Devin CLI

Shiba integrates Cognition’s official Devin CLI, allowing teams with Devin accounts to run Devin’s frontier software engineer model (devin/swe-2) inside account-owned Cloudflare Sandbox containers.

Agent Overview

Attribute	Specification
Harness ID	`devin`
CLI Binary	`devin`
Pinned Version	`3000.10.31` (in Dockerfile)
Default Model	`devin/swe-2`
Supported Models	`devin/swe-2`
Credential Model	Cloudflare Worker Secret (`DEVIN_API_KEY`) injected via Bearer header at egress
Invocation	`devin -p "<task>"`

How It Works

Unlike model-direct harnesses, Devin is a service CLI that communicates with Cognition’s orchestration cloud:

Developer Request
  └── CodingOrchestrator generates delegate_coding_task({ harness: "devin" })
      └── Human signs off on exact prompt & repository
          └── Ephemeral Cloudflare Sandbox boots
              ├── Provisioned with dummy credentials.toml in $XDG_DATA_HOME/devin/
              ├── Egress proxy intercepts api.devin.ai & server.codeium.com
              ├── Swaps real DEVIN_API_KEY into Authorization: Bearer header
              └── Executes: devin -p "<task>"

Configuration & Deployment

1. Store the DEVIN_API_KEY Secret

Obtain your Devin service API key from Cognition, then configure it as a Wrangler secret on your Worker:

pnpm wrangler secret put DEVIN_API_KEY --config apps/backend/wrangler.jsonc

2. Verify Credential Status

Open your Shiba dashboard at /app/ and navigate to Agents. The Devin CLI row will display a green Configured status chip when the secret is present (the secret value itself is never exposed).

Selecting Devin in a Task

Specify harness: "devin":

{
  "repoUrl": "https://github.com/acme/core-engine",
  "task": "Investigate and resolve flaky worker heartbeat tests",
  "baseBranch": "main",
  "harness": "devin",
  "codingModel": "devin/swe-2"
}

Event Parsing & Exit Codes

Devin streams progress lines via stdout:

Welcome and organization banner lines are cleanly filtered.
Action and file edit lines are parsed into the unified run timeline.
Unauthenticated or session errors immediately reject the run with honest diagnostic feedback.