ABCD Agent

AI coding agent for ABCD Edu, running inside VS Code (or any compatible fork — Antigravity, Cursor, Codium, Code Server, …).

Chat with an agent that touches your repo. Reads files, runs ripgrep, applies patches, runs shell commands — all gated by a per-action approval card so you stay in control.
Plan mode. The model drafts a markdown plan in a side-by-side editor, you review/edit/comment like a PR, then approve to execute.
Inline edit (⌘K). Highlight code, describe the change, get a streaming diff with hunk-level approval.
Reasoning-model streaming. GPT-5 / o-series "thinking" deltas surface in a collapsible block so long reasoning turns don't look like the agent has hung.
Backed by ABCD Cloud. OpenAI / Anthropic upstreams routed through a Cloud Run service in Seoul; auth uses your portal.abcdedu.com account.

Install the extension.
Run ABCD Agent: Sign In from the command palette.
A browser tab opens at portal.abcdedu.com/extension/connect. Log in with your ABCD account, give the new token a name (e.g. "MacBook"), and authorize.
The browser redirects to a local loopback URL the extension is listening on — token lands in your editor's encrypted secret storage automatically. Close the tab and chat.

You can manage tokens (revoke, see last-used time) any time at https://portal.abcdedu.com/account/api-tokens.

Commands

Command	Default keybinding	Purpose
`ABCD Agent: Open Chat`	—	Reveals the chat sidebar.
`ABCD Agent: Sign In`	—	Browser-based OAuth-style sign-in.
`ABCD Agent: Inline Edit`	⌘K	Edit the current selection / line with the agent.
`ABCD Agent: Toggle Mode Picker`	⇧⌘M	Switch between Ask / Accept edits / Plan / Auto.
`ABCD Agent: New Chat`	—	Start a fresh conversation.
`ABCD Agent: Chat History`	—	Quick-pick over previous chats.
`ABCD Agent: Set API Base URL`	—	Override the Cloud Run endpoint (leave blank to revert to production default).
`ABCD Agent: Clear Token`	—	Forget the stored PAT.
`ABCD Agent: Use Local Cloud`	—	Point at a locally-running cloud-api (dev only).

Mode	What auto-runs without approval
Ask permissions	Nothing — every patch and shell command pops an approval card.
Accept edits	File patches auto-apply; shell commands still require approval.
Plan mode	Read-only tools only; the agent drafts a plan you review side-by-side before any write.
Auto mode	Both patches and shell commands auto-run. Use only on a workspace where you can throw away everything.

Key	Default	What it does
`abcdAgent.apiBaseUrl`	production Cloud Run URL	Override the cloud-api endpoint. Leave blank to use the default.
`abcdAgent.model`	`gpt-5.3-codex`	Model name forwarded to the cloud router. `gpt-` / `o1` / `o3` / `o4` → OpenAI, `claude-*` → Anthropic, `dev` → offline mock.
`abcdAgent.signInUrl`	`https://portal.abcdedu.com/extension/connect`	Override the page opened by the sign-in command (set to `http://localhost:3000/extension/connect` for portal dev).

Workspace tools (read files, search, apply patches, run terminals) execute on your machine — the cloud only sees the conversation transcript needed for the next LLM call.
Tokens are stored in your editor's encrypted SecretStorage. Logging out revokes the row server-side.
Telemetry: none in the extension. The cloud-api logs (user_id, model, token counts) for usage tracking; no message content is persisted.