opencodezen4dsh

---

opencodezen4dsh is a maintained fork of FishBottle7/opencode2dsh (upstream 0.3.11, MIT), developed at defaver04/opencodezen4dsh. It keeps the upstream's native-adapter architecture and adds three fork-owned improvements — see CHANGELOG.md (fork.2) and CHANGELOG.md (fork.1):

1. xhigh and max thinking levels for every reasoning model. reasoningEfforts() unions the models.dev-declared ladder with FORCED_EFFORT_LEVELS = ['xhigh', 'max'], so models whose metadata declares no ladder (big-pickle, mimo-v2.6-flash-free, …) or a short one (step-5-preview-free, muse-spark-*-contributor-free) no longer hide levels the Zen gateway accepts (reasoning_effort: max and reasoning.effort: xhigh both answer 200 live).
2. muse-spark-1.2-contributor-free and muse-spark-1.3-contributor-free are unblocked. The 403 RegionError recorded upstream in issue #56 does not reproduce from every exit IP; both ids answer HTTP 200 on POST /zen/v1/responses, so they are back in the picker (7 → 9 exposed models on the live catalog). Still blocked, with re-probe evidence: jev-1.13-free (400 ModelProtocolUnsupported, SystemOne-only), exo-free and ling-3.0-flash-fin-free (gone from GET /v1/models).
3. Identity-less tool calls can no longer corrupt the session log (fork.2). When a free-lane model answers a tool turn with a text-shaped <tool_call><parameter=…> imitation instead of a native call, pi-ai assembles an id:""/name:"" tool call that used to be persisted verbatim — poisoning the whole history (tool call id requires a nonempty string, session unopenable). Such calls are now demoted to text with their arguments preserved, so the turn continues instead of corrupting the log.

Install from a build artifact instead of patching by hand:

# from this repo (builds lib/ via prepack)
git clone https://github.com/defaver04/opencodezen4dsh.git && cd opencodezen4dsh/packages/plugin
pnpm install && pnpm pack
dsh plugin --profile web add ./opencodezen4dsh-dsh-plugin-<version>.tgz

or grab the .tgz attached to the latest Release (the release workflow runs typecheck + the full test suite, packs, and uploads it). After installing, restart dsh web — Cordis imports plugin modules once per process, so a picker refresh alone never picks up a new bundle.

---

opencodezen4dsh registers a native DSH LlmAdapter that streams directly from OpenCode Zen's anonymous free lane — the same models OpenCode's own CLI uses without an account, served to your DSH model picker as a regular provider called opencodezen4dsh.

Requests leave your machine looking exactly like traffic from the OpenCode CLI (same user agent, same correlation headers), and the model catalog stays fresh through a three-tier fallback chain. There is nothing to log into and nothing to host.

Highlights

• Zero credential, zero setup — the anonymous lane needs no key; install, restart, chat
• Native adapter, no sidecar — one npm package, no child process, no binary, no local port (the legacy Go sidecar is not part of the published package; see legacy/)
• CLI-identical disguise — requests carry the OpenCode CLI user agent and its session/request/project header set, derived per conversation
• Selectable thinking levels — reasoning-capable free models expose an effort picker in DSH's model selector (declared ladders where the model metadata provides them, Off/Minimal/Low/Medium/High otherwise, plus xhigh/max forced by this fork); Responses-only muse-spark-* models expose only upstream-supported effort levels, and no selection keeps the provider default
• Image input where it is declared — image attachments are accepted for every model whose models.dev entry lists image in modalities.input, and refused for the ones that declare text only; models.dev being silent falls back to the live-verified vision family
• Live catalog with a fallback chain — live upstream list ∩ free-by-metadata, falling back to offline cache and a verified static list
• Self-healing — fast startup retries, periodic refresh, and a written health snapshot for diagnostics
• Proper error surfaces — upstream failures (rate limit, auth, timeout, transport) arrive in DSH as classified finish reasons, and retries stay owned by DSH
• Session-log safe — identity-less tool-call lapses from free-lane models are demoted to text instead of poisoning the durable history (fork.2)

Install

From the plugin market (recommended, once this repo is listed there): in DSH open Settings → Plugin Market, search opencodezen4dsh, one-click install.

From a GitHub Release (recommended today):

# download the .tgz attached to the latest Release, then:
dsh plugin --profile web add ./opencodezen4dsh-dsh-plugin-<version>.tgz

From source (build the tarball yourself):

git clone https://github.com/defaver04/opencodezen4dsh.git
cd opencodezen4dsh/packages/plugin
pnpm install && pnpm pack
dsh plugin --profile web add ./opencodezen4dsh-dsh-plugin-<version>.tgz

Verify: restart dsh web, open the model picker, and pick a model from the opencodezen4dsh group.

Requires DSH (DeepSeek Harness) 0.1.7 or newer with a web profile (the IP 池 settings page rides the 0.1.7 Plugins page contract); Node.js ≥ 20 (already present if DSH runs); outbound HTTPS to opencode.ai and models.dev.

Configuration

Defaults work out of the box. Override via the profile's cordis.patch.yml:

- id: opencodezen4dsh
  name: '@opencodezen4dsh/dsh-plugin'
  config:
    mode: adapter        # adapter (default) | sidecar
    providerId: opencodezen4dsh
    refreshSeconds: 300  # catalog refresh cadence

How it works

DSH session
   │  harness chunks (block-start / text-delta / usage / finish …)
   ▼
ZenAdapter (registered LlmAdapter)
   │  pi-ai openai-completions stream (chat models)
   │  pi-ai openai-responses stream (`muse-spark-*`, Responses-only on Zen)
   ▼
https://opencode.ai/zen/v1        ← Authorization: Bearer public
   with CLI-identical headers:
     user-agent: opencode/…
     x-opencode-client, x-opencode-session, x-session-affinity,
     X-Session-Id, x-opencode-request, x-opencode-project

• Tool bridge — the free-tier gate accepts exactly the 11 official OpenCode tools, so the model only ever sees OpenCode schemas. adapter/tool-bridge.ts rewrites each finished call to the host's tool name and schema before it reaches DSH, and replays history in the official shape. Host tools outside the 11 are tunnelled: the model calls task{subagent_type:"dsh:job_output", prompt:"{\"job_id\":\"j1\"}"} (the gate never inspects those two free-text fields) and the adapter turns it into a real job_output{job_id:"j1"}, so DSH permissions and sandboxing still apply. The catalog (name, compact parameters, one sentence) is built from the host's tool list on every request — new and MCP tools need no code change; priority tools (subagent control, jobs, goals, then MCP) survive the size cap first and dropped names stay listed while space allows. Weak models may format a tunnel call wrongly; the host's error goes back to the model. Switches: toolBridge, toolTunnel.
• Session correlation — session/project ids are SHA-256 derived from the conversation's first user turn (stable per conversation, non-reversible), and each request gets a fresh random id, mirroring the CLI.
• Catalog fallback chain — S1: live GET /v1/models; S2: models.dev pricing metadata decides "free"; S3: a compile-time verified static list. A disk cache (~7-day TTL) covers upstream outages.
• Resilience — the adapter registers immediately at startup; if the first catalog fetch races your network (VPN/TUN reconnects, DNS), the plugin retries on a short cadence (~1 min) before settling into the periodic refresh.
• Sidecar mode (mode: sidecar, legacy) — spawns a local Go agent (a single-tenant port of opencode2api) on 127.0.0.1:<random>, token-authenticated, and registers a standard llm-pi-ai route. Not part of the published package; build it from legacy/agent (go build ./cmd/agent) and point agentPath at the binary.

Health & troubleshooting

The plugin writes a health snapshot after every refresh round:

~/.opencodezen4dsh/adapter-status.json

{
  "status": "ready",
  "total": 64,
  "exposed": 9,
  "lastError": "",
  "writtenAt": "2026-08-29T07:01:54.915Z"
}

Security

• No secrets involved: the anonymous lane's key is the literal string public; nothing is stored, nothing telemetry.
• Install paths restricted to lib/ only; no build scripts run from dependencies.
• All requests go directly from your machine to opencode.ai / models.dev.

Development

git clone https://github.com/defaver04/opencodezen4dsh.git
cd opencodezen4dsh/packages/plugin
pnpm install
pnpm typecheck && pnpm test
pnpm build                    # bundle to lib/

The legacy Go sidecar lives in legacy/agent (go test ./...). Architecture notes and the porting record live in docs/.

Releasing: pnpm pack in packages/plugin (prepack builds and syncs docs); push a v* tag and the release workflow typechecks, tests, packs, and attaches the .tgz to the GitHub Release.

Acknowledgments

• opencode2dsh by @FishBottle7 — the upstream project this fork builds on (0.3.11, MIT). Fork-owned changes are listed at the top of this file and in CHANGELOG.md.
• opencode2api by @jasonxu114514 — the legacy Go sidecar in legacy/agent is a port of its anonymous-lane implementation, and the catalog fallback chain and request-disguise details are derived from it.
• OpenCode — for running the free anonymous Zen lane.
• @earendil-works/pi-ai — the wire layer used by adapter mode.
• DeepSeek Harness and the dsh-market community.

License

MIT © FishBottle7 (upstream) and defaver04 (fork). See LICENSE.