@miller-tech/uap 2.21.3 → 2.36.1

package/README.md

@@ -6,17 +6,25 @@
 
 [![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-450%2B_suites-brightgreen)](#testing)
+[![Tests](https://img.shields.io/badge/tests-500%2B_suites-brightgreen)](#testing)
 [![License](https://img.shields.io/badge/license-MIT-black)](LICENSE)
 
-`v2.16.6` · 367 TypeScript modules across 26 subsystems · 459 vitest suites (+ a ~1,200-test Python enforcer/proxy suite) · 9 agent harnesses
+447 TypeScript modules across 34 subsystems · 530 vitest suites (~6,250 tests, + a ~1,400-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)
+[Quickstart](#quickstart) · [Why UAP?](#why-uap) · [Origin & principle](#origin--principle) · [The delivery pipeline](docs/guides/DELIVERY_PIPELINE.md) · [`uap deliver`](#the-deliver-harness) · [Docs](docs/INDEX.md)
 
 </div>
 
 ---
 
+## Origin & principle
+
+UAP is a **Miller-Tech** tool built for real delivery floors. The failures it targets are the ones that keep recurring in agentic development — "done" code that never ran, edits on `main`, forgotten decisions, prose architecture rules that blocked nothing, agents that stall on any question or silently decide for you. The fix that works, every time, is the same: **turn the software-engineering practice into an executable check the agent cannot route around. Enforce, don't suggest.**
+
+That principle became UAP — a discipline layer installed *underneath* the agent harness, putting a verification station at every point where agentic development actually breaks. It is hardened under its own gates (worktree isolation, parallel review, evidence-bound shipping), which makes the tool the proof of its own thesis. Design decisions are recorded in the [ADR set](docs/architecture/OVERVIEW.md); inspirations and prior art are credited in [docs/ATTRIBUTION.md](docs/ATTRIBUTION.md).
+
+---
+
 ## Why UAP?
 
 Shipping software with an AI agent is a lot like running a small factory floor. Intent comes in one end; working, verified, merged code should come out the other. In between are stations — understand the job, set up a bench, build it, **check it actually works**, ship it — and a jam at any one of them quietly ruins everything downstream.
@@ -32,6 +40,9 @@ Coding agents are capable but undisciplined line workers. They forget yesterday'
 | Two agents clobber each other | Worktree isolation + live file coordination | Conflict-free parallel work |
 | "Guidelines" get ignored | Policy gates as executable hooks, not prose | Violations are **blocked**, not politely suggested |
 | Same mistake, every session | Memory promotion + pattern learning | The floor gets better every run |
+| The agent stalls on any question — or silently decides for you | Attention queue + AFK auto-decider | Asks are recorded with a **safe default** and work continues; unattended adoption is logged, never silent |
+| Fast tasks idle behind slow siblings | Completion-triggered dispatch | A finished task's slot refills instantly; dependents start the moment their last dependency settles |
+| Plans the agent can't follow; it guesses its own tooling | Plan readability gate + self-knowledge card | Legible plans; the agent answers "how do I X" from its own version-pinned card |
 
 **→ Take the full station-by-station tour: [The UAP Delivery Pipeline](docs/guides/DELIVERY_PIPELINE.md).**
 
@@ -76,7 +87,7 @@ 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.
+Hardened for shared floors: 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.
 
 ---
 
@@ -89,11 +100,11 @@ Hardened for shared floors (v1.224): rollback is scoped to the run's own write-s
 | **Isolation** | Editing `main`, clobbering files | [Worktrees](docs/guides/WORKTREE_WORKFLOW.md), live file coordination, delivery gate |
 | **Build** | Plausible-but-wrong code, stubs, empty output | [`uap deliver`](docs/guides/DELIVER.md), serving-layer recipes, [local-model guardrails](docs/guides/LOCAL_MODELS.md) |
 | **QC / verify** | "Done" on code that never ran | Completion gates, `uap verify`, acceptance judge, generator≠evaluator |
-| **Coordination** | Parallel agents colliding/deadlocking | [Coordination](docs/guides/COORDINATION.md), model-slot concurrency, [deploy batching](docs/guides/DEPLOY_BATCHING.md) |
+| **Coordination** | Parallel agents colliding/deadlocking | [Coordination](docs/guides/COORDINATION.md), model-slot concurrency, [deploy batching](docs/guides/DEPLOY_BATCHING.md), attention queue (asks instead of stalls) |
 | **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** (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)**.
+Running the whole length of the floor: **policy gates** (34 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)**.
 
 ---
 
@@ -118,10 +129,10 @@ UAP installs hooks into your agent harness, then mediates every tool call throug
 └──┬─────────┬──────────┬──────────┬──────────┬───────────────┘
    ▼         ▼          ▼          ▼          ▼
  Memory   Policy    MCP Router   Delivery   Coordination
- 4 tiers  32 gates  compression  + verify   + deploy batch
+ 4 tiers  34 gates  compression  + verify   + deploy batch
 ```
 
-- **60+ CLI commands** across 26 source subsystems (367 TypeScript modules).
+- **58 top-level CLI commands** (~210 registrations including subcommands) across 34 source subsystems (434 TypeScript modules).
 - Deep dive: **[docs/architecture/OVERVIEW.md](docs/architecture/OVERVIEW.md)** · protocol spec: **[docs/architecture/PROTOCOL.md](docs/architecture/PROTOCOL.md)**.
 
 ---
@@ -177,6 +188,7 @@ Install into all detected harnesses with `uap hooks install`; audit coverage wit
 | | |
 |---|---|
 | **[The Delivery Pipeline](docs/guides/DELIVERY_PIPELINE.md)** | The station-by-station tour — start here for the big picture |
+| **[ATTRIBUTION](docs/ATTRIBUTION.md)** | Prior art, inspirations, and credits |
 | **[Getting Started](docs/getting-started/)** | Installation, quickstart, configuration |
 | **[Guides](docs/guides/)** | deliver, memory, MCP router, worktrees, policies, multi-model, local models |
 | **[Architecture](docs/architecture/)** | System overview + the UAP protocol |
@@ -193,8 +205,8 @@ Start at the **[documentation index](docs/INDEX.md)**.
 ```bash
 npm install
 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 test                # vitest — 518 suites (~6,050 tests)
+npm run test:enforcers  # Python policy-enforcer suite (~1,400 tests)
 npm run bench           # benchmark suite
 ```
 

package/config/model-profiles/qwen38-coder390.json

@@ -0,0 +1,46 @@
+{
+  "_profile": "qwen38-coder390",
+  "_description": "Qwen3.8-27B Coder390 EfficientThink (LynnStyle Q4 + Q8 MTP) — community post-train of the official Qwen3.8-27B (dense qwen35 arch, Apache-2.0; nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2). Served by uap-coder390-server (llama.cpp continuity script, native draft-mtp from the checkpoint's blk.64 head, Q8 mmproj for vision). The tune's headline: EfficientThink — reasoning P50 down 44-53% and 94K-cap truncations down 13→3 on LCB vs the stock 27B, with LCB 89-90. No DFlash2 draft ships for GGUF (mutually exclusive upstream).",
+  "model": "qwen3.8-27b-coder390",
+  "max_tokens": 57344,
+  "_max_tokens_comment": "Same 50%-of-window ratio as qwen38. The tune is TRAINED to stop early (that is its whole point), but reasoning still spends completion budget in reasoning_content; keep the cap generous so a long-necessary derivation is never truncated into an empty answer — the failure this tune exists to fix must not be reintroduced by the server config. The proxy's PROXY_TOOL_TURN_MAX_TOKENS clamp still applies to tool turns.",
+  "temperature": 0.3,
+  "top_p": 0.9,
+  "top_k": 20,
+  "min_p": 0.05,
+  "repetition_penalty": 1.0,
+  "_sampling_comment": "UAP's tool-call-tuned values, carried from qwen38 (never vendor chat defaults). EfficientThink changes the model's reasoning LENGTH, not its sampling; these stay as-is until a `uap tune` pass on real agent traffic says otherwise.",
+  "stop_sequences": ["<|im_end|>"],
+  "timeout_ms": 300000,
+  "context_window": 114688,
+  "_context_window_comment": "LLAMA_CTX_SIZE=131072 is the single-rail TOTAL (one 24 GiB card holding 18.5 GiB weights + 0.6 mmproj + KV); the per-session cap sits 87.5% below the pool exactly like the strata precedent, leaving the engine its own scratch room. If the measure pass shows the q4_0 KV pool smaller than 131k cells, lower this to 0.875x of the measured pool.",
+  "optimize_for_tool_calls": true,
+  "mode_switch_buffer_tokens": 500,
+  "enable_thinking": true,
+  "_thinking_comment": "Top-level; the proxy probes the backend once and adapts (PROXY_CHAT_TEMPLATE_KWARGS=auto). The GGUF carries the chat template; the continuity script's qwen-sharp.jinja override is NOT used for this profile's unit — set LLAMA_CHAT_TEMPLATE_FILE=embedded in coder390-server.env if the qwen-sharp template proves incompatible (the tune was trained with its own template).",
+  "dynamic_temperature": {
+    "enabled": true,
+    "decay": 0.5,
+    "floor": 0.2,
+    "_comment": "On retry, temp = max(floor, base_temp * decay^attempt)."
+  },
+  "tool_call_batching": {
+    "enabled": true,
+    "system_prompt_suffix": "When multiple tools are needed, call ALL of them in a single response. Do not call one tool and wait - emit all tool calls together.",
+    "_comment": "Reduces round-trips by encouraging multi-tool emission in one turn."
+  },
+  "concurrency": {
+    "max_parallel_requests": 1,
+    "_comment": "One rail, one 131k pool: a second hot session would overcommit the KV pool. Matches the strata single-rail posture (slots, proxy concurrency and admission move together).",
+    "kv_capacity_shared": true
+  },
+  "measured": {
+    "date": null,
+    "gpu": "NVIDIA GeForce RTX 3090, 24576 MiB",
+    "_gpu_comment": "Footprint not yet measured — run `uap models measure qwen3.8-27b-coder390/coder390-q4-lynn-mtp` after the first boot and record it here and in the registry (the admit gate refuses unmeasured models, fail-closed)."
+  },
+  "routing": {
+    "endpoint": "http://127.0.0.1:4000/v1",
+    "_comment": "Route through the anthropic-proxy, never the llama-server port directly. Clients select this profile with the header 'x-uap-model-profile: qwen38-coder390' (or a uap_model_profile body param); model-id routing uses the registry's advertises aliases."
+  }
+}

package/config/model-profiles/qwen38.json

@@ -1,9 +1,10 @@
 {
   "_profile": "qwen38",
   "_description": "Qwen3.8 Flash-Next — local Strata serve-layer deployment for agentic tool-calling workloads. Successor to qwen36 and to the 2026-09-20 GSQ-RCO llama.cpp deployment; the default local executor in UAP routing. Routes through the anthropic-proxy for the tool/finalize guardrails, NOT the inference server directly.",
-  "_engine": "strata serve layer 0.1.40.1 (python server 0.1.40.1 + engine binary 0.1.40, compiled locally with a user-space CUDA 13.4 toolkit after the Debian 12.4 headers proved too old for 0.1.40) on :8080, launched by the systemd user unit uap-strata-server.service with /home/cogtek/dev/strata/strata-iq3_s.json. Serves llama.cpp-shaped /props, /slots and /v1, plus a JSON /metrics (engine/live/requests — NOT Prometheus text; `uap inference health` auto-detects it). Weights: strata-data pack iq3_s (Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S since the 2026-10-07 swap per the IQ3_S speed≈IQ3_XXS reports; predecessors Swift 1.5 IQ3_XXS 2026-10-06, Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS, native experts + PLE shard) with MTP speculative drafts (spec 8, spec-min-p 0.8 since the 2026-10-06 tuning; previously spec 4, spec-min-p 0.5, mtp_max 4) and an auto-sized expert cache. Model id advertised at /v1/models: qwen3.8-flash-next-iq3_s; the server ignores requested model ids entirely, which is why stale client pins (e.g. a resumed session still naming qwen36 ids) keep working.",
+  "_engine": "strata serve layer 0.1.44.1 (build_info 'Strata 0.1.44.1', compiled locally with the user-space CUDA 13.4 toolkit) on :8080, launched by the systemd user unit uap-strata-server.service with /home/cogtek/dev/strata/strata-swift-iq3_s.json since the 2026-10-17 Swift 1.5 A/B trial (revert: repoint the unit's --config at strata-iq3_s.json, daemon-reload, restart; the unit file carries the dated note). Serves llama.cpp-shaped /props, /slots and /v1, plus a JSON /metrics (engine/live/requests — NOT Prometheus text; `uap inference health` auto-detects it). Weights: strata-data pack swift-iq3_s (Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S, the UkisAI reasoning-efficient post-train, since the 2026-10-17 trial; predecessor base Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S since the 2026-10-07 swap per the IQ3_S speed≈IQ3_XXS reports; earlier predecessors Swift 1.5 IQ3_XXS 2026-10-06, Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS, native experts + PLE shard) with MTP speculative drafts (spec 8 since the 2026-10-06 tuning; spec-min-p 0.70 after the 0.1.44.1-era calibration, previously 0.8, before that 0.5; previously spec 4, mtp_max 4) and an auto-sized expert cache. Model id advertised at /v1/models: swift-1.5-iq3_s; the server ignores requested model ids entirely, which is why stale client pins (e.g. a resumed session still naming qwen36 or qwen3.8-flash-next ids) keep working.",
   "_engine_history": "SUPERSEDED 2026-10-04: this profile described buun-llama-cpp behind uap-gsq-rco-server.service (2 rails sharing a 229376-cell VBR pool, DFlash2 draft). That unit is now INACTIVE — the box serves the Strata layer on :8080 instead (tuned 2026-10-03: +15% over the llama.cpp fork on agent workloads). The old description was itself a 2026-09-20 correction of a ninfer-serve description. Verify against /metrics before trusting engine claims here; the dead unit's ExecStart still reports -c 229376 to systemctl, and on 2026-10-04 `uap inference health` printed that stale pool size for the live (131072) backend.",
-  "model": "qwen3.8-flash-next-iq3_s",
+  "model": "swift-1.5-iq3_s",
+  "_model_comment": "The alias the live backend advertises at /v1/models; the rails suite drift-checks this field against the serving process (test/model-profiles-qwen38-rails.test.ts). 2026-10-17: the Swift 1.5 IQ3_S A/B trial swapped the unit from the base pack, which advertised qwen3.8-flash-next-iq3_s. The server ignores requested model ids, so clients pinning the old id are unaffected; on revert, restore this field alongside the unit's --config.",
   "max_tokens": 57344,
   "_max_tokens_comment": "Deliberately generous: 50% of the 114688 window, the same ratio the previous 65536/131072 pairing used. The engine returns chain-of-thought in a SEPARATE reasoning_content field, so reasoning spends completion budget without ever appearing in content. A tight cap therefore does not produce a short answer, it produces an EMPTY one with finish_reason=length — measured previously as ~90% first-turn decode failure that read as bad output format and was really budget truncation. Do not lower this without re-measuring decode compliance. (The proxy additionally clamps tool turns to PROXY_TOOL_TURN_MAX_TOKENS=32768 regardless.)",
   "temperature": 0.3,

package/config/model-registry.json

@@ -0,0 +1,110 @@
+{
+  "version": 1,
+  "_comment": "Repo defaults for the model placement registry. Shape only: unit/service names, engines, profiles, affinity, and the config inventory. Measured footprints (resident_gpu_mib, host_rss_mib, device totals) are machine state and live in ~/.uap/model-registry.json, deep-merged over this file. A config without a measurement is 'unmeasured' and fails closed: it may be served if already resident, but it can never be loaded or evicted. See docs/specs/operator-model-placement.md §4.1.",
+  "devices": {
+    "gpu0": {
+      "kind": "gpu"
+    },
+    "cpu0": {
+      "kind": "cpu"
+    }
+  },
+  "models": {
+    "qwen3.6-27b": {
+      "display": "Qwen3.6 27B",
+      "engine": "dflash",
+      "task_affinity": ["general", "ugc"],
+      "affinity": { "device": ["gpu0"] },
+      "endpoint": "http://127.0.0.1:8080",
+      "unit": "uap-dflash-server",
+      "advertises": [
+        "qwen3.6-27b",
+        "Qwen3.6-27B",
+        "luce-dflash",
+        "claude-haiku-4-5-20251001",
+        "claude-sonnet-4-6",
+        "claude-sonnet-5-20250514",
+        "claude-fable-5"
+      ],
+      "configs": {
+        "dflash-q3_k_m": {
+          "status": "unmeasured",
+          "engine": "dflash",
+          "unit": "uap-dflash-server",
+          "note": "lucebox-hub DFlash speculative-decoding server (Qwen3.6-27B Q3_K_M target + draft head); machine profile: ~/.config/uap/dflash-server.env. The ugc-factory brain (pay2u tools/ugc-factory) requests 'qwen3.6-27b' through the UAP proxy, so placement auto-loads this config on demand — run `uap models measure` to record this machine's footprint before it can load (fail-closed). The Claude contract ids are advertised here too: the proxy serves them locally (ANTHROPIC_PASSTHROUGH_MODELS=__local_only__), and without a registry mapping the placement gate would 409-park every SDK client that checks /v1/models and then asks for one."
+        }
+      }
+    },
+    "qwen3.8-flash-next": {
+      "display": "Qwen3.8 Flash Next",
+      "engine": "strata",
+      "unit": "uap-strata-server",
+      "service": "strata-server",
+      "launch": "/home/cogtek/dev/strata/run-iq3_s.sh",
+      "endpoint": "http://127.0.0.1:8080",
+      "task_affinity": ["code", "general"],
+      "affinity": { "device": ["gpu0"] },
+      "configs": {
+        "strata-iq3_s": {
+          "status": "unmeasured",
+          "note": "active backend on this host; run `uap models measure` to record its footprint"
+        },
+        "strata-swift-iq3_s": {
+          "status": "unmeasured",
+          "engine": "strata",
+          "unit": "uap-strata-server",
+          "note": "Swift 1.5 (UkisAI reasoning-efficient Flash-Next post-train) GSQ-RCO IQ3_S on the strata engine; A/B trial vs strata-iq3_s since 2026-10-17 with identical tuning (spec 8/0.70, vram-reserve 3072, int8 KV resident 32768, 131k). Weights: ukisai Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S, sha256-verified, PLE table in shard 1 (setup-detected); pack strata-data/packs/swift-iq3_s; config strata-swift-iq3_s.json behind uap-strata-server.service (revert: point the unit --config back at strata-iq3_s.json, daemon-reload, restart). Run `uap models measure` to record this machine's footprint before it can load (fail-closed). Trial evidence: /home/cogtek/dev/strata/bench/results/2026-10-17-swift-iq3s-trial/."
+        }
+      },
+      "advertises": ["qwen3.8-flash-next-iq3_s", "swift-1.5-iq3_s", "gpt-5.1-codex"]
+    },
+    "qwen3.8-27b": {
+      "display": "Qwen3.8 27B",
+      "engine": "llama.cpp",
+      "task_affinity": ["code", "general"],
+      "affinity": { "device": ["gpu0"] },
+      "endpoint": "http://127.0.0.1:8080",
+      "advertises": ["Qwen3.8-27B"],
+      "configs": {
+        "llama-mtp": {
+          "status": "unmeasured",
+          "engine": "llama.cpp",
+          "unit": "uap-llama-server",
+          "profile": "config/llama-profiles/qwen38-27b-mtp.env"
+        },
+        "exl3-4bit": {
+          "status": "unmeasured",
+          "engine": "exl3",
+          "unit": "uap-exl3-server",
+          "profile": "config/exl3-profiles/qwen38-27b-exl3.env"
+        },
+        "flashnext": {
+          "status": "unmeasured",
+          "engine": "llama.cpp",
+          "unit": "uap-flashnext-server"
+        },
+        "signal": {
+          "status": "unmeasured",
+          "engine": "llama.cpp",
+          "unit": "uap-signal-server"
+        },
+        "gsq-rco": {
+          "status": "unmeasured",
+          "engine": "llama.cpp",
+          "unit": "uap-gsq-rco-server",
+          "service": "gsq-rco-server"
+        },
+        "ampere": {
+          "status": "unmeasured",
+          "engine": "llama.cpp",
+          "unit": "uap-ampere-server"
+        },
+        "swift": {
+          "status": "unmeasured",
+          "engine": "llama.cpp",
+          "unit": "uap-swift-server"
+        }
+      }
+    }
+  }
+}