qwen.md

Qwen with DSH

Abyss observes Qwen through Alibaba Cloud Model Studio's OpenAI-compatible Responses and Chat Completions APIs and Anthropic-compatible Messages API. The shared protocol parsers handle all three; the native DashScope generation API is outside this integration.

Harness, provider, and model remain independent. For example, a DSH request produces agent.name = "deepseek-harness", llm.provider = "dashscope.aliyuncs.com", and llm.model = "qwen3.8-flash". Workspace endpoints and compatible gateways retain their own hostname.

Client configuration

Keep the key in the client environment as QWEN_API_KEY; Abyss does not need the key in its configuration. This patch uses official DSH 0.2.0-rc.2 and qwen3.8-flash with a general Beijing-region Model Studio key:

[
  {
    "id": "agent-default-model",
    "config": {
      "provider": "qwen",
      "model": "qwen3.8-flash",
      "reasoningEffort": "low"
    }
  },
  {
    "id": "llm-pi-ai",
    "config": {
      "providers": {
        "qwen": {
          "apiKeyEnv": "QWEN_API_KEY",
          "api": "openai-completions",
          "baseURL": "https://dashscope.aliyuncs.com/compatible-mode/v1",
          "models": [
            {
              "id": "qwen3.8-flash",
              "contextWindow": 1048576,
              "maxTokens": 2048,
              "reasoningEfforts": {"low": "low"}
            }
          ],
          "compat": {
            "thinkingFormat": "qwen",
            "maxTokensField": "max_tokens",
            "supportsDeveloperRole": false
          }
        }
      }
    }
  }
]

Save as qwen-patch.json. Configure the Abyss CA and TLS policy, then launch:

export NODE_EXTRA_CA_CERTS='/absolute/path/to/abyss-root-ca.pem'
abyss run -- dsh --profile headless --patch /absolute/path/to/qwen-patch.json \
  --json 'Read probe.txt with the read tool and report its contents.'
abyss run -- dsh --profile web --patch /absolute/path/to/qwen-patch.json
abyss run -- dsh-tui --patch /absolute/path/to/qwen-patch.json

For Responses, change api to openai-responses and remove compat. The same base URL yields /compatible-mode/v1/responses.

For Messages, use the official llm-deepseek adapter with a Qwen endpoint:

[
  {
    "id": "agent-default-model",
    "config": {
      "provider": "deepseek-official",
      "model": "qwen3.8-flash",
      "reasoningEffort": "low"
    }
  },
  {
    "id": "llm-deepseek",
    "config": {
      "apiKeyEnv": "QWEN_API_KEY",
      "baseURL": "https://dashscope.aliyuncs.com/apps/anthropic",
      "maxTokens": 2048,
      "models": [
        {"id": "qwen3.8-flash", "contextWindow": 1048576, "maxTokens": 2048}
      ]
    }
  }
]

deepseek-official selects this adapter's registered route; the configured URL determines the provider. Both llm-deepseek and llm-pi-ai belong to the official DeepSeek Harness project.

Select an endpoint matching the key's region and account type. Workspace endpoints use https://{WorkspaceId}.{region}.maas.aliyuncs.com with the same paths. The documented legacy regional hosts remain supported. Coding Plan uses separate keys and endpoints, such as https://coding.dashscope.aliyuncs.com/v1 and https://coding.dashscope.aliyuncs.com/apps/anthropic; these credentials are not interchangeable with general Model Studio keys. See the Coding Plan documentation. Live verification here uses the general Beijing endpoint, not Coding Plan.

Observed content and usage

Chat Completions reasoning_content, Responses reasoning events/items, and Messages thinking blocks map to the reasoning conversation role. Responses can send response.reasoning_text.delta while its final reasoning item contains summary; the snapshot reconciles the streamed text.

Chat Completions argument-only deltas may carry an empty tool id. The parser preserves the previously received nonempty ID and name while assembling the arguments, so the next request's tool result still matches the call.

Messages can correct its provisional message_start usage in message_delta, including reducing input tokens after cache accounting. The parser keeps the latest reported counter, preserves omitted counters, and accepts explicit zero corrections. It merges partial usage before partitioning compatible input totals.

OpenAI-compatible input includes cached tokens; native Messages reports uncached input separately from its cache counters. Normalized usage follows:

total_tokens = input_tokens + cache_read_tokens + cache_write_tokens + output_tokens

Reasoning tokens are a subset of output, not an additional contribution to the total. A missing reasoning counter does not mean no thinking text was returned.

Session attribution and regression

With the tested DSH version, Responses sends session_id and native Messages sends x-deepseek-harness-session-id; both retain the real DSH session. The pi Chat Completions route sends neither, so Abyss uses a per-request flow-* fallback. Model response IDs are not substituted for client sessions.

Credential-free JSON/SSE fixtures cover the three protocols, reasoning, tool roundtrips, empty tool metadata, corrected cumulative usage, cache accounting, provider attribution, and content capture disabled:

cargo test --locked -p abyss-agent-hook --test qwen
cargo test --locked -p abyss-broker --test broker_cli_blackbox qwen

Broker tests exercise a real explicit proxy and plugin socket. Live verification uses DSH headless, dsh-tui with the DSH backend, and DSH Web UI through HTTPS interception, delivery, backend storage, and dashboard display.

Protocol references: Chat Completions, Responses, Messages.