V16: Rules 22-25 (trust calibration, error-feedback, attribution gate, backtest integrity) + land prior uncommitted Rule 15 two-phase exit rewrite

alpaca: add P-012 — internal metric integrity self-validates, self-heals, fail-closed
Detects snapshot poisoning by positive re-derivation (pre-inception, raw-balance leak, adaptive sanity bound); quarantines (never deletes); tiered fail-closed health (HEALTHY commits, DEGRADED report-only, CRITICAL holds) with age-adaptive MIN_CLEAN; self-heals to HEALTHY via natural clean-snapshot refill. Backs the 2026-07-02 audit finding (36/37 snapshots poisoned, OSS steering live params off a phantom drawdown). Impl: scripts/metric_integrity.py, scripts/migrate_p010_metric_integrity.py, compute_oss.py clean-filter+tier, apply_adaptations.py tier gate. Tests: tests/test_p010_metric_integrity.py (10).
2026-07-24 16:46:15 -04:00 · 2026-07-24 16:46:15 -04:00 · 2026-07-24 16:46:15 -04:00 · 2026-07-23 18:47:17 +00:00 · 2026-07-23 18:11:06 +00:00 · 2026-07-23 18:10:36 +00:00
23 changed files with 2347 additions and 150 deletions
@@ -128,3 +128,10 @@ safe-mutate --verify "CMD" [--expect "PATTERN"] --mutate "CMD" [--reason "WHY"]
 Read the [Authoring Guide](docs/AUTHORING-GUIDE.md) before writing any new contract.
 It covers the full process: verify → draft → lint → review → ship, with templates
 and style rules.
 ## Maintaining this file
 Keep this file for knowledge useful to almost every future agent session in this project.
 Do not repeat what the codebase already shows; point to the authoritative file or command instead.
 Prefer rewriting or pruning existing entries over appending new ones.
 When updating this file, preserve this bar for all agents and keep entries concise.
@@ -0,0 +1 @@
 AGENTS.md
@@ -22,6 +22,7 @@ owners:
 - abiba
 - mumuni
 - kwame
 - ops
 trigger_types:
 - scheduled
 - event_driven
@@ -58,6 +59,7 @@ index:
    - memory-audit-maintenance
    - gpu-fleet
    - infrastructure-update
    - infrastructure-maintenance
    reference:
    - infrastructure-control
    - ra-h-os-custodianship-contract
@@ -90,6 +92,7 @@ index:
    - infrastructure-control
    - infrastructure-monitoring
    - infrastructure-update
    - infrastructure-maintenance
    - pm2-self-heal
    - disk-gc-threat-response
    gpu:
@@ -130,7 +133,6 @@ index:
    - build-zulip-plugin
    - stirling-pdf-agent-access
    - gpu-fleet
    - infrastructure-update
    - infrastructure-control
    - zulip-adapter-lessons
    - pi-approval-architecture
@@ -144,6 +146,9 @@ index:
    - mumuni-delegation
    kwame:
    - hello-world
    ops:
    - infrastructure-maintenance
    - infrastructure-update
  by_trigger:
    scheduled:
    - hermes-key-enforcement
@@ -156,6 +161,7 @@ index:
    - litellm-health
    - memory-audit-maintenance
    - infrastructure-update
    - infrastructure-maintenance
    event_driven:
    - litellm-self-heal
    - pm2-self-heal
@@ -197,6 +203,7 @@ index:
    - hermes-zulip-plugin
    - build-zulip-plugin
    - infrastructure-update
    - infrastructure-maintenance
    - ra-h-os-custodianship-contract
    - mumuni-delegation
    normal:
@@ -1256,13 +1263,108 @@ contracts:
  last_run: null
  last_status: null
  drift_alerts: []
 - name: infrastructure-maintenance
  file: infrastructure-maintenance.prose.md
  kind: responsibility
  category: maintenance
  sensitivity: high
  status: active
  owner: ops
  version: 1.0.0
  trigger:
    type: scheduled
    cadence: 0 2 * * 0
    description: Weekly host-level maintenance Sunday at 2am ET (replaces infrastructure-update
      build-phase role; infra-update moves to ops)
    cron_job_id: null
  execution:
    agent: ops
    timeout: 3600
    requires:
    - infrastructure-monitoring run within last 30 minutes (pre-update health baseline)
    - Proxmox snapshot of primary host OR /tmp backup dir created this run
    - LiteLLM master key from Infisical vault for health verification
    protocol:
    - Load contract from prose-contracts/main
    - Phase 0 preflight — capture health baseline, backup check, record image baseline
    - Phase 1 apt update && apt upgrade -y on primary host
    - Phase 2 docker compose pull for LiteLLM, SearXNG, and other running containers
    - Phase 3 restart stacks one at a time with per-stack health verification
    - Phase 4 verify every critical service (LiteLLM, SearXNG, Zulip, Gitea, PM2, Hermes gateways)
    - On failure — rollback per protocol, escalate, do not loop beyond circuit breaker
    - Log actions to ~/.hermes/runs/infrastructure-maintenance/
  verification:
    postconditions:
    - check: all critical services running after update
      verify: 'curl -sf http://192.168.68.116/litellm/v1/models && curl -sf https://chat.sysloggh.net/api/v1/server_settings && curl -sf https://git.sysloggh.net/api/v1/version && curl -sf http://192.168.68.7:8888 && pm2 jlist'
      expect: all probes 200 OK / processes online
    - check: no regressions from pre-update health baseline
      verify: diff Phase 0 health-baseline against Phase 4 results
      expect: no GREEN service turned RED
    - check: docker containers on latest stable tags
      verify: docker inspect --format '{{.Config.Image}}' <container> per service matches image-baseline.pulled_tag
      expect: all containers running pulled tags
    - check: APT packages up to date with no held broken packages
      verify: apt list --upgradable 2>/dev/null | wc -l and apt-get -s upgrade | grep -ci broken
      expect: upgradable == 0, broken == 0
    artifact: maintenance run report with phase results and any rollback/escalation
    verify_commands:
    - curl -sf http://192.168.68.116/litellm/v1/models
    - curl -sf http://192.168.68.7:8888
    - curl -sf https://chat.sysloggh.net/api/v1/server_settings
    - curl -sf https://git.sysloggh.net/api/v1/version
    - pm2 jlist
  receipt:
    format: json
    storage: ~/.hermes/runs/infrastructure-maintenance/
    graph_node: true
    schema:
      contract: string
      run_id: string
      timestamp: ISO 8601
      agent: string
      status: pass|fail|escalated
      phase: preflight|apt|images|restarts|verify|rollback|done|failed
      actions_taken: array
      postconditions: array
      drift_alerts: array
      evidence_path: string
  escalation:
    info:
      action: log_to_receipt
      notify: []
    warning:
      action: relay_alert
      notify:
      - abiba
      - mumuni
    critical:
      action: relay_alert
      notify:
      - abiba
      - mumuni
    fatal:
      action: relay_alert + pause + human_required
      notify:
      - abiba
      - mumuni
      - kwame
  circuit_breaker:
    max_retries: 2
    window: 7200
    trip_action: escalate_to_fatal
  depends_on:
  - infrastructure-monitoring
  last_run: null
  last_status: null
  drift_alerts: []
 - name: infrastructure-update
  file: infrastructure-update.prose.md
  kind: responsibility
  category: maintenance
  sensitivity: high
  status: active
-  owner: abiba
+  owner: ops
  version: 1.0.0
  trigger:
    type: scheduled
@@ -0,0 +1,391 @@
 # CHECKPOINT — Alpaca Trading System Remediation Plan
 **Status:** GATE 0 COMPLETE · **Created:** 2026-07-01 · **Resumed:** 2026-07-01T14:35Z
 **Pick-up instruction for a new session:** Read this file top-to-bottom. Gate 0 is done; resume at Gate 1.
 ---
 ## ✅ GATE 0 — COMPLETE (2026-07-01T14:46Z)
 ### OSS Before/After
 | Metric | Before (broken) | After (fixed) | Delta |
 |--------|-----------------|---------------|-------|
 | `oss` | 1.094 (above target!) | **-0.605** (below floor) | −1.699 |
 | `oss_trend` | "improving" | **"degrading"** | ✓ fixed |
 | `calmar` | 2.174 (sign destroyed) | **-2.174** | sign restored |
 | `spy_30d_return` | 0.58 (hardcoded) | **-1.44** (live) | real benchmark |
 | `rolling_alpha` | -1.01 (wrong) | **1.01** (portfolio beat SPY) | flipped |
 | `portfolio_beta` | -0.74 (wrong) | **0.30** (correct) | sign fixed |
 | `win_rate_adaptations` | 0.5 (dead code) | **removed** | dead code excised|
 | `benchmark_error` | — | **null** | fetch succeeded |
 ### Changes Made
 | Gate | File | Changes |
 |------|------|---------|
 | 0.1 | `scripts/compute_oss.py:65-66` | `abs(cagr)` → `cagr`; floor `0.0` → `-5.0` |
 | 0.2 | `scripts/compute_oss.py:344-362` | Removed dead `_compute_adapt_win_rate()` + call site |
 | 0.3 | `scripts/compute_oss.py:45-61` | Added `_fetch_spy_30d_return()` (yfinance via ^GSPC) |
 | 0.3 | `scripts/compute_oss.py:293-301` | Replaced hardcoded 0.58 with live fetch; alpha=None on failure |
 | 0.3 | `scripts/compute_oss.py:334` | `_compute_beta` now accepts live spy_return param |
 | 0.3 | `scripts/compute_oss.py:328-329` | JSON output now includes `spy_30d_return` + `benchmark_error` |
 | 0.4 | `tests/test_oss_monotonicity.py` | NEW — 3 property tests (Calmar sign, monotonicity, known values) |
 | 0.4 | `tests/test_benchmark_integrity.py` | NEW — 3 tests (no 0.58 literal, fetch signature, alpha sign mock) |
 | 0.4 | `~/.hermes/scripts/contract_invariant_watchdog.py` | Added 2 new test files to TEST_FILES list (now 7 files) |
 ### Verification
 - `compute_oss.py` runs clean: OSS=-0.605, trend=degrading, live SPY=-1.44%
 - `tests/`: 36 passed (30 original + 6 new)
 - Watchdog: "✅ All contract invariants hold (7 test files passed)"
 ### Next: Gate 1 — Make the optimization loop load-bearing
 ---
 ## ✅ GATE 1 — COMPLETE (2026-07-01T15:00Z)
 ### Changes Made
 | Sub-gate | File | Changes |
 |----------|------|---------|
 | 1.1 | `scripts/migrate_gate1_schema.py` | NEW — creates `adaptations` + `parameter_state` tables, seeds 5 params |
 | 1.1 | `trade_journal.db` | `parameter_state` (5 params) + `adaptations` (0 rows, clean) |
 | 1.2 | `scripts/apply_adaptations.py` | NEW — reads OSS+regime, proposes bounded changes, writes DB |
 | 1.3 | `scripts/apply_adaptations.py:auto_unwind()` | Auto-reverts params after 3 consecutive non-positive forward_oss_delta |
 | 1.4 | `scripts/compute_oss.py:W_ADAPT=0.10` | Re-weighted: W_CALMAR=0.35, W_WIN_RATE=0.10, W_ADAPT=0.10 |
 | 1.4 | `scripts/compute_oss.py:_compute_adapt_win_rate()` | Queries `adaptations.result='improved'` (not dead `analysis_runs`) |
 | 1.5 | `tests/test_adaptation_persistence.py` | NEW — 3 tests (seeded params, write+unwind, migration idempotent) |
 | 1.5 | `~/.hermes/scripts/contract_invariant_watchdog.py` | Added test_adaptation_persistence.py (now 8 test files) |
 ### Adaptation Logic
 | OSS Range | Trend | Regime | Action |
 |-----------|-------|--------|--------|
 | < 0.5 | degrading | bearish | **Tighten**: lower STOP_LOSS_PCT, raise CONFIDENCE_FLOOR, lower HEDGE_CAP |
 | > 0.8 | improving | bullish | **Loosen**: raise STOP_LOSS_PCT, lower CONFIDENCE_FLOOR, raise HEDGE_CAP |
 | Any | — | neutral | **Hold** — no change needed |
 - Step size: 10% of (max−min) range
 - Bounds enforced from `parameter_state`
 - Auto-unwind after 3 consecutive non-positive forward_oss_delta
 ### Verification
 - `apply_adaptations.py` runs clean: proposes STOP_LOSS_PCT adaptation in bearish regime
 - Auto-unwind test: 3 failed adaptations → reversion to original value
 - `compute_oss.py` shows `AdaptWR=0.5` (neutral default, wired to real data)
 - **39 tests pass** (30 original + 6 Gate 0 + 3 Gate 1)
 - **Watchdog: "✅ All contract invariants hold (8 test files passed)"**
 - LLM cron `2ca03134ea62` is now reporter, not actor — calls `apply_adaptations.py`
 ### Next: Gate 2 — Make hedging capable
 ---
 ## 0. ORIENTATION (read first)
 **Repo:** `/opt/alpaca-trading-agent`  ·  **venv:** `.venv/bin/python`
 **Contract (source of truth):** `/opt/openprose-contracts/cron-contracts/alpaca-trading-system.prose.md`
 **Full audit narrative:** `/opt/openprose-contracts/cron-contracts/alpaca-adversarial-audit-2026-07-01.md`
 **Watchdog:** `~/.hermes/scripts/contract_invariant_watchdog.py` (cron job `85522a081e55`, daily 08:00, no_agent, silent on pass)
 **Adaptation cron (LLM-mode):** job `2ca03134ea62` (DeepSeek-v4-pro, pinned)
 **DBs:**
 - `trade_journal.db` — perf, orders, trades, `analysis_runs`, `config_store`. **This is the DB `compute_oss.py` reads** (`DB_PATH`, line 18).
 - `consensus/signals.db` — `signals`, `consensus`.
 **Governing design principle (Theo, non-negotiable):** the contract must be **declarative** — declare desired *outcome states* + expose *levers* + define *feedback signals*, IaC-style. NOT prescriptive step lists. Every fix below converts a frozen constant into a tuned lever or adds a missing feedback signal. Wherever you must hardcode, justify it as an irreducible exception.
 **[fv] discipline for this work:** internal claims (code lines, DB columns) are verified by direct inspection (stronger than search — it's ground truth). External claims (financial definitions, SOTA findings) require independent `web_search` per claim. Label every assertion CONFIRMED / DISPUTED / UNVERIFIED. Add a property test that would have caught each defect before shipping the fix.
 ---
 ## 1. VERIFIED CURRENT STATE (the evidence base)
 ### Live OSS output (2026-07-01T14:19Z) — the smoking gun
 ```
 oss=1.094 (target 1.0, floor 0.5)  oss_trend="improving"
 calmar=2.174  profit_factor=0.679  win_rate=0.5
 total_return_pct=-0.43%  sharpe=-1.38  realized_pnl_30d=-$1.67 over 24 trades
 ```
 A losing, negative-Sharpe, PF<1 portfolio scores ABOVE its aspirational target and reads "improving."
 **VERIFY:** `cd /opt/alpaca-trading-agent && .venv/bin/python scripts/compute_oss.py | python3 -m json.tool`
 ### Confirmed defects (all [fv] CONFIRMED by direct inspection)
 | ID | Defect | Exact location | Evidence |
 |----|--------|---------------|----------|
 | D1 | Calmar uses `abs(cagr)` — sign destroyed | `scripts/compute_oss.py:65` | `compute_calmar(-2.35,-0.43,30)=2.174` vs `(+0.43)=2.28` |
 | D1b | Calmar clamp floor `max(calmar,0.0)` erases negatives even if sign fixed | `scripts/compute_oss.py:66` | `return round(min(max(calmar,0.0),5.0),3)` |
 | D2 | Benchmark hardcoded | `scripts/compute_oss.py:274` (`portfolio_return - 0.58`), `:319` (`spy_30d_return = 0.58`) | never fetched; alpha+beta are self-referential |
 | D3 | Win-rate query references 2 non-existent columns | `scripts/compute_oss.py:352` (`analysis_type`), `:349` (`result`) | `PRAGMA table_info(analysis_runs)` has NEITHER → bare `except:pass` at :360 → returns 0.5 forever |
 | D4 | Hedge structurally incapable | `.env.prod` + `consensus/ross/config.py:84-86` | `MAX_HEDGE_POSITIONS=1`, `HEDGE_POSITION_SIZE_PCT=0.01` → max ~1% of equity; contract Rule 8 assumes 25% (`PARKER_HEDGE_CAP_PCT=0.25`, config.py:113) |
 | D5 | P-004/005/006 unimplemented | grep entire tree | `hedge_strategy`, `discovery_lens`, `source_provenance` = 0 hits in *.py, 0 DB columns |
 | D6 | No adaptation persistence | DB introspection | tables `parameter_state`/`adaptations`/`regime_state` do not exist; `config_store` untouched since 2026-05-25 |
 | D7 | Discovery is static | `config_store.tickers` | 7 fixed symbols, last write 2026-05-22; frank=1, janet=1 distinct tickers ever |
 | D8 | Position size flat, `POSITION_SIZE_PCT` dead | `consensus/ross/sizing.py:7`, `config.py:138` | `calculate_position_size` returns `cfg.MAX_POSITION_NOTIONAL` which `= MIN_NOTIONAL` ($5); the 0.025 pct is never used |
 | D9 | Watchdog covers only P-002/P-003 | `contract_invariant_watchdog.py:19-25` | 5 test files = dust filter + equity source + pipeline; P-004/5/6 never tested |
 | D10 | "Consensus" is single-track | `consensus/sophie.py` + signals.db | all executed decisions score=1; universes disjoint by design |
 ### Confirmed ASSETS (things that already exist — reuse, don't rebuild)
 | Asset | Location | Use for |
 |-------|----------|---------|
 | Live market fetch (VIX/SPX/NDX/BTC via yfinance) | `consensus/track_parker.py:76-101` (`_fetch_market_data`) | D2 fix — reuse `spx.history(period="1mo")` for real SPY 30d return |
 | Bounded tunable levers (already documented w/ min-max) | `config.py:99-113`: STOP_LOSS_PCT[-0.03,-0.07], PROFIT_TAKE_PCT[0.03,0.10], STALE_EXIT_DAYS[3,7], CONFIDENCE_FLOOR[0.65,0.85], PARKER_HEDGE_CAP_PCT[0.10,0.30] | D6 — levers EXIST; they need a *writer*, not a redesign |
 | Property-test harness (hypothesis, temp-DB, module-reload) | `tests/test_perf_equity_property_invariants.py` | template for all new invariant tests |
 | Regime detection (VIX-trend + SPX-vs-SMA20) | `track_parker.py:152+` (`_detect_regime`) | D4 — regime signal exists to drive dynamic hedge sizing |
 | slippage_bps / intended_price columns | broker_orders schema | P-008 cost-aware gate (data already captured, unused) |
 | Risk-gate framework (Gates 0-7) | `consensus/ross/gates.py:14` (`check_risk_gates`) | insert cost-aware gate here |
 ---
 ## 2. EXTERNAL EVIDENCE (July 2026 SOTA) — [fv] CONFIRMED against sources
 - **Calmar = signed CAGR ÷ |maxDD|** — negative return MUST yield negative Calmar. (Investopedia, Wikipedia, QuantifiedStrategies — CONFIRMED, ≥3 sources.) → D1/D1b are unambiguous bugs.
 - **FINSABER (arXiv:2505.07078v5, KDD'26):** LLM edge deteriorates over long horizons/broad universe; agents "overly conservative in bull, overly aggressive in bear." Prescription: **prioritize trend detection + regime-aware risk controls over framework complexity.** CONFIRMED (abstract).
 - **StockBench (arXiv:2510.02209):** most LLM agents don't beat buy&hold over 4mo, but ALL limit drawdown (-11/-14% vs -15.2%). → realistic capturable edge = **drawdown control, not alpha.** CONFIRMED.
 - **Agent Market Arena (arXiv:2510.11695):** performance driven by **architecture (memory/risk/sizing), not LLM choice.** CONFIRMED.
 - **CLQT (arXiv:2606.29771):** return leaderboards mislead; apparent alpha dissolves once look-ahead controlled → "mostly passive factor exposure." Evaluate capability, not single score. CONFIRMED. → OSS is exactly the single-number leaderboard being warned against.
 - **Signal half-life/capacity (OpenAlgo):** turnover set by half-life; every trade pays half-spread+impact; net-of-cost is the only honest test. CONFIRMED.
 - **Leveraged inverse decay (StockTitan 2026):** daily-reset -3x (SQQQ/SPXU) decays even when directionally right → prefer 1x inverse (SH/PSQ) for multi-day holds. CONFIRMED. Contract Rule 18 already says this; `MAX_HEDGE_HOLD_DAYS=3` + HEDGE_TICKERS still lists SQQQ/SPXU fight it.
 **Synthesis implication:** the honest declared outcome state is **"match benchmark net of costs while bounding drawdown,"** NOT "maximize growth." Re-weight the objective accordingly (P-011).
 ---
 ## 3. THE PLAN (dependency-ordered gates; LLM-paced, no human timelines)
 > Each task: **DO** (exact edit) → **VERIFY** (command + expected result) → **TEST** (property/regression test to add). No task is "done" until its VERIFY passes and its TEST is committed.
 ### GATE 0 — Restore metric truth (BLOCKING; everything downstream is meaningless until this lands)
 **0.1 — Fix Calmar sign + clamp**
 - DO: `scripts/compute_oss.py:65` → `calmar = cagr / (abs(max_drawdown_pct) / 100.0)` (drop `abs(` on cagr).
 - DO: `:66` → allow negatives: `return round(min(max(calmar, -5.0), 5.0), 3)` (floor -5.0, not 0.0).
 - CONSIDER: the OSS aggregate weight math (`:278-280`) assumes non-negative components — after this, re-derive so a negative Calmar drags OSS below floor. Verify OSS can now go below 0.5 on a losing month.
 - VERIFY: `.venv/bin/python -c "import sys;sys.path.insert(0,'scripts');from compute_oss import compute_calmar as c;print(c(-2.35,-0.43,30), c(-2.35,0.43,30))"` → first value must be **negative**, second positive.
 - VERIFY: `.venv/bin/python scripts/compute_oss.py` → with current -0.43% return, `oss` must now be **< 0.5** and `oss_trend` must NOT be "improving".
 - TEST: add `tests/test_oss_monotonicity.py` (hypothesis): `∀ dd, days, r1<0<r2 ⇒ compute_calmar(dd,r1,days) < compute_calmar(dd,r2,days)` AND `compute_calmar(dd, r<0, days) < 0`.
 **0.2 — Fix or remove adaptation win-rate**
 - Root cause: query needs `analysis_type` AND `result` columns that don't exist in `analysis_runs`.
 - OPTION A (preferred, aligns with Gate 1): create the `adaptations` table (Gate 1.1) and repoint this query at it.
 - OPTION B (interim): remove the `W_ADAPT * 0.5` term from OSS and renormalize weights, so OSS isn't diluted by a frozen constant. Document as temporary.
 - VERIFY: grep confirms no silent `except: pass` returns a magic 0.5 that feeds OSS.
 - TEST: assert `_compute_adapt_win_rate` raises or returns None (not 0.5) when the backing table/columns are absent — no silent defaults.
 **0.3 — Live benchmark (kill the 0.58 literal)**
 - DO: add `_fetch_spy_30d_return()` reusing Parker's pattern (`import yfinance as yf; yf.Ticker("^GSPC").history(period="1mo")`; 30d pct change). Thread into `:274` (alpha) and `:319` (beta).
 - DO: on fetch failure, FAIL LOUD (return error dict / mark UNVERIFIED) — never silently default to a constant.
 - VERIFY: `.venv/bin/python scripts/compute_oss.py` shows a `spy_30d_return` that changes day-to-day and ≠ 0.58; `rolling_alpha` flips sign when portfolio crosses SPY.
 - TEST: `tests/test_benchmark_integrity.py` — alpha sign must equal sign(portfolio_return − live_spy_return); mock the fetcher.
 **0.4 — Gate 0 acceptance**
 - Re-run `compute_oss.py`; confirm it now reflects reality (the -$1.67 loss).
 - Add 0.1/0.2/0.3 tests to the watchdog TEST_FILES list; run watchdog → must now FAIL if any regression reintroduces the bugs.
 ### GATE 1 — Make the optimization loop load-bearing (depends on Gate 0)
 **1.1 — Persistence schema.** Create tables in `trade_journal.db`:
 - `adaptations(id, fired_at, param_name, old_value, new_value, regime, trigger_reason, oss_before, oss_after, forward_oss_delta, result TEXT, unwound INTEGER DEFAULT 0)`
 - `parameter_state(param_name PRIMARY KEY, current_value, min_bound, max_bound, updated_at, last_adaptation_id)`
 - seed `parameter_state` from `config.py:99-113` bounds (STOP_LOSS_PCT, PROFIT_TAKE_PCT, STALE_EXIT_DAYS, CONFIDENCE_FLOOR, PARKER_HEDGE_CAP_PCT).
 **1.2 — Move adaptation from LLM prompt to code.** Create `scripts/apply_adaptations.py`: reads OSS + regime, proposes bounded param change, writes `adaptations` + `parameter_state` rows, and config now READS from `parameter_state` (falls back to env). LLM cron `2ca03134ea62` becomes reporter, not actor.
 - Declarative framing: contract declares target OSS ≥ 1.0 and the bounded levers; this code is the controller that tunes toward it.
 **1.3 — Auto-unwind (Rule 8-14 guardrail).** After N evaluations (contract says 3), if `forward_oss_delta ≤ 0`, revert param from `parameter_state` and mark `unwound=1, result='reverted'`.
 - TEST: property test — any adaptation with non-positive forward delta over N evals is reverted; reversibility holds (state returns to pre-adaptation value).
 **1.4 — Wire win-rate.** Repoint `_compute_adapt_win_rate` at `adaptations.result='improved'`. Now OSS weight 0.15 is real.
 **1.5 — Watchdog.** Add P-009 (monotonicity) + P-010 (persistence/reversibility) tests. Watchdog fails if loop goes inert.
 ### Next: Gate 2 — Make hedging capable (depends on Gate 0 regime truth)
 ---
 ## ✅ GATE 2 — COMPLETE (2026-07-01T15:15Z)
 ### Changes Made
 | Sub-gate | File | Changes |
 |----------|------|---------|
 | 2.1 | `consensus/ross/hedge.py` | **NEW** — regime-driven hedge sizing buckets, cap enforcement, hold-day tiers |
 | 2.1 | `consensus/ross/sizing.py` | Hedge-aware: routes hedge tickers through dynamic sizing, non-hedge unchanged |
 | 2.1 | `consensus/ross/gates.py` | **Gate 8 (Hedge Gate)**: enforces per-ticker count + total notional cap from regime |
 | 2.1 | `consensus/ross/executor.py` | Extracts `bearish_score` from signal metadata, passes to sizing + gates |
 | 2.2 | `trade_journal.db:broker_orders` | Added `hedge_strategy TEXT` column |
 | 2.2 | `scripts/compute_hedge_attribution.py` | **NEW** — per-strategy profit factor, auto-retire PF < 0.5 for 30+ days |
 | 2.3 | `consensus/ross/gates.py` | **Gate 2.3 (Cost-Aware)**: advisory warning when avg slippage > 50 bps + filled < intended×0.995 |
 | test | `tests/test_hedge_sizing.py` | **NEW** — 6 tests (monotonicity, leverage tiers, bucket bounds, non-hedge passthrough) |
 | wdog | `contract_invariant_watchdog.py` | Added `test_hedge_sizing.py` (now 9 test files) |
 ### Hedge Sizing Buckets
 | Bearish Score | Bucket | Max Positions | Size/Position | 1x Hold Days | 3x Hold Days |
 |--------------|--------|--------------|---------------|-------------|-------------|
 | < 0.20 | no hedge | 0 | 0% | — | — |
 | 0.20–0.39 | moderate | 1 | 1.0% | 5 | 2 |
 | 0.40–0.69 | heavy | 2 | 1.5% | 5 | 2 |
 | ≥ 0.70 | extreme | 3 | 2.0% | 5 | 2 |
 - Capped at `PARKER_HEDGE_CAP_PCT` from `parameter_state` (current: 0.25)
 - 1x inverses (SH, BITI, PSQ): 5-day max hold (Rule 18 + StockTitan evidence)
 - 3x inverses (SQQQ, SPXU): 2-day max hold (crisis-only, high decay)
 ### New Gates
 | Gate | Scope | Enforcement |
 |------|-------|------------|
 | **Gate 8** | Hedge position count + notional cap | Hard block: returns False if at limit |
 | **Gate 2.3** | Cost-aware execution (slippage_bps) | Advisory warning (non-blocking — data is retrospective) |
 ### Verification
 - Sizing: SH (hedge, bearish=0.8) → $4 on $200 equity; SH (bearish=0.3) → $2; AAPL → $5 flat
 - Hold days: SH (1x) → 5 days, SQQQ (3x) → 2 days ✓
 - **45 tests pass** (30 + 6 Gate 0 + 3 Gate 1 + 6 Gate 2)
 - **Watchdog: "✅ All contract invariants hold (9 test files passed)"**
 - OSS = −0.475 (unchanged — Gate 2 doesn't affect OSS computation)
 ---
 ## ✅ GATE 3 — COMPLETE (2026-07-01T15:40Z)
 ### Changes Made
 | Sub-gate | File | Changes |
 |----------|------|---------|
 | 3.1a | `consensus/ross/discovery.py` | **NEW** — DiscoveryLens framework: 3 built-in lenses (ParkerTrend, VolumeSurge, SectorRotation), registry, Source Health Score, correlation-admission gate, probationary sizing |
 | 3.1b | `consensus/signals.db:signals` | Added `discovery_lens TEXT` + `source_provenance TEXT` columns |
 | 3.1b | `trade_journal.db:broker_orders` | Added `discovery_lens TEXT` + `source_provenance TEXT` columns |
 | 3.1c | `consensus/ross/sizing.py` | Probationary sizing: 50% notional for tickers with <3 trades or PF < 1.0 |
 | 3.1c | `consensus/ross/executor.py` | Passes `db_path` to sizing for probationary check |
 | 3.2 | `scripts/update_ticker_universe.py` | **NEW** — Replaces static config_store.tickers with lens-proposed, correlation-gated, aggregated universe (≥2 lenses → priority) |
 | 3.3 | `scripts/update_ticker_universe.py` | Coverage-gap streaks + auto-broaden lens criteria after 3 consecutive gaps (Rule 19) |
 | test | `tests/test_discovery_lens.py` | **NEW** — 9 tests (≥3 lens invariant, SHS empty-DB, correlation gate edge cases, probationary sizing, script idempotency, gap tracking) |
 | wdog | `contract_invariant_watchdog.py` | Added `test_discovery_lens.py` (now 10 test files) |
 ### Lens Architecture
 | Lens | Source | Max Tickers | Criteria |
 |------|--------|-------------|----------|
 | `parker_trend` | Parker macro regime signals (signals.db) | 10 | min 3 active tickers |
 | `volume_surge` | Volume anomaly detection | 15 | min 3 active tickers |
 | `sector_rotation` | Maya gold miners / critical minerals | 12 | min 3 active tickers |
 - **Source Health Score**: yield_rate × avg_PF × 100 × decay (0.85^months) — per lens, 90-day lookback
 - **Correlation gate**: rejects proposed tickers with r ≥ 0.80 vs existing positions (20-day min history)
 - **Probationary sizing**: 50% notional for tickers with <3 trades or PF < 1.0
 - **≥3 lens invariant**: universe update aborts if <3 lenses active
 - **Rule 19 hindsight**: coverage-gap streaks tracked; criteria auto-broaden (1.3× factor) after ≥3 gaps
 ### Dynamic Universe Pipeline
 ```
 Active Lenses → propose_candidates() → correlation-gate → aggregate votes
 (≥2 lenses = priority) → merge with existing proven tickers → write config_store
 ```
 ### Verification
 - Lens registry: 3 active lenses (`parker_trend`, `volume_surge`, `sector_rotation`) ✓
 - SHS on empty DB: 0.0 (no crash) ✓
 - Pearson: 1.0 perfect, -1.0 perfect negative ✓
 - Probationary: SH/untraded tickers → True ✓
 - Dynamic universe script: idempotent, preserves existing tickers ✓
 - Gap tracking: streaks initialized ✓
 - **54 tests pass** (30 + 6 Gate 0 + 3 Gate 1 + 6 Gate 2 + 9 Gate 3)
 - **Watchdog: "✅ All contract invariants hold (10 test files passed)"**
 - OSS = −0.477 (unchanged — Gate 3 doesn't affect OSS computation)
 ---
 ## ✅ GATE 4 — COMPLETE (2026-07-01T15:50Z)
 ### Contract Edits (human-approved)
 | Section | Change |
 |---------|--------|
 | §Goal "Return Maximization" → "Return Optimization" | Primary outcome = match benchmark net of costs, bound drawdown ≤25%; outperformance is secondary |
 | OSS weights | Calmar 35→**40%**, Alpha 20→**15%**, Win Rate 10%, Adapt WR 10% (unchanged) |
 | `parameter-state` | Added `oss_target` (0.5), `oss_floor` (0.3), `oss_weights` (per-weight bounds), `hedge_regime_buckets` (3 tiers) |
 | Invariant list | Added **P-007** (benchmark integrity), **P-008** (cost-aware execution), **P-009** (metric monotonicity), **P-010** (adaptation reversibility), **P-011** (drawdown-first objective) |
 ### Code Changes
 | File | Change |
 |------|--------|
 | `scripts/compute_oss.py` | W_CALMAR 0.35→**0.40**, W_ALPHA 0.20→**0.15** |
 | `alpaca-trading-system.prose.md` | Full contract update: 3 sections edited + 5 invariants added |
 ### OSS Impact
 | Metric | Before (Gate 3) | After (Gate 4) |
 |--------|----------------|----------------|
 | OSS | −0.477 | **−0.590** |
 | Trend | degrading | degrading |
 | Calmar weight | 35% | **40%** |
 | Alpha weight | 20% | **15%** |
 The shift to drawdown-first weighting amplifies the honest signal: Calmar at −2.174 now drives the composite more heavily, producing a truer picture.
 ### Verification
 - Contract updated per Theo's approval ✓
 - **54 tests pass** (unchanged — contract edits don't affect test count)
 - **Watchdog: "✅ All contract invariants hold (10 test files passed)"**
 - OSS −0.590 | Calmar −2.174 | SPY −1.1% | Beta 0.39
 ### Next: Gate 5 — Wire the loop end-to-end (cron + adaptation pipeline integration)
 ### GATE 3 — Real field of view (parallelizable with Gate 2)
 **2.1 — Dynamic hedge sizing.** Replace frozen `MAX_HEDGE_POSITIONS=1`/`HEDGE_POSITION_SIZE_PCT=0.01` with regime-driven sizing bounded by `PARKER_HEDGE_CAP_PCT` (up to the 0.25 already authorized). Sizing reads `parameter_state`. Reconcile `MAX_HEDGE_HOLD_DAYS=3` with Rule 18 (allow longer 1x-inverse holds).
 **2.2 — P-004 hedge attribution.** Add `hedge_strategy` column to orders/trades + tagging; compute per-strategy profit factor; retire strategies <0.5. Demote SQQQ/SPXU (decay) to crisis-only; default SH/PSQ (1x) per Rule 18 + StockTitan evidence.
 **2.3 — P-008 cost-aware gate.** New gate in `gates.py` using existing `slippage_bps`/`intended_price`: reject trades where expected edge < modeled round-trip cost (half-spread+commission+slippage). Directly addresses small-account cost drag.
 - TEST: hedge sizing scales with regime severity; cost gate blocks a sub-cost-edge trade; P-004 attribution sums correctly.
 ### GATE 3 — Real field of view (parallelizable with Gate 2)
 **3.1 — P-005/P-006.** Add `discovery_lens` + `source_provenance` columns + tagging; per-lens yield + Source Health Score; correlation-admission gate; probationary sizing for new tickers.
 **3.2 — Dynamic universe.** Replace static `config_store.tickers` with lens-proposed candidates each cycle (declare: "≥3 active lenses, correlation-gated"; expose lens criteria as lever).
 **3.3 — Rule 19 hindsight.** Write coverage-gap counts; broaden lens criteria when gaps detected.
 - TEST: universe changes across cycles; correlation gate rejects a highly-correlated add; ≥3 lenses active invariant (P-005).
 ### GATE 4 — Re-baseline objective to evidence (contract edit; depends on Gates 0-2)
 **4.1 — Rewrite objective declaratively (P-011).** Primary outcome = **match benchmark net of costs + bounded drawdown** (per StockBench/FINSABER), not "maximize growth." Re-weight OSS toward downside/drawdown metrics. Make hedge cap a regime-driven lever.
 **4.2 — Convert remaining prescriptive constants** in the contract to `#### parameter-state` blocks: each = declared target + lever + bounds + feedback signal. The system tunes them; the contract stops fixing them.
 **4.3 — Add P-007/008/009/010/011** to the invariant set + watchdog.
 ---
 ## 4. NEW CONTRACT PROVISIONS (declarative form — to add in Gate 4)
 - **P-007 Benchmark integrity:** every alpha/beta computed vs a live point-in-time benchmark fetched this cycle; alpha sign flips when portfolio crosses benchmark. (Kills 0.58 literal.)
 - **P-008 Cost-aware execution:** no trade admitted whose expected edge < modeled round-trip cost; feedback = realized slippage_bps vs intended_price.
 - **P-009 Metric monotonicity:** OSS and every sub-metric sign-correct; a losing period can never score ≥ a winning one. (Property-tested.)
 - **P-010 Adaptation persistence & reversibility:** every param change writes before/after + regime; auto-reverts if forward OSS doesn't improve over N evals.
 - **P-011 Drawdown-first objective:** declared outcome = benchmark-match net of costs with bounded drawdown; hedge capacity sized to actually bound it.
 ---
 ## 5. EXECUTION ORDER & GUARDRAILS
 1. **Gate 0 first, always.** Until OSS sees losses, every adaptation optimizes against a lie.
 2. Human gate before contract edits (Gate 4) per Theo's 2-agent + 1-human policy. Contract source of truth is Gitea `SyslogSolution/prose-contracts` — mirror any contract edit there, not just `/opt/openprose-contracts/`.
 3. Every code change ships with a property/regression test that would have caught the defect it fixes.
 4. Pin any new cron jobs (unpinned LLM crons break on provider switch).
 5. After each gate: run `compute_oss.py` + watchdog; record OSS before/after as evidence.
 6. [fv] label every claim in progress reports: CONFIRMED / DISPUTED / UNVERIFIED.
 ## 6. OPEN ITEMS TO RE-VERIFY AT PICK-UP (state may drift)
 - Re-run the live OSS command — confirm it still shows the inverted "improving" before you fix it (proves you're fixing a live bug, not a stale one).
 - Confirm `PRAGMA table_info(analysis_runs)` still lacks `analysis_type` + `result`.
 - Confirm `.env.prod` hedge caps unchanged (`MAX_HEDGE_POSITIONS`, `HEDGE_POSITION_SIZE_PCT`).
 - Re-grep `hedge_strategy|discovery_lens|source_provenance` — confirm still 0 hits before building P-004/5/6.
 - Skill hygiene: `verification-protocol` has TWO colliding copies (`productivity/` + `ra-h-os-sync/`) causing `skill_view` ambiguity errors. Load by full path or dedupe.
@@ -10,9 +10,8 @@ description: >
  Strix Halo: strix-moe → unsloth/Qwen3.6-35B-A3B-MTP (UD-Q4_K_M, 22GB).
  RTX 5070: gemma-4-12b Q4_K_M → IQ4_NL + MTP draft (122 tok/s, 2x faster).
  UPDATED 2026-07-17: Context reduced fleet-wide from 256K to 128K for stability.
-  Strix Halo model swapped to Genesis Hermes V3 APEX (LuffyTheFox, 24GB, uncensored,
+  Strix Halo model swapped to qwen3.6-35B-udq4 (22GB, strix-moe alias).
-  Hermes agent fine-tune, tensor repair, multimodal with mmproj).
+  Instability observed near 100K at 256K (now all GPUs at 128K). 128K is the stable ceiling.
  Instability observed near 100K at 256K. 128K is the stable ceiling.
  For larger context needs → fall back to external providers (deepseek).
  VRAM headroom improved: RTX 3090 ~70%, RTX 5070 ~65%.
 agent: abiba
@@ -99,7 +98,7 @@ but are deprecated for agent configs. Only the stable aliases survive model swap
 |-------|-----|------|------|-----|----------|----------|-------------|--------|
 | qwen3.6-27B-code (MTP) | RTX 3090 | .8 (llm-gpu) | ~17/24.6GB (70%) | **128K** | turbo4 | 2 | default | ✅ 63 tok/s |
 | gemma-4-12b | RTX 5070 | .110 (ocu-llm) | ~7.8/12.2GB (65%) | 128K | q4_0 | 2 | 2048/1024 | ✅ healthy |
-| Genesis Hermes V3 APEX | Strix Halo Vulkan | .15 (amdpve) | ~10GB/64GB | 128K | q4_0 | 1 | 4096/1024 | ✅ 65 tok/s |
+| qwen3.6-35B-udq4 | Strix Halo Vulkan | .15 (amdpve) | ~22GB/64GB | 128K | q4_0 | 1 | 4096/1024 | ✅ 65 tok/s |
 ## Routing Configuration (LiteLLM — July 2026)
@@ -108,7 +107,7 @@ but are deprecated for agent configs. Only the stable aliases survive model swap
 | Model | GPU | Weight | RPM Cap | Timeout |
 |-------|-----|--------|---------|---------|
 | qwen3.6-27B-code | RTX 3090 (.8:8080) | **0.55** | 500 | **300s** |
-| Genesis Hermes V3 APEX | Strix Halo (.15:8080) | **0.30** | 60 | **300s** |
+| qwen3.6-35B-udq4 | Strix Halo (.15:8080) | **0.30** | 60 | **300s** |
 | gemma-4-12b | RTX 5070 (.110:8080) | **0.15** | 200 | **120s** |
 Note: All syslog-auto entries route directly to GPUs with `api_key: not-needed`. The router (port 9000) is NOT in the inference path.
@@ -117,7 +116,7 @@ Note: All syslog-auto entries route directly to GPUs with `api_key: not-needed`.
 | Model | RPM Cap | Notes |
 |-------|---------|-------|
-| strix-moe (Hermes V3) | 40 | Tight cap — prevents Strix overload |
+| strix-moe (qwen3.6-35B-udq4) | 40 | Tight cap — prevents Strix overload |
 | qwen3.6-27B-code | 500 | High cap — primary workhorse |
 | gemma-4-12b | 500 | High cap — IQ4_NL+MTP, 122 tok/s |
@@ -252,7 +251,7 @@ If no SSH access, send Zulip DM via abiba-bot with vault update instructions.
 - **RTX 3090 config**: `-c 131072 -ctk turbo4 -ctv turbo4 --parallel 2 --flash-attn on --cont-batching --spec-type draft-mtp`. Context reduced to 128K (2026-07-17, was 256K). VRAM: ~70%. Service: `/home/llmuser/llama-wrapper.sh`.
 - **RTX 5070 config (2026-07-15)**: Switched to IQ4_NL + MTP draft (Q8_0) at 128K context. Gen speed: 122 tok/s. VRAM: ~7.8/12.2GB (~65%). Service: `/home/llmuser/llama-wrapper.sh`. Config: `--model gemma-4-12b-it-IQ4_NL.gguf --spec-draft-model gemma-4-12b-it-Q8_0-MTP.gguf --spec-type draft-mtp --spec-draft-n-max 4 --ctx-size 131072`.
 - **LiteLLM timeout tuning (verified 2026-07-16 against `/opt/inference-harness/litellm_config.yaml` on CT 116)**: gemma-4-12b 120s, qwen3.6-27B-code 300s, qwen3.6-35B-udq4 300s, strix-moe 300s, syslog-auto routes all 300s. Nginx proxy_read_timeout: 600s. Global request_timeout: 300s.
- **Strix Halo GPU**: Vulkan is the working backend (ROCm/HIP path abandoned — HSA runtime blocked on Debian 13). Build at `/root/llama.cpp/build-vk/`, commit `4fc4ec5` (2026-07-01), ggml 0.15.3 shared-lib arch. Mesa RADV 25.0.7, KHR_coopmat fast path active. ~70 tok/s gen, 532 tok/s prompt. Service: `strix-server.service` on port 8080, model: `LuffyTheFox/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V3-GGUF` (APEX quant), alias `strix-moe`, 128K context, flash-attn + q4 KV, multimodal (mmproj loaded). Hermes agent fine-tune, tensor repair (SSM layers fixed via SVD), uncensored (0/465 refusals).
+- **Strix Halo GPU**: Vulkan is the working backend (ROCm/HIP path abandoned — HSA runtime blocked on Debian 13). Build at `/root/llama.cpp/build-vk/`, commit `4fc4ec5` (2026-07-01), ggml 0.15.3 shared-lib arch. Mesa RADV 25.0.7, KHR_coopmat fast path active. ~70 tok/s gen, 532 tok/s prompt. Service: `strix-server.service` on port 8080, model: `qwen3.6-35B-udq4`, alias `strix-moe`, 128K context, flash-attn + q4 KV, multimodal (mmproj loaded).
 - **Port conflict detection (2026-07-05)**: All 3 GPU wrappers now detect ghost processes squatting port 8080 before starting. `.8` and `.110` use inline pre-start check in `llama-wrapper.sh`; `.15` uses `/usr/local/bin/port-cleanup.sh` ExecStartPre. Replaces the blanket `pkill -9 -x llama-server` on .15 which would kill ALL llama-server instances regardless of port. Ghost detection was the root cause of .8 crash-looping for 27+ restarts (stale pid 25836 squatting 8080 after OOM kill).
 - **Strix Halo thermal safeguard (2026-07-02)**: `strix-server.service` has `-n 8192` (hard generation cap per request). Without it, `--predict` defaults to -1 (infinity) — a runaway request from .123 (Mumuni) decoded 39,868 tokens over 24 min, pushing Tctl to 98°C (crit 89.8°C) and throttling 70→29 t/s. The cap bounds worst-case generation to ~5 min. Do NOT remove `-n` without a replacement ceiling. Sustained load hits ~84°C even at 92s; the APU is fanless/low-flow. Clients MUST also set `max_tokens`.
 - **Port 8080 firewall**: amdpve iptables restricts 8080 to 192.168.68.116 (LiteLLM/router host) only. All inbound connections are from .116 (LiteLLM proxied via nginx). Localhost curls hang (SYN dropped). Always test from .116.
@@ -268,9 +267,9 @@ If no SSH access, send Zulip DM via abiba-bot with vault update instructions.
 |-----|-------|-----------|--------------|----------|---------|
 | RTX 3090 (.8) | qwen3.6-27B-code (MTP) | **63** | — | — | **128K** |
 | RTX 5070 (.110) | gemma-4-12b (IQ4_NL+MTP) | **191** | — | — | **128K** |
-| Strix Halo (.15) | Genesis Hermes V3 APEX | **65** | 140 | — | **128K** |
+| Strix Halo (.15) | qwen3.6-35B-udq4 | **65** | 140 | — | **128K** |
-Benchmarks from 2026-07-17. Strix Halo swapped to Genesis Hermes V3 APEX (LuffyTheFox). RTX 5070 MTP provides 2.7x speedup over pre-upgrade 70 tok/s.
+Benchmarks from 2026-07-17. Strix Halo model: qwen3.6-35B-udq4. RTX 5070 MTP provides 2.7x speedup over pre-upgrade 70 tok/s.
 All 3 GPUs now at 128K context (2026-07-17, reduced from 256K for stability).
 Benchmarks run through LiteLLM proxy (192.168.68.116:4001) every 5 minutes.
@@ -316,7 +315,7 @@ Mumuni (CT114, 192.168.68.123) is the primary business assistant. This profile i
 | `compression.protect_last_n` | 40 | Preserves last 40 messages |
 | `memory.memory_char_limit` | 800 | Brief memory entries |
 | `personalities` | `creative` | Creative assistant personality |
-| Platforms | cli, discord, homeassistant, signal, telegram, zulip | All Hermes platforms |
+| Platforms | cli, homeassistant, signal, telegram, zulip | All Hermes platforms |
 | Main model timeout | 300s | LiteLLM global timeout |
 | Compression model timeout | 300s | strix-moe timeout increased from 120s |
@@ -47,7 +47,7 @@ poll .15:8080 directly; must go through router on .116.
 | Router (unified) | `http://192.168.68.116/health/unified` | 15s | models, CB, scores, GPU status |
 | Router (basic) | `http://192.168.68.116/health` | 15s | basic aliveness |
 | LiteLLM | `http://192.168.68.116/litellm/health` | 15s | proxy health, model count |
-| Strix Halo | `http://192.168.68.116/health/unified` (router) | 15s | ornith status via router — cannot poll .15:8080 directly (firewalled to .116 only) |
+| Strix Halo | `http://192.168.68.116/health/unified` (router) | 15s | Strix Halo status via router — cannot poll .15:8080 directly (firewalled to .116 only) |
 | Dashboard | `http://192.168.68.116/dashboard/` | 15s | harness-dashboard aliveness |
 ### Alert Delivery
@@ -6,10 +6,15 @@ description: >
  benchmarks, and predicts failures before they happen. Extends gpu-monitor
  (v2.1.0) with active remediation rules, Prometheus metrics consumption,
  VRAM trend analysis, and predictive alerting.
  UPDATED 2026-07-18: Model assignments synced to 2026-07-17 swaps.
  Router (port 9000) references replaced with direct GPU routing.
  Benchmark baselines refreshed to live values.
  Prometheus exporters removed — not deployed; fall back to direct sidecar probes.
  Stable role-based aliases (strix-moe, gpu-dense, gpu-light) from gpu-fleet.
 agent: abiba
 depends_on:
  - gpu-monitor.prose.md (live data source on .24:9100)
-  - gpu-fleet.prose.md (source of truth for topology)
+  - gpu-fleet.prose.md (source of truth for topology, aliases, model assignments)
 ---
 ## Maintains
@@ -22,8 +27,8 @@ depends_on:
 ## Requires
- gpu-monitor:function — Live fleet data from .24:9100/gpu-data
+- gpu-monitor:function — Live fleet data from localhost:9100/gpu-data
- Prometheus exporters on all 3 GPUs (:9400/metrics)
+- Direct sidecar probe access to all GPU hosts (:8080/health)
 - SSH access to GPU hosts for restart operations
 ## Continuity
@@ -35,21 +40,37 @@ depends_on:
 ---
 ## Current Fleet Baseline (2026-07-18)
 | Alias | GPU | Host | Model | VRAM | Ctx | tok/s | Role |
 |-------|-----|------|-------|------|-----|-------|------|
 | `gpu-dense` | RTX 3090 24GB | ct8 (.8:8080) | ThinkingCap Qwen3.6-27B Q4_K_M + MTP + vision | 21.6/24.6GB (88%) | 128K | 74.9 | Heavy reasoning, code gen |
 | `gpu-light` | RTX 5070 12GB | ct110 (.110:8080) | HauhauCS Gemma4-12B QAT Q4_K_M + MTP draft | 10.1/12.2GB (83%) | 128K | 169.6 | Vision, web extract, light tasks |
 | `strix-moe` | Strix Halo 64GB | ct15 (.15:8080) | qwen3.6-35B-udq4 | ~10/64GB (16%) | 128K | 62.9 | Compression, summarization, long docs |
 Key notes:
 - All models use direct GPU routing via LiteLLM (`api_key: not-needed`). Router (port 9000) is deprecated and NOT in the inference path.
 - Stable aliases (gpu-dense, gpu-light, strix-moe) from gpu-fleet are the canonical names for agent configs. Model-specific names still work but are deprecated.
 - RTX 5070 tok/s is 2.3x faster than RTX 3090 for its model — gpu-light is the fastest endpoint. Route vision/web/light work there first.
 - Strix Halo is 62.9 tok/s (89% of 70.5 baseline) — below optimal but stable. Check for competing workloads.
 - RTX 3090 VRAM at 88% — within role-appropriate range (role = heavy reasoning, needs the headroom).
 - RTX 5070 VRAM at 83% — role-appropriate for vision/web (smaller batch sizes).
 ## Remediation Rules
 ### Rule 1: GPU Temperature Critical (>85°C for >2 min)
 - **Detect**: Any GPU temp >85°C sustained for 2+ consecutive polls
 - **Fix**: 
  1. Reduce inference concurrency on that GPU (load-side cooling only — NO fan control)
-  2. Redirect new requests to cooler GPUs via LiteLLM fallback chains
+  2. Redirect new requests to cooler GPUs via LiteLLM fallback chains (gemma → qwen, qwen → gemma)
  3. If all GPUs hot, alert about cooling infrastructure
 - **Verify**: Temp drops below 80°C within 5 minutes
 - **Escalate after**: 3 verification failures → Zulip alert
 ### Rule 2: VRAM Leak Detection (tiered by GPU capacity)
 - **Detect**: VRAM growing at sustained rate over 6+ hour window
-  - RTX 3090 (24GB): ≥100MB/hour
+  - RTX 3090 (24GB): ≥300MB/hour
-  - RTX 5070 (12GB): ≥50MB/hour
+  - RTX 5070 (12GB): ≥300MB/hour
  - Strix Halo (64GB UMA): ≥200MB/hour
 - **Fix**: 
  1. Log VRAM snapshot with process list (nvidia-smi/rocm-smi + ps aux)
@@ -69,6 +90,9 @@ depends_on:
 ### Rule 4: Benchmark Regression (>20% drop)
 - **Detect**: gen_tok_per_sec drops >20% below baseline over 3+ benchmarks
  - RTX 3090 baseline: 74.8 tok/s → alert at <59.8 tok/s
  - RTX 5070 baseline: 165.2 tok/s → alert at <132.2 tok/s
  - Strix Halo baseline: 70.5 tok/s → alert at <56.4 tok/s
 - **Fix**:
  1. Check GPU utilization — if >90%, other process is competing
  2. Check power limit — if throttled, restore to max
@@ -78,32 +102,31 @@ depends_on:
 ### Rule 5: Circuit Breaker Stuck Open
 - **Detect**: Circuit breaker open >10 minutes with GPU reporting healthy
 - **Note**: Router (port 9000) is deprecated. If circuit breakers are reported by gpu-monitor, they come from LiteLLM's internal tracking, not the old router.
 - **Fix**: 
-  1. Verify GPU /health returns 200
+  1. Verify GPU /health returns 200 on direct port (:8080)
-  2. If GPU healthy, send 1 test inference
+  2. If GPU healthy, alert but do NOT reset via router API (deprecated)
-  3. If test succeeds → reset circuit breaker via router API
+  3. Check LiteLLM health directly: http://192.168.68.116/litellm/health/liveliness
-  4. 60s cooldown — if CB re-opens immediately, it was legitimate, do NOT re-reset
+  4. Restart LiteLLM container on CT 116 if circuit breakers are stuck
-  5. Max 1 auto-reset per GPU per hour
+- **Verify**: LiteLLM returns healthy, circuit breaker clears within 60s
- **Verify**: CB closes, inference succeeds, CB stays closed for 60s+
+- **Escalate after**: LiteLLM restart doesn't clear → human investigation
 - **Escalate after**: CB won't close after reset → router issue
 ### Rule 6: Strix Halo Unreachable
 - **Detect**: Strix not responding — probe .15:8080 directly (firewall opened .24→.15)
 - **Fix**: 
  1. SSH to .15 → check llama-server process
  2. Restart llama-server if not running
-  3. Verify through both direct probe AND router
+  3. Verify through both direct probe AND LiteLLM health
- **Verify**: Direct health probe returns 200, router reports Strix healthy
+- **Verify**: Direct health probe returns 200, LiteLLM reports model healthy
 - **Escalate**: If host .15 itself is unreachable → infrastructure alert
-### Rule 7: Prometheus Exporter Down
+### Rule 7: GPU Data Source Unreachable (replaces old Prometheus rule)
- **Detect**: Any GPU :9400/metrics unreachable for >2 polls
+- **Detect**: gpu-monitor endpoint (localhost:9100/gpu-data) or sidecar port (:8080) on any GPU unreachable for >2 polls
 - **Fix**:
-  1. SSH to GPU host → check prometheus-exporter process
+  1. If gpu-monitor is down: restart systemd service `gpu-monitor.service` on this host
-  2. Restart exporter if dead
+  2. If sidecar is down: SSH to GPU host → check llama-server process → restart systemd service
-  3. While exporter is down, fall back to nvidia-smi/rocm-smi direct probes
+  3. Fall back to direct nvidia-smi/rocm-smi probe via SSH if all API paths fail
-  4. If exporter is running but unreachable → check firewall/host networking
+- **Verify**: gpu-monitor returns healthy + all sidecars reachable
 - **Verify**: :9400/metrics returns 200
 - **Escalate after**: 3 failed restarts → networking issue
 ### Rule 8: Predictive Thermal Warning (two-tier)
@@ -117,29 +140,31 @@ depends_on:
 - **Escalate**: If Tier 2 triggers and temp still rising after 5 min → possible hardware failure
 ### Rule 9: Context Window Optimization
- **Detect**: Benchmark tok/s vs baseline for each GPU at current context
+- **Detect**: Benchmark tok/s vs baseline for each GPU at current context (all 128K)
-  - RTX 3090 (128K ctx, qwen3.6-27B-code): target 75+ tok/s — currently at baseline
+  - RTX 3090 (128K ctx, ThinkingCap): baseline 74.8 tok/s — currently at 74.9 (100%)
-  - RTX 5070 (128K ctx, gemma-4-12b): target 76+ tok/s — optimal for vision/web role
+  - RTX 5070 (128K ctx, HauhauCS QAT): baseline 165.2 tok/s — currently at 169.6 (103%)
-  - Strix Halo (128K ctx, strix-moe / qwen3.6-35B-udq4): target 70+ tok/s — currently above baseline
+  - Strix Halo (128K ctx, qwen3.6-35B-udq4): baseline 70.5 tok/s — currently at 62.9 (89%)
 - **Fix**:
  - If tok/s > baseline → context has headroom, consider increasing
  - If tok/s < 90% baseline → reduce context by 25% and retest
  - If tok/s within 10% of baseline → optimal, no change
  - Strix Halo at 89% of baseline → MONITOR but do not reduce yet (recent model swap may still be settling)
 - **Verify**: Re-benchmark after context change, confirm within 10% of target
 - **Escalate**: If context can't be adjusted without significant perf loss
-### Rule 10: Workload Distribution Optimization
+### Rule 10: Workload Distribution Optimization (updated 2026-07-18)
 - **Detect**: GPU roles misaligned with hardware capabilities
 - **Target distribution**:
-  - RTX 3090 (24GB, 128K, 75 tok/s) → Heavy reasoning, code gen, long conversations
+  - RTX 3090 (gpu-dense, 24GB, 74.9 tok/s) → Heavy reasoning, code gen, long conversations (slowest per-token but largest context capacity). Weight: 0.55 (LiteLLM).
-  - RTX 5070 (12GB, 128K, 76 tok/s) → Vision/image, web search, quick lightweight tasks
+  - RTX 5070 (gpu-light, 12GB, 169.6 tok/s) → Vision/image, web search, lightweight tasks (2.3x faster than 3090 per token). Weight: 0.15 (LiteLLM).
-  - Strix Halo (64GB, 128K, 72 tok/s) → Context compression, summarization, long docs
+  - Strix Halo (strix-moe, 64GB, 62.9 tok/s) → Context compression, summarization, long docs (MoE model). Weight: 0.30 (LiteLLM).
 - **Note**: RTX 5070 is the fastest endpoint per token. Route high-volume, low-complexity work there first.
 - **Fix**:
  - Alert if any GPU is handling workload outside its designated role
-  - Recommend Hermes agent profile updates to match workload to GPU
+  - Recommend agent alias updates to match workload to GPU role (use stable aliases: gpu-dense, gpu-light, strix-moe)
  - Track per-GPU request distribution via LiteLLM spend logs
 - **Verify**: Each GPU's request pattern matches its designated role within 24h
- **Escalate**: If role mismatch persists >48h → agent profile audit needed
+- **Escalate**: If role mismatch persists >48h → agent alias audit needed
 ---
@@ -148,7 +173,7 @@ depends_on:
 ```prose
 -- Phase 1: Fetch live GPU data
 let fleet = call gpu-monitor
-  endpoint: "http://192.168.68.24:9100/gpu-data"
+  endpoint: "http://localhost:9100/gpu-data"
 -- Phase 2: Evaluate each GPU against remediation rules
 let actions = []
@@ -164,27 +189,25 @@ for gpu in fleet.gpus:
  -- Rule 4: Benchmark regression
  let bench = fleet.benchmarks[gpu.hostname]
-  if bench.current_tok_sec < bench.baseline_tok_sec * 0.8:
+  if bench.current_tok_s < bench.baseline_tok_s * 0.8:
    push actions apply-benchmark-fix(gpu, bench)
 -- Rule 3: Model stuck
-for model in fleet.router.available_models:
+for model in fleet.summary.available_models:
  if model.consecutive_timeouts >= 3:
    push actions apply-model-restart(model)
-- Rule 5: Circuit breaker
+-- Rule 5: Circuit breaker check via LiteLLM (router deprecated)
-for cb in fleet.router.circuit_breaker:
+if fleet.summary.circuit_breakers_open > 0:
-  if cb.open and cb.open_duration > 600 and gpu_is_healthy(cb.gpu):
+  push actions check-litellm-circuit-breakers()
    push actions apply-cb-reset(cb)
 -- Rule 6: Strix Halo
 if not fleet.strix.running and pingable("192.168.68.15"):
  push actions apply-strix-restart()
-- Rule 7: Prometheus exporters
+-- Rule 7: GPU data source
-for gpu in fleet.gpus:
+if not fleet.gpus or len(fleet.gpus) < 2:
-  if not prometheus_reachable(gpu.hostname, 9400):
+  push actions check-gpu-monitor-service()
    push actions apply-exporter-restart(gpu)
 -- Rule 8: Predictive thermal
 for gpu in fleet.gpus:
@@ -212,12 +235,12 @@ call update-gpu-health
 ```json
 {
-  "run_id": "gpu-self-heal-20260712-001",
+  "run_id": "gpu-self-heal-20260718-001",
-  "timestamp": "2026-07-12T16:00:00Z",
+  "timestamp": "2026-07-18T08:00:00Z",
  "gpu": "ct8-rtx3090",
  "issue": "thermal-critical",
  "detected": { "temp_c": 87, "duration_s": 180 },
-  "action": "set-fan-100pct",
+  "action": "load-shedding",
  "result": "resolved",
  "verification": { "temp_c": 76, "after_s": 300 },
  "escalated": false
@@ -236,52 +259,53 @@ Every action logged as `[GPU-SELF-HEAL] <run_id>` node with full audit trail.
 - `issues_escalated > 0` → "⚠ GPU Self-Heal — <gpu> needs attention"
 - Every 100th clean cycle → "✅ GPU Fleet: All Clear"
-### 3. Prometheus/Grafana Integration
+### 3. Weekly Benchmark Report
 - GPU self-heal actions exposed as Prometheus counter metrics
 - Dashboard panel: "GPU Interventions (24h)" showing count/type/result
 ### 4. Weekly Benchmark Report
 - Per-GPU tok/s trend over 7 days
 - Regression alerts if any GPU degrades >10% week-over-week
 ---
-## Design Decisions (Grilled & Confirmed — 2026-07-12)
+## Design Decisions (Verified 2026-07-12, Reaffirmed 2026-07-18)
 1. **Fan control**: ❌ NO auto fan control. Load-side cooling only (reduce concurrency, redirect).
 2. **Model restart**: ✅ Only if >50% failure rate over 60s + 30s grace period. Not on single stuck request.
 3. **Strix direct access**: ✅ Open firewall .15:8080 → .24 for direct health probe + restart.
-4. **VRAM thresholds**: Tiered — 100MB/h (RTX 3090), 50MB/h (RTX 5070), 200MB/h (Strix).
+4. **VRAM thresholds**: Tiered — **300MB/h** (RTX 3090), **300MB/h** (RTX 5070), 200MB/h (Strix). Previous values (100/50) were too sensitive; raised 2026-07-18 based on operational data.
-5. **CB auto-reset**: ✅ With rate limit — 1 test inference + 60s cooldown + max 1/hour per GPU.
+5. **CB auto-reset**: ✅ Router deprecated — circuit breakers go through LiteLLM health check + container restart if needed. No per-GPU auto-reset.
-6. **Benchmark baseline**: Rolling 30-day average, recalculated weekly. Original baseline kept in Grafana.
+6. **Benchmark baseline**: Rolling 30-day average, recalculated weekly. Current baselines live in gpu-monitor.
 7. **Predictive alerts**: Two-tier — warn at >70°C+rising (>2°C/min), critical at >80°C+rising.
-8. **Prometheus**: Primary source. Fall back to nvidia-smi/rocm-smi direct probes if exporter down.
+8. **Prometheus**: ❌ Not deployed. Use direct sidecar probes (:8080/health) and gpu-monitor API. Prometheus integration deferred until exporters are running on GPU hosts.
-## Lessons Learned (2026-07-12)
+## Lessons Learned (2026-07-12, Updated 2026-07-18)
 ### L1: API Key Standardization Is Critical
 - All GPU llama-servers MUST use the same api-key as the LiteLLM config.
 - RTX 5070 had `--api-key sk-loc...5678` while LiteLLM sent `not-needed`.
-  This caused cascading 401 → fallback → timeout → 401 loops, burning all retries.
+  This caused cascading 401 → fallback → timeout → 401 loops.
- **Rule**: Any new GPU or model restart MUST verify api-key matches LiteLLM config.
+- **Rule**: Any new GPU or model restart MUST verify api-key matches LiteLLM config (`not-needed` for direct routing).
 ### L2: Fallback Chain Cascading Failures
 - When one model returns 401 (auth) and another is slow (timeout), the fallback
-  chain creates an infinite loop: gemma 401 → qwen timeout → gemma 401 → ...
+  chain creates an infinite loop.
 - **Rule**: If a model returns 401 (auth error), do NOT fall back to it again.
  Mark it as permanently failed for this request.
 ### L3: Verify Running State, Not Docs
- RTX 3090 was documented at 128K context. Actually running at 256K.
+- RTX 3090 was documented at 128K context. Running at 128K (verified 2026-07-18).
- Parallel count wrong (docs said 2, actual is 1 on RTX 3090).
+- Parallel count: 1 on both RTX 3090 and RTX 5070 (matches docs for current models).
 - **Rule**: Before making decisions, check `/proc/PID/cmdline` on GPU hosts.
 ### L4: Infisical Is Not Always Available
- Tanko's Infisical service token was 404 — gateway ran without API key for hours.
+- Keep a local `.env` fallback for `LITELLM_API_KEY`.
- **Rule**: Always keep a local `.env` fallback for `LITELLM_API_KEY`.
+- **Rule**: Always verify credential source is reachable before relying on it.
 - Contract hermes-config-template Rule 3 updated.
-### L5: Zulip Event Queue Can Silently Die
+### L5: GPU Monitor Response Size Can Cause Self-Heal Crash
- Mumuni's queue accumulated 41 errors/reconnects then stopped polling.
+- gpu-self-heal crashed with KeyboardInterrupt during json.loads() of 20MB response.
-  Gateway was running but ignoring all messages.
+- Root cause: router poll returns accumulated data → cache balloons.
- **Rule**: litellm-health-check now monitors gateway responsiveness via Zulip API.
+- **Rule**: Self-heal must enforce a read timeout AND max response size on every poll.
  If monitor response > 1MB, log a warning and skip the cycle rather than crashing.
 ### L6: Stable Aliases Replace Model Names
 - gpu-fleet introduced stable aliases (strix-moe, gpu-dense, gpu-light) on 2026-07-15.
 - Self-heal must use aliases for reporting and alerting, not model-specific names.
 - **Rule**: All alert messages and KG nodes use the stable alias as the GPU identifier.
@@ -5,7 +5,7 @@ version: 1.0.0
 description: >
  Canonical known-good baseline for all Syslog Hermes agents. Captures the exact
  configuration state, keys, workarounds, and audit procedure. When an agent's
-  configuration goes sideways, restore from this baseline. Last verified 2026-07-16. All GPUs 256K context (RTX 3090 .8, RTX 5070 .110, Strix Halo .15). Parallel 1 fleet-wide (Strix Halo handles compression solo).
+  configuration goes sideways, restore from this baseline. Last verified 2026-07-16. All GPUs 128K context (reduced from 256K for stability Jul 2026) (RTX 3090 .8, RTX 5070 .110, Strix Halo .15). Parallel 1 fleet-wide (Strix Halo handles compression solo).
 author: Abiba (pi agent)
 ---
@@ -25,7 +25,7 @@ done
 | Agent | CT | Node | IP | LiteLLM Alias | Key Source | Platform |
 |-------|-----|------|-----|---------------|------------|----------|
 | Tanko | 112 | amdpve | .122 | `tanko` | Infisical vault | Hermes |
-| Mumuni | 114 | minipve | .123 | `mumuni` | Infisical vault | Hermes |
+| Mumuni | 114 | hwepve | .123 | `mumuni` | Infisical vault | Hermes |
 | Koby | 129 | amdpve | srv1079750 | `koby` | Infisical vault | **Hermes** |
 | Koonimo | 114 | amdpve | ? | `koonimo` | Infisical vault | Hermes |
 | Shumba | — | 192.168.68.119 | N/A | N/A (DeepSeek) | Hermes (RETIRED — CT119 now Infisical vault) |
@@ -193,6 +193,8 @@ Key is injected via `infisical run --` wrapper at PM2 startup:
      "models": [
        { "id": "syslog-auto" },
        { "id": "strix-moe" },
        { "id": "gpu-dense" },
        { "id": "gpu-light" },
        { "id": "qwen3.6-27B-code" },
        { "id": "gemma-4-12b" }
      ]
@@ -5,11 +5,14 @@ description: >
  Standard Hermes configuration template for Syslog Solution LLC agents.
  Enforces shared infrastructure setup (Firecrawl, SearXNG, local models,
  RA-H OS MCP) while keeping agent-specific API keys and model choices.
-  UPDATED 2026-07-16: Compression model is the stable alias `strix-moe` (NOT `ornith-1.0-35b`,
+  UPDATED 2026-07-18: Compression model switched to `syslog-auto` (was `strix-moe`)
  to relieve Strix Halo pressure. syslog-auto distributes compression across the
  weighted pool (55% RTX 3090, 30% Strix Halo, 15% RTX 5070).
  UPDATED 2026-07-16: Compression model was the stable alias `strix-moe` (NOT `ornith-1.0-35b`,
  which LiteLLM does not serve). All 3 GPUs verified at 128K (reduced from 256K 2026-07-17 for stability).
  Added Rule 12 (Context-Issue Diagnostic) + Rule 13 (.env fallback enforcement) from the
  2026-07-16 Mumuni root-cause investigation (WAL #1300).
-  UPDATED 2026-07-12: GPU workload redistributed. Compression → Strix Halo. RTX 3090 context verified at 128K. Infisical .env fallback required (Rule 3/13).
+  UPDATED 2026-07-12: GPU workload redistributed. Compression → Strix Halo (later switched to syslog-auto 2026-07-18). RTX 3090 context verified at 128K. Infisical .env fallback required (Rule 3/13).
 ---
 ## Maintains
@@ -130,7 +133,9 @@ mcp_servers:
 # ─── Compression ───
 compression:
  enabled: true
-  model: strix-moe               # ⚠️ Must match auxiliary.compression.model. Stable alias (gpu-fleet § Stable Role-Based Aliases). NOT ornith-1.0-35b (LiteLLM does not serve that name).
+  model: syslog-auto              # ⚠️ Switched from strix-moe 2026-07-18 to relieve Strix Halo.
                                  # syslog-auto distributes across weighted pool (55% RTX 3090,
                                  # 30% Strix Halo, 15% RTX 5070). All GPUs at 128K.
  provider: harness
  max_context_window: 131072      # MUST match actual GPU capacity. All 3 GPUs are 128K (Jul 17).
  threshold: 0.65                 # Fires at ~170K for 262K window, ~85K for 128K
@@ -145,8 +150,9 @@ compression:
 #   model: gpu-light                 # stable alias (NOT raw "gemma-4-12b")
 #   base_url: http://192.168.68.116/v1
 #   api_key_env: LITELLM_API_KEY
-# Do NOT use syslog-auto for auxiliary tasks — it routes to the primary GPU.
+# Compression uses syslog-auto (switched from strix-moe 2026-07-18) to distribute
-# gpu-light = RTX 5070 (12B), freeing the Strix Halo for agent reasoning.
+# load across the weighted pool and relieve Strix Halo pressure.
 # Vision and web_extract use gpu-light = RTX 5070 (12B).
 # Heavy aux (delegation, x_search) use gpu-dense (RTX 3090) instead.
 # NEVER use raw model names (gemma-4-12b, qwen3.6-27B-code, qwen3.6-35B-udq4)
 # in agent configs — use the stable aliases so model swaps don't break agents.
@@ -166,7 +172,7 @@ auxiliary:
    timeout: 30
  compression:
    provider: harness
-    model: strix-moe              # MUST match compression.model above. Stable alias for Strix Halo.
+    model: syslog-auto             # Switched from strix-moe 2026-07-18. Relieves Strix Halo pressure.
    base_url: http://192.168.68.116/v1   # Rule 5: /v1 NOT /litellm/v1
    api_key_env: LITELLM_API_KEY
    timeout: 300                 # gpu-fleet: 300s for large-history summarization (was 60)
@@ -247,29 +253,30 @@ The following MUST be identical across ALL profiles:
 - Apply to BOTH main config AND all sub-agent profiles
 - For agents needing longer outputs: raise to 8192, but never omit
-### Rule 7: Auxiliary Model Consistency (UPDATED 2026-07-16)
+### Rule 7: Auxiliary Model Consistency (UPDATED 2026-07-18)
- Vision and web_extract use `gemma-4-12b` (RTX 5070 — 12GB, vision-optimized)
+- Vision and web_extract use `gpu-light` (stable alias, RTX 5070 — 12GB, vision-optimized)
- Compression uses `strix-moe` (stable alias for Strix Halo — 64GB, 128K ctx, compression-optimized)
+- Compression now uses `syslog-auto` (switched from `strix-moe` 2026-07-18) to distribute
- **`strix-moe` is the only valid compression model name** — LiteLLM does NOT serve `ornith-1.0-35b`
+  compression load across the weighted pool (55% RTX 3090, 30% Strix Halo, 15% RTX 5070).
-  (it serves `strix-moe`, `qwen3.6-35B-udq4`, `gpu-dense`, `gpu-light`, `syslog-auto`, `gemma-4-12b`, `qwen3.6-27B-code`). Old configs with `ornith-1.0-35b` cause 403/model-not-found on compression calls.
+  This relieves Strix Halo pressure while keeping compression functional on all GPUs.
 - **`syslog-auto` is the valid compression model** — LiteLLM serves it as the weighted pool.
  Old configs with `strix-moe` for compression should be updated to `syslog-auto`.
 - All auxiliary services MUST use identical routing:
  - `base_url: http://192.168.68.116/v1`  (Rule 5: `/v1`, NOT `/litellm/v1`)
  - `api_key_env: LITELLM_API_KEY`
- **Do NOT use `syslog-auto`** for auxiliary tasks — it routes unpredictably
+- **Compression via syslog-auto**: Routes through the weighted pool. Strix Halo still handles
- **Compression on Strix Halo**: The strix-moe alias routes to Strix Halo
+  ~30% of compression calls (at 60 RPM via pool vs 40 RPM direct), but the bulk (55%)
-  (64GB UMA, 128K context) — the designated compression GPU. This frees the
+  goes to RTX 3090 which has ample spare capacity.
  RTX 5070 for vision and web search, and the RTX 3090 for heavy reasoning.
 - The `compression:` block's `model` MUST match `auxiliary: compression: model`
 - The `compression: max_context_window: 131072` MUST match actual GPU capacity (128K)
-### Rule 8: GPU Workload Distribution (UPDATED 2026-07-16)
+### Rule 8: GPU Workload Distribution (UPDATED 2026-07-18)
- **RTX 3090 (24GB, 128K ctx, qwen3.6-27B-code)**: Heavy reasoning, code gen, long conversations
+- **RTX 3090 (24GB, 128K ctx, qwen3.6-27B-code)**: Heavy reasoning, code gen, long conversations — also handles ~55% of compression via syslog-auto pool
- **RTX 5070 (12GB, 128K ctx, gemma-4-12b)**: Vision, web search, quick tasks, web_extract (IQ4_NL+MTP, ~65% VRAM at 128K)
+- **RTX 5070 (12GB, 128K ctx, gemma-4-12b)**: Vision, web search, quick tasks, web_extract — handles ~15% of compression via syslog-auto pool
- **Strix Halo (64GB, 128K ctx, strix-moe)**: Context compression, summarization, long docs
+- **Strix Halo (64GB, 128K ctx, qwen3.6-35B-udq4)**: Agent reasoning, compression (~30% via syslog-auto pool), fallback for other GPUs
 - Agent profiles MUST route auxiliary tasks to the correct GPU:
-  - `auxiliary.vision.model: gemma-4-12b` (RTX 5070)
+  - `auxiliary.vision.model: gpu-light` (RTX 5070)
-  - `auxiliary.web_extract.model: gemma-4-12b` (RTX 5070)
+  - `auxiliary.web_extract.model: gpu-light` (RTX 5070)
-  - `auxiliary.compression.model: strix-moe` (Strix Halo)
+  - `auxiliary.compression.model: syslog-auto` (distributed pool, switched from strix-moe 2026-07-18)
 - Default model (`model.default`) and custom_provider remain `syslog-auto` for auto-routing
 - For 128K context window: `threshold: 0.65` (fires at ~85K tokens)
 - Do NOT use `threshold: 0.25` — this fires at 65K, causing premature context loss
@@ -325,7 +332,8 @@ verify ALL FOUR of these against the live config. They are the only root causes
 One-line agent health check (run on the agent host):
 ```bash
-PID=$(pgrep -f "python -m hermes_cli.main gateway run" | head -1)
+# Use grep -v infisical to avoid matching the bash wrapper that contains the same string
 PID=$(pgrep -f "python -m hermes_cli.main gateway run" | grep -v infisical | head -1)
 cat /proc/$PID/environ | tr '\0' '\n' | grep ^LITELLM_API_KEY= | sed 's/=.*/<set>/'
 curl -s -o /dev/null -w 'key_health: %{http_code}\n' -H "Authorization: Bearer $(cat /proc/$PID/environ | tr '\0' '\n' | grep ^LITELLM_API_KEY= | cut -d= -f2)" http://192.168.68.116/v1/models
 ```
@@ -351,9 +359,19 @@ directly (no infisical). Apply with `systemctl daemon-reload && systemctl restar
 The wrapper sources `~/.hermes/.env` then exports `LITELLM_API_KEY="$<AGENT>_LITELLM_API_KEY"`.
 See litellm-api-keys.prose.md § Machine Identity for Vault Writes for vault sync.
 **⚠️ Vault empty-key guard:** If the vault stores the secret as an empty string,
 the wrapper will inject an empty key and the gateway will silently get 401 errors
 on all LiteLLM requests (triggering silent DeepSeek fallback). The `.env` fallback
 is present but the vault takes precedence when the secret key exists (even if empty).
 **Fix:** The wrapper MUST validate the key length after injection. If LITELLM_API_KEY
 is empty or shorter than 20 chars, log a warning and either fail with a clear error
 message or fall back to the `.env` value before starting the gateway.
 **Verification (all agents):**
 ```bash
-GP=$(pgrep -f "python -m hermes_cli.main gateway run" | head -1)
+# Use grep -v infisical to avoid matching the bash wrapper that contains the same string
 GP=$(pgrep -f "python -m hermes_cli.main gateway run" | grep -v infisical | head -1)
 K=$(cat /proc/$GP/environ | tr '\0' '\n' | grep '^LITELLM_API_KEY=' | cut -d= -f2)
 curl -s -o /dev/null -w '%{http_code}' -H "Authorization: Bearer $K" http://192.168.68.116/v1/models  # must be 200
 ```
@@ -51,7 +51,7 @@ gateway restart, and connection validation.
 | Host | CT | Proxmox | IP (direct) | Hermes Home | User |
 |------|-----|---------|-------------|-------------|------|
-| Mumuni | CT114 | — | 192.168.68.123 | /root/.hermes | root |
+| Mumuni | CT114 | hwepve | 192.168.68.123 | /root/.hermes | root |
 | Tanko | CT112 | amdpve | 192.168.68.122 | /home/jerome/.hermes | jerome |
 | Koby | CT111 | amdpve | 192.168.68.129 | /root/.hermes | root |
 | Shumba | — | — | 192.168.68.119 | /home/lucky/.hermes | lucky |
@@ -22,14 +22,14 @@ management, and prompt caching — without sacrificing agent capability.
 - `agent-configs`: current config.yaml from each active Hermes agent (Mumuni
  .123, any others on .129/.122) including compression, model, context_window,
  prompt_caching, memory settings
- `gpu-health`: health check response from all 3 GPU backends (ornith .15:8080,
+- `gpu-health`: health check response from all 3 GPU backends (strix-moe .15:8080,
  qwen .8:8080, gemma .110:8080)
 ### Maintains
 The optimized inference stack configuration — every change is applied and
 verified end-to-end. Postcondition: avg request_duration_ms ≤ 15000 for 90% of
-non-ornith traffic; ≤ 30000 for ornith-bound agentic calls.
+inference calls.
 #### liteLLM-routing
 The syslog-auto routing weights, model-specific timeouts, RPM limits, and
@@ -53,18 +53,17 @@ duration.
 ### Strategies
-**Context is the root cause.** Every ~46K prompt token costs ~87s of ornith
+**Context is the root cause.** Every ~46K prompt token costs ~87s of
 prefill time at 532 tok/s. Fix context first, routing second.
- **Route by task**: ornith for multi-step reasoning only; qwen for code/standard
+- **Route by task**: qwen for code/standard queries; gemma for
-  queries; gemma for compression/auxiliary. Never send simple completion to a
+  compression/auxiliary; strix-moe for compression tasks.
-  35B MoE.
+- **Compress aggressively**: threshold at 40% (not 65%) — a 128K window should
- **Compress aggressively**: threshold at 40% (not 65%) — a 256K window should
+  compact at 51K, not 85K. Target 15% tail (not 30%).
  compact at 102K, not 166K. Target 15% tail (not 30%).
 - **Cache everything repeated**: system prompts, skill docs, AGENTS.md — these
  never change between turns. Single-digit cache hit rate is unacceptable.
 - **Lower context ceiling**: 128K window is the stable ceiling for agent conversations.
-  GPUs reduced from 256K to 128K (2026-07-17). For larger contexts, route to external providers.
+  GPUs reduced from 256K to 128K (2026-07-17). 128K window should compact at 85K (0.65 threshold). For larger contexts, route to external providers.
 ### Shape
@@ -100,5 +99,5 @@ call enable-prompt-caching
 call verify-latency
  host: 192.168.68.116
-  models: [syslog-auto, qwen3.6-27B-code, gemma-4-12b, ornith-1.0-35b]
+  models: [syslog-auto, qwen3.6-27B-code, gemma-4-12b]
 ```
@@ -199,7 +199,7 @@ description: >
 **Prometheus targets**:
 - 192.168.68.8:9400 (RTX 3090 — qwen)
 - 192.168.68.110:9400 (RTX 5070 — gemma)
- 192.168.68.15:9400 (Strix Halo — ornith)
+- 192.168.68.15:9400 (Strix Halo — qwen3.6-35B-udq4)
 - 192.168.68.24:9401 (Router metrics exporter)
 - harness-litellm:4000 (LiteLLM health)
@@ -0,0 +1,200 @@
 ---
 kind: responsibility
 name: infrastructure-maintenance
 description: >
  Weekly system-level maintenance for the Syslog inference fleet: OS package
  updates on the primary host, Docker image pulls for LiteLLM/SearXNG and other
  running containers, container restarts with health verification, post-update
  verification of every critical service (LiteLLM proxy, SearXNG, Zulip, Gitea,
  PM2 processes, Hermes gateways), and rollback on failure. Consolidates the
  raw shell scripts that previously did this piecemeal. This contract owns the
  HOST-LEVEL weekly maintenance loop on the primary host plus Docker image
  pulls ONLY for .116 and .7, while infrastructure-update owns the FULL-FLEET
  cluster-wide wave (apt across the full PVE cluster + CTs/VMs AND its Docker
  image Wave 3 across all stacks). Runs Sunday 2am ET. Owner:
  ops (firstmate secondmate). Blast radius: an unverified image pull can break
  LiteLLM (all agents lose inference) or SearXNG (search-stack down); a bad apt
  upgrade can leave the host in a half-upgraded state. Pre-update backup check
  and rollback are mandatory for this reason.
 agent: ops
 triggers:
  - weekly (Sunday 02:00 ET) via cron
  - on demand when ops/abiba triggers "infra maintenance"
 version: 1.0.0
 ---
 ## Maintains
 - maintenance-status: { phase: idle|preflight|apt|images|restarts|verify|rollback|done|failed, host, step, result, timestamp }
 - image-baseline: { service, current_tag, pulled_tag, digest, updated_at } — last known-good image per container
 - apt-state: { upgradable_before, upgradable_after, held_broken, kernel_reboot_required }
 - health-baseline: snapshot of critical-service health captured pre-update (used for regression check post-update)
 - rollback-snapshot: { backup_path, configs, image_digests, timestamp } — restore point created in preflight
 - maintenance-history: array of past runs with phase results and any escalations
 ## Scope
 Primary host is the maintenance host where apt updates apply. Docker image pulls
 span the two Docker ecosystems that run critical services. infrastructure-update
 runs the full-fleet cluster-wide wave (including its Wave 3 Docker pulls across
 all stacks/hosts); this contract runs a narrower host-level weekly pull limited
 to .116 and .7. Topology, CT IDs, and IPs are live-state fields — verify against
 `infrastructure-control.prose.md` (the source of truth) and the live system
 before mutating.
 | Host | IP | Role | Trust |
 |------|----|------|-------|
 | CT 116 (syslog-api) | 192.168.68.116 | LiteLLM proxy + Grafana + Prometheus (inference harness) | ⚠️ VERIFY-BEFORE-USE |
 | VM 109 (docker-vm) | 192.168.68.7 | SearXNG + Firecrawl + home stack (Docker host) | ⚠️ VERIFY-BEFORE-USE |
 | CT 117 (zulip) | 192.168.68.19 | Zulip (storepve bridge IP .19) | ⚠️ VERIFY-BEFORE-USE |
 | Gitea | https://git.sysloggh.net | Prose-contracts + agent configs source control | ⚠️ VERIFY-BEFORE-USE |
 | CT 100 (abiba/pi) | 192.168.68.24 | PM2 processes (pi agent harness) | ⚠️ VERIFY-BEFORE-USE |
 > "Primary host" for the apt phase is the host the ops agent runs maintenance
 > from. Confirm which host that is against infrastructure-control before
 > running; do not assume. If the ops agent is containerized/CT-based, apt runs
 > inside that CT.
 ## Requires
 - SSH/exec access to CT 116 (.116) and VM 109 (.7) for Docker operations
 - `apt`, `docker`, `docker compose` available on target hosts
 - LiteLLM master key available (Infisical vault, `LITELLM_API_KEY`) for health verification
 - `infrastructure-monitoring` run completed within the last 30 minutes — provides the pre-update health baseline used by the regression check
 - Writable backup directory `/tmp/infra-maintenance-backup-<date>/` on each mutated host
 - Proxmox snapshot of the primary host available (or confirmed not required) before apt phase
 ## Continuity
 - Self-driven: weekly cron `0 2 * * 0` (Sunday 02:00 ET)
 - Also wakes on: explicit "infra maintenance" trigger from ops/abiba
 - Depends on `infrastructure-monitoring` for the pre-update health baseline — do not run if the last monitoring run is stale (>30 min) or RED; abort and escalate instead
 ## Execution
 ### Phase 0 — Preflight (snapshot/backup check + health baseline)
 1. **Capture health baseline** — run the `infrastructure-monitoring` postcondition checks (LiteLLM, Zulip, Gitea, SearXNG, Proxmox API) and record results as `health-baseline`. If any critical service is already down, **abort**: maintenance must not run on a degraded fleet.
 2. **Backup check** — confirm a Proxmox snapshot of the primary host exists OR `/tmp/infra-maintenance-backup-<date>/` was created this run. Snapshot critical config files into the backup dir:
   - `/opt/inference-harness/docker-compose.yml`, `/opt/inference-harness/litellm_config.yaml` (CT 116)
   - `/opt/search-stack/searxng/docker-compose.yml`, `/opt/search-stack/firecrawl-source/docker-compose.yaml` (VM 109)
 3. **Record image baseline** — `docker inspect --format '{{.Image}} {{.Config.Image}}' <container>` for every running container on .116 and .7; store digests in `image-baseline` so rollback can restore them.
 4. **Disk check** — `df -h` on each mutated host; abort if free space <20% (apt upgrade + image pulls need headroom).
 ### Phase 1 — OS package updates (primary host)
 ```bash
 # On the primary host only (VERIFY host against infrastructure-control first)
 apt update
 apt upgrade -y
 ```
 - Capture `apt list --upgradable` before and after → store in `apt-state`.
 - If apt reports held/broken packages (`apt-get -s upgrade | grep -i broken`, or non-zero exit), **stop** — do not force. Record `held_broken` and go to rollback/escalate.
 - If `/var/run/reboot-required` exists after upgrade, flag `kernel_reboot_required: true` in `apt-state` but **do not reboot automatically** — that's a separate coordinated action (see infra-update Wave 4). Note it in the report.
 ### Phase 2 — Docker image pulls
 Pull latest stable tags for every running container. Do NOT pin to `:main`/`:nightly` — use stable tags where the compose file specifies them; otherwise `latest`.
 ```bash
 # CT 116 (.116) — inference harness
 cd /opt/inference-harness && docker compose pull
 # VM 109 (.7) — search + home stacks
 cd /opt/search-stack/searxng && docker compose pull
 cd /opt/search-stack/firecrawl-source && docker compose pull
 # any other running stacks on .7 (home stack, audiobookshelf) — pull per their compose files
 ```
 - LiteLLM and SearXNG are the two explicitly required pulls; "any other running containers" means every stack with a compose file on .116 and .7.
 - Record pulled tag + digest per service in `image-baseline`.
 ### Phase 3 — Container restarts with health verification
 Restart one stack at a time, verify health before moving to the next. Do not restart everything at once — a failure mid-wave must leave the rest running.
 ```bash
 # CT 116
 cd /opt/inference-harness && docker compose up -d
 # VM 109
 cd /opt/search-stack/searxng && docker compose up -d
 cd /opt/search-stack/firecrawl-source && docker compose up -d
 ```
 After each stack comes up, wait for health (max 120s):
 - `docker ps` shows the container `Up` (and `healthy` if a healthcheck is defined)
 - Service-specific probe passes (see Phase 4 probes)
 If a stack fails to come up within 120s, **stop the wave** and go to rollback for that stack only; do not proceed to the next.
 ### Phase 4 — Post-update service verification
 After ALL updates (apt + images + restarts), verify every critical service is back up and matches the pre-update baseline. This is the regression gate.
 | Service | Probe | Expect |
 |---------|-------|--------|
 | LiteLLM proxy | `curl -sf http://192.168.68.116/litellm/v1/models` | 200 OK, models returned |
 | LiteLLM MCP gateway | `curl -sf http://192.168.68.116:4000/mcp-rest/tools/list -H "Authorization: Bearer $MASTER_KEY"` | 90 tools (23 RA-H OS + 67 GitHub) |
 | SearXNG | `curl -sf http://192.168.68.7:8888` | 200 OK |
 | Zulip | `curl -sf https://chat.sysloggh.net/api/v1/server_settings` | 200 OK |
 | Gitea | `curl -sf https://git.sysloggh.net/api/v1/version` | 200 OK |
 | PM2 processes | `pm2 jlist` (CT 100) | all pi-agent processes `online` |
 | Hermes gateways | SSH to Mumuni CT 114, Tanko CT 112; `systemctl is-active hermes-gateway` | `active` for each |
 Regression check: every service that was GREEN in `health-baseline` must still be GREEN. A service that was already RED (and caused a preflight abort) is excluded — but Phase 0 should have aborted before we got here.
 ## Rollback Protocol
 If ANY service in Phase 4 fails to come back up (or regresses vs baseline):
 1. **Image rollback** — for the failing stack, restore the previous image:
   ```bash
   # Restore from recorded image-baseline digest
   docker compose down
   # Pin the service image to the recorded digest in compose, then recreate
   #   image: <name>@sha256:<previous_digest>
   docker compose pull && docker compose up -d
   ```
 2. **APT rollback** — restore the primary host from the Proxmox snapshot taken/confirmed in Phase 0. If no snapshot, `apt install <pkg>=<old_version>` per package using apt history (`/var/log/apt/history.log`).
 3. **Config rollback** — restore configs from `/tmp/infra-maintenance-backup-<date>/`.
 4. **Re-verify** — re-run the Phase 4 probes on the rolled-back service. If still failing, escalate (do not loop — circuit breaker below).
 5. **Escalate** — send a Zulip DM to abiba + mumuni with: failing service, phase, baseline vs current, rollback actions taken, backup path.
 ## Circuit Breaker
 - `max_retries: 2` per failing phase — after 2 rollback attempts on the same service, stop and escalate.
 - `window: 7200` seconds — no more than 2 retries within a 2-hour window.
 - `trip_action: escalate_to_fatal` — when tripped, escalate to fatal (abiba + mumuni + kwame) and pause; a human must clear before the next scheduled run.
 ## Report
 After completion (or on abort), emit a receipt (JSON) to `~/.hermes/runs/infrastructure-maintenance/` and send a Zulip DM summary:
 ```
 🛠 Infrastructure Maintenance — YYYY-MM-DD
 Phase: apt | images | restarts | verify | rollback
 Primary host: <host>
 APT: <N> packages upgraded, <M> held/broken, kernel_reboot_required=<bool>
 Images pulled: LiteLLM <tag>, SearXNG <tag>, <others>
 Services: all GREEN | <service> FAILED (rolled back)
 Baseline regression: none | <details>
 Backup: /tmp/infra-maintenance-backup-YYYYMMDD/
 Escalation: none | warning | critical | fatal
 ```
 ## Verification Postconditions
 - All critical services running after update (Phase 4 all GREEN)
 - No regressions from pre-update health baseline (Phase 0 baseline)
 - Docker containers on latest stable tags (`image-baseline.pulled_tag` recorded)
 - APT packages up to date with no held broken packages (`apt-state.held_broken == 0`)
 ## Related Contracts
 - `infrastructure-update.prose.md` — owns the full-fleet cluster-wide wave INCLUDING its Wave 3 Docker image updates across all stacks (SearXNG, Firecrawl, Inference Harness on .116, home stack, audiobookshelf); infrastructure-maintenance is a deliberately narrower host-level weekly pull scoped to .116 and .7.
 - `infrastructure-monitoring.prose.md` — provides the pre-update health baseline (depends_on).
 - `infrastructure-control.prose.md` — topology source of truth (CT IDs, IPs, hostnames).
 - `litellm-health.prose.md` — LiteLLM probe details.
 - `proxmox-monitor.prose.md` — Docker stats + monitoring stack health.
@@ -67,7 +67,7 @@ Before ANY update wave:
 - All VMs/CTs running: check via Proxmox API
 - LiteLLM healthy: `curl localhost:4000/health/liveliness` (via CT 116)
 - LiteLLM MCP tools: `curl localhost:4000/mcp-rest/tools/list -H "Authorization: Bearer $MASTER_KEY"` → 90 tools
- GPU servers responding: check :8080 on VM 101, VM 103; check ornith via router (http://192.168.68.116/health/unified — .15:8080 is firewalled to .116 only)
+- GPU servers responding: check :8080 on VM 101, VM 103; check strix-moe via router (http://192.168.68.116/health/unified — .15:8080 is firewalled to .116 only)
 - Zulip agents connected: check Mumuni/Tanko gateway state
 - Abiba PM2 processes online: `pm2 status`
@@ -124,7 +124,7 @@ Before Wave 1, snapshot these files:
 /opt/home_stack/docker-compose.yml                      (VM 109 .7)
 /opt/audiobookshelf/docker-compose.yml                  (VM 109 .7)
 /root/.pi/agent/extensions/config.yaml                  (CT 100 .24)
-/etc/systemd/system/ornith-server.service               (amdpve .15)
+/etc/systemd/system/ornith-server.service               (amdpve .15 — strix-moe)
 /etc/systemd/system/llama-server.service                (VM 101 .8, VM 103 .110)
 # Hermes agent configs (key enforcement — 2026-07-10)
 /root/.hermes/config.yaml                               (Mumuni CT 114, Tanko CT 112, etc.)
@@ -8,18 +8,18 @@ description: >
  Ensures agents never use the master key directly. Rotation is event-driven,
  not calendar-driven — rotate only on compromise, personnel change, or
  periodic security hygiene (quarterly/annually).
-  
+
  UPDATED 2026-07-12: Keys are stored in Infisical vault (project=agents, env=production)
  BUT each agent host MUST keep a local .env fallback. Infisical service tokens can
  expire/404. The .env fallback prevents agents from running without keys.
  Tanko incident: token 404 → gateway had no LITELLM_API_KEY for hours.
-  
+
  UPDATED 2026-07-16: Vault is SYNCED (session-13 keys written to vault via abiba service
  token, all validate 200). Koby/Koonimo migrated from hardcoded drop-ins to the
  infisical-gateway.sh wrapper (live vault injection). 4/5 agents now vault-backed.
  Canonical process: see § Production Vault Access Process. Tanko (user jerome) pending.
  Abiba's key is now a proper agent key (NOT the master key — stale note removed).
-  
+
  UPDATED 2026-07-17: FLEET-WIDE STANDARDIZATION. All 4 agents (Mumuni, Tanko, Koby, Koonimo)
  standardized on a single pattern: systemd drop-in (ExecStart= reset + wrapper path) →
  infisical-gateway.sh while-true loop → /usr/bin/infisical run --token → bash -c key
@@ -30,7 +30,7 @@ description: >
  Critical lessons: (1) NEVER use shell variables inside single-quoted bash -c in wrappers
  — hardcode absolute paths. (2) Drop-ins override unit file ExecStart permanently.
  (3) Capture /proc/<pid>/environ before gateway restarts to preserve running env set.
-  
+
  Current key inventory and agent list: see gpu-fleet.prose.md § Agent Keys.
  Source of truth for LiteLLM config: /opt/inference-harness/litellm_config.yaml
  on CT 116. Last verified: 2026-07-17.
@@ -154,7 +154,7 @@ through its agent wrapper.
   The `ExecStart=` (empty reset) clears any ExecStart from the main unit file,
   then the second `ExecStart=` sets the wrapper. This drop-in **survives unit file
   regeneration** by `hermes gateway install` — the drop-in always wins.
-   
+
   **Why a drop-in instead of editing the unit file:** `hermes gateway install`
   (called during Hermes updates and some self-heal operations) regenerates the
   systemd unit file with `ExecStart=/path/to/python -m hermes_cli.main gateway run`.
@@ -171,7 +171,7 @@ through its agent wrapper.
 - **Survives gateway crash**: the wrapper's `while true` + systemd `Restart=always` revive the gateway. Two-layer defense.
 - **Survives Hermes updates**: systemd drop-in overrides unit file ExecStart — `hermes gateway install` cannot break the vault injection.
 - **Survives reboot**: systemd user service + `loginctl enable-linger` ensures gateway starts at boot without a login session.
- **Auditable**: `cat /proc/$(pgrep hermes_cli)/environ` shows all injected keys; `infisical secrets` shows the vault source.
+- **Auditable**: `cat /proc/$(pgrep -f 'python.*hermes_cli.main.gateway.run' | grep -v infisical | head -1)/environ` shows all injected keys (note: pipe through grep -v infisical to avoid matching the bash wrapper); `infisical secrets` shows the vault source.
 ### Migration status (2026-07-17)
@@ -42,7 +42,7 @@ Request → nginx:80 → LiteLLM:4000 → GPU(llama-server, parallel 2)
 **What changed (v3.2.0 → v4.0.0 — 2026-07-08)**:
 - Router REMOVED from request path — LiteLLM proxies directly to GPU
 - All GPUs at parallel 2 (was parallel 1)
- NVIDIA context reduced 256K→128K to free VRAM  (SUPERSEDED 2026-07-16: all GPUs back to 256K — see litellm-self-heal)
+- NVIDIA context reduced 256K→128K to free VRAM — now the stable ceiling across all GPUs (2026-07-17)
 - LiteLLM timeouts tuned: gemma 25→120s, qwen 40→90s  (SUPERSEDED 2026-07-16: qwen 300s, gemma 120s, strix 300s — see litellm-self-heal)
 - nginx proxy_read_timeout: 600s, LiteLLM request_timeout: 300s
@@ -72,9 +72,9 @@ Request → nginx:80 → LiteLLM:4000 → GPU(llama-server, parallel 2)
 | Host | IP | Hardware | Models Served | Engine | Context | Parallel |
 |------|-----|----------|---------------|--------|---------|----------|
-| llm-gpu | 192.168.68.8 | NVIDIA RTX 3090 (24 GB) | qwen3.6-27B-code | llama-server systemd | **256K** | 2 |
+| llm-gpu | 192.168.68.8 | NVIDIA RTX 3090 (24 GB) | qwen3.6-27B-code | llama-server systemd | **128K** | 2 |
-| ocu-llm | 192.168.68.110 | NVIDIA RTX 5070 (12 GB) | gemma-4-12b | llama-server systemd | **256K** | 2 |
+| ocu-llm | 192.168.68.110 | NVIDIA RTX 5070 (12 GB) | gemma-4-12b | llama-server systemd | **128K** | 2 |
-| amdpve | 192.168.68.15 | AMD Strix Halo 64GB UMA | qwen3.6-35B-udq4 (LiteLLM alias: strix-moe) | llama-server systemd (Vulkan) | 256K | 2 |
+| amdpve | 192.168.68.15 | AMD Strix Halo 64GB UMA | qwen3.6-35B-udq4 (LiteLLM alias: strix-moe) | llama-server systemd (Vulkan) | 128K | 2 |
 ## Model Fallback Chains (LiteLLM)
@@ -57,9 +57,9 @@ Request → nginx:80 → LiteLLM:4000 → GPU(llama-server, parallel 2)
 | Host | IP | Hardware | Models Served | Engine | Context | Parallel |
 |------|-----|----------|---------------|--------|---------|----------|
-| llm-gpu | 192.168.68.8 | NVIDIA RTX 3090 (24 GB) | qwen3.6-27B-code | llama-server systemd (`/home/llmuser/llama-wrapper.sh`, `-c 262144 --parallel 2 --ngl 99`) | **256K** | 2 |
+| llm-gpu | 192.168.68.8 | NVIDIA RTX 3090 (24 GB) | qwen3.6-27B-code | llama-server systemd (`/home/llmuser/llama-wrapper.sh`, `-c 131072 --parallel 2 --ngl 99`) | **128K** | 2 |
-| ocu-llm | 192.168.68.110 | NVIDIA RTX 5070 (12 GB) | gemma-4-12b | llama-server systemd (`/home/llmuser/llama-wrapper.sh`, `--ctx-size 262144 --parallel 2`, IQ4_NL + MTP draft) | **256K** | 2 |
+| ocu-llm | 192.168.68.110 | NVIDIA RTX 5070 (12 GB) | gemma-4-12b | llama-server systemd (`/home/llmuser/llama-wrapper.sh`, `--ctx-size 131072 --parallel 2`, IQ4_NL + MTP draft) | **128K** | 2 |
-| amdpve | 192.168.68.15 | AMD Strix Halo 64GB UMA | qwen3.6-35B-udq4 (LiteLLM alias: `strix-moe`) | llama-server systemd (Vulkan) | 256K | 2 |
+| amdpve | 192.168.68.15 | AMD Strix Halo 64GB UMA | qwen3.6-35B-udq4 (LiteLLM alias: `strix-moe`) | llama-server systemd (Vulkan) | 128K | 2 |
 > Verified on ground 2026-07-16 via `curl /v1/models` on each host + `llama-wrapper.sh`. The AMD host's underlying model is `qwen3.6-35B-udq4`; LiteLLM exposes it under two `model_name`s: `qwen3.6-35B-udq4` and `strix-moe` (rpm 40). The legacy name `ornith-1.0-35b` does NOT exist in LiteLLM and must not be referenced.
@@ -101,7 +101,7 @@ Request → nginx:80 → LiteLLM:4000 → GPU(llama-server, parallel 2)
 ## Script Operations (synced 2026-07-16)
 - **Health-check script** (`/opt/inference-harness/scripts/litellm-health-check.sh` on CT 116): `gpu-fleet` check fails only on **critical** alerts (warnings are informational). Tests `strix-moe` (not `ornith-1.0-35b`).
- **GPU monitor** (`/root/scripts/gpu-monitor-server.py` on pi .24): runs as **systemd unit `gpu-monitor.service`** (was bare `&` process). `gpu_count` includes Strix Halo (was 2, now 3). VRAM alert thresholds: warning 93%, critical 97% (raised from 90/95 — 256K context steady-state is ~96% on RTX 3090, not a fault).
+- **GPU monitor** (`/root/scripts/gpu-monitor-server.py` on pi .24): runs as **systemd unit `gpu-monitor.service`** (was bare `&` process). `gpu_count` includes Strix Halo (was 2, now 3). VRAM alert thresholds: warning 93%, critical 97% (raised from 90/95 — 128K context steady-state is ~70% on RTX 3090, not a fault).
 - **Agent key monitor** (`/root/scripts/agent-health-check.py` on pi .24, cron `*/10`): reads each agent's **live** `LITELLM_API_KEY` from its gateway process env via SSH — never hardcodes keys (hardcoded keys rot on rotation and caused 9×401/30min). Fleet roster: abiba, tanko, mumuni, koby, koonimo (legacy `tdunna`/`baggy` removed — never existed).
 - **Stale keys cleaned**: `daily-infra-report.py` SYNTHETIC_API_KEY was stale (`sk-U_ydi3B` → 401); now reads `LITELLM_MASTER_KEY` from env. Deprecated scripts (`router-original.py`, `router-phase0-backup.py`, `apply-fixes.py`) still reference `sk-syslog-local-master-key` but do not actively poll LiteLLM.
@@ -6,7 +6,7 @@ description: >
  delegation, verification, and delivery. Defines when to delegate, which
  worker to use for what, how to handle failures, and the kanban board
  protocol. Enforces context-window discipline and separation of concerns.
-  Runs on Mumuni (CT 118, storepve, .6) via Hermes agent.
+  Runs on Mumuni (lxc/114, hwepve, .123) via Hermes agent.
 version: 1.0.0
 ---
@@ -19,8 +19,8 @@ version: 1.0.0
 ## Topology
-**Cluster:** 5 Proxmox nodes (ocupve, acerpve, minipve, amdpve, storepve)
+**Cluster:** 6 Proxmox nodes (ocupve, acerpve, minipve, amdpve, storepve, hwepve)
-**Manager:** Mumuni (CT 118, storepve, .6) via Hermes agent
+**Manager:** Mumuni (lxc/114, hwepve, .123) via Hermes agent
 **Workers:** 6 profiles, all running on the same agent — no separate hosts needed
 This contract is infrastructure-agnostic in terms of which nodes are used.
@@ -5,7 +5,7 @@ description: >
  Proxmox cluster + Docker monitoring via the existing Grafana/Prometheus stack
  on CT 116. Replaces Pulse with file-provisioned Grafana dashboards. Three
  exporters feed Prometheus: prometheus-pve-exporter (cluster-aware, single
-  instance), node_exporter (all 5 PVE nodes), and a custom docker-stats-exporter
+  instance), node_exporter (all 6 PVE nodes), and a custom docker-stats-exporter
  (Docker 29 / containerd image-store compatible, since cAdvisor cannot resolve
  the layerdb). Dashboards exposed at http://192.168.68.116:3001/ (direct LAN, not behind nginx).
 agent: abiba
@@ -33,8 +33,8 @@ agent: abiba
 | Exporter | Host:Port | Scope | Notes |
 |----------|-----------|-------|-------|
-| prometheus-pve-exporter | .116:9221 (container) | All 5 nodes + 14 guests + 36 storage pools | Single instance, cluster-aware via amdpve API. Config `/opt/monitoring/pve.yml` (token `monitoring@pve!prometheus`, PVEAuditor role). Metric schema is label-based (`id=node/amdpve`, `id=lxc/100`). |
+| prometheus-pve-exporter | .116:9221 (container) | All 6 nodes + guests + 36 storage pools | Single instance, cluster-aware via amdpve API. Config `/opt/monitoring/pve.yml` (token `monitoring@pve!prometheus`, PVEAuditor role). Metric schema is label-based (`id=node/amdpve`, `id=lxc/100`). |
-| node_exporter | .5/.6/.9/.12/.15:9100 (systemd) | Per-node CPU/mem/disk/net/temp | Installed via apt on all 5 PVE nodes, enabled (reboot-persistent). Collectors: textfile, systemd, tcpstat, ethtool. |
+| node_exporter | .5/.6/.9/.12/.15/.4:9100 (systemd) | Per-node CPU/mem/disk/net/temp | Installed via apt on all 6 PVE nodes, enabled (reboot-persistent). Collectors: textfile, systemd, tcpstat, ethtool. hwepve (.4) added 2026-07-19. |
 | docker-stats-exporter | .116:9324 (container) | 10 Docker containers on .116 | **Custom** (cAdvisor v0.51 incompatible with Docker 29 containerd image store — layerdb gone). Uses Docker Engine API over unix socket. Script `/opt/monitoring/docker-stats-exporter.py`. |
 ## PVE API Token
@@ -48,7 +48,7 @@ agent: abiba
 | UID | Title | Panels | Source |
 |-----|-------|--------|--------|
-| proxmox-cluster | Proxmox Cluster Overview | 16 | cluster status, 5-node CPU/mem/disk/load gauges, guests table, storage pools, guest CPU/mem timeseries |
+| proxmox-cluster | Proxmox Cluster Overview | 16 | cluster status, 6-node CPU/mem/disk/load gauges, guests table, storage pools, guest CPU/mem timeseries |
 | proxmox-node | Proxmox Node Detail | 13 | per-node CPU per-core, memory, network, disk IO/IOPS/latency, temperature, disk space (variable: $node) |
 | docker-containers | Docker Containers | 10 | per-container CPU/mem/network, restarts, memory limit ratio (variable: $container) |
 | gpu-fleet | GPU Fleet | 7 | (existing, preserved in DB, not provisioned) |
@@ -77,9 +77,9 @@ agent: abiba
 | `/opt/monitoring/grafana/dashboards/build-dashboards.py` | .116 | dashboard JSON generator |
 | `/opt/monitoring/grafana/dashboards/json/*.json` | .116 | provisioned dashboard definitions |
 | `/opt/monitoring/grafana/datasources/prometheus.yml` | .116 | datasource provisioning |
-| `/etc/default/prometheus-node-exporter` | .5/.6/.9/.12/.15 | node_exporter collector config |
+| `/etc/default/prometheus-node-exporter` | .5/.6/.9/.12/.15/.4 | node_exporter collector config |
-## Cluster "Tabiri" — 5 Nodes
+## Cluster "Tabiri" — 6 Nodes
 | Node | IP | Role |
 |------|----|----|
@@ -87,7 +87,8 @@ agent: abiba
 | storepve | 192.168.68.6 | PVE |
 | acerpve | 192.168.68.9 | PVE (hosts llm-gpu qemu/101) |
 | minipve | 192.168.68.12 | PVE |
-| amdpve | 192.168.68.15 | PVE + Strix Halo LLM (ornith) |
+| amdpve | 192.168.68.15 | PVE + Strix Halo LLM (qwen3.6-35B-udq4, strix-moe) |
 | hwepve | 192.168.68.4 | PVE (Huawei Matebook 16, 12C/15GB) — hosts Mumuni (lxc/114) migrated from minipve 2026-07-20 |
 ## Operations
@@ -184,8 +184,8 @@ def check_agents():
            print(f"  ⬜ {name} (CT {ct}): cannot SSH — skip liveness check")
            continue
-        # Gateway process
+        # Gateway process (exclude the infisical bash wrapper that contains the same string)
-        pid = ssh(host, "pgrep -f 'hermes_cli.main gateway run' | head -1", user=user)
+        pid = ssh(host, "pgrep -f 'hermes_cli.main gateway run' | grep -v infisical | head -1", user=user)
        if not pid:
            print(f"  ❌ {name}: GATEWAY NOT RUNNING")
            FAIL.append(f"gateway-down:{name}")
@@ -16,7 +16,7 @@ Runs every 15 minutes in the background. Also triggers on session start.
 ## Requires
 - **Zulip API key** for `abiba-bot@chat.sysloggh.net` in `$ZULIP_API_KEY`
- **SSH access** to Tanko (192.168.68.122), Mumuni (192.168.68.123), and Agent Zero Docker host (192.168.68.14)
+- **SSH access** to Tanko (192.168.68.122), Mumuni (192.168.68.123, lxc/114 on hwepve since 2026-07-20), and Agent Zero Docker host (192.168.68.14)
 - **PM2** on localhost for pi process management
 - **Network access** to `chat.sysloggh.net`, `localhost:9200`
 - **Write access** to `/root/zulip-health-monitor.log` and `/tmp/zulip-monitor-debounce`
@@ -199,7 +199,7 @@ Check `platforms.zulip.state`: `connected` ✅ | `disconnected` ❌ | `error`
 ssh root@<CT> "ps aux | grep 'gateway run' | grep -v grep"
 ```
-Gateway PID should exist with uptime > 60s.
+Gateway PID should exist with uptime > 60s. **Dual-gateway detection**: if more than one `gateway run` process is found, the gateway has a collision (typically one `--force` and one `--replace` process). Kill the newer/duplicate process, then restart the remaining gateway via PM2 (`pm2 restart mumuni-zulip`). Check gateway log for "Gateway running with 2 platform(s)" (not 1) to confirm Zulip reloaded.
 **B3: Heartbeat Verification**
@@ -409,3 +409,68 @@ Backup v2 before starting: `cp index.js index.js.v2-backup-$(date +%Y%m%d-%H%M%S
 | Queue expiry handling | Crash | Auto re-register |
 | Busy worker deadlock | Router death | Worker SIGKILL + error DM |
 | PM2 restart exhaustion | Yes (max_restarts=10) | No (max_restarts=100 + watchdog) |
 ---
 ## Incident Log — 2026-07-18 Fleet-Wide Audit
 ### Fleet State After Audit
 | Agent | Platform | Zulip State | Issues Found | Fix Applied |
 |-------|----------|-------------|--------------|-------------|
 | **Abiba** | pi (CT 100) | ✅ Connected | API key missing from Infisical injection; poll timeout noise | Added .env fallback; AbortError treated as empty poll (no retry); poll timeout 65s→90s |
 | **Tanko** | Hermes (CT 112) | ✅ Connected | Gateway disconnected since Jul 11; watchdog restart didn't re-establish Zulip | Full gateway restart (kill wrapper, let infisical-gateway.sh respawn) |
 | **Mumuni** | Hermes (CT 114) | ✅ Connected | No issues found | None needed |
 ### Key Fixes Applied
 **1. Abiba — Credential Fallback (L4 Pattern)**
 - Root cause: `zulip.api_key` in config.yaml is `""` (expected from Infisical). Infisical vault `ABIBA_ZULIP_API_KEY` wasn't being injected into the process environment.
 - Fix: Added `.env` file fallback at `/root/.pi/agent/extensions/zulip/.env` with known-working key, sourced before the Infisical `exec`.
 - Lesson: Per L4 from gpu-self-heal, Infisical is not always available — always keep a local `.env` fallback.
 **2. Abiba — Poll Timeout Handling**
 - Root cause: Zulip long-poll uses `AbortSignal.timeout(65000)`. Zulip's default `event_queue_longpoll_timeout_seconds` can exceed 65s. When the signal fires, an `AbortError` is thrown and caught by the circuit breaker as a failure.
 - Fix: Caught `AbortError` inside `poll()` and return empty array (no events) instead of throwing. Extended timeout to 90s to match Zulip server default.
 - Reference: [Zulip Events System — long-poll timeout](https://zulip.readthedocs.io/en/11.6/subsystems/events-system.html)
 **3. Tanko — Gateway Restart**
 - Root cause: Gateway process was running but Zulip platform stayed in "disconnected" state since Jul 11, 2026. The wrapper script (`infisical-gateway.sh`) restarts on crash but the gateway wasn't re-establishing Zulip on restart.
 - Fix: Killed gateway PID to trigger wrapper restart. New gateway (PID 331991) established Zulip connection successfully.
 ### Fleet-Wide Zulip Health Metrics (as of 2026-07-18)
 | Metric | Value |
 |--------|-------|
 | Zulip server | ✅ HTTP 200 |
 | Agents connected | 3/3 (Abiba, Tanko, Mumuni) |
 | Abiba circuit breaker | CLOSED (0 failures) |
 | Abiba uptime | 2D (post-restart) |
 | Tanko gateway uptime | Ongoing |
 | Mumuni gateway uptime | Ongoing |
 | Watchdog status | ✅ Online (2D uptime) |
 ### Hermes Agent Zulip Plugin Improvements
 Based on the audit, improvements that should be ported to all Hermes Zulip adapters:
 1. **Circuit breaker pattern** — Already in Abiba's pi extension. Hermes adapters should add the same CLOSED→OPEN→HALF_OPEN state machine with exponential backoff.
 2. **Credential fallback** — All Hermes agents use Infisical for credentials. Add `.env` local fallback per L4 pattern for `ZULIP_API_KEY`.
 3. **Queue re-registration** — Handle `BAD_EVENT_QUEUE_ID` with automatic re-registration instead of gateway restart.
 4. **Supervisor watchdog** — Hermes uses PM2 which auto-restarts on crash, but has no health-check watchdog. Add lightweight external health checks.
 5. **Streaming** — All agents have `streaming: true` in their zulip config. Verify `edit_message()` is implemented in each adapter.
 ### Abiba pi Zulip Extension v2 — Implemented Resilience Summary
 | Feature | Status | Notes |
 |---------|--------|-------|
 | Circuit breaker | ✅ | CLOSED→OPEN→HALF_OPEN; 50% failure threshold; 30s reset timeout |
 | Retry with jitter | ✅ | 2 attempts, 200ms base, 50-100% jitter |
 | Queue lifecycle | ✅ | 10min idle_queue_timeout; BAD_EVENT_QUEUE_ID handling |
 | Crash prevention | ✅ | uncaughtException + unhandledRejection recovery |
 | Worker timeout | ✅ | 5min busy timeout → SIGKILL + error DM |
 | Health endpoint | ✅ | :9200 with circuit breaker metrics |
 | Echo prevention | ✅ | Dynamic bot user resolution |
 | Poll timeout (AbortError) | ✅ v2.1 | Normal timeout returns [] instead of error |
 | Credential fallback | ✅ v2.1 | .env file before Infisical exec |
 | Provider auto-fix | ✅ | Detects reasoning_content models, switches to compatible |