@miller-tech/uap 1.221.0 → 2.0.0

package/README.md

@@ -6,10 +6,10 @@
 
 [![npm](https://img.shields.io/npm/v/@miller-tech/uap?color=blue&label=npm)](https://www.npmjs.com/package/@miller-tech/uap)
 [![Node](https://img.shields.io/badge/node-%3E%3D18-brightgreen)](https://nodejs.org)
-[![Tests](https://img.shields.io/badge/tests-170%2B_suites-brightgreen)](#testing)
+[![Tests](https://img.shields.io/badge/tests-450%2B_suites-brightgreen)](#testing)
 [![License](https://img.shields.io/badge/license-MIT-black)](LICENSE)
 
-`v1.163` · 220+ modules · 200+ test suites · 9 agent harnesses
+`v1.224.0` · 367 TypeScript modules across 26 subsystems · 459 vitest suites (+ a ~1,200-test Python enforcer/proxy suite) · 9 agent harnesses
 
 [Quickstart](#quickstart) · [Why UAP?](#why-uap) · [The delivery pipeline](docs/guides/DELIVERY_PIPELINE.md) · [`uap deliver`](#the-deliver-harness) · [Docs](docs/INDEX.md)
 
@@ -76,6 +76,8 @@ What happens under the hood:
 
 It works with frontier models *and* local models (llama.cpp / Qwen) served over the Anthropic Messages API. See **[docs/guides/DELIVER.md](docs/guides/DELIVER.md)**.
 
+Hardened for shared floors (v1.224): rollback is scoped to the run's own write-set (a shared-worktree rollback can't revert another agent's work), projects can declare extra gates in `delivery.gates[]`, the execution gate is polyglot (Python packages, native binaries), and `delivery.model`/`routing`/`criticality` in `.uap.json` steer model choice — with root-owned, expiring operator overrides for emergencies.
+
 ---
 
 ## The line, station by station
@@ -91,7 +93,7 @@ It works with frontier models *and* local models (llama.cpp / Qwen) served over
 | **Shipping** | Regressions, red CI, skipped version bumps | Worktree→PR flow, version gates, CI feedback watcher |
 | **Feedback** | The same mistake every session | Memory promotion, pattern learning, session analysis |
 
-Running the whole length of the floor: **policy gates** (24 executable enforcers that *block* non-compliant tool calls — worktree, test, schema-diff, expert-review, delivery-enforcement…) and the **MCP Router** (keeps the context window lean). Full catalog: **[docs/reference/FEATURES.md](docs/reference/FEATURES.md)**.
+Running the whole length of the floor: **policy gates** (32 executable enforcers that *block* non-compliant tool calls — worktree, test, schema-diff, expert-review, delivery-enforcement, quality-metrics, design-token, visual-verification…), the **quality-metrics gate** (deterministic complexity/coverage/mutation budgets with a ratchet baseline, so standards can only tighten), and the **MCP Router** (keeps the context window lean). Full catalog: **[docs/reference/FEATURES.md](docs/reference/FEATURES.md)**.
 
 ---
 
@@ -109,14 +111,16 @@ UAP installs hooks into your agent harness, then mediates every tool call throug
 ┌─────────────────────────────────────────────────────────────┐
 │                       UAP CLI (uap)                         │
 │  setup · memory · deliver · verify · worktree · policy      │
-│  task · coord · droids · model · mcp-router · design …      │
+│  quality · plan · fidelity · interaction · task · coord ·   │
+│  droids · model · mcp-router · proxy · orchestrator ·       │
+│  self-harness · tune · design · principles · bench …        │
 └──┬─────────┬──────────┬──────────┬──────────┬───────────────┘
    ▼         ▼          ▼          ▼          ▼
  Memory   Policy    MCP Router   Delivery   Coordination
- 4 tiers  24 gates  compression  + verify   + deploy batch
+ 4 tiers  32 gates  compression  + verify   + deploy batch
 ```
 
-- **30+ CLI commands** across 18 source subsystems (220+ TypeScript modules).
+- **60 CLI commands** across 26 source subsystems (367 TypeScript modules).
 - Deep dive: **[docs/architecture/OVERVIEW.md](docs/architecture/OVERVIEW.md)** · protocol spec: **[docs/architecture/PROTOCOL.md](docs/architecture/PROTOCOL.md)**.
 
 ---
@@ -187,11 +191,17 @@ Start at the **[documentation index](docs/INDEX.md)**.
 
 ```bash
 npm install
-npm run build      # TypeScript compile
-npm test           # vitest — 170+ suites
-npm run bench      # benchmark suite
+npm run build           # TypeScript compile
+npm test                # vitest — 459 suites (~5,000 tests)
+npm run test:enforcers  # Python policy-enforcer suite (~1,200 tests)
+npm run bench           # benchmark suite
 ```
 
+Every ship is gated: `npm test` and `npm run build` must be green, the change
+must come from a worktree, and the version bump goes through
+`npm run version:patch|minor|major` (no manual edits). Details:
+[CONTRIBUTING.md](CONTRIBUTING.md).
+
 ---
 
 ## Acknowledgements

package/config/exl3-profiles/qwen38-27b-exl3.env

@@ -0,0 +1,68 @@
+# =====================================================================
+# Qwen 3.8 27B EXL3 3.5bpw + DFlash2 EXL3 5.0bpw draft, exllamav3
+# (MiaAI-Lab fork 1.4.2) @ 131072 ctx/request, batch-1 -- 2026-09-04.
+# THIS IS THE PRODUCTION PROFILE. Service: uap-exl3-server.service ->
+# /home/cogtek/Qwen3.8-EXL3-kit/start.sh
+# Install: this is a REFERENCE copy; the live file is the kit's .env at
+#   /home/cogtek/Qwen3.8-EXL3-kit/.env
+#
+# Supersedes config/llama-profiles/qwen38-27b-dflash2.env (kept as the
+# rollback profile). Same base model, same per-request context -- the
+# serving engine changes: GGUF/llama.cpp -> EXL3/exllamav3.
+#
+# MEASURED on the RTX 3090 (sm_86) after cutover 2026-09-04:
+#   decode, short ctx, warm:   ~82 tok/s  (vs ~55 llama.cpp baseline,
+#                              same 600-word essay probe, 1192 tokens)
+#   decode, deep long ctx:     ~41-51 tok/s (KV cache growth falloff)
+#   VRAM resident:             20964 MiB / 24576 MiB
+#   model load:                < 90 s
+#   tool calling:              native OpenAI-style tool_calls, verified
+#
+# 3090 (sm_86) CAVEATS -- do not copy the 4090 recipe blindly:
+#   * CACHE_QUANT=nvfp4 / fp8 need CC >= 8.9 (Ada/Hopper). sm_86 is fixed
+#     in hardware; CACHE_QUANT=4 (Hadamard int4) is the same ~4.5
+#     bits/elem and is what runs.
+#   * nvcc is 12.4 (deb layout at /usr/bin, no /usr/local/cuda), so torch
+#     is pinned to cu124 wheels and CUDA_HOME=/usr.
+#   * The engine fork needs triton >= 3.3 (ASTSource constexprs kwarg)
+#     while torch 2.6 pins triton 3.2 and its inductor breaks on 3.3
+#     (AttrsDescriptor removed). Working pair: torch 2.6.0+cu124 +
+#     triton 3.3.1 + the torch.compile->identity shim in the kit's
+#     tools/serve_openai.py. pip's dependency-conflict warning about
+#     this is expected; do not "fix" it by downgrading triton.
+#   * start.sh re-runs its bootstrap if its health check fails; the
+#     unshimmed import check fails for the reason above, so the kit
+#     carries tools/healthcheck.py (shimmed) as the gate. Without it,
+#     every start downgraded triton 3.3.1 -> 3.2.0 and broke the engine.
+#
+# PROXY RAILS (unchanged, still correct): PROXY_CONTEXT_WINDOW=131072
+# (--cache_size is per-request here, matching llama.cpp's per-slot 131072
+# at --parallel 1) and PROXY_CONCURRENCY_LIMIT=1 (the server is batch-1
+# and queues concurrent callers internally). The server accepts ANY
+# client-sent model id, so presets (qwen3.8-27b, ...) need no alias
+# maintenance -- the llama.cpp std::set alias-ordering trap is gone.
+# =====================================================================
+
+# --- engine install ------------------------------------------
+TORCH_INDEX_URL=https://download.pytorch.org/whl/cu124
+
+# --- models ---------------------------------------------------
+MODEL_DIR=models/Qwen3.8-27B-EXL3-3.5bpw
+
+# --- speculative decoding -------------------------------------
+DRAFT=dflash2
+DRAFT_DIR=models/Qwen3.8-27B-DFlash2-EXL3-5.0bpw
+
+# --- HTTP ------------------------------------------------------
+PORT=8080
+HOST=0.0.0.0
+
+# --- cache / context -------------------------------------------
+CONTEXT_SIZE=131072
+CACHE_QUANT=4
+GPU_MEM_GB=22
+CPU_CACHE_GB=0
+
+# Box layout: nvcc is /usr/bin/nvcc (deb), no /usr/local/cuda
+CUDA_HOME=/usr
+CUDA_PATH=/usr

package/config/llama-profiles/qwen38-27b-dflash2.env

@@ -1,6 +1,11 @@
 # =====================================================================
 # Qwen 3.8 27B (dense, UD-IQ4_XS) @ 131072 ctx/slot / 2 slots, DFlash2
-# speculative decoding -- 2026-08-20. THIS IS THE DEFAULT PROFILE.
+# speculative decoding -- 2026-08-20.
+# SUPERSEDED 2026-09-04 by the EXL3/exllamav3 stack (see
+# docs/guides/QWEN38_EXL3.md, config/exl3-profiles/qwen38-27b-exl3.env) --
+# ~1.5x faster decode at the same 131072 context. This profile is kept as
+# the ROLLBACK path: systemctl --user stop uap-exl3-server && systemctl
+# --user enable --now uap-llama-server.
 # Service: uap-llama-server.service -> scripts/run-llama-server-continuity.sh
 # Install: cp to ~/.config/uap/llama-server.env
 #