@miller-tech/uap 1.36.1 → 1.38.0

package/README.md

@@ -89,7 +89,7 @@ uap setup -p all
 | Deploy Batching    | 1 module       | Squash, merge, parallelize deploy actions across agents                          |
 | Policy Enforcement | 8 modules      | Store, evaluate, and enforce operational policies with audit trail               |
 | Browser            | 1 module       | Stealth web automation via CloakBrowser (Playwright drop-in)                     |
-| MCP Router         | 11 modules     | 2-tool meta-router + expert-consultation registry (98% token savings)            |
+| MCP Router         | 12 modules     | 3-tool meta-router (discover/execute/deliver) + expert-consultation registry (98% token savings) |
 | Models             | 10 modules     | Multi-model routing, planning, execution, validation, 13 model profiles          |
 | Delivery Harness   | 14 modules     | `uap deliver`: convergence loop, best-of-N explorer, critic, practice recall, escalation, ideation seeds, HALO tracing, coordination + deploy queueing |
 | Patterns           | 23 patterns    | Battle-tested workflows from Terminal-Bench 2.0                                  |
@@ -295,18 +295,19 @@ await browser.close();
 
 ## MCP Router
 
-Replaces N tool definitions with 2 meta-tools for 98% token reduction.
+Replaces N tool definitions with 3 meta-tools for 98% token reduction: `discover_tools`, `execute_tool`, and `deliver` (auto-routes a coding task into the `uap deliver` convergence loop, which classifies complexity and drives a model to verified completion against real gates).
 
-### Components (10 modules)
+### Components (11 modules)
 
 | Component         | File                                  | Purpose                                     |
 | ----------------- | ------------------------------------- | ------------------------------------------- |
-| MCP Server        | `src/mcp-router/server.ts`            | Exposes `discover_tools` and `execute_tool` |
+| MCP Server        | `src/mcp-router/server.ts`            | Exposes `discover_tools`, `execute_tool`, `deliver` |
 | Config Parser     | `src/mcp-router/config/parser.ts`     | Loads MCP configs from standard paths       |
 | Fuzzy Search      | `src/mcp-router/search/fuzzy.ts`      | Tool discovery with fuzzy matching          |
 | Client Pool       | `src/mcp-router/executor/client.ts`   | Manages connections to MCP servers          |
 | Tool Execute      | `src/mcp-router/tools/execute.ts`     | Tool execution with policy gate             |
 | Tool Discover     | `src/mcp-router/tools/discover.ts`    | Tool discovery definitions                  |
+| Tool Deliver      | `src/mcp-router/tools/deliver.ts`     | Routes a task into `uap deliver` (sandbox-confined subprocess) |
 | Output Compressor | `src/mcp-router/output-compressor.ts` | Compresses tool output                      |
 | Session Stats     | `src/mcp-router/session-stats.ts`     | Per-tool token consumption tracking         |
 
@@ -322,7 +323,7 @@ Tier 2: ModelRouter    -- Assigns optimal model per subtask
 Tier 3: TaskExecutor   -- Executes with validation, dynamic temperature, rate limiting
 ```
 
-### Components (10 modules)
+### Components (11 modules)
 
 | Component          | File                               | Purpose                                      |
 | ------------------ | ---------------------------------- | -------------------------------------------- |
@@ -374,6 +375,7 @@ the model's say-so.
 9. **Deploy batching** (`--deploy`) — on success, queues a commit of the applied files into the deploy batcher; execute with `uap deploy flush`.
 10. **`--optimize`** — one switch for every convergence aid: 4 candidates/turn + critic + practices + escalation + ideation + HALO + coordination (deploy stays explicit).
 11. **Test protection (default)** — pre-existing test/spec files are snapshotted at loop start (case-folded, symlink-alias-aware) and the applier refuses model writes to them, with steering feedback and a prompt warning; test-runner/compiler configs (`vitest.config.*`, `tsconfig*.json`, `jest.config.*`, `pytest.ini`, …) are blocked too, closing gate-rigging by indirection. Protection extends to the spec's **transitive oracle material** — helpers/fixtures/mocks the tests import (by convention or data extension, including through tsconfig path aliases like `@fixtures/*` and `baseUrl` bare imports), quoted fixture paths, reserved missing goldens, and recursive helper chains (the unit under test stays writable). A **runtime integrity guard** hashes every protected file and re-verifies after each gate run: tampering from test code executing during the gates is restored and the gate result discarded. New test files remain allowed. Opt out with `--no-protect-tests`.
+12. **Mission autonomy (default)** — the loop tells the driven model to complete the *entire* task without stopping to ask questions or pausing between phases (state assumptions and proceed; emit a one-line progress note; don't invent unrequested scope). The real gates remain the arbiter. Opt out per-run with `ConvergenceConfig.autonomous: false`. An **operator-guidance channel** lets you steer a running, unattended mission without stopping it: `--guidance-file <path>` is polled each turn and its text is injected as high-priority guidance — write to the file to redirect, clear it to drop the steer. This complements the execution-level `onIteration` directive channel (stop/escalate/switch-model).
 12. **Dynamic optimization (default)** — every instruction is classified for complexity (simple / moderate / complex); non-trivial requests automatically get the aids that improve outcomes (moderate → exploration ×3 + critic + practices + HALO + coordination; complex → the full `--optimize` stack). Any explicit aid flag, `--no-auto`, or `UAP_DELIVER_AUTO=0` disables auto mode. Deploy queueing is never auto-enabled.
 
 ### Components (14 modules)
@@ -443,6 +445,7 @@ uap deliver "..." --ideate --candidates 4 --deploy
 | `--optimize`               | Enable every convergence aid (deploy excluded)                         |
 | `--no-auto`                | Disable dynamic optimization (auto-classified aids are the default)     |
 | `--no-protect-tests`       | Allow modifying pre-existing test files (protected by default)          |
+| `--guidance-file <path>`   | Poll this file each turn; steer a running mission without stopping it    |
 | `--endpoint <url>`         | Override the model endpoint (OpenAI-compatible `/v1`)                  |
 | `--dry-run` / `--json`     | Show the plan only / emit machine-readable result                     |