@vllm-sr/dsh 0.0.0-stage → 0.3.1

package/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

package/README.md

@@ -1,3 +1,129 @@
-# Temporary Holding Version
+# @vllm-sr/dsh
 
-This version is a temporary placeholder for this package. An operational version to replace this has been submitted for review and is awaiting a staged release.
+**System Duo** for [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness): *Thinking, fast and slow. One
+key.*
+
+An agent's LLM thinks slowly, in tokens. System One, a decision model, answers the many small questions around it in
+about half a second each, with calibrated probabilities: is this command dangerous, does this file try to give the
+agent orders, is the work really done, which team should get each of these 230 tickets.
+
+- **System Two**, the LLM, plans and acts: `vllm-sr/auto` from [vLLM Semantic Router](https://github.com/vllm-project/semantic-router),
+  which sends each step to the right open model.
+- **System One**, the decision model, decides: `vllm-sr/d3-ultra`. It answers typed questions (choice, yes/no, score,
+  box) with probabilities, from one forward pass per question instead of generated text.
+
+Both come from [System Duo](https://app.vllm-sr.ai) with one free key, and the plugin gets that key for you: System One
+answers at `https://app.vllm-sr.ai/v1/systemone`, and `vllm-sr/auto` at `https://app.vllm-sr.ai/v1/chat/completions`.
+
+## Quick start
+
+```bash
+npm install -g @deepseek-ai/dsh                 # DeepSeek Harness 0.2.1-rc.5 or a later 0.2 release, Node 24.9 or later
+dsh plugin --profile web add @vllm-sr/dsh
+dsh web
+```
+
+The package can also come straight from GitHub: `dsh plugin --profile web add github:Xunzhuo/dsh-system-one-plugin`.
+
+On its first start the plugin gets one free System Duo key, with the same small proof of work the site's "Get your
+API key" button runs in the browser, and saves it in DeepSeek Harness's credential store and in `~/.dsh/.env` (readable
+by you only). A notice in the session says so. `vllm-sr/auto` becomes the default model; choose another model
+in the app and System One keeps working beside it.
+
+Headless:
+
+```bash
+dsh plugin --profile headless add @vllm-sr/dsh
+dsh headless "Triage tickets.jsonl: route each ticket, score its urgency, flag churn risk"
+```
+
+A headless run reads its key when it starts. If there is none yet, the first run fetches and saves it, and the LLM
+answers from the next run (or start `dsh web` once first). In CI, set `VLLM_SR_API_KEY` to a key from
+[app.vllm-sr.ai](https://app.vllm-sr.ai) instead; a key you set yourself is always used and never replaced.
+
+Had the earlier `dsh-system-one` 0.2 from GitHub or a local path? Replace it with
+`dsh plugin --profile web remove dsh-system-one`, then add `@vllm-sr/dsh` as above. Your keys and the `system-one` row
+of your profile patch stay as they are.
+
+## What it does
+
+| Reflex | Hook | What System One answers | What happens |
+| --- | --- | --- | --- |
+| Triage | `agent/pre-step` | kind of work, effort, unclear?, risky? | a one-line brief joins the request, with hints (use `system_one` for batch work, ask first when unclear) |
+| Guard | `tools/pre-execute` | risk (Safe, Normal, Dangerous, Critical), leaks secrets?, on task? | the call runs, waits for your approval (Web UI), or is denied with the reason (headless) |
+| Shield | `tools/post-execute` | does this passage give orders to an AI agent? | flagged passages are scored again line by line, and only the injected lines are removed |
+| Done-check | `agent/turn-stopping` | complete?, checked?, waiting for you? | an unfinished turn goes back once or twice to finish and verify |
+| Tool | `system_one` | any rubric the LLM writes, over a file of items or images | one result row per item (JSONL or CSV) and a summary |
+
+## Your data
+
+System One sends what it decides about: the user's request (clipped), each tool call it reviews, the passages of tool
+results it screens, the end of the turn, and the items given to the `system_one` tool. System Two receives the
+conversation as any LLM provider does. System Duo logs requests to improve the services and to train models; read its
+[terms](https://app.vllm-sr.ai/terms) and [privacy](https://app.vllm-sr.ai/privacy) pages before you use it with
+private code or data. Each reflex can be turned off.
+
+The local trace (`~/.dsh/system-one/trace.jsonl`, readable by you only) records every decision with its probabilities
+and redacts secret-shaped strings; it never records keys. Set `trace: off` to disable it.
+
+## Configuration
+
+Every field has a default. Set any of them on the `system-one` row of a profile patch
+(`~/.dsh/profiles/<name>/cordis.patch.yml`):
+
+```yaml
+- id: system-one
+  config:
+    model: vllm-sr/d3-ultra        # vllm-sr/d3-flash is a little faster; d3-ultra is the most accurate
+    autoKeys: true                 # fetch the free System Duo key on first use
+    guardAskMode: ask              # what a flagged call does: ask (wait for approval) or deny
+    guardAskAt: 1.8                # risk score (0 Safe to 3 Critical) that asks
+    guardDenyAt: 2.5               # risk score that always denies
+    guardExfiltrationAt: 0.7       # "leaks secrets" probability that denies
+    guardTools: [bash, write, edit, apply_patch, web_fetch]
+    shieldTools: [web_fetch, read, bash]
+    triage: true
+    doneCheck: true
+    tool: true
+    trace: ~/.dsh/system-one/trace.jsonl   # or off
+```
+
+The key comes from the credential reference `VLLM_SR_API_KEY` (`llmKeyEnv`), which both systems use. A separate
+System One key in `DECISION_API_KEY` (`apiKeyEnv`), as plugin 0.2 kept it, still wins for System One; a 0.2 key for
+`vllm-sr/auto` keeps working too.
+If the decision service cannot be reached, a guarded call falls back to `guardOnError` (default `allow`); the
+harness's own sandbox and approvals still apply.
+
+## Measured
+
+On our recorded runs, from one workstation with the network included:
+
+| What | Result |
+| --- | --- |
+| A decision | p50 0.49 s, p90 0.52 s (53 guard decisions) |
+| Guard | 35 of 36 right on our 36-command probe set: 16 of 16 risky commands stopped, 19 of 20 safe ones allowed |
+| Shield | 22 of 22 right on our 22-passage probe set: 10 of 10 injections caught, 12 of 12 clean passages kept |
+| Triage of 230 tickets | 121 s and 117,877 LLM prompt tokens, against 543 s and at least 510,986 for the LLM alone |
+| UI screenshots | 5 of 5 broken screens found and boxed, no false alarms, 3.96 s for 12 |
+
+The probe sets are ours and small: read them as a smoke test, not a benchmark.
+
+## Develop
+
+```bash
+npm test                                         # unit tests with a fake decision service and a fake key site
+VLLM_SR_API_KEY=... node probes/guard-probe.mjs vllm-sr/d3-ultra      # live calibration of the guard
+VLLM_SR_API_KEY=... node probes/shield-probe.mjs vllm-sr/d3-ultra     # live calibration of the shield
+```
+
+A `v*` tag publishes the package to npm from GitHub Actions with trusted publishing, with no npm token in the
+repository; see [RELEASING.md](https://github.com/Xunzhuo/dsh-system-one-plugin/blob/main/RELEASING.md).
+
+## Trademarks
+
+DeepSeek Harness is a trademark of DeepSeek, named here only to identify the harness the plugin works with.
+The plugin is not a DeepSeek product, and DeepSeek does not endorse System Duo.
+
+## License
+
+Apache-2.0

package/cordis.patch.yml

@@ -0,0 +1,29 @@
+# System Duo for DeepSeek Harness: System One (decision models) on the agent loop, and System Two, the free LLM
+# vllm-sr/auto from vLLM Semantic Router, both at app.vllm-sr.ai with one key fetched on first use. A model chosen in
+# the app's settings still wins, and a profile patch can replace either row.
+- insert:
+    - id: system-one
+      name: '@vllm-sr/dsh'
+
+- id: llm-pi-ai
+  config:
+    providers:
+      vllm-sr:
+        displayName: System Duo (vLLM Semantic Router)
+        apiKeyEnv: VLLM_SR_API_KEY
+        api: openai-completions
+        baseURL: https://app.vllm-sr.ai/v1
+        compat:
+          supportsDeveloperRole: false
+          maxTokensField: max_tokens
+        timeoutMs: 600000
+        models:
+          - id: vllm-sr/auto
+            name: vllm-sr/auto
+            contextWindow: 131072
+            maxTokens: 32768
+
+- id: agent-default-model
+  config:
+    provider: vllm-sr
+    model: vllm-sr/auto

package/package.json

@@ -1,6 +1,76 @@
 {
   "name": "@vllm-sr/dsh",
-  "version": "0.0.0-stage",
-  "stub": true,
-  "description": "Temporary package placeholder for staged publishing"
-}
+  "version": "0.3.1",
+  "description": "System Duo for DeepSeek Harness. Thinking, fast and slow. One key: System One decision models guard, shield, triage and check your agent around System Two (vllm-sr/auto), with one free app.vllm-sr.ai key.",
+  "keywords": [
+    "deepseek-harness",
+    "dsh",
+    "dsh-plugin",
+    "agent",
+    "decision-model",
+    "guardrails",
+    "prompt-injection",
+    "vllm-semantic-router",
+    "vllm-sr",
+    "system-duo",
+    "system-one"
+  ],
+  "author": "Xunzhuo Liu",
+  "license": "Apache-2.0",
+  "homepage": "https://github.com/Xunzhuo/dsh-system-one-plugin#readme",
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/Xunzhuo/dsh-system-one-plugin.git"
+  },
+  "bugs": {
+    "url": "https://github.com/Xunzhuo/dsh-system-one-plugin/issues"
+  },
+  "type": "module",
+  "main": "src/index.js",
+  "exports": {
+    ".": "./src/index.js",
+    "./package.json": "./package.json"
+  },
+  "files": [
+    "src",
+    "cordis.patch.yml",
+    "README.md",
+    "LICENSE"
+  ],
+  "engines": {
+    "node": ">=24.9",
+    "dsh": ">=0.2.1-rc.5 <0.3.0"
+  },
+  "dsh": {
+    "bundle": {
+      "patch": "./cordis.patch.yml"
+    }
+  },
+  "peerDependencies": {
+    "@deepseek-ai/cordis": "^4.0.5",
+    "@deepseek-ai/dsh-llm": ">=0.2.1-rc.5 <0.3.0",
+    "@deepseek-ai/dsh-tools": ">=0.2.1-rc.5 <0.3.0",
+    "@deepseek-ai/schemastery": "^3.18.5"
+  },
+  "peerDependenciesMeta": {
+    "@deepseek-ai/cordis": {
+      "optional": true
+    },
+    "@deepseek-ai/dsh-llm": {
+      "optional": true
+    },
+    "@deepseek-ai/dsh-tools": {
+      "optional": true
+    },
+    "@deepseek-ai/schemastery": {
+      "optional": true
+    }
+  },
+  "publishConfig": {
+    "access": "public",
+    "provenance": true
+  },
+  "scripts": {
+    "test": "node --test test/*.test.js"
+  }
+}

package/src/client.js

@@ -0,0 +1,225 @@
+/**
+ * Decision API client. A decision model answers typed questions (choice, noul,
+ * score, box) about an input with calibrated probabilities, from one forward
+ * pass per question. Connections are reused: a fresh TLS handshake costs more
+ * than the decision itself.
+ */
+import { readFileSync } from 'node:fs'
+import http from 'node:http'
+import https from 'node:https'
+import { homedir } from 'node:os'
+import { setTimeout as sleep } from 'node:timers/promises'
+
+/** Inputs answered by one request: the top-level state plus `limits.max_states` more. */
+export const ITEMS_PER_REQUEST = 9
+
+export class DecisionError extends Error {
+  constructor(message, status = 0, code = undefined) {
+    super(message)
+    this.name = 'DecisionError'
+    this.status = status
+    this.code = code
+  }
+}
+
+function expandHome(path) {
+  return path?.startsWith('~/') ? homedir() + path.slice(1) : path
+}
+
+/**
+ * @param {object} options
+ * @param {string} options.baseURL
+ * @param {string} options.apiKeyEnv - environment variable holding the key
+ * @param {string} [options.apiKeyFile] - mode-600 file holding the key, read when the variable is unset
+ * @param {() => Promise<string | undefined>} [options.resolveKey] - the key from DeepSeek Harness's credential store, asked first
+ * @param {number} [options.maxConcurrency]
+ * @param {number} [options.timeoutMs]
+ * @param {number} [options.retries]
+ */
+export function createDecisionClient(options) {
+  const baseURL = options.baseURL.replace(/\/+$/, '')
+  const maxConcurrency = options.maxConcurrency ?? 12
+  const timeoutMs = options.timeoutMs ?? 30000
+  const retries = options.retries ?? 3
+  const endpoint = new URL(baseURL)
+  const transport = endpoint.protocol === 'http:' ? http : https
+  // Agent steps are seconds apart; fetch's default pool would close the
+  // socket in between and pay a new TLS handshake on every decision.
+  const agent = new transport.Agent({ keepAlive: true, keepAliveMsecs: 15000, maxSockets: maxConcurrency, scheduling: 'lifo' })
+  let active = 0
+  const waiting = []
+
+  function send(path, method, payload, headers, signal) {
+    return new Promise((resolve, reject) => {
+      const request = transport.request(new URL(baseURL + path), { method, agent, headers, signal }, response => {
+        const chunks = []
+        response.on('data', chunk => chunks.push(chunk))
+        response.on('end', () => resolve({ status: response.statusCode ?? 0, headers: response.headers, text: Buffer.concat(chunks).toString('utf8') }))
+        response.on('error', reject)
+      })
+      request.on('error', reject)
+      request.end(payload)
+    })
+  }
+
+  async function apiKey() {
+    const stored = await options.resolveKey?.()
+    if (stored) return stored.trim()
+    const fromEnv = process.env[options.apiKeyEnv]
+    if (fromEnv) return fromEnv.trim()
+    if (options.apiKeyFile) {
+      try {
+        return readFileSync(expandHome(options.apiKeyFile), 'utf8').trim()
+      } catch {
+        return undefined
+      }
+    }
+    return undefined
+  }
+
+  async function acquire(signal) {
+    while (active >= maxConcurrency) {
+      signal?.throwIfAborted()
+      // A cancelled caller leaves the queue at once: a wakeup it took without taking the slot would be lost for good.
+      await new Promise((resolve, reject) => {
+        const wake = () => {
+          signal?.removeEventListener('abort', cancel)
+          resolve()
+        }
+        const cancel = () => {
+          const at = waiting.indexOf(wake)
+          if (at >= 0) waiting.splice(at, 1)
+          reject(signal.reason)
+        }
+        waiting.push(wake)
+        signal?.addEventListener('abort', cancel, { once: true })
+      })
+    }
+    active++
+  }
+
+  function release() {
+    active--
+    waiting.shift()?.()
+  }
+
+  async function post(path, body, signal, limits = {}) {
+    const deadlineMs = limits.timeoutMs ?? timeoutMs
+    const maxRetries = limits.retries ?? retries
+    const key = await apiKey()
+    if (!key) throw new DecisionError(`no System One key: set ${options.apiKeyEnv}, or get a free System Duo key at https://app.vllm-sr.ai`, 401, 'missing_api_key')
+    const payload = JSON.stringify(body)
+    const headers = {
+      authorization: `Bearer ${key}`,
+      'content-type': 'application/json',
+      'content-length': Buffer.byteLength(payload),
+    }
+    for (let attempt = 0; ; attempt++) {
+      const deadline = AbortSignal.timeout(deadlineMs)
+      let response
+      try {
+        response = await send(path, 'POST', payload, headers, signal ? AbortSignal.any([signal, deadline]) : deadline)
+      } catch (error) {
+        if (signal?.aborted) throw error
+        if (attempt < maxRetries) {
+          await sleep(300 * 2 ** attempt)
+          continue
+        }
+        throw new DecisionError(`Decision API unreachable: ${error.message}`, 0, 'network')
+      }
+      if (response.status >= 200 && response.status < 300) {
+        return { json: JSON.parse(response.text), requestId: response.headers['x-request-id'] }
+      }
+      let code
+      try {
+        code = JSON.parse(response.text).error?.code
+      } catch {
+        code = undefined
+      }
+      const retryable = response.status === 429 || response.status === 502 || response.status === 503
+      if (retryable && attempt < maxRetries) {
+        const after = Number(response.headers['retry-after'])
+        const wait = Number.isFinite(after) && after > 0 ? Math.min(after, 10) * 1000 : 400 * 2 ** attempt
+        await sleep(wait + Math.random() * 200)
+        continue
+      }
+      throw new DecisionError(`Decision API ${response.status}${code ? ` ${code}` : ''}`, response.status, code)
+    }
+  }
+
+  /**
+   * Ask questions about one input, or about several inputs at once.
+   * @param {object} request
+   * @param {string} request.model
+   * @param {object} request.questions - shared by every input
+   * @param {Array<{state: string, images?: string[]}>} request.inputs - at most {@link ITEMS_PER_REQUEST}
+   * @param {AbortSignal} [request.signal]
+   * @param {number} [request.timeoutMs] - per-call deadline (default: the client's)
+   * @param {number} [request.retries] - per-call retries (default: the client's)
+   * @returns {Promise<{answers: object[], ms: number, model: string, inputTokens: number, requestId?: string}>}
+   *   answers in input order
+   */
+  async function ask({ model, questions, inputs, signal, timeoutMs: callTimeoutMs, retries: callRetries }) {
+    if (inputs.length === 0 || inputs.length > ITEMS_PER_REQUEST) {
+      throw new RangeError(`one request answers 1 to ${ITEMS_PER_REQUEST} inputs, got ${inputs.length}`)
+    }
+    const [first, ...rest] = inputs
+    const body = { model, state: first.state, questions }
+    if (first.images?.length) body.images = first.images
+    if (rest.length > 0) {
+      body.states = Object.fromEntries(rest.map((input, index) => [
+        `i${index + 1}`,
+        { state: input.state, questions, ...input.images?.length ? { images: input.images } : {} },
+      ]))
+    }
+    await acquire(signal)
+    const started = performance.now()
+    try {
+      const { json, requestId } = await post('/v1/decisions', body, signal, { timeoutMs: callTimeoutMs, retries: callRetries })
+      const answers = [json.answers, ...rest.map((_, index) => json.states?.[`i${index + 1}`]?.answers ?? {})]
+      let inputTokens = json.usage?.input_tokens ?? 0
+      for (const entry of Object.values(json.states ?? {})) inputTokens += entry.usage?.input_tokens ?? 0
+      return { answers, ms: performance.now() - started, model: json.model ?? model, inputTokens, requestId }
+    } finally {
+      release()
+    }
+  }
+
+  /** Open the connection ahead of the first decision; failures are ignored. */
+  function warm() {
+    send('/api/info', 'GET', undefined, {}, AbortSignal.timeout(timeoutMs)).catch(() => {})
+  }
+
+  return { ask, warm, hasKey: async () => Boolean(await apiKey()), close: () => agent.destroy() }
+}
+
+/** Probability that a noul answer is yes, or undefined when the question failed. */
+export function yes(answer) {
+  return typeof answer?.noul === 'number' ? answer.noul : undefined
+}
+
+/** Compact, readable copy of one answer for traces and model-facing text. */
+export function compactAnswer(answer) {
+  if (!answer || answer.error) return { error: answer?.error ?? 'missing', reason: answer?.reason }
+  const round = value => Math.round(value * 1000) / 1000
+  switch (answer.type) {
+    case 'choice':
+      return {
+        choice: answer.choice,
+        confidence: round(answer.confidence),
+        p: Object.fromEntries(Object.entries(answer.probabilities ?? {}).map(([k, v]) => [k, round(v)])),
+      }
+    case 'noul':
+      return { yes: round(answer.noul) }
+    case 'score':
+      return {
+        score: round(answer.score),
+        confidence: round(answer.confidence),
+        p: Object.fromEntries(Object.entries(answer.probabilities ?? {}).map(([k, v]) => [answer.legend?.[k] ?? k, round(v)])),
+      }
+    case 'box':
+      return { box: answer.box, confidence: round(answer.confidence) }
+    default:
+      return answer
+  }
+}