@hecer/yoke 1.10.0 → 1.12.0

package/docs/QWEN-MODEL-SUPPORT.md

@@ -0,0 +1,142 @@
+# Qwen Code, DeepSeek and Kimi
+
+Validated against Qwen Code 0.23.0 on 2026-09-08. Yoke's execution provider remains
+`qwen`: Qwen Code supplies the coding tools and connects to the selected model API.
+DeepSeek and Kimi do not require imaginary `deepseek` or `kimi` executables.
+
+## Setup
+
+Install Qwen Code and configure/authenticate the account you want to use. For an
+existing Qwen Code model configuration:
+
+```sh
+yoke setup . --yes --agent=qwen
+```
+
+The default `qwen-standard` routing worker uses Qwen Code's configured model. It
+no longer assumes an account can access four particular Qwen API models or calls
+the same model different intelligence tiers. Existing custom workers are kept.
+Use `--routing-preset` to explicitly reset old generated workers; otherwise edit
+`.yoke/config.yaml` to retain your own measured profiles. A single standard profile
+cannot satisfy stronger prepared assessments: add suitable explicit workers before
+running those tasks, or choose an appropriate configured routing fallback.
+
+For the opt-in DeepSeek and/or Kimi API presets:
+
+```sh
+yoke setup . --yes --agent=qwen --runner=qwen --model-provider=deepseek,kimi
+```
+
+Provide `DEEPSEEK_API_KEY` and `MOONSHOT_API_KEY` in the environment of the Qwen
+process. Setup stores only their **variable names**, never key values, in
+`.qwen/settings.json`. These are separate API accounts/billing; the Kimi preset
+uses Moonshot's platform API, not Kimi Code subscription credentials. Select only
+providers whose credentials you have configured. Setup does not contact either API
+or prove authentication. Commit the generated project configuration and skills
+before using isolated Git worktrees so the worker receives them.
+
+| Preset | Models installed | Endpoint | Key variable |
+| --- | --- | --- | --- |
+| `deepseek` | `deepseek-v4-flash`, `deepseek-v4-pro` | `https://api.deepseek.com/v1` | `DEEPSEEK_API_KEY` |
+| `kimi` | `kimi-k2.6`, `kimi-k2.7-code`, `kimi-k3` | `https://api.moonshot.ai/v1` | `MOONSHOT_API_KEY` |
+
+New API workers use `openai::MODEL` in Yoke. The adapter translates this to Qwen's
+`--auth-type openai --model MODEL`, selecting the model's endpoint and environment
+key together even when Qwen's default login uses another protocol. Plain model IDs
+keep their existing behavior. The same explicit selector syntax supports `anthropic`,
+`gemini`, `vertex-ai` and `qwen-oauth` when configured in Qwen Code; only DeepSeek and
+Kimi have new automatic API presets.
+
+Example runner or routing worker:
+
+```yaml
+runner:
+  agent: qwen
+  model: openai::kimi-k3
+routing:
+  enabled: true
+  strategy: capability
+  maxCandidates: 3
+  workers:
+    - id: deepseek-standard
+      agent: qwen
+      model: openai::deepseek-v4-flash
+      tier: standard
+      costTier: low
+      capabilities: [implementation]
+    - id: kimi-frontier
+      agent: qwen
+      model: openai::kimi-k3
+      tier: frontier
+      costTier: high
+      capabilities: [implementation]
+```
+
+Setup preserves existing endpoint/key/generation settings for a matching model ID,
+existing custom worker IDs and an existing runner unless explicitly changed. It
+adds missing preset entries and creates backups for changed settings. Regional or
+self-hosted endpoints can be configured directly in Qwen's `modelProviders.openai`.
+Avoid ambiguous duplicate model IDs across endpoints when selecting with `--model`.
+Worker tiers and cost tiers are editable starting hypotheses, not measured rankings
+or prices. Yoke accepts up to 32 configured profiles; this does not increase the
+parallel execution limit.
+
+## Reasoning and permissions
+
+Yoke does not pass a nonexistent Qwen `--effort` flag. Configure model-specific
+reasoning in the matching Qwen `modelProviders` entry's `generationConfig` using
+Qwen's supported parameters. Qwen Code owns streaming, tool execution and replay
+of `reasoning_content` between tool calls. Supported thinking controls differ by
+model and endpoint; no identical thinking levels or benchmark quality is implied.
+
+- `safe`: `--approval-mode auto-edit --sandbox --allowed-tools run_shell_command`.
+  The explicit shell allowance enables headless test/build commands inside the
+  required Qwen sandbox. It is not a command-by-command human approval flow.
+  A working Qwen sandbox backend must be installed; Yoke never silently removes it.
+- `read-only`: plan approval mode plus sandbox, without the shell allowance.
+- `unsafe`: explicit `--yolo`, with no sandbox requested by Yoke.
+- When Yoke owns worker slots, the Qwen `agent`/legacy `task`, `create_sub_session`,
+  `team_create` and `send_message` tools are excluded. This controls native
+  delegation tools, not arbitrary programs invoked from a permitted shell.
+- `bare`, `reasoningEffort` and enabling native multi-agent execution remain
+  explicitly unsupported selections for Qwen.
+
+The RTK hook uses native `PreToolUse` denial plus a retry instruction for simple,
+supported commands. Qwen's documented `updatedInput` is not consumed by the
+0.23.0 execution path, so the hook does not claim transparent argument rewriting.
+It never executes commands or grants permission. Existing RTK commands and complex
+shell expressions are left alone; the generated context also supplies RTK guidance.
+RTK must be installed in the execution environment, including the sandbox.
+Re-running setup/retrofit removes only the obsolete Yoke `BeforeTool` entry,
+preserves unrelated hooks and backs up the original settings.
+
+## Validation and limits
+
+Regression tests cover native result/text/structured-result envelopes, nested
+results, failures, cumulative token/model statistics, safe invocation and native
+worker exclusions, manual skill policy, host/project detection, automatic review
+selection, hook migration, all-agent configuration and idempotent API setup.
+
+Reproduce the manual transport check after building:
+
+```sh
+node scripts/qwen-contract-smoke.mjs /path/to/installed/qwen-code/cli-entry.js
+```
+
+A manual contract smoke test used the **real Qwen Code 0.23.0 CLI** against a local
+synthetic OpenAI-compatible SSE server. Qwen, DeepSeek and Kimi model selections
+completed a read-file tool roundtrip, preserved reasoning content in the subsequent
+request and returned parseable results with cumulative token usage. This local
+transport check used explicit unsafe mode only in a scratch fixture; it does not
+validate a real provider account, model quality, billing, rate limits, Windows or
+a production sandbox backend. No authenticated provider benchmark was performed.
+
+## Primary references
+
+- [Qwen headless output and permissions](https://qwenlm.github.io/qwen-code-docs/en/users/features/headless/)
+- [Qwen model providers and generation configuration](https://qwenlm.github.io/qwen-code-docs/en/users/configuration/model-providers/)
+- [Qwen source](https://github.com/QwenLM/qwen-code): native output adapters,
+  `toolHookTriggers.ts`, `coreToolScheduler.ts`, configuration and provider presets.
+- [DeepSeek API models and endpoints](https://api-docs.deepseek.com/)
+- [DeepSeek thinking tool calls](https://api-docs.deepseek.com/guides/thinking_mode/)
+- [Kimi API models and endpoints](https://platform.kimi.ai/docs/overview)

package/gemini-extension.json

@@ -1,6 +1,6 @@
 {
   "name": "yoke",
-  "version": "1.10.0",
+  "version": "1.12.0",
   "description": "Cross-agent coding harness: curated skill canon, mechanical safety gates, autonomous loop with proof artifacts. CLI: npm i -g @hecer/yoke",
   "contextFileName": "GEMINI-EXTENSION.md"
 }

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@hecer/yoke",
-  "version": "1.10.0",
-  "description": "One harness, three agents, zero trust in \"done\" — cross-agent coding harness for Claude Code, Codex CLI, and Gemini CLI: one skill canon, mechanical safety gates, an autonomous loop with screenshot/video proofs.",
+  "version": "1.12.0",
+  "description": "One harness, four agents, zero trust in \"done\" — cross-agent coding harness for Claude Code, Codex CLI, Gemini CLI, and Qwen Code: one skill canon, mechanical safety gates, an autonomous loop with screenshot/video proofs.",
   "type": "module",
   "bin": {
     "yoke": "dist/cli.js"