dsh-dario-sidecar

Harness-owned lifecycle for dario — a DeepSeek Harness (DSH) profile plugin.

In plain terms

You probably pay for a ChatGPT or Claude subscription. That plan works in the vendor's own app — but most other tools, including the DeepSeek Harness this plugin is written for, expect an API key. An API key bills you per token, separately, on top of the subscription you already pay for.

dario closes that gap. It's a small program that runs on your machine and lets any OpenAI- or Anthropic-compatible tool use your subscription instead of an API bill.

One catch: dario only works while it's running. Normally that means opening a terminal, starting it yourself, and hoping you remember after every reboot — and when you forget, your tools fail with confusing errors.

This plugin wires dario into your harness's own startup: harness starts → dario starts. Harness stops → dario stops. Install it once, stop thinking about it. The rest of this README covers the details.

What it does

• Spawns dario proxy through DSH's subprocess seam when the plugin mounts. Default args ['proxy'] serve every plan dario holds credentials for; override args for a single-plan posture (see Configuration).
• Readiness-gates on GET /v1/models answering 2xx before reporting ready (default 120 s budget, 500 ms polls), so a slow proxy surfaces as "starting", not as a model-route credential error.
• Adopts an already-healthy listener instead of spawning a doomed duplicate. Adoption is not ownership: teardown kills only what this plugin started.
• Tears down effect-scoped: SIGTERM → grace (5 s default) → SIGKILL, tree-scoped, on unload/HMR — the proxy never outlives the harness.
• Surfaces mid-life exits in the harness log with exit code/signal.

Prerequisites

• A working DSH deployment with a profile (e.g. web).

• dario installed and logged in on the host the harness runs on:

  npm install -g @askalf/dario
  dario add altman   # attach a ChatGPT Plus/Pro plan (paste the redirect URL back)
  dario login        # and/or a Claude Pro/Max plan
  

  This plugin manages the proxy process only — OAuth login is a human step it deliberately never performs.

Install

Managed route (wires dependency + bundle entry; the package's own patch supplies the loader row, so defaults apply):

dsh plugin --profile web add github:GrannyProgramming/dsh-dario-sidecar

Manual route — add the dependency to your profile's package.json:

{
  "dependencies": {
    "dsh-dario-sidecar": "github:GrannyProgramming/dsh-dario-sidecar"
  }
}

install profile dependencies, then pick one of the two ways to mount it:

• Bundle entry (defaults, no config): add "dsh-dario-sidecar" to dsh.profile.bundles. The package's own patch inserts the loader row.

• Manual insert (you want to override config, e.g. args): leave it out of the bundles list and add the row to the profile's cordis.patch.yml user patch layer instead:

  - insert:
      - id: dsh-dario-sidecar
        name: 'dsh-dario-sidecar'
        config:
          args: ['proxy', '--no-claude-auth']   # your override
  

⚠️ These two are mutually exclusive. A bundle entry and a manual insert for the same id is a duplicate loader entry id boot failure — the loader refuses the whole tree, not just the duplicate row.

Configuration

Everything lives in the insert row's config:

- insert:
    - id: dsh-dario-sidecar
      name: 'dsh-dario-sidecar'
      config:
        args: ['proxy', '--no-claude-auth']   # example: ChatGPT-only posture

Field	Default	Meaning
command	dario	Executable to spawn (absolute, or resolved on PATH at load).
args	['proxy']	Arguments, no shell. ['proxy','--no-claude-auth'] never loads the Claude OAuth path — useful with a dead/stale Claude credential or a ChatGPT-only deployment.
host / port	127.0.0.1 / 3456	Readiness-probe target; passed to the child as DARIO_HOST / DARIO_PORT.
env	{}	Extra env vars merged onto the scrubbed ambient env (PATH/HOME survive).
readyTimeoutMs	120000	Readiness budget before the spawn is torn down and load fails.
pollIntervalMs	500	Interval between readiness polls.
killGraceMs	5000	SIGTERM → SIGKILL escalation grace.

Point a model route at it

With the sidecar up, add dario as an llm-pi-ai custom provider in $DSH_HOME/settings.yaml:

llm-pi-ai:
  providers:
    dario:
      displayName: Dario (subscription)
      api: openai-completions
      baseURL: http://127.0.0.1:3456/v1
      apiKeyEnv: DARIO_API_KEY   # any non-empty value; dario accepts a literal "dario"
      compat:
        supportsDeveloperRole: true
        maxTokensField: max_completion_tokens
      models:
        - id: gpt-5.6-sol
          name: GPT-5.6 Sol
          contextWindow: 272000
          maxTokens: 128000
          input: [ text, image ]
          reasoningEfforts: { off: none, low: low, high: high, max: max }

Verify what your account serves before declaring models: curl -s localhost:3456/v1/models. The listing is static — a live /v1/chat/completions round-trip is the only real auth proof.

Troubleshooting

• Load fails not ready within …ms; stderr tail: … — dario started but never served: check dario status, purge stale credentials (dario logout), or raise readyTimeoutMs.
• resolveExecutable failure — dario is not on the harness process's PATH. Set command to the absolute binary path.
• Port conflict — set port in the insert row and the matching baseURL in your model route.
• WSL + Windows both have dario — the two installs shadow each other and version-split (hash -r to re-resolve). Keep exactly one on the harness host, or pin command to the right binary.
• Parallel agents — subscription tokens flag under concurrency; dario is strictly single-flight. Route parallel fan-out to per-token API providers instead.

Origin

Extracted from a working tiered-routing deployment; the design was fixed and verified in council_of_agents #7 (lifecycle mechanism) and #5 (integration + verification). Shape follows DSH's own @deepseek-ai/dsh-lsp-stdio subprocess-lifecycle pattern.

License

MIT