dsh-aiservice-deeplink

AI Service Deeplink (aiservice://) import and export for DSH model configuration. Implements the transport and configuration layers of the protocol draft v1 against the composition's llm-pi-ai provider routes.

A single aiservice://model, aiservice://provider, or aiservice://bundle link carries an endpoint, a wire protocol, a credential, model capabilities, and inference and transport policy. This package turns such a link into DSH provider routes, and turns the configured routes back into a link.

Surfaces

Layout

Where it is mounted

~/.dsh/profiles/web/cordis.patch.yml inserts one row pointing at lib/index.js by absolute path. The package is deliberately dependency-free: the row is not part of the profile's pnpm project, so it cannot import @deepseek-ai/*, and every host service is reached through the declared inject list or through ctx.get at request time.

Mapping

Two adapters own routes in this composition, and import and export treat them differently.

Export reads both namespaces, so the built-in DeepSeek route shares like any other. llm-pi-ai is a registry keyed by route; llm-deepseek owns exactly one fixed route (deepseek-official) with a flat section, and is projected onto the same shape — endpoint, credential reference, thinking policy, and the resolved catalog, which the adapter supplies as schema defaults. The exported endpoint.provider is deepseek, a catalog identifier a receiver can infer defaults from, rather than the harness's own route name.

Two limits are worth knowing:

• The adapter falls back to DEEPSEEK_BASE_URL from the trusted launch environment when no baseURL is configured, and that value is not readable from a plugin. An explicitly configured endpoint travels; otherwise the link carries DeepSeek's public endpoint.
• The adapter's own reasoningEffort default is applied per request and never appears in the resolved settings, so an unconfigured effort is omitted rather than guessed. An explicit one travels.

Reasoning is the one field that can break a whole route

providers.<route>.reasoning is a hard default applied to every model on the route, and a request naming a level a model does not support is refused with UNSUPPORTED_REASONING_EFFORT. A hand-declared model declares no reasoning at all, so it supports only off. Hoisting one model's effort onto the route therefore makes every other model unusable — which is what an unconditional mapping from inference.reasoning.effort does.

So the effort is handled per model:

• mode: disabled → reasoningEfforts: false, the harness's own way of declaring a non-reasoning model.
• An explicit effort the payload also equips for the wire (it names compatibility.thinking_format or supports_reasoning_effort) → reasoningEfforts: { off: null, <level>: <level> } on that model, so the level exists and the session picker offers it.
• An explicit effort with nothing saying how the level reaches the wire → the model is left alone and the loss is reported. Declaring the level anyway would advertise a control that sends nothing, which is the same fabrication as inventing a wire spelling.
• A route-wide reasoning default is set only when every model on the route can take the same level. Note this is about what each model supports, not about the models agreeing: provider defaults make every model resolve to the same effort, and that is exactly the case that would otherwise force a level no model was equipped to send.

Import always writes llm-pi-ai. The native DeepSeek route is a single-vendor adapter, not a place to point at an arbitrary endpoint, and route names are unique across the whole llm registry — so an imported link never claims a name another adapter owns, even when its endpoint matches.

Everything the mapping cannot carry (transport.user_agent, inference.temperature, inference.reasoning.history and budget_tokens, retry.max_retry_after_seconds, a completions-only switch on an Anthropic route) is reported in the preview's notes rather than dropped silently. On export the same applies to a thinking level above every level the protocol names: the harness's max is shared as xhigh and the substitution is reported beside the link.

Guarantees

• Parsing never has side effects. No network, no writes; discovery and credential storage happen only inside import.
• Failure is atomic. The whole change goes through one settings.update('llm-pi-ai', …) call, which validates before persisting, so a rejected import leaves the previous configuration untouched (§7).
• Discovery failure does not erase configuration. A provider that declares models keeps them when its endpoint is unreachable (§4.2.1).
• Credentials stay out of logs and errors. Responses and diagnostics carry error codes and paths, never the URI or a secret (§6, §7).
• Sharing includes credentials by default, and the panel offers the credential-free mode, which writes credential.type = "prompt" instead (§6.1).
• An import never replaces a credential that is already there. The preview reports whether the reference is new, already stored, or supplied by the launching environment, and replacing a stored value takes an explicit overwriteCredentials.
• An import writes only the routes it names. Nothing else is restated, so a route the user edited concurrently, and the stored formatting of every other entry, both survive.

Credentials

The harness owns two disjoint key spaces, and an imported API key belongs to the first:

A pi-ai route names a reference, so that is what this plugin writes: the same <PROVIDER_UPPER_SNAKE>_API_KEY derivation the Models page uses, through the same credentials.set. References are layered against the environment, and the inherited process environment outranks the managed file — so a write the launching shell would shadow is refused by the provider, and this plugin checks credentials.describe first and reports it instead of failing mid-import.

Because a reference is a shared name rather than a private slot for one link, the derived name is never written over a value that is already there. A link naming anthropic would otherwise replace whatever key the user had stored for that provider.

Tests

node --test 'test/*.test.mjs'

test/protocol.test.mjs reproduces the protocol repository's published model.uri, provider.uri, and bundle.uri byte for byte and checks the rejections the spec requires. Point AISERVICE_EXAMPLES at another checkout of that repository to run against a different revision.

Changing it

The profile mounts the host half by absolute path, and Node caches an ES module by URL: edits to lib/*.js reach the running process only after dsh web restarts. The client bundle additionally needs a page refresh.