HHostingByteSoftware
Ctrl K
ComunidadGuías…

CHANGELOG.md

Rama main · réplica de lectura del repositorio original.

Archivo original ↗Preguntar a KiBot
# Changelog

All notable changes to Cortex are documented here. Format follows
[Keep a Changelog](https://keepachangelog.com/en/1.1.0/).

## [3.5.0] — 2026-09-29

Cortex on your phone. Android and iPhone get the full workflow (chat, approvals,
projects, diffs, checkpoints, routines, replay, usage) against the desktop or a
headless `cortex-serve`, over Tailscale or the LAN, for free: an APK on the
releases page and an installable web app for iPhone. See docs/MOBILE.md.

### Added
- Mobile client rebuilt (`mobile/`): chat-first with streaming markdown,
  tool cards, a sticky approval dock with Approve/Deny/Remember and risk
  badges, an inbox of what needs you, thread list with rename/delete,
  projects with git status and per-file diffs, hold-to-restore checkpoints,
  routines (create, edit, run now, history), quota rings and the
  reliability card, run timelines, model and plan-mode defaults, devices,
  light/dark themes, a demo mode, an offline outbox and a resync after
  reconnect. Hash routes, pull-to-refresh, safe-area insets, haptics.
- Pairing: Settings → Mobile shows a QR and a 6-digit code (10 minutes,
  single use). Paired phones hold a per-device bearer token, stored hashed
  in `~/.cortex/mobile-devices.json`, listable and revocable from Settings.
  Requests that are not from the local machine (Tailscale Serve and
  reverse proxies included) need the token; the WebSocket takes it as a
  query parameter or a first frame.
- `/api/v2`: capabilities, threads shared with the desktop message store,
  send through the same chat pipeline as the desktop (routing, hooks,
  tracing, failover, push, 16 ms token coalescing), runs and timelines
  from the tracing store, approvals, projects with discover/add, git
  status and diff (200 KB cap), checkpoints with a confirm header on
  restore, reliability, usage, routines, models, mobile settings and push
  status. Phone runs stream into the desktop window too.
- Web Push without a third party: RFC 8291 encryption and RFC 8292 VAPID
  built on the crates already in the tree, key kept in the vault, egress
  limited to the browser vendors' push services. Fires on approval needed,
  run finished, run failed and quota above 90 %, each with a link that
  opens the right screen.
- Native shell (`mobile/native`, Capacitor 8): `cortex://` deep links,
  camera QR pairing, haptics, secure storage, Android back handling. The
  Mobile workflow builds a sideloadable APK on every push and attaches it
  to tagged releases; the iOS job proves the Xcode project compiles and
  ships to TestFlight when signing secrets exist.
- Desktop Settings → Mobile lists paired devices with revoke.

### Fixed
- Thread pagination returned the cursor of the first row that did not fit
  instead of the last row that did, so the second page skipped a thread.
- Repo-relative path checks refused `/abs` on Linux but not on Windows.

## [3.4.0] — 2026-09-29

Speed and daily-driver polish. Everything below is verified by the E2E job,
which launches the release build on Ubuntu and Windows and now records
startup timings on every push.

### Added
- Needs attention: a pending approval raises a toast with an Open action,
  a desktop notification when the window is hidden, an inbox row under a
  new Approvals filter, and amber count badges on the Threads rail item
  and per thread. Today opens with a Needs-attention card: approvals,
  running threads with Stop, recent failures. Ctrl+Shift+A jumps to the
  oldest approval.
- Finished runs show "Done — n edits · m tools" with Diff, Review, Replay
  and Copy actions.
- Command palette v2 (Ctrl/Cmd+K): fuzzy search with recents over actions,
  panels, threads, sessions, settings sections (deep-links open the tab and
  highlight the card), projects, profiles and slash commands. Prefixes
  `>` `@` `#` `/`. Platform-correct key hints.
- The six shortcuts that were documented but never bound now work
  (new chat, quick open, theme, compact, new window, jump to approval);
  the shortcuts sheet is generated from the keymap.
- Add a project folder in one click from the sidebar, the empty state,
  the chat header and the palette.
- The onboarding wizard detects installed agent CLIs and their login
  state, asks for a project folder, and folds theme into one control.
- Right-click in the file explorer: open in editor (VS Code family, Zed,
  JetBrains, or the OS default, with line numbers where supported),
  reveal in file manager, open a terminal here, copy path. The terminal
  pane shows its working directory and can open an external terminal in
  the project.
- Routines can run on a local agent CLI without a gateway, pin an agent
  and a project, run daily at a set time, and show their next run.

### Changed
- Startup and responsiveness:
  - 41 panels, the settings modal, the onboarding wizard, the markdown
    renderer and the xterm login modal load on demand; codemirror, xterm,
    mermaid, highlight.js and markdown are separate vendor chunks.
  - Streaming no longer re-renders the whole shell per token; token
    fragments are merged every 16 ms before reaching the UI and SQLite.
  - Keychain seeding, Tailscale and WSL probes and the sidecar spawn
    leave the pre-window path; the gateway and vault keys are cached
    after one read (list_agents did 13 keychain reads every 30 s).
  - SQLite indexes for the Reliability Dashboard, Run Replay, usage and
    semantic search; project discovery and file listing run off the
    async runtime; twelve polling loops pause while the window is hidden.
  - Release builds use thin LTO, one codegen unit and stripping.
- Settings (3,800 lines) and the chat pane (2,500 lines) are split into
  per-tab and per-part files with a shared section/toggle primitive.
  Every toggle has a description; headings are sentence case; Esc closes
  settings and Ctrl/Cmd+, toggles it; focus returns to the composer.
- New chat empty state with project-aware example prompts and clickable
  @-token and slash chips; approval, tool and error cards share one
  frame; one design-token block; settings styles ship with the lazy
  settings chunk.
- Mobile: 44 px touch targets, safe-area padding, inbox header with a
  pending count, clearer approval cards.

### Fixed
- An editor given as a Windows path was not recognised on Linux.
- Crash reporting had no caller; window and render errors now reach it.

### Build and CI
- The E2E job records mount, first paint, window-shown, invoke count,
  DOM size, heap and bundle size per OS on every run, with soft budgets
  and a job-summary table.

## [3.3.0] — 2026-09-28

New ways to reach Cortex from outside the app, cheaper Claude Code turns,
and proof that the real app starts on Linux and Windows on every push.

### Added
- `/review [base] [--model <slug>]`: AI code review of the uncommitted diff
  (or the branch against a base), grouped by severity with click-to-open
  `file:line`. By default the reviewer is a model from a different provider
  than the one you're using, so code isn't graded by the model that wrote it.
  Diffs are secret-redacted and size-capped.
- Cortex as an MCP server. `POST /mcp` on the local server lets Claude Code,
  Codex and Gemini CLI search the Brain, ask it questions, list recent
  sessions, read reliability stats and create checkpoints (restore only if
  you allow destructive tools). Off by default, bearer-token protected, with
  paste-ready client config under Settings → Connections.
- Phone push via ntfy or Gotify (Settings → General → Phone push): approval
  needed, run finished or failed, failed routine, quota above 90%. Taps open
  the mobile app's inbox. Token kept in the OS keychain; a self-hosted server
  on your LAN or tailnet needs an explicit opt-in.
- Native session resume for Claude Code: each turn resumes the CLI's own
  session and sends only the new message instead of replaying the whole
  transcript, which saves quota and keeps the CLI's tool memory. A model or
  project switch starts fresh; a missing session falls back once to the full
  history. `CORTEX_CLI_NATIVE_RESUME` controls it; Codex support is opt-in.
- Quota-aware failover (Settings → Providers, off by default): a turn that
  fails on quota or a transient error before any output is re-sent once to
  the next agent in your chain, and Run Replay shows why.
- UI for backend features that had none: `.cortexignore` chip on the project,
  stale Brain notes check, past-chat semantic search with reindex, and
  retrieval-quality evals.

### Fixed
- The mobile app's deep links (such as `/inbox`) were served with a 404
  status, which broke service-worker caching and push links.
- JavaScript crashes were never recorded; window errors and render errors now
  reach the crash log.
- The E2E probe patched `console.error` in normal sessions and accepted
  snapshot writes without `CORTEX_E2E`.
- The Brain folder follows OneDrive-redirected Documents on Windows.
- Cookbook RAM detection works on Windows and macOS.
- Test-run timeouts kill the whole process tree; removing a worktree also
  removes its branch; the gateway client no longer panics on a broken TLS
  store.

### Build and CI
- New E2E workflow builds the release app on Ubuntu and Windows, launches it
  (Xvfb on Linux, WebView2 on Windows) and checks from the renderer's own
  heartbeat that it mounted, painted and threw no errors. It also smoke-tests
  the headless server: web app, API, WebSocket origin checks and `/mcp`.
- The codebase is formatted with rustfmt and Prettier, and CI enforces it.

### Removed
- Four unused frontend modules, about 60 unreferenced exports, duplicated
  path and formatting helpers, and two dead backend commands.

## [3.2.0] — 2026-09-27

A hardening release: Windows and Linux are now both built and tested on every
push, and a full audit fixed the bugs that turned up.

### Fixed — Windows
- Agent CLIs installed through npm (`claude.cmd`, `codex.cmd`, …) are spawned
  without hand-built `cmd /C` command lines, so prompt text containing `&`,
  `|` or `"` can no longer break out into cmd.exe. No console window flashes
  per run, and stopping a run kills the CLI's whole process tree.
- `npm`, `npx`, `pnpm`, `pytest`, MCP servers and hooks resolve their `.cmd`
  shims, so tests, monitors, dependency audit, TypeScript diagnostics
  (`@problems`) and PRP validation gates work on Windows instead of reporting
  "not installed".
- Canonical `\\?\C:\` paths are normalised, so trusted projects, the last
  opened project and chat sharing into a project folder are recognised.
- `CORTEX_PROJECTS_ROOT` uses the OS path-list separator; `:` split drive
  letters apart.
- The terminal defaults to PowerShell 7, then Windows PowerShell, then cmd
  (`CORTEX_SHELL` overrides; on Linux/macOS it uses `$SHELL`).
- Browser session detection checks every Chromium profile and gives a clear
  message for App-Bound (v20) cookies.
- git pull/push/clone and Gitea backup can no longer hang on a hidden
  credential prompt.

### Fixed — everywhere
- A keyring hiccup (locked Secret Service, Credential Manager error) could
  overwrite the vault master key and make every stored provider key
  unreadable. Only a missing key now creates a new one, and vault writes are
  serialised.
- `@web` and `@websearch` crashed the send with a nested-runtime panic.
- The mobile web app is bundled into installers; installed builds used to
  serve a 404 for the phone UI.
- The mobile server's WebSocket checks `Origin`, so another website open in
  the same browser can't read streamed chat tokens.
- Gitea backup no longer mirrors the file holding its own access token.
- Clicking a note in the vault manager called a command that didn't exist.
- Event listeners leaked when a panel unmounted before `listen()` resolved
  (chat, settings, batch runner, routines, safe mode, terminal).
- The editor's file reads were always denied by the Tauri capability set.
- The SSH usage poller sent a literal `<gateway-ct>` placeholder; the
  container id is now configurable in `infra.json`.
- SQLite run history uses WAL and a busy timeout, so the desktop app and
  `cortex-serve` can share it.
- Test runs have a 20-minute timeout instead of hanging forever.

### Build and CI
- CI runs typecheck, ESLint, the frontend and mobile builds, `cargo check
  --all-targets` and `cargo test` on both Ubuntu and Windows.
- A fresh clone could not compile until the Go sidecar was built by hand;
  CI and the docs now cover it.
- Windows release builds failed on a maintainer-local signing certificate
  thumbprint. The release workflow imports a certificate from secrets when
  present and otherwise builds unsigned. Linux releases build on Ubuntu
  22.04 for a lower glibc floor.

### Removed
- The unused status bar and its six child components, the setting that
  toggled it, and about 430 lines of CSS no longer referenced anywhere.

## [3.1.0] — 2026-07-03

A release about run observability: seeing what agents actually did, how
reliable each model has been, and reaching any model you host. (Versioning
jumped from 0.2.x to match the app's marketing version.)

### Added
- Agent Reliability Dashboard (under Observability): per-provider and
  per-model success rate, p50/p95 latency, token totals, and estimated cost,
  aggregated from local run history. Time-range filter, CSV/JSON export, and
  failing rows link to the run that failed. Metrics are a local view — no
  visibility into gateway-internal retries — and cost is an estimate; the UI
  says so.
- Run Replay ("agent black box"): play back any past run as a timeline — the
  prompt, the routing decision, each tool call and approval, file edits,
  errors, result, and per-run cost. Export as redacted JSONL. Read-only.
- Homelab Model Fabric: register any OpenAI-compatible endpoint (vLLM,
  llama.cpp, LM Studio on LAN/tailnet, or a hosted API) from Settings →
  Providers → Model fabric. Health check, model discovery, latency test, and
  chat via a `fabric-<name>` agent. Reachability probes never send the API
  key. Routing is explicit for now: pick the endpoint's model in the composer.

### Fixed
- Gateway runs now record real token usage (previously zero), so cost and
  usage rollups reflect the primary path.
- The routing reason for each turn is captured and shown in Run Replay.

## [0.2.9] — 2026-06-28

### Fixed
- Layout mangle: chat pane collapsed to 0px when opening a project or chat
  (scrollIntoView + 1fr race). Replaced with scroll-margin + deferred scroll.
- Auto-update: manifest URL, Gitea asset detection, and signature validation
  all wired correctly for self-hosted Gitea releases.
- Duplicate user messages appearing in chat on send.
- ~30 additional HIGH/MEDIUM/LOW bugs from a full-app audit (scrollbar
  bleed, provider discovery, theme consistency, a11y, tooltip z-index, etc.).

### Added
- Image/file upload in chat composer: drag-drop, clipboard paste, and
  attach button. Images sent as base64 data URIs; text files as fenced
  code blocks.
- Voice-to-chat (mic button): browser SpeechRecognition primary, whisper-cli
  fallback. Works in Edge/Chromium webview out of the box.
- Thinking/working indicator shows the selected model name for all providers
  (e.g. "GPT 5.5 is thinking", "CLAUDE SONNET 4 is thinking").
- Project persistence: last active project restored on app restart.
- Native Windows project discovery: repos directly under ~ are found
  alongside ~/projects/*; WSL UNC paths de-duped in favor of native paths.

## [0.2.8] — 2026-06-25

### Added
- Ultimate multi-model agent: all 7 CLI agents + 13 OpenAI-compatible APIs +
  6 local runtimes routable from one chat. Model picker with slug search.
- Mobile web access: companion HTTP server for phone/tablet browsers.
- Cost-aware auto-routing: capability-gated first, then cheapest-capable wins.
- Gateway made fully optional: local CLI agents work standalone without a
  gateway deployment. `infra.json` configures only Ollama base URL.

## [0.2.5] — 2026-06-14

### Added
- Brain / Knowledge system: semantic search over chat history and Obsidian
  vault via local embeddings + vector store. `@brain` context token.
- Unified RAG: "chat with your brain" — ask questions across all stored
  knowledge, not just the current session.
- Save-to-Brain: save any message or selection to the knowledge base.

## [0.2.0] — 2026-06-03

### Added
- Teams: manager + specialist worker agents with automatic task decomposition.
- Lanes: same task across providers in isolated git worktrees.
- Arena: head-to-head A/B model comparison with persistent ELO leaderboard.
- MCP catalog: JSON-RPC MCP client + one-click server installation.
- Checkpoints + `/undo`: git-independent workspace snapshots with diff preview.
- Plan/Act mode toggle (Ctrl+M): plan mode blocks write/exec tools.
- `@` context tokens: `@brain`, `@diff`, `@recent`, `@status`, `@grep`, `@web`,
  `@summary`, `@file`, and more.
- Session auto-condense: long conversations automatically summarized to stay
  within context limits.
- Custom agent roles/profiles and mode customization.
- Architect mode (`/architect`): two-phase planner + editor split.

## [0.1.0] — 2026-05-20

### Added
- Initial release: Tauri 2 desktop app (Rust + React).
- Streaming chat with Claude CLI, Codex CLI, Gemini CLI, Ollama, and
  OpenAI-compatible gateway.
- Orchestrator with `@mention` routing, capability scoring, fan-out.
- Memory layer: project runbooks, CLAUDE.md, Obsidian vault integration.
- SQLite observability store (OpenTelemetry-shaped spans).
- Project discovery under ~/projects/*, file tree, command palette (Ctrl+K).
- Search across memory and chat history (Ctrl+Shift+F).
- Ed25519-signed auto-updates from Gitea.
- Audit log (JSONL, 90-day retention).