feat(tui): complete code session agent loop

Wire the TUI to the live Code session API with device login,
streaming tokens, tool rows, Plan/Spec locks, and cancel.

Co-authored-by: Mathis <echobt@users.noreply.github.com>

Author: cursoragent

.rules/api.md

@@ -1,9 +1,9 @@
 # API and product
 
 - Default API origin: `https://api.cortex.foundation`.
-- Auth origin: `https://auth.cortex.foundation` (WorkOS session / device code / existing `/v1` auth).
+- Device login: `POST /v1/auth/device` and `POST /v1/auth/device/token` on the API origin. Do not call `/v1/auth/device/code` (404). `auth.cortex.foundation` is not the device endpoint.
 - Software / updates: `https://software.cortex.foundation`.
 - The harness is a client. Chat/Code turns, tools, computers, plugins, and snapshots go through that API. Do not embed a second model provider.
-- Keep `/v1/models`, `/v1/responses` (or the current Chat/Code route in `cortex-engine`), and `/v1` auth paths. If the backend adds a route, update the client and the user-facing error map together.
+- Code turns: `POST /v1/code/sessions/{id}/turns` with `{ message, mode: "chat"|"code" }`. Plan/ask are TUI/harness locks, not API modes.
 - When the API is unreachable, the product fails closed with a product-facing error. Do not fall back to a local mock model.
 - Environment overrides (`CORTEX_API_URL`, `CORTEX_API_KEY`, `CORTEX_AUTH_TOKEN`) are for operators and tests. Defaults must stay production Cortex URLs.

CHANGELOG.md

@@ -7,6 +7,26 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ---
 
+## Unreleased
+
+### Added
+- Code session turns against `POST /v1/code/sessions/{id}/turns` (streaming tokens, tool rows, Plan/Spec harness locks)
+- Device login against `POST /v1/auth/device` on `api.cortex.foundation` (no silent guest substitution)
+- This PC / Cloud / SSH computer label on the welcome card
+- Workspace Code session id cache (`~/.cortex/code-sessions.json`)
+- Guest session as an explicit TUI login choice
+
+### Fixed
+- Cancel aborts the local SSE stream and best-effort POSTs cancel (API route is still 404)
+- Task without a live ModelClient reports failure instead of a fake success
+- Remote tool rows keep their label instead of dropping arguments
+
+### Changed
+- Sign-in docs point at `api.cortex.foundation` device login
+- Stream timeouts use the product-facing coding-service error
+
+---
+
 ## 0.0.5
 
 ### Added