Doctopus.ai

Editor front-end for the gateway. You stay in your editor; the gateway fans your prompt out to every AI coding tool, verifies and judges the candidates, and returns one optimized answer. How you treat that answer (save / edit / undo) becomes the implicit feedback that trains the router and the RAG memory.

What it does

• Ask / Refactor (Ctrl/Cmd+Alt+A): select code (optional), describe the task. The extension calls POST /v1/ask with your developer_id, team, and the file's language.

• Result panel: the winning answer with confidence, the routing mode, the verification result (syntax/lint/tests), why it was chosen, the learned patterns of yours that shaped it, and an Alternatives escape hatch.

• Apply: replaces your selection with the answer (code fences stripped).

• Implicit feedback → POST /v1/feedback, recovering the "lost click":

  You do this after Apply	Outcome sent	Effect on backend
  Save with the code intact	committed	reward winner arm · learn the pattern · charge metering
  Edit the code, then save	edited	neutral arm · charge metering
  Undo (or Reject), then save	reverted	penalize winner arm · no charge

  The save/edit/undo mapping is a deliberately simple heuristic (see src/feedbackTracker.ts); a stronger signal (passed CI / survived in git) is the documented next step.

Run it

# 1. start the gateway (from the repo root)
uvicorn gateway.app2:app --reload          # serves http://localhost:8000

# 2. build & launch the extension
cd ide/vscode
npm install
# press F5 in VS Code -> "Run Doctopus.ai Extension" (opens an Extension Dev Host)

Then in the dev-host window: open a file, select some code, press Ctrl/Cmd+Alt+A, type an instruction.

Settings

Layout

Other IDEs

The client/feedback logic is small and HTTP-only; a JetBrains plugin would reimplement the same three calls (ask, apply, feedback) against the same endpoints.