kilo-zen2dsh registers native DSH LlmAdapters backed by two independent OpenAI-compatible gateways:

• kilo2dsh → Kilo Gateway (/api/gateway/models and /api/gateway/chat/completions)
• opencode2dsh → OpenCode Zen (/zen/v1/models and the model-specific /chat/completions or /responses endpoint)

Kilo requests are genuinely keyless: the plugin suppresses the generated Authorization header on the wire. Zen's compatibility lane uses the public Bearer public placeholder and OpenCode-compatible request markers by default; the gateway may still require its own anonymous eligibility, quota, or a user account. Neither provider is an authentication or billing bypass.

Highlights

• Native adapter, no child process or local port in the published package.
• Dynamic free-model discovery using Kilo's isFree/is_free flag, with kilo-auto/free and :free suffix compatibility fallbacks.
• Independent OpenCode Zen catalog and adapter (opencode2dsh) with the documented free model IDs, a separate cache, and Responses API routing for muse-spark-1.2-contributor-free.
• Text-output and tool-capable models only, suitable for DSH agent turns.
• Startup retry, periodic refresh, seven-day disk cache, and a health snapshot.
• Optional authenticated Kilo or Zen token for compatible deployments; no account token is enabled by default.

Install

dsh plugin --profile web add @huanx/kilo-zen2dsh

When using this checkout before the npm package is published, build a local tarball instead:

cd packages/plugin
pnpm install
pnpm pack
dsh plugin --profile web add ./huanx-kilo-zen2dsh-0.3.0.tgz

Restart dsh web, open the model picker, and choose a model in either the kilo2dsh (Kilo) or opencode2dsh (Zen) provider group. Node.js 20 or newer is required.

Configuration

The defaults are keyless and point at the public Kilo gateway:

- id: kilo2dsh
  name: '@huanx/kilo-zen2dsh'
  config:
    mode: adapter
    providerId: kilo2dsh
    gatewayBaseUrl: https://api.kilo.ai/api/gateway
    refreshSeconds: 300
    requireTools: true
    zenEnabled: true

To use an authenticated gateway token, set upstreamApiKeyEnv explicitly. Leaving it empty is intentional; the plugin will not pick up KILO_API_KEY (or another ambient secret) by accident.

Set zenEnabled: false when you only want Kilo, or when the Zen gateway is returning an anonymous-lane 429. sidecar mode is the legacy Go bridge and currently exposes Kilo only; the two-provider setup is implemented in native adapter mode.

Wire behavior

Kilo:
  GET  https://api.kilo.ai/api/gateway/models
  POST https://api.kilo.ai/api/gateway/chat/completions
       (no Authorization header for the default free lane)

OpenCode Zen:
  GET  https://opencode.ai/zen/v1/models
  POST https://opencode.ai/zen/v1/chat/completions
       (Bearer public + OpenCode compatibility headers)
  POST https://opencode.ai/zen/v1/responses
       (used for Responses-only free models)

The Kilo adapter sends ordinary OpenAI-compatible JSON/SSE plus Kilo correlation headers (X-KILOCODE-EDITORNAME, X-KILOCODE-TASKID, and X-KILOCODE-PROJECTID). The Zen adapter sends the OpenCode compatibility marker and correlation headers required by the current anonymous lane. This is an upstream compatibility requirement, not a guarantee that Zen will accept third-party clients.

The model filter is deliberately conservative:

1. isFree or is_free is authoritative when present.
2. If the flag is absent, kilo-auto/free, openrouter/free, and IDs ending in :free/-free are accepted.
3. Image-output models and models without advertised tool support are hidden by default.

Automatic context and output-limit reconciliation

Kilo model-directory records are not always internally consistent: a limit may be present on the top-level record, top_provider, or a nested limit(s) object, and compatible gateways use several spellings. The adapter merges those declarations and uses the smallest positive context limit. It keeps the advertised context window for prompt history, while calculating a separate safe output budget.

The current gateway catalog advertises minimax/minimax-m3:free with a 943,718-token completion limit, but the backend rejects requests above 524,288. The adapter therefore applies a 524,288-token gateway compatibility ceiling (or the smaller model-specific limit) to defaultMaxTokens and to both max_tokens and max_completion_tokens immediately before the wire request. An oversized default supplied by DSH is consequently reduced automatically; callers do not need to edit their model configuration. This keeps a 1M-token context declaration available for input while preventing the 400 error shown by MiniMax-M3. The compatibility ceiling is Kilo-specific; OpenCode Zen keeps the limits reported by its own catalog.

For Zen, the public model directory currently contains minimal OpenAI records, so the adapter accepts the documented big-pickle exception and IDs ending in -free/:free, while treating an explicit future isFree/is_free flag as authoritative. The live catalog replaces the bootstrap list after a successful refresh.

Health and troubleshooting

Kilo adapter status is written to ~/.kilo2dsh/adapter-status.json; Zen status is written to ~/.kilo2dsh/zen-adapter-status.json. Their caches are kilo-models.json and zen-models.json respectively.

Kilo's anonymous free usage is controlled by Kilo and is rate-limited by IP (the current documentation describes 200 requests per hour per IP). A 429 or an unavailable model is an upstream service decision; wait, choose another free model, or configure an authenticated token.

Zen's free models are promotional and can be withdrawn or rate-limited. The current OpenCode documentation describes them as limited-time models, and the gateway has rejected non-OpenCode user agents with 429 FreeUsageLimitError in some deployments. Treat Zen as best-effort, avoid sending confidential data to free models, and use zenApiKeyEnv if your Zen account requires a token.

For the optional legacy bridge:

cd legacy
go build ./cmd/agent
go test ./...

It keeps a local authenticated /v1 endpoint for DSH, but its upstream Kilo requests use the same keyless behavior and /api/gateway paths. Use native adapter mode for the independent Zen provider.

Adaptation source

This project is adapted from the FishBottle7/opencode2dsh DSH/OpenCode integration prototype. The Kilo adapter is the migration target; the original Zen free-tier behavior is retained as a separate optional native adapter rather than being mixed into Kilo's keyless transport. The Kilo, Zen, and QwenPaw protocol references are listed below.

Development

cd packages/plugin
pnpm install
pnpm typecheck
pnpm test
pnpm build

References

• Kilo Gateway authentication
• Using Kilo for Free
• Kilo Gateway API reference
• OpenCode Zen documentation
• OpenCode provider source (apiKey: public free lane)
• OpenCode issue: anonymous Zen user-agent behavior
• QwenPaw Kilo provider
• @earendil-works/pi-ai

License

MIT © FishBottle7