Glina

Model backend

The model backend (pipeline/llm.js) is the sculpt loop's one transport: complete({ system, messages, maxTokens }) → { text, stopReason }. There are exactly two backends, both OpenAI-compatible /v1/chat/completions:

  • Brama — the org model router. The default and the only fleet path.
  • OpenRouter — an operator-sanctioned alternative, enabled only by explicit config (models.openrouter present, models.backend: "openrouter" to prefer it). The key still comes from Skarbiec.

Direct provider APIs (Anthropic, OpenAI, …) intentionally do not exist in this package. Configuring one is refused outright:

direct provider APIs are not supported in this package — model access goes through Brama (models.brama) or, when the operator sanctions it in the config, OpenRouter (models.openrouter)

That sentence fires when models.anthropic, models.openai, or models.direct appears in the config. The code paths were deleted at the owner's demand, not gated.

Selection

models.backend (default "brama") chooses a preference, with the following actual selection order:

  • Brama is ready only when models.brama.url, key, and bearer are all present. With the default/brama preference, a ready Brama is selected.
  • Otherwise, a configured models.openrouter.key selects OpenRouter — even when the preference says brama. This lets an incomplete Brama tuple fall back to the sanctioned alternative.
  • With openrouter preferred but no OpenRouter key, a ready Brama is used. If Brama is incomplete too, setup fails with backend=brama requested but models.brama.url/key/bearer are not configured (the wording is inherited from the implementation).
  • With neither backend usable: no model backend configured — set models.brama.url+key+bearer or models.openrouter.key (skarbiec:// references in pipeline.config.json).

Brama transport

  • URL: models.brama.url + /v1/chat/completions; model models.brama.model (default any).
  • Auth: authorization: Bearer <bearer> plus the HMAC agent identity — x-agent-id, x-agent-timestamp, and x-agent-signature = HMAC-SHA256(key, "<agent_id>:<ts>:<sha256(body)>").
  • Retries: attempts (default 4) on transport failures and HTTP 502/503, with linear backoff; a signed refusal (401/403) or routing error never retries. Each attempt has a hard deadline (timeoutMs, default 120 s) so a black-holed connection can never hang the sculpt loop.
  • A 200 with empty content (reasoning routes burning the token cap on hidden thinking) retries like a 502.
  • Exhaustion: brama unreachable after N attempts: <reason>; a final HTTP error is brama HTTP <status>: <message>.

OpenRouter transport

  • URL: models.openrouter.url (default https://openrouter.ai/api/v1) + /chat/completions; plain bearer key; timeoutMs default 180 s; attempts default 4.
  • 429 and 5xx retry; other 4xx are final (openrouter HTTP N: <message>). Empty content retries. Exhaustion: openrouter unreachable after N attempts: … / openrouter exhausted N attempts: ….

Message shape

Anthropic-style content blocks are flattened to plain OpenAI messages; image blocks (viewport screenshots) are replaced by the note [viewport screenshot attached] — neither backend receives raw image bytes.

Reply parsing

parseJsonFrom(text) extracts the sculpt step: a fenced ```json block first, else a balanced-brace scan that takes the first object that parses. Refusals: model reply JSON does not parse: <reason> and model reply contained no JSON object: <prefix>.