Hyperbrief Trigger Check

---
name: hyperbrief-trigger-check
version: 0.9.2
description: ALWAYS run BEFORE composing any message that asks the user for a decision, approval, or choice. Cheap escalation rubric (4-score + 5 MUST-trigger conditions) that returns one of {AUTONOMOUS_DECIDE, SUMMARY_BRIEF, FULL_HYPERBRIEF, MINIMAL_BRIEF, BLOCK_FRAMING}. v0.8 adds the brief-tier toggle (Hyperbrief.md §2.5) — the sub-threshold output is no longer always a one-liner; the tier floor (off | summary | full, default summary, set via HB.<tier> command / HYPERBRIEF_BRIEF_TIER env / .hyperbrief/config.json brief_tier) is resolved and the verdict is max(rubric_tier, floor) over BLOCKED_STUB < SUMMARY_BRIEF < FULL_HYPERBRIEF, so no setting can ever lower the tier the rubric demands. v0.9 adds the request tier (Hyperbrief.md §2.6) — a bare request for a brief resolves to SUMMARY_BRIEF rather than the heaviest tier, named tiers 요약/상세/심층 map to SUMMARY_BRIEF/FULL_HYPERBRIEF/DEEP_BRIEF, and DEEP_BRIEF is request-only and the only tier permitted to fan out. Triggered by message-intent patterns ('괜찮을까요','할까요','should we','which option','approve','confirm','choose between','OK to') OR by Superscalar opening a write/deploy/send lane OR by inbound Constellation DECISION_REQUEST. Also routes audience-profile commands (tone L<n>.<n>.<n> + term_pairing L<n>.{E|I|N}.{C|D|B|R|A}[!|?]) to the hyperbrief skill for AudienceProfileFallback population. Invokes the full hyperbrief skill ONLY when outcome != AUTONOMOUS_DECIDE. Skip for pure read-only fan-outs.
---

# Hyperbrief Trigger Check — the escalation gate

You are about to ask the user a decision question, OR Superscalar just opened a write/deploy/send lane, OR an inbound `DECISION_REQUEST` arrived. **Before composing the message**, run this 30-second rubric.

## 1. Trigger detection

Run this skill if ANY of these apply:

- Next assistant message will contain `괜찮을까요` / `할까요` / `should we` / `which (one|option)` / `approve` / `confirm` / `choose between` / `OK to` / similar decision-soliciting patterns.
- Superscalar fan-out just opened a `write`/`deploy`/`send` lane (Superscalar's irreversibility barrier touches a default-forbidden action).
- Inbound A2A is `CUSTOM/DECISION_REQUEST` requiring response.
- About to call `AskUserQuestion` tool.

**Skip** if:

- All fan-out lanes are read-only (no side effects).
- This is a continuation of an already-decided plan within the same approved Phase (Constellation §4: "no confirming planned dispatch").

## 1.5 Audience-profile command routing (v0.6+)

Independent of the escalation rubric, certain user inputs are **audience-profile commands** that must be forwarded to the `hyperbrief` skill (which owns `AudienceProfileFallback` IR population) for parsing + persistence, regardless of whether a decision is being made.

Recognized command shapes (parse BEFORE the escalation rubric; if matched, route AND continue):

- **Tone profile**: `L<a>.<b>.<c>` (e.g. `L1.1.1`, `L2.2.2`) — sets `audience_profile_fallback.tone_profile`.
- **term_pairing command** (v0.6, additive): `L<a>.<mode>.<scope>[suffix]` where:
  - `mode` ∈ `{E, I, N}` (every / initial / none).
  - `scope` ∈ `{C, D, B, R}` (conversation / document / board / review) — multi-scope via `+` (e.g. `C+B`), or shortcut `A` (= `C+D+B+R`).
  - optional `suffix`: `!` (force retroactive_apply = Y), `?` (force retroactive_apply = prompt).
  - Examples: `L1.I.C` / `L1.E.A` / `L2.N.D` / `L1.I.C+B` / `L2.E.D!` / `L1.I.A?`.
- **Combined** (single token, dot-chained): tone command followed by term_pairing command in the same utterance — both apply.

Routing behavior:

1. Capture the matched command token(s) verbatim.
2. Invoke the `hyperbrief` skill with `{audience_profile_command: "<token>", caller_context: "trigger-check-routed"}` regardless of the escalation verdict below.
3. The `hyperbrief` skill is responsible for: command parsing → `AudienceProfileFallback` population (incl. `term_pairing.mode/scope/retroactive_apply`) → low-frequency override (≤3 occurrences in scope unit → I mode still pairs every occurrence) → C-scope auto-forward semantics (conversation scope skips retroactive prompt; documents/board/review honor `retroactive_apply`) → dictionary resolution (`dictionary_ref` or `dictionary_inline`, placeholder if absent).
4. Then proceed with the escalation rubric (§2-§6) for whatever else the user message contains.

If the user message is **only** an audience-profile command (no decision content), the verdict is `AUTONOMOUS_DECIDE` (profile updates are reversible, in-scope, agent-decider).

## 2. Escalation 4-score (each 0-3)

| Indicator | 0 | 1 | 2 | 3 |
|---|---|---|---|---|
| **Irreversibility** | trivially reversible | reversible with effort | reversible with migration cost | one-way |
| **Blast radius** | single file / local | one module | cross-module | cross-system / external |
| **Time horizon** | minutes | hours | days | weeks+ |
| **Reversal cost** | free | low | moderate | severe |

`escalation_sum = irreversibility + blast_radius + time_horizon + reversal_cost`.

## 3. MUST-trigger conditions (any one → FULL_HYPERBRIEF regardless of sum)

1. `irreversibility_score >= 2`.
2. Blast radius crosses a module boundary (`touched_contracts` non-empty OR `touched_external_consumers` non-empty).
3. Any `touched_external_consumers` requires out-of-band notification.
4. Resource estimate exceeds threshold (tokens > 200k OR money > $50 OR time > 4h OR new external API/service).
5. Decision supersedes a `parent_decision_id` of an existing decision.

## 4. Cynefin domain classification (4 quick questions)

| Question | Yes → |
|---|---|
| Is the cause-effect relationship known and well-documented (best practice exists)? | `clear` |
| Is it knowable through expert analysis (good practice exists)? | `complicated` |
| Are cause-effect only knowable in retrospect (probe-sense-respond)? | `complex` |
| Is the situation rapidly destabilizing (act-sense-respond)? | `chaotic` |
| If none of the above clearly applies | `confused` |

## 5. RAPID Decider identification (1 question)

"Who has formal authority + accountability for this decision's outcome?" Common answers:
- `user` — default for product/governance/identity decisions.
- `agent` — for low-stakes, fully reversible, in-scope-of-plan actions.
- `external` — escalation needed beyond user (legal, compliance, partner).

If Decider != `user` AND escalation_sum >= 4, route is still `FULL_HYPERBRIEF` but `rapid.decider` records the actual authority.

## 6. Routing decision

```
if cynefin_domain == "confused":
    return BLOCK_FRAMING
    # Emit DECISION_REJECT_FRAMING to self / user with reason "domain unclear; clarify first".
    # Do NOT proceed to a brief — domain confusion must be resolved before option enumeration.

if cynefin_domain == "chaotic":
    return MINIMAL_BRIEF
    # Single action card. §7 domain_format = "single_action_card".
    # Skip §3 / §5 details. Auto-schedule 24h retrospective via hyperbrief-revisit.

any_must_trigger = (
    irreversibility_score >= 2
    OR blast_radius_crosses_module
    OR external_consumer_needs_notification
    OR resource_above_threshold
    OR supersedes_prior_decision
)

# --- what the RUBRIC asks for, on its own ---
rubric_tier = FULL_HYPERBRIEF if (escalation_sum >= threshold OR any_must_trigger) else BLOCKED_STUB

# --- what the TIER FLOOR asks for (v0.8.0, Hyperbrief.md §2.5) ---
floor = {off: BLOCKED_STUB, summary: SUMMARY_BRIEF, full: FULL_HYPERBRIEF}[brief_tier]

# --- what the REQUEST asks for (v0.9.0, Hyperbrief.md §2.6) ---
# A bare request for a brief is a SUMMARY. Asking for a brief is not asking for
# the heaviest one — before v0.9 it was, and that surcharge was never chosen.
# 심층/deep is request-only: neither the rubric nor the floor may produce it,
# because its added cost is fan-out and fan-out is paid in someone's tokens.
request_tier = (
    DEEP_BRIEF        if user named 심층 / deep
    else FULL_HYPERBRIEF if user named 상세 / full / detailed
    else SUMMARY_BRIEF   if user asked for a brief at all (bare or 요약/summary/short)
    else BLOCKED_STUB    # no request — the rubric and floor decide alone
)

# --- resolve as a MAX over the lattice. never min, never assignment ---
#     BLOCKED_STUB  <  SUMMARY_BRIEF  <  FULL_HYPERBRIEF  <  DEEP_BRIEF (request-only)
verdict = max(rubric_tier, floor, request_tier)

if verdict == DEEP_BRIEF:
    return DEEP_BRIEF
    # Full 9-section pipeline + a verification annex: adversarial refutation of the
    # recommendation's load-bearing claims. This is the only tier that may fan out.
    # Emitted as FullBrief IR + optional `verification_annex` — NOT a new schema shape.

if verdict == FULL_HYPERBRIEF:
    return FULL_HYPERBRIEF
    # Invoke the `hyperbrief` skill with the staged 9-section generation pipeline.
    # Do NOT fan out here. Depth of evidence is the deep tier's job, and a fan-out
    # nobody asked for is what made "brief me" read as expensive.

if verdict == SUMMARY_BRIEF:
    return SUMMARY_BRIEF
    # Invoke the `hyperbrief` skill with the 3-stage summary pipeline.
    # The summary MUST carry full_brief_fallback so the reader can escalate in one action.

return AUTONOMOUS_DECIDE
    # Only reachable at brief_tier == "off".
    # Decide autonomously + post-notify in ONE line (no brief, no question).
    # Format: "[decided autonomously, sum=<n>] <one-line summary of action + reversal path>"
```

### 6.1 Resolving `brief_tier` (v0.8.0)

Read it in this order and stop at the first hit:

1. **This session's in-conversation setting** — `HB.off` / `HB.summary` / `HB.full`, or an explicit instruction to that effect. Distinguished from the §1.5 tone commands by the `HB.` prefix; the two are orthogonal and may both be in force.
2. **`HYPERBRIEF_BRIEF_TIER` env** — `off` | `summary` | `full`.
3. **`.hyperbrief/config.json` → `brief_tier`** — the project's pinned policy. The PreToolUse hook resolves this same value and names the resolved floor in its advisory line, so when you see that line you already have the answer.
4. **Default `summary`.**

If the file is unreadable or the value unrecognized, use `summary` **and say so once** in the surface you emit. A silent fallback is indistinguishable from the setting having been honored, which matters most for the operator who pinned `full` and did not get it.

### 6.2 What the max is protecting

`max` — not assignment, not min. `brief_tier` raises the floor and can never lower the tier the rubric asked for: an irreversible or cross-module decision gets the nine sections at `off`, `summary`, and `full` alike. If you find yourself emitting a summary for something that fired a MUST-trigger because the setting said `summary`, you have inverted the feature into a bypass (AF-27), and the correct output is the full brief.

The §7 anti-triggers below short-circuit **ahead** of this resolution and are not overridden by any setting — `confused` still blocks framing and `chaotic` still gets the single action card at `full`. A frame nobody can classify does not become classifiable at greater length, and a chaotic domain's constraint is elapsed time, which a long brief spends.

One consequence worth stating because it is easy to implement wrongly: at `brief_tier: full` the §7 self-throttle's lever does nothing (raising the threshold 4→5 cannot matter when the floor is already full), but it MUST still emit its self-warning card and name `summary` as the remedy. A fatigue guard that stops speaking because it lost its lever looks exactly like one that found nothing wrong (AF-28).

## 7. Anti-trigger / suppression rules

- **Alert-fatigue self-throttle**: if rolling 20-cycle stats show `user_acceptance_rate > 70% AND user_premortem_input_avg_length < 30 chars`, raise the threshold from 4 to 5 for the next 10 cycles and emit a self-warning card to the user.
- **Time-pressure fallback**: if `deadline - now < 1h` AND `FULL_HYPERBRIEF` triggered AND full IR generation would exceed remaining time → emit `DECISION_DEFER` with reason "insufficient time for proper hyperbrief; recommend deadline extension OR autonomous-decision-with-post-notify acceptance" and surface the trade-off. **Do NOT shortcut sections** — partial briefs are worse than no brief.
- **Frame-rejection routing**: if user previously chose `reject_framing` on a parent decision, do NOT silently retry with a slightly-modified frame (anti-pattern AF-1 in Hyperbrief.md §6). Surface the parent's `reframing_reason` first.

## 8. Output contract

Return a structured handoff to the caller:

```jsonc
{
  "verdict": "AUTONOMOUS_DECIDE" | "SUMMARY_BRIEF" | "FULL_HYPERBRIEF" | "DEEP_BRIEF" | "MINIMAL_BRIEF" | "BLOCK_FRAMING",
  "brief_tier": "off" | "summary" | "full",              // v0.8: resolved per §6.1
  "brief_tier_source": "session" | "env" | "config" | "default" | "default-after-parse-failure",
  "rubric_tier": "BLOCKED_STUB" | "FULL_HYPERBRIEF",     // v0.8: what the rubric alone asked for, before the floor
  "request_tier": "BLOCKED_STUB" | "SUMMARY_BRIEF" | "FULL_HYPERBRIEF" | "DEEP_BRIEF",  // v0.9 §2.6: what the ASK asked for
  "request_tier_phrase": "<the words that named it>" | null,   // null when the request was bare (→ SUMMARY_BRIEF)

  "escalation_sum": <int 0-12>,
  "scores": { "irreversibility": <0-3>, "blast_radius": <0-3>, "time_horizon": <0-3>, "reversal_cost": <0-3> },
  "must_triggers_fired": [ /* names of any conditions in §3 that fired */ ],
  "cynefin_domain": "clear" | "complicated" | "complex" | "chaotic" | "confused",
  "reversibility_class": "two_way" | "one_way_with_migration_path" | "one_way",
  "rapid_decider": "user" | "agent" | "external",
  "autonomy_refusal_reason": "<1 line — why this is not autonomous>", // null if verdict == AUTONOMOUS_DECIDE
  "audience_profile_command": "<token>" | null  // v0.6+: captured tone/term_pairing command if matched in §1.5; null otherwise
}
```

If `verdict` is any brief tier (`SUMMARY_BRIEF` · `FULL_HYPERBRIEF` · `DEEP_BRIEF` · `MINIMAL_BRIEF`), immediately invoke the `hyperbrief` skill with this handoff as context — it selects the pipeline from the verdict. **Only `DEEP_BRIEF` may fan out**; at every other tier the brief is produced in-turn, because a fan-out the operator did not name is a cost they did not choose (§2.6). Emitting `request_tier` alongside `verdict` makes the §2.6 invariant auditable the same way `rubric_tier` makes AF-27 auditable: a `verdict` below `max(rubric_tier, floor, request_tier)` is always a defect, and a `DEEP_BRIEF` verdict whose `request_tier` is not `DEEP_BRIEF` means the request-only rule was violated. If `verdict == AUTONOMOUS_DECIDE`, proceed with the decision and emit the one-line post-notify (reachable only at `brief_tier: off`). Emitting `rubric_tier` alongside `verdict` is what makes AF-27 auditable after the fact: `verdict` below `rubric_tier` is always a defect, and the pair is the only place that comparison is recoverable. If `verdict == BLOCK_FRAMING`, surface domain confusion to the user before any option enumeration. If `audience_profile_command != null`, the `hyperbrief` skill must also be invoked for command parsing + AudienceProfileFallback population (orthogonal to the verdict path).

## 9. Back-compat (v0.6 cut)

- v0.5.6 callers that do NOT emit an `audience_profile_command` field: unchanged behavior. The §1.5 routing simply does not fire (no match → no command captured → null in output).
- v0.5.6 IR shapes remain valid against the v0.6 schema (all 4 new slots are optional): `section_0_decision_header.evaluation_lenses`, `section_8_recommendation.recommended_methodology`, `maturity_anchor` (top-level), `audience_profile_fallback.term_pairing`. This skill does not depend on those slots being populated.
- New `version` field in frontmatter (`0.6.0`) is informational; callers that pin `name: hyperbrief-trigger-check` resolve unchanged.

## Reference

Full spec: https://github.com/SoliEstre/EstreGenesis/blob/main/Hyperbrief.md (§2 Trigger Rubric, §5.6 Audience Profile, §11.5 Readiness Rubric)