Add Devin (Cognition) as a first-class provider

[el-analista]

Summary

Please add Devin (Cognition) as a built-in provider in Manifest. It is not in the current list of 31 providers and cannot be added via the Custom provider because its API shape is not OpenAI- or Anthropic-compatible.

• Provider: Devin
• API docs: https://docs.devin.ai/api-reference/overview
• Auth: API key (service user with RBAC), Authorization: Bearer <key> header
• Model: Cognition autonomous-agent stack, surfaced through a single session model rather than a model catalog

Why it cannot be a Custom provider

Manifest's Custom slot is explicitly limited to endpoints that speak /v1/chat/completions or /v1/messages. Devin's API is session-based and org-scoped:

• Base URL: https://api.devin.ai/v3/organizations/{org_id}/
• Lifecycle: POST .../sessions to create, then POST .../sessions/{id}/messages to send a prompt, then poll GET .../sessions/{id} (or stream via SSE on the messages endpoint) for the assistant result
• Responses include a structured messages[] array plus a status_enum (working, finished, etc.) rather than an OpenAI choices[].message shape
• Service users are required for programmatic access (RBAC, attribution)

A wrapper executor therefore has to translate the OpenAI-style request Devin does not understand, and translate Devin session/stream output back into the OpenAI response shape Manifest clients expect. That adapter is exactly the kind of thing a first-class provider entry in credentials/ + packages/ should provide (the same pattern already used for Kiro, OpenCode Go, NousResearch, etc., which also expose non-OpenAI APIs).

Proposed behavior

• Add a Devin tile in the providers list, parallel to Kiro / OpenCode Go / NousResearch.
• Auth flow: paste a service-user API key + organization ID. Store them as a Manifest credential.
• When a request hits the Devin provider, the executor:
  • creates (or reuses) a Devin session,
  • posts the user's last message into it,
  • streams messages back as OpenAI-compatible SSE deltas while the session is working,
  • closes the loop and returns a final chat.completion / chat.completion.chunk once status flips to finished.
• Show Devin under the same routing/picker UI as the other providers, with cost/quota tracking and the existing fallback chain (e.g. fall back from Devin to a paid API model on rate limit or session failure).

Reference material

• Devin v3 API overview: https://docs.devin.ai/api-reference/overview
• Teams quick start (service user + first call): https://docs.devin.ai/api-reference/getting-started/teams-quickstart
• Documented HTTP status codes: 200 / 201 / 400 / 401 / 403 / 404 / 429 / 500

A prior-art adapter shape can be borrowed from the OmniRoute Devin feature analysis, though the executor here would target the Manifest package layout rather than OmniRoute.

Use case

Several Manifest users (e.g. those on Teams / Enterprise Devin plans) want to keep using a single /auto endpoint from their agent/harness but route some traffic to Devin for long-running, autonomous coding tasks while still mixing in cheaper API models for short-turn completion, without writing a custom OpenAI-compatible shim for Devin in every harness.