-hold-all-exit"
  - "interview-gated-choices"
---

# Muse-Choices Deny/Escalate Policy — DECISION RECORD (Final)

Topic: add deny/escalate decisions to the `muse-choices` auto-approve daemon
(`bin/muse_choice_watcher.py`), which today only approves (top choice per
prompt kind). Interviewed 2026-10-06/07 per grill contract; accepted
verbatim below, which flipped this record from Draft to Final.

## Standing constraints (settled by user)

- All prompts must resolve: no stuck states are acceptable in any outcome.
- The full-auto top-choice flow must always exist as a path.
- Model review of choices is a FUTURE layer. Deferred out of this interview.

## Settled decisions

- D0 (helper form): a checked-in repo rules file informs decisions.
  Source: user's structured answerquared 2026-10-06 ("Repo rules file
  (Recommended)" for "which helper should inform approve/deny/hold
  decisions"). Rationale recorded at selection time: deterministic,
  versioned, sub-second, unit-testable; no agent round-trip latency.

## Settled during interview

- D0b (approve path needs no helper): straightforward prompts resolve
  locally with top-choice keys (the five matcher kinds, already
  implemented and live). Helpers (D0 rules file) govern deny/hold
  judgments only. Source: user direction 2026-10-06 ("the watcher
  itself should be able to input 1"; "always have the flow for full
  auto just top choice").
- D1 (rule match dimensions): command pattern first, plus kind and
  text pattern. Source: user selected option 1, 2026-10-06.
  Rationale: the observed risk lives in the `$ command` of approval
  dialogs; kind/text add precision around it.
- D2 (deny mechanics): deny exists only for permission kinds --
  `muse-approval` dialogs receive `2` + Enter, `y/n` prompts receive
  `n` + Enter. Question kinds (interview, letter, numbered) always
  resolve top-choice and are never denied. Source: user selected
  option 1, 2026-10-06. Rationale: deny is only meaningful where a
  permission is refused; questions stay total.
- D3 (hold mechanics): hold leaves the dialog untouched, suppresses
  auto-answer, raises a HELD entry in `box muse-choices status` plus
  an audit record; the operator resolves via a box command, otherwise
  a SHORT window expires back to top-choice approve. Source: user
  selected option 1 with "short window", 2026-10-06. Exact duration
  proposed below (2 minutes, tunable); accepted or amended with the
  scope text in D5.

- D4 (unmatched default): approve top-choice, exactly today's
  behavior. Source: user selected option 1, 2026-10-06. Rationale:
  follows from the standing constraints; rules carve out only
  deny/hold exceptions, so an empty rules file changes nothing.

## Open questions (unresolved)

None. All interview questions resolved and the scope accepted.

## Scope contract (ACCEPTED)

Artifact boundary, IN:

- `docs/MUSE-CHOICES-POLICY.md`: this record (Draft -> Final on acceptance).
- New `muse-choices-rules.json` at repo root (beside
  `keepalive-config.json`): the checked-in deny/hold rules.
- `bin/muse_choice_watcher.py`: rule evaluation, deny/hold paths, HELD
  state with short-window expiry, resolve plumbing.
- `bin/super-cli.py`: `box muse-choices resolve` command + HELD display
  in status.
- `tests/test_muse_choice_watcher.py`: rule eval, per-kind deny keys,
  hold/suppress/expiry, resolve flow.

Artifact boundary, OUT (rejected or deferred, each needs its own
interview to re-enter):

- Model review of choices (deferred future stage).
- Peer-agent consultation (rejected in D0).
- New matcher shapes (matcher suite's lane).
- `box runtime` work (adjacent lane, untouched).
- Timer cadence / daemon supervision changes.

Done means (all observable):

- [ ] This record marked Final with the acceptance quoted.
- [ ] Rules file loads; empty rules == today's behavior exactly.
- [ ] Deny sends `2`+Enter / `n`+Enter per D2: unit tests + one live
      scratch proof per permission kind.
- [ ] Hold suppresses + shows HELD + resolves via box + expires to
      approve: unit tests + one live scratch proof of hold and one of
      expiry.
- [ ] Audit records for deny/hold/resolve/expire in `box-ctl.jsonl`.
- [ ] Full suite green; fleet reloaded; desired state left as found.

Acceptance (quoted verbatim, chat, 2026-10-07T00:19:57Z): "ACCEPT".
Accepted as written, including the 2-minute tunable hold window. Per the
grill scope contract, later work outside the IN boundary needs explicit
owner approval or its own follow-up interview; "go" authorizes only this
boundary. No owning issue exists in this workflow, so this record is the
lane-coordination evidence.

## Non-goals (accepted with the scope)

- Model-based review of choices (deferred future layer).
- Peer-agent consultation over sidechat (rejected in favor of D0).
- Changes to approval matching shapes (covered by the matcher test suite).

## agy hold-all exit criteria — DECISION RECORD (Final)

Interview opened 2026-10-07. Topic selected by user (option 1, chat):
what proof flips agy panes from hold-all to auto-answering. The D0–D5
record above is Final and untouched by this interview. Numbering
continues as E1+ to avoid collision.

### Settled during interview

- E1 (proof bar): supervised live proof per kind, fixed count.
  Source: user selected option 1, 2026-10-07. Rationale: the risk is
  behavioral (how agy's renderer treats digit keys) — only live
  observation on real agy prompts proves it; synthetic tests cannot.
  A correct observation means a resolve-approve (or equivalent
  human-gated send of the kind's key) produces the intended selection
  with no stray keys. The count itself is E1b.
- E1b (count): three consecutive correct observations per kind,
  uniform across kinds; a miss resets the count. Source: user
  selected option 1, 2026-10-07. Rationale: one success could be
  luck (cursor already placed); three in a row across separate
  prompts proves the key binding, not the moment.
- E2 (rollout grain): independent per-kind flip — each kind exits
  hold-all the moment its 3 observations complete. Source: user
  selected option 1, 2026-10-07. Rationale: matches the uniform bar;
  rarely-seen kinds do not block proven ones.
- E3 (flip authority): automatic on completed proof — the 3
  observations are themselves human-gated (each a resolve-approve),
  so completion is the approval. The flip applies automatically and
  is recorded in the audit log plus a dated note in this doc.
  Source: user selected option 1, 2026-10-07. Rationale: no extra
  ceremony beyond the supervision already done; the audit trail
  shows all 3 proofs.
- E4 (rollback): single miss re-holds the kind — any post-flip
  wrong-send (wrong selection or stray keys), however observed,
  immediately re-holds that kind and wipes its proof count; re-flip
  needs 3 fresh observations. Source: user selected option 1,
  2026-10-07. Rationale: fail-closed and symmetric with the bar.
- E5 (applicability): fleet-wide — proof on any agy pane flips the
  kind on all agy panes. Source: user selected option 1, 2026-10-07.
  Rationale: the key binding is a property of the agy binary,
  identical on every pane.
- E6 (grill exclusion): numbered-plain stays coordinator-gated on
  agy; the E-rules never flip it. Source: user selected option 1,
  2026-10-07. Rationale: the E-rules prove the technical question
  (keys work); the grill hold answers the governance question
  (interviews need sign-off). Proof cannot lift a governance gate.

### Open questions (unresolved)

None. All interview questions resolved and the scope accepted.

### Scope contract (ACCEPTED)

- IN: this record section only (agy exit criteria, now Final).
- OUT: the flip mechanism (code), tests, daemon restarts, any
  implementation. Accepting this record never approves those stages;
  each returns for its own go-ahead.
- Done means: every E-question settled below, section marked Final,
  user acceptance quoted verbatim with channel and time.
- "Go" / "do it all" authorize only the IN boundary. No owning issue
  exists in this workflow, so this record is the lane-coordination
  evidence.

Acceptance (quoted verbatim, chat, 2026-10-07T05:55:03Z): "ACCEPT".
Accepted as written. Per the grill scope contract, later work outside
the IN boundary (the flip mechanism, tests, restarts) needs explicit
owner approval or its own follow-up interview; "go" authorizes only
this boundary.