nc://google-gemini-cli

A Nodecode cell for Google Cloud Code Assist, the backend the Gemini CLI signs in to: Gemini models on a Google account's free tier, or on a Google Cloud project. With it, /models lists the models under google-gemini-cli/, /google-gemini-cli login signs in, and a turn on one of them goes to https://cloudcode-pa.googleapis.com the way omp sends it.

No lane of the organism speaks this wire, so the cell registers its own, google-gemini-cli. Its stream is the organism's own Gemini fold with what differs put around it:

• the request is the Gemini body wrapped with the account's project and the model it names: {"project", "model", "request": {contents, systemInstruction, tools, toolConfig, generationConfig}}, POSTed to /v1internal:streamGenerateContent?alt=sse with the Google token as bearer and the Gemini CLI's user agent and client metadata
• the history is converted the way omp converts its own: user text and images, the model's turns with the thought signatures it made (only to the model that made them; the cell keeps them, never the message), Gemini 3's bypass sentinel on an unsigned first call, every result of a step in one user turn, reasoning another model wrote as a fenced text
• tools ride parametersJsonSchema, with what Google's schema has no field for left out and its constraint said in the description
• effort follows omp: Gemini 3 as a thinking level, Gemini 2.5 as a token budget on its -thinking model, a model that cannot think off running at its own lowest level
• each event is {"response": <Gemini chunk>}; an in-band {"error"} and a blocked prompt are said as errors, the planning text a Flash model leaks into its answer ({"thought": ...}) is held back and dropped, and a stream that ends without a finish reason is a truncated one

Ported from oh-my-pi's google-gemini-cli provider. See NOTICE.

Sign-in

/google-gemini-cli login                 open Google's consent page
/google-gemini-cli code ADDRESS          the address the browser landed on
/google-gemini-cli status
/google-gemini-cli logout

login answers at once with Google's consent page, and listens at http://127.0.0.1:8085/oauth2callback (another free port when that one is taken) for the browser to come back. When the browser runs on another machine than Nodecode, paste the address it lands on with /google-gemini-cli code. Nodecode exchanges the code, finds the account's Cloud Code Assist project (the one it has, else the free tier provisioned; a paid tier names its project with GOOGLE_CLOUD_PROJECT or GOOGLE_CLOUD_PROJECT_ID) and says once how it went. A failure stands as a notice until a sign-in succeeds.

The sign-in is kept in ~/.nodecode/auth.json under oauth_tokens.google-gemini-cli:

{"access_token": "ya29...", "refresh_token": "1//...", "expires_at": 1791234567,
 "project_id": "...", "email": "me@example.com"}

expires_at is Google's expiry less five minutes, as omp keeps it. A round whose token expires within a minute refreshes it first and writes it back; a refresh Google refuses stands as a notice naming /google-gemini-cli login. No other provider's key (GOOGLE_API_KEY among them) is ever sent to Cloud Code Assist.

Install

Copy this folder into ~/.nodecode/cells/nodecode-google-gemini-cli/ and restart Nodecode, or run (restart-cells).

Configure

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

{
  "google-gemini-cli": {
    // where Cloud Code Assist is served
    "base_url": "https://cloudcode-pa.googleapis.com"
  }
}

Files

• models.json: omp's bundled rows as the catalog reads them (tools/omp-models.py)
• wire.json: the same rows as the wire reads them: the model a request names, its lineage, how it thinks, omp's compat flags. Written by python3 -I wire-rows.py OMP_CHECKOUT google-gemini-cli wire.json

Test

From this repository's root:

tools/test-cell.sh nodecode-google-gemini-cli

Gaps

The wire is carried on the core's Gemini fold, with no core change. What is not carried, and the seam it would need:

• Call ids. CALL-GOOGLE-STREAMING's fold mints its own tool-<n> per round and drops the wire's functionCall.id; the cell sends ids unique within each request instead (Gemini takes none). The fold would need to keep a call's wire id when the part carries one.
• Signatures. The assembled message (DEFINE-PROVIDER-LANE's MESSAGE) keeps one reasoning_content and one content, and a lane may add no field of its own to it (another lane sends a message's fields as they are). So the cell keeps a round's thought signature, its answer signature and the model that made them in a table of its own, keyed by the session and the message's reasoning, answer and calls; each call's signature rides the core's own thought_signatures. Three things follow: several thought blocks in one round replay as one block under the last signature; the table lives in this process, so after a restart an earlier round's thought replays unsigned (as text, or dropped for Claude) and a Gemini 3 first call carries the bypass sentinel; and a message the core rewrites (a continuation merged after a cut stream) is no longer found. The seam would be per-block reasoning on the message that names its producing model, as the Responses lane's reasoning_items does for its own wire.
• The first-event watchdog. omp gives a Flash model 60 s and every other 300 s to send its first event; WALK-PROVIDER-STREAM has one idle deadline (the config's request timeout) and takes no first-event deadline. The seam would be a :first-event-seconds key on WALK-PROVIDER-STREAM.
• omp replays an empty STOP twice itself and names a thought-only answer; the core's own PROVIDER-EMPTY-RESPONSE retry stands in for both.
• Model discovery (omp asks retrieveUserQuota for the account's models): not ported. /models lists the bundled roster and asks no one; /connect's key check is told a key is not checked, since the credential is a sign-in.
• The schema normalizer is a compact port of omp's (2,400 lines of utils/schema/normalize.ts): unsupported keywords, references, type lists, constants and enums, required names; not every combiner shape.
• omp's markup healing of <thinking> written inside an answer: the core routes a leading thinking tag only.

MIT licensed.