@bibobonking/opensol 3.1.6 → 3.2.1

package/CHANGELOG.md

@@ -1,12 +1,14 @@
 # Changelog
 
-## 3.1.5 - Fix Bug cannot get property "sessions" without inject
-
-- fixed bug cannot get property "sessions" without inject!
-
-## 3.1.5 - Update Log
-
-- Agent loop: removed the 16-round cap (also for subagents). Provider errors now retry up to 5 times with backoff, only for transient errors (429 bursts, 5xx, overload, network); quota, credit and "limit reached" errors are never retried.
+## Realeased 3.2.0 - Final Version
+
+- **Shell on every OS.** Commands run in bash (Linux), a login zsh/bash (macOS) or PowerShell/cmd (Windows) instead of a hard-coded `/bin/bash`; the agent is told which. `OPENSOL_SHELL` overrides. Plan mode has a read-only allowlist per shell family, process trees are killed correctly on Windows, and deploy builds use the same shell.
+- **Smarter compaction.** Keeps the newest turns verbatim, merges earlier summaries, tracks changed files from tool calls, handles huge sessions in chunks, retries, falls back safely, also runs between tool rounds, and has `/compact [focus]`. Summaries show as a divider you can expand.
+- **Better subagents.** Types `explore`/`research`/`review`, parallel calls, live progress in the tool card, round budget with forced report, context pruning, transient-error retry, and approval now actually applies to them. Subagents can no longer start subagents or receive tools outside their type.
+- **New composer and home.** Compact box with the folder picker above it, a **+** menu (files, web search, MCP, skills link), plain-text Build/Plan, permission, effort and model controls whose panels flip and scroll to stay on screen, and a simpler new-session screen.
+- Mascot is now optional: Settings → General → Show mascot (off by default).
+- Cleanup: removed the legacy `src/agent/{compact,subagent,tools,prompt,model,approval}.ts` duplicates, a stale `server/package.json`, unused composer/ChatView code, and merge-conflict markers in `.gitignore` and `.npmignore`. Docs synced (`~/.opensol`, build workflow name, architecture).
+- Agent loop: removed the 16-round cap (subagents now have their own round budget, see above). Provider errors now retry up to 5 times with backoff, only for transient errors (429 bursts, 5xx, overload, network); quota, credit and "limit reached" errors are never retried.
 - Model deck: hovering no longer scrolls the list, which made rows jump.
 - Plugins button disabled for maintenance.
 - Skills: real skill folders (SKILL.md + scripts/references) installed by folder upload or npm, loaded on demand through the new `skill` tool. Built-in skills removed.

package/README.md

@@ -2,36 +2,39 @@
 
 A local-first AI coding agent by OpenSoftX. OpenSol pairs a warm React chat interface with a zero-dependency Node daemon that owns the real workspace, shell, git, search, deploy and provider credentials on your machine.
 
-The daemon binds to `127.0.0.1` only. It serves the prebuilt app from `dist/` and stores everything under `~/.maxcode`. API keys and hosting tokens are encrypted with AES-256-GCM using a local `master.key`; the browser only ever receives a masked hint.
+The daemon binds to `127.0.0.1` only. It serves the prebuilt app from `dist/` and stores everything under `~/.opensol`. API keys and hosting tokens are encrypted with AES-256-GCM using a local `master.key`; the browser only ever receives a masked hint.
 
-## Run the built app or install with npm
+## Run it
 
-Requires Node.js 20 or newer. No `node_modules` is needed to run the built app.
-
-## Install globally via npm
-
-```bash
-npm install -g @bibobonking/opensol
+Requires Node.js 20 or newer (the daemon itself has no runtime dependencies, except `fflate` for plugin packages).
 
 ```bash
-npm start run  
+npm install -g @bibobonking/opensol   # once
+opensol start --open                  # or: npm start, from a checkout
 ```
+
 ```
-  opensol start [--port=4174] [--open] [--foreground]
-  opensol stop
-  opensol status
-  opensol open
-  opensol logs
+opensol start [--port=4174] [--open] [--foreground]
+opensol stop
+opensol status
+opensol open
+opensol logs
 ```
 
 On first launch, choose a real project folder, then add a provider key in **Settings → Models** and pick a live model. Everything else — workspace tree, editor, git, search, build/plan modes, subagents and deploys — runs through the local daemon.
 
+Works on **Windows, macOS and Linux**. See *Shell* below for how commands run on each.
+
 ## What is in the app
 
 - **Build and Plan modes.** Build can edit files and run commands; Plan is read-only investigation.
 - **Live models and reasoning.** Model lists come from each provider; the models.dev catalog is only a fallback. Thinking is the provider's job: an effort control appears only when the selected model advertises effort levels or a token budget, and it is sent automatically.
 - **Themes.** Light, Dark, Night Owl, Claude Code and Flame, plus System, which follows the OS preference. Syntax highlighting follows the theme.
-- **Real tools.** Read/write/edit files, grep, list files, shell, git, web search, URL reading and a read-only SubAgent. Each user turn renders as one answer, with consecutive tool rounds collapsed into a single “Used N tools” group.
+- **Real tools.** Read/write/edit files, grep, list files, shell, git, web search, URL reading, skills and read-only subagents. Each user turn renders as one answer, with consecutive tool rounds collapsed into a single “Used N tools” group.
+- **New composer.** A compact box: the project folder sits above it, **+** adds files, toggles web search and Plan mode, opens MCP and switches skills. Permissions and effort are plain text at the bottom left (a panel opens on click); the model name is at the bottom right and opens the Model Deck (`Ctrl/⌘ M`).
+- **Shell.** The `shell` tool runs in your default shell: bash/zsh (login shell) on Linux and macOS, PowerShell 7 or Windows PowerShell on Windows (`cmd` as a last resort). The agent is told which one in its system prompt. Override with `OPENSOL_SHELL=/path/to/shell`. Plan mode accepts only read-only commands in the matching syntax.
+- **Smart compaction.** Near 85% of the context window (checked before a turn and between tool rounds), older turns are summarised while the newest ones stay verbatim. Earlier summaries are merged, not re-summarised; changed files are tracked from tool calls; if the model is unavailable a deterministic fallback keeps the session going. Run `/compact [focus]` to do it yourself.
+- **Subagents.** `subagent` delegates read-only work to a separate context: `explore` (codebase), `research` (web + code) and `review` (bugs in a diff). Several calls in one message run in parallel, progress streams into the tool card, and each has a round budget and context pruning.
 - **Turn queue.** A prompt sent while a turn is streaming waits its turn and runs next, instead of splitting into a second reply.
 - **Local plugins.** Two plugin modes:
   - **API v1 (sandboxed):** Install a package, review its declared workspace capabilities and HTTPS domains, preview in a sandbox, mark it Safe. Plugin code gets a restricted API (`registerCommand`, `registerPanel`, `registerTheme`, workspace read/write, HTTPS-only network to declared hosts).
@@ -41,7 +44,7 @@ On first launch, choose a real project folder, then add a provider key in **Sett
 - **Deploy.** Cloudflare Pages, Netlify and Vercel, for the open workspace or another existing folder. Tokens stay in the daemon and are passed to CLIs through environment variables.
 - **Local history.** Conversations, settings, workspaces and encrypted keys survive restarts.
 
-Slash commands include `/search`, `/plan`, `/build`, `/new`, `/workspace`, `/deploy`, `/settings` and `/clear`.
+Slash commands: `/search`, `/plan`, `/build`, `/compact`, `/new`, `/workspace`, `/deploy`, `/settings` and `/clear`.
 
 ## Development
 
@@ -71,8 +74,9 @@ OpenCode free-tier models only work inside OpenCode itself, so they are hidden h
 TypeScript, React 19, Vite, Tailwind CSS, Zustand, direct provider adapters (OpenAI, Anthropic, Gemini, OpenCode and OpenAI-compatible endpoints), and a Node HTTP daemon with no runtime dependencies.
 
 ```
-src/        React app, stores, agent loop, adapters and tools
-server/     local daemon, encrypted store, workspace tools and deploy
-dist/       verified production build
-public/     favicon
+src/        React app, Cordis kernel (src/core), stores, adapters and UI
+server/     local daemon, encrypted store, workspace tools, shell, MCP and deploy
+desktop/    Tauri wrapper
+shared/     plugin package contract used by both sides
+dist/       production build served by the daemon
 ```

package/desktop/README.md

@@ -33,7 +33,7 @@ cd desktop && npx @tauri-apps/cli@2 build
 ```
 
 Installers appear in `desktop/src-tauri/target/release/bundle/`. Each OS has to be built on that
-OS (Windows installers on Windows, macOS on macOS); `.github/workflows/desktop.yml` does all four
+OS (Windows installers on Windows, macOS on macOS); `.github/workflows/build.yml` does all four
 on GitHub Actions when you push a `v*` tag or run it manually.
 
 For a quick run from source: `cd desktop/src-tauri && cargo run` (needs `dist/` built).