nc://factory-droid

A Nodecode cell for Factory Droid, Factory's model subscription. With it, /factory-droid login signs in with a WorkOS device code, /models lists Factory's roster (Claude, GPT, Grok, Gemini, Kimi, GLM, DeepSeek, Qwen, MiniMax, ...), and a turn on factory-droid/<model> goes to Factory's LLM proxy the way oh-my-pi sends it, which is the way the Factory CLI does:

• each model rides the wire its upstream speaks, on its own path: /api/llm/o/v1/chat/completions (Kimi, GLM, DeepSeek, Qwen, Inkling, Mistral, Nemotron), /api/llm/o/v1/responses (GPT, Grok), /api/llm/a/v1/messages (Claude, MiniMax M2.7) and /api/llm/g/v1/generate (Gemini), on https://api.factory.ai, or https://api.eu.factory.ai for an account whose residency is the EU
• every request carries the WorkOS token as a bearer and the Factory CLI's identity: User-Agent: factory-cli/0.230.0, X-Client-Version, X-Factory-Client: cli, the org (X-Factory-Org-Id), the upstream that serves the model in the account's inference region (x-api-provider, the first of the model's rotation that region serves), how it was chosen (x-provider-routing-source), the session (x-session-id, the same id at every round of a session) and a fresh x-assistant-message-id; the OpenAI and Anthropic wires add their SDK's X-Stainless-* fingerprint, the Anthropic one x-api-key: placeholder, and a route to OpenAI OpenAI-Platform
• the system prompt opens with Droid's identity line
• a model's own output ceiling is asked for, and its own default effort when the session picks none (off for Opus 4.5, Sonnet 4.5 and Haiku 4.5)
• chat completions: reasoning as each upstream takes it (Fireworks: reasoning_effort and reasoning_history: preserved, or interleaved for DeepSeek; Baseten and Databricks: the effort alone, DeepSeek's off as low, Nemotron's as chat_template_args.enable_thinking; Mistral: the effort), temperature: 1, no cache key
• Responses: GPT with reasoning.summary: auto, text.verbosity: low, the session as prompt_cache_key and safety_identifier, a 24 h cache on OpenAI, service_tier: priority on the fast tiers, tool_choice: auto and parallel_tool_calls with tools, and no output cap; Grok with the effort alone and its output cap; every tool strict: false; no temperature
• Messages: adaptive thinking with a summary on the Claude models, plain adaptive on Opus 4.6 and Sonnet 4.6, a token budget (4096, 12288, 24576) on Sonnet 4.5 and Haiku 4.5 with the interleaved-thinking beta, a budget and an effort on Opus 4.5 and MiniMax M2.7; off stays adaptive on the always-thinking models; budget thinking on a history no longer led by thinking drops the thinking field and replays the history without its thinking blocks; the effort beta where the upstream wants it, fine-grained tool streaming on Anthropic routes, speed: fast on the fast tiers
• Gemini: the model in the body, temperature: 1, topP: 0.95, topK: 64, a thinkingLevel (MEDIUM only where the model has it), no output cap, tool names in the CLI's [a-zA-Z0-9_-] shape (turned back on the calls that return) and schemas projected onto the shape Factory's Gemini takes, consecutive tool results in one turn under result, and an unsigned call in the newest turn marked so Gemini's signature check passes it
• a refusal for the network's region says which model and what to do

Ported from oh-my-pi's factory-droid provider. See NOTICE.

Sign-in

/factory-droid login     answers at once with the address to open and the code to enter
/factory-droid status    signed in, as whom, the org, and when the token expires
/factory-droid logout    forgets the sign-in

The device code is asked of WorkOS (api.workos.com/user_management) with the Factory CLI's client. The login polls on a background thread; then Factory's /api/cli/whoami names the org, the residency region and the inference region, and the outcome comes as a notice. A login that resolves no org is refused. The token is kept in the shared auth.json under oauth_tokens.factory-droid (access_token, refresh_token, expires_at from the token's exp, email, account_id, org_id from the token's external_org_id claim or whoami, active_organization_id (WorkOS's own, sent back on a refresh), region, inference_region) and refreshed before a round when it expires within a minute; a refresh asks whoami again and keeps the regions it had when that fails and the org is the same.

There is no key variable: Factory's API keys are for its control plane and do not reach the LLM proxy. A round without a sign-in is refused with the words to sign in.

Install

Copy this folder into ~/.nodecode/cells/nodecode-factory-droid/ and restart Nodecode, or run (restart-cells).

Configure

Optional. The section is on unless it says "enabled": false:

{
  "factory-droid": {
    // Factory's host; an EU account moves to https://api.eu.factory.ai on
    // its own while this is Factory's host, a relay's address is kept as is
    "base_url": "https://api.factory.ai"
  }
}

An operator's providers.factory-droid.sdk (or a model's sdk) pins a wire, as for any provider.

The roster

oh-my-pi bundles no Factory rows: Factory has no model listing, so its discovery (packages/catalog/src/discovery/factory-droid.ts) builds the roster per account from the registry in compat/rules/providers/factory-droid.kdl and narrows it with live feature flags and org policy. models.json is what that discovery answers with no live answer: the global region, no flags, no policy. That leaves out the ten models behind a feature flag (Opus 5.5 Fast, Sonnet 5.5, GPT-6.1 Sol, GPT-6 Luna, Garnet, GLM-5.2 Fast, DeepSeek V4.1 Flash, MiniMax M3, Atlas, Aster) and the two behind an explicit opt-in (Fable 5 and 5.1): 48 models. Each row carries its wire, window, output ceiling, effort ladder and default, the upstreams that serve it in each inference region in rotation order, the EU limits where the registry names them, and the upstream's list price (Factory itself bills credits).

discovery-models.py wrote it, from an oh-my-pi checkout:

python3 -I discovery-models.py ~/.cache/nodecode-cells/omp-src models.json

The cell never runs the script. The picker's listing for factory-droid is this roster, asked of no endpoint.

Test

From this repository's root:

tools/test-cell.sh nodecode-factory-droid

Not ported

• Credit rates and billing pools (Factory bills Standard Credits; rounds are priced here at the upstream's list price).
• The Mistral thinking content parts (GLM and Mistral Medium on the EU's Mistral route replay thinking as typed parts), and the " " reasoning fallback DeepSeek on Fireworks and Baseten wants on a tool-call turn that streamed no reasoning.
• The Gemini stream's own finish reasons (a content-filter stop as an error with its category) and $ref dereferencing before a schema is projected; the core's GenAI fold reads the stream.
• Fable's server-side fallbacks, since both Fable models are opt-in and off the offline roster.
• The billing limits report (/api/billing/limits).

Gaps

• Live discovery. oh-my-pi asks GET /api/feature-flags and GET /api/organization/managed-settings with the token, then shows a flag-gated model once its flag is on, hides one the org blocks or did not allow, honours live provider routing (x-provider-routing-source: configured_order), and treats that roster as authoritative. A cell can add rows through list-provider-models, but Nodecode's picker (model-choices) shows the catalog's rows and the listing's together, so nothing can take a catalog row away per account, and the catalog (models-catalog-table, hooked without the credential) is one for every account. The seam would be a catalog row member (say listing_authoritative) that makes the picker show only the listing's rows for that provider once a listing has answered, with the listing's rows able to carry the catalog fields (window, efforts, price) a turn reads.
• Usage from response headers. Fireworks can report cached prompt tokens only in the fireworks-cached-prompt-tokens response header. The core's walk-provider-stream reads the response headers and hands the lane's fold frames only, so no hook sees them; the seam would be the response headers passed to the fold (or a :response-headers key on the lane's usage record).

MIT licensed.