nc://devin · Nodecode

nodecode-devin

A Nodecode cell for Devin: Codeium's Cascade backend, the models a Devin account is entitled to (SWE-1.6 and SWE-1.7, Claude, GPT, Gemini, GLM and the rest of the account's roster). With it, /devin login signs in with the browser, /models lists the account's roster, and a turn on devin/<model> goes to https://server.codeium.com the way the Devin CLI sends it.

Cascade speaks none of the organism's four wires: it is the Connect protocol over HTTP/1.1 with protobuf messages. The cell registers a lane of its own, devin, which makes the exchanges itself:

GetUserJwt (unary) turns the session token into a user JWT, and names the account's own chat host when it has one
AssignModel (unary) resolves a router model (adaptive) into a concrete uid and an assignment JWT
GetChatMessage (server streaming) sends the conversation as one gzipped Connect frame and folds the answer's frames into text, thinking (with its signature, replayed next round to the model that made it), tool calls and usage; the end-of-stream frame's error fails the round
GetCliModelConfigs (unary) lists the account's roster for /models

The protobuf messages (exactly the fields omp sets and reads), the Connect envelope and gzip are written by hand in proto.lisp.

A Cascade roster names one uid per reasoning effort (claude-opus-5-low ... claude-opus-5-max). As omp does, the cell folds each such family into one model whose effort picks the uid: from the family metadata the server ships, and from omp's reviewed Devin families (wire.json).

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

Sign-in

/devin login        answers the address to open; Nodecode listens at
                    http://127.0.0.1:59653/callback for Devin's answer
/devin code ADDRESS the address the browser landed on, when it runs elsewhere
/devin status
/devin logout

The sign-in is Devin's CLI grant (authorization code with PKCE) at https://app.devin.ai/auth/cli/continue, its code exchanged at https://api.devin.ai/auth/cli/token for a session token. auth.json keeps it under oauth_tokens.devin:

{"access_token": "...", "refresh_token": "...", "expires_at": 1767225600,
 "api_endpoint": "https://api.devin.ai", "enterprise_url": "https://app.devin.ai"}

expires_at is the token's own exp less five minutes, or a year from the sign-in when it names none. omp has no refresh for Devin, and neither does the cell: an expired sign-in is set aside, a round asks for /devin login (a standing notice says so) and DEVIN_API_KEY answers in its place when it is set.

DEVIN_API_KEY, or a key /connect saved, works without a browser sign-in.

Install

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

Configure

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

{
  "devin": {
    // where Cascade is served
    "base_url": "https://server.codeium.com"
  }
}

Test

From this repository's root:

tools/test-cell.sh nodecode-devin

Regenerating the data

python3 -I tools/omp-models.py ~/.cache/nodecode-cells/omp-src devin nodecode-devin/models.json
python3 -I nodecode-devin/tools/wire-rows.py ~/.cache/nodecode-cells/omp-src nodecode-devin/wire.json

Where this differs from omp

Each of these is a choice made here, said so it is not mistaken for omp's:

The request is gzipped without compression. The image links no compressor (dexador's chipz only inflates), so the request frame is a gzip member of stored deflate blocks: valid gzip, as large as the protobuf it carries. Answers are inflated with chipz, compressed or not.
The kept request has no secrets. A lane's fourth value is the request the organism keeps (its sha256 on the provider_request fact, its bytes in request_bodies). Cascade carries the session token and the user JWT inside the body's Metadata, so the cell answers the GetChatMessageRequest with both left empty, not the exact bytes sent.
Another model's thinking is demoted as plain text. omp renders it in the target model's own thinking dialect (renderDemotedThinking); the dialect registry is not carried, so the text goes ahead of the answer as it is, followed by a newline (omp's anthropic-dialect form).
A trailer error keeps Connect's status. omp raises every error an end-of-stream trailer carries as a validation error, never retried; here the trailer's Connect code maps to the HTTP status Connect pairs with it (invalid_argument 400, unauthenticated 401, unavailable 503, ...), so the organism's retry policy retries the 5xx and 429 ones. The large history invalid_argument ... internal error case omp flags as a context overflow says context in its detail, which is how the organism's eviction recognizes an overflow.
The Cascade id is derived from the session. omp passes its session id as cascade_id; the cell passes a UUID derived from the Nodecode session id, so every round of a session shares it and the id has the shape the server issues.
A clean end without the end-of-stream frame ends the round, as omp's reader does; an end inside a frame is a truncated stream.
Discovery is kept in memory. omp caches the discovered roster; the cell asks again after a restart, on the first /models listing or the first round of a model the seeds do not name. Until then omp's reviewed families still route an effort to its uid.
The tool schema normalization for Devin's Gemini backend is compact: references resolved, a type list made one type and nullable, constants and enums as strings, unsupported keywords dropped with their constraints said in the description; not every option of omp's normalizeSchemaForGoogle.
The X / X-thinking pair rule is not carried. omp's global deriveThinkingPairFamilies folds any same-priced X and X-thinking pair into one model, its effort ladder read from omp's model policy (resolveModelPolicy), which this cell does not carry. Such a pair lists as two models here; the server's own families and omp's reviewed ones are folded.

Gaps

What omp does that the cell cannot carry without a core change:

Which model wrote a message. omp replays an assistant message's thinking, its signature and the id Cascade gave it only to the model that wrote it (isNativeDevinMessage, which reads the message's own api, provider and model). The organism's assistant message carries no author, and a field the lane added would ride every later round to whatever lane serves it (the chat lane replays a message's keys verbatim), so the cell keeps the author beside the message, in memory: a table keyed by the digest of the message's text, reasoning and tool-call ids, bounded at 4096 messages and cleared at the limit and when the cell stops. After a restart (or past the bound) a message's author is unknown, and what is lost is exactly this: the round replays that message's thinking demoted to text instead of as thinking, drops its signature, and names it by the derived bot-<uuid> id instead of Cascade's own message id. It needs the stored assistant message to carry the provider and model that wrote it (a field the lanes do not send), or a per-message side record the store keeps for a lane.
Credits. Cascade reports credit_cost, committed_credit_cost and committed_acu_cost per round, which omp keeps on its usage. The cell decodes them and drops them: nle::provider-usage has token slots only. Carrying them needs a slot there (or a usage plist a lane may extend) and a reader that shows it.
Tool results that failed. omp sends tool_result_is_error on a failed tool's result. The organism's tool message (nle::message "tool") carries no failure mark, so the field is never set. It needs the mark on the message the turn loop builds (RUN-TOOL-CALL's refusal path).
Selector aliases. omp resolves Devin's short names (opus, sonnet, swe, gpt-5.6-sol, ...) to logical models when a selector names the provider. The cell honours them on the wire (MODEL-SPEC), but the organism's model selection has no per-provider alias seam, so /models and (select-model ...) know only the listed ids.

MIT licensed.