Model Actions

An agent's model has one API, whatever serves it. Link Harness uses it for inference and model discovery. An outside harness can keep its provider's native carrier; provider requests and credentials stay in the model bridge.

The Grant

The proxy supplies LNK_MODEL_URL, including its URL secret. Append an action to it. A request cannot supply a provider, address or key: its proved grant selects those. Model names must be plain names and fit the agent spec's allow list. A fallback has a separate grant and URL, LNK_FALLBACK_MODEL_URL.

MethodActionRequestAnswer
GETshowOptional ?name=<name>Model details
GETlistNoneSorted plain model names
GETstatusNoneready, and plain loaded names when known
POSTchatchat::ChatNewline-delimited chat::Stream events
POSTembedtexts, optional namevectors
POSTtranscribename, audiotext
POSTcounttext, optional nametokens
GETusageNoneRequests, unmetered requests and token totals
POSTcontextOptional chatKept chats and one chat's parts (Context Reporting)

show also includes known numeric facts, including answer limits, thinking support and price, when available. The types are in link_model_contract::chat and api. Unknown fields are refused. chat takes a system string, messages, optional tools, answer and window token limits, and optional thinking settings. A message has a user or assistant role, text, and optional calls, tool answers, thinking blocks and media. Media bytes are base64, with an explicit MIME type; they are checked before a provider sees them. The bridge applies its known price instead of trusting a caller's.

Streaming

A chat stream contains text events followed by one answer, or an error with a closed failure kind. Its completed answer includes tool calls, usage and thinking. A truncated stream has no completed answer. Cancelling a client request ends its provider request; no inference is retried automatically. Unsupported operations, including token counting where the provider exposes no tokenizer, return HTTP 501. Counts are never guessed. Usage is the calling agent's measured lifetime total; it does not invent a daily budget or remaining tokens.

Failure kinds

A chat that fails says one of four kinds, in Link's own words with what shows or fixes it. unreached means nothing answered, such as a local runtime that isn't running. too_long means the prompt is past the model's window, overloaded that the model answered 429 or a 5xx, and refused that its provider or runtime turned the request down. The Link Harness falls back to its second model on the first three.

Refused before the stream

A call refused before its stream keeps what the sandbox or bridge said, with its status. A 429 or 503 is overloaded, and any other refusal is refused. A 502 or 504 is unreached: the sandbox couldn't reach or start the model module's bridge, and lnk module list is named as the next step.

Limits

Requests are at most 32 MiB. Provider answers and chat streams are bounded at 8 MiB; discovery and small reads have tighter registry budgets. The bridge refuses redirects and uses action-specific time limits. Provider errors returned through these routes do not include host addresses, paths or credentials.

CLI and Commands

lnk model show, usage, context, chat, embed, transcribe, and count use the same API at LNK_MODEL_URL. Request actions read typed JSON from standard input, or --request <file>. A person reads each answer as fields, and chat's text as it streams, then the rest of its answer. With --json they print the JSON, chat's a stream event a line. Each lnk agent model command uses the grant of the agent its --agent names. Agent commands spell the actions /model:<action> and take JSON for structured requests. The registry generates their discovery names.

Request Inference

lnk model chat, lnk model embed, lnk model transcribe, and lnk model count submit the typed request for their action. lnk agent model chat, lnk agent model embed, lnk agent model transcribe, and lnk agent model count select a named agent’s model grant instead of requiring the caller’s model endpoint.

Inspect the Grant

lnk model usage reads measured usage, and lnk model context reads retained context. lnk agent model usage, lnk agent model context, lnk agent model status, lnk agent model list, and lnk agent model show inspect those values through the selected agent’s grant. Context details are explained in Context Reporting.

Owner Discovery

lnk model list --json prints plain names, as the route does. The owner runtime inventory, including addresses, is lnk model runtime list --json; lnk agent start uses it. lnk model show <name> returns the same Shown type as the route to an agent, granted a runtime. Run by its owner, it prints the model's known facts, including their overrides, as blocks for a person and the public Shown object with --json. lnk issue create reads that JSON for a model named in a failure.