mure-advisor

Tiered per-turn advisor for main session + WhatsApp. Cheap watcher + heavy escalation. Pure annotation, two surfaces.

SKILL.md

GitHub

mure-advisor — Per-Turn Watchdog for Main Session + WhatsApp

Purpose

A second pair of eyes on every main-lane response. NOT a self-critique step inside the lane — a structurally independent model that reads the lane's draft + recent context + the user's input, then emits an annotation. Provides adversarial-ally coverage in conversational flow without slowing the main response.

Tiered architecture (mandatory)

Tier 1 — cheap watcher (always-on)

Runs every main-lane turn. Default prime: Sonnet 5. Eligible pool:

  • anthropic/claude-sonnet-5 (prime)
  • anthropic/claude-haiku-4-5 (recon + cheap research lanes; eligible as watcher when Sonnet 5 is capped)
  • ollama-cloud/deepseek-v4-flash (when subscription active)
  • deepseek-v4-flash:direct (via direct API; subject to daily cap — see §Budget)

Haiku is NOT a heavy-advisor fallback. Per owner (2026-07-09): Haiku's lane is recon + cheap research. Heavy-tier failure never falls through to Haiku; Haiku stays in the cheap / recon lanes.

Output: a per-turn note in the canonical schema (see §Note schema). The watcher LLM-judges escalation; no flag-config file maintained.

Tier 2 — heavy advisor (escalation-only)

Only runs when the watcher escalates. Eligible pool:

  • anthropic/claude-opus-4-8 (anchor — heaviest)
  • deepseek-v4-pro:direct (Think Max mode; subject to daily cap — see §Budget)
  • zai/glm-5.2 (Z.ai quota pool)
  • minimax-portal/MiniMax-M3 (Minimax Ultra tier; vision-capable cross-modal at 1M ctx)

Heavy-tier fallback chain (binding): Opus 4.8 → DeepSeek V4-Pro Think Max → GLM-5.2 → MiniMax-M3. The dispatcher advances on quota / failure. Haiku is NOT in this chain.

Hard security rule: on security-relevant turns, heavy MUST be Opus 4.8. Sonnet 5 is explicitly disqualified for cybersecurity per model-audit-2026-07-09.md. The dispatcher enforces this; the heavy variant pick for sentinel/adjudicator/architect turns is non-negotiable.

Out of scope for now

  • Fable-5 EXCLUDED — limited access until 2026-07-12, then PPT, then unavailable. When access returns, evaluate against the auditor's notes before re-adding.
  • GLM-5.2 gated by Z.ai weekly quota — checking on 2026-07-10 reset before relying on it for sustained load.

Budget (binding)

DeepSeek direct API daily cap

Per owner (2026-07-09): max $1.0-1.5 USD per day on DeepSeek direct API (V4-Pro Think Max + V4-Flash combined).

Dispatcher enforcement:

  • Track daily spend in _SYSTEM/state/deepseek-burn.jsonl (one row per dispatch with $amount).
  • At 80% of cap: surface a deepseek-cap-warning to operator via the advisor notes channel (severity: watch).
  • At 100% of cap: refuse new DeepSeek dispatches, fall back to the next eligible model in the chain (GLM-5.2 → M3 for heavy; Haiku → M3 for watcher cheap).
  • Failed dispatches emit a deepseek-cap event to the operator dashboard.

Default cap: $1.25/day (mid-band; operator can adjust in _SYSTEM/state/mure-advisor.json). The cap is operator-tunable, hard-stop is non-negotiable.

Per-tier burn reporting

Each tier surfaces daily / weekly / monthly burn totals to the operator dashboard. Tunable thresholds warn at 80%, stop at 100%.

Note schema (binding)

Every emitted note MUST conform. Source of truth: _SYSTEM/research/advisor-note-schema-2026-07-09.md.

{
  "id":         "<uuid v4>",
  "turnId":     "<chat turn id>",
  "laneId":     "<dispatched lane id>",
  "model":      "<model string>",
  "tier":       "watcher|heavy",
  "severity":   "info|watch|warn|block",
  "summary":    "<one-line, <=140 chars>",
  "nextMove":   "<concrete action>",
  "confidence": 0.0-1.0,
  "evidenceRef": [{"kind":"file:line|url|trace","value":"..."}],
  "scope":      "this-turn|next-turn|both"
}

Notes below injectionFloor (default 0.5) are NOT injected into next-turn context — operator signal hygiene.

Trigger

  • Watcher: every turn.
  • Heavy: watcher escalates on: logical contradiction, missed branch, protected-path risk, overconfident unverified claim, scope creep, security-sensitive content.

Authority

Pure annotation. Post-hoc. Never blocks delivery. The chat model has already replied; the operator sees the note and the next turn's injection arrives one turn later. Origin: rene-branch-patterns:29-32 (directive-guard.mjs is observe-only, never blocks — same posture at higher fidelity).

Surface (dual)

  1. Visible annotation in main chat / WhatsApp reply.
  2. One-turn-lag injection into chat-model context for the NEXT turn via additionalContext (precedent: directive-guard.mjs).

The note's scope field controls: this-turn (visible only), next-turn (injected only), both.

Scope (channels)

Main session + WhatsApp only by default. NOT subject to advisor annotation:

  • Autonomous cron-triggered tasks
  • Dedicated MURE lane dispatches (engineer / scout / etc. run with their own result-packet discipline, not advisor notes)
  • Sub-sessions without the operator in the loop

Advisory for sub-session lanes is a future feature (operator-configurable).

Arming

  • Watcher lane DISARMED by default. Arm via either:
    • Session env: YURI_ADVISOR=1
    • Persistent flag: touch _SYSTEM/state/mure-advisor.enabled (gitignored)
  • Heavy escalation auto-armed when watcher armed.
  • Config changes (pool eligibility, schema version, injectionFloor, budget cap) — owner-gated.

Integration with MURE

  • mure.mjs --advisor-status → live watcher tier + last-N notes + per-tier escalation rate + deepseek-burn today.
  • mure.mjs --variants advisor → current variant picks per tier.
  • Telemetry: every emission logged to _SYSTEM/state/advisor-ledger.jsonl with full schema; per-line: turn id, lane, model, tier, severity, confidence, scope.
  • The mure-calibrator reads the ledger for Brier scoring; outputs feed the variant matrix's MLP re-ranker (see mure-role-variant-matrix skill).

Reinforcement / calibration

  • Watcher and heavy outputs feed mure-calibrator continuously.
  • Tracked over time:
    • Watcher escalation rate — target 5-15% of turns.
    • Heavy note acceptance rate — operator adopts the nextMove how often?
    • Per-model precision/recall — does the model producing the most notes correlate with rightness?
    • Confidence calibration — notes with confidence=0.9 should be right ~90% of the time.
    • DeepSeek burn vs cap.
  • Outputs route to MLP re-ranker (cheap-tier selection) and to the operator dashboard for trust calibration.

Failure modes

  • Watcher over-escalation (> 30% rate): the cheap tier is too cautious; tune the LLM-judge prompt, lower its default escalation threshold.
  • Watcher under-escalation (< 3% rate): the cheap tier is rubber-stamping; swap prime or downgrade from Sonnet 5 to a more skeptical model.
  • Heavy always-on by accident: DO NOT DO. Apex-blowout. The tiering is the load-bearing architectural choice — disable via the disarm flag if a regression silently lifts the gate.
  • Tier-1 model hallucinating heavy escalation signal: confidence < 0.6 on heavy-promote → drop the promotion. Confidence ≥ 0.85 → strong promotion.
  • Note injection polluting context: raise injectionFloor; or for a single noisy model, swap it out of the watcher pool.
  • DeepSeek cap exceeded mid-day: dispatcher auto-falls back; operator sees deepseek-cap-warning first, then deepseek-cap event.

Anti-patterns (in code OR docs)

  • Self-critique step in the main lane (no — must be a separate model).
  • Heavy advisor running every turn (no — apex blowout; tiering is binding).
  • Block-on-severity (no — authority is pure annotation; severity-block is declarative for the record only).
  • Flag-config file maintained by hand (no — escalation is LLM-judged).
  • Notes without schema compliance (no — silent failures rot the ledger).
  • Falling through to Haiku for heavy-advisor failure (no — Haiku is recon only; heavy chain is Opus → DVP → GLM-5.2 → M3).
  • DeepSeek spend unbounded (no — daily cap is binding).

Related

  • Source schema: _SYSTEM/research/advisor-note-schema-2026-07-09.md (v1.0.0).
  • Recon: _SYSTEM/research/rene-branch-patterns-2026-07-09.md (precedent: directive-guard.mjs).
  • Audit: _SYSTEM/research/model-audit-2026-07-09.md (model eligibility).
  • Adversarial review: invoked against _SYSTEM/research/adversarial-review-advisor-design-2026-07-09.md (4-fix reconciliation).
  • Fleet doctrine: skills/fleet-economy/SKILL.md (apex-reservation rule, ~35% orchestrator budget after the 20→35 update).
  • Variant matrix: skills/mure-role-variant-matrix/SKILL.md (the watcher prime + heavy picks are seeds here).

Versioned authority

v1.0.0 — created 2026-07-09. Revision incorporated: DeepSeek daily cap + M3 in Tier 2 + Haiku tier clarification. Pending Skill Workshop approval.

Change log

  • 2026-07-09 — initial draft per grill round + Opus 4.8 adversarial review (HV03).
  • 2026-07-09 (rev 1) — added MiniMax-M3 to Tier 2 heavy pool; added DeepSeek direct-API daily budget cap ($1.0-1.5/day); clarified Haiku is recon/cheap (NOT heavy fallback); added heavy-tier fallback chain Opus → DVP → GLM-5.2 → M3.