release: v1.7.0 — the multi-agent system, rebuilt

docs/NEWS.md

@@ -2,6 +2,10 @@
 
 Full news history for ClawCodex. The [README News section](../README.md#-news) keeps only the 10 most recent items.
 
+- **2026-09-23 (v1.7.0):** **The multi-agent system, rebuilt — persistent teams, workers that really resume, and one supervisor for every agent (#915, #950)** — ClawCodex used to spawn subagents down two paths that never met, and only one of them was observable. A foreground delegation registered nowhere, so nothing could list or stop it; the TUI's agents overlay called three RPCs that had no backend; a follow-up sent to a finished background worker flipped it back to `running` without ever starting a model loop; and the team tools existed with no production path that actually ran a teammate. v1.7.0 rebuilds that layer. **One supervisor (#915, #950):** a session-scoped supervisor admits every worker — foreground, background, team and workflow — behind two backstops (`CLAWCODEX_MAX_CONCURRENT_AGENTS`, default 32; `CLAWCODEX_MAX_AGENT_DEPTH`, default 3), with live status, per-agent interrupt and a session-wide pause on new spawns in the TUI's agents overlay and the web client's subagent list (#933); a refused spawn comes back as a tool error the model can act on. **Persistent teams (#950):** `TeamCreate` makes your session the leader; each named `Agent` call adds a teammate that stays alive between assignments with its context intact; `SendMessage` routes findings to a named peer or to the leader; a shared, locked task board hands out work in dependency order; plan approvals and shutdowns follow a matched request/response protocol; and interrupting a worker withdraws its pending permission prompt. **Lifecycles that hold (#950):** resume reloads the worker's history under the same ID, a correction accepted while it was finishing is no longer left unread, notifications reach the session and parent that own them, isolation runs in a real Git worktree or fails before the model runs, and ending a session interrupts every worker it owns and waits a bounded time for them to stop. Verified by end-to-end suites that drive real query loops, tools, worktrees and WebSocket connections against a scripted provider, plus a live DeepSeek smoke run — and scoped plainly: teams are in-process (no tmux/iTerm pane or remote backends), one per workspace, a worker resumes only within the session that started it, and there is no automatic crash recovery ([verification notes](multi-agent-runtime-verification.md)). **Also in v1.7.0:** `clawcodex --nano`, the pi-style minimal harness — a ≈2K-token fixed payload instead of ≈17K, and level with the pi harness on the full Terminal-Bench 2.1 suite (`deepseek-v4-flash`, k=1): 64/89 and 63/89 in nano's two latest runs, against pi's 63/89 (#879–#885, #889–#891, #894, #896–#902); a Bash fix for every mode, where a command printing more than ~64 KB stalled until the timeout and came back cut off (#900); DeepSeek-V4.1-Flash as the DeepSeek default, with `/cost` following its new peak/off-peak card (#905, #925); model and effort picks saved as your default on every interface (#930); cost-aware auto-compaction (#903); ChatGPT-subscription model discovery (#913, #917); and a round of web-client work — subagents in the header with a child view per run (#922), a sidebar of tabs that reads the workspace (#918), attachments of any file type (#949), and saved sessions that open in milliseconds instead of ~45 s (#947). **Upgrade notes:** `clawcodex serve` now refuses a non-loopback bind without `--allow-remote` (#921), and `clawcodex agent-server` refuses one without `--token` (#924).
+- **2026-08-15 (v1.6.0):** **ClawCodex Web — the whole agent in a browser tab (`ui-web/`, `clawcodex web`)** — one command serves the full agent at a stable, typeable address: **`http://127.0.0.1:8081`**. It is deliberately *not* a second server: the browser drives the **same in-process agent, JSON-RPC gateway, and durable session store as the TUI and the Desktop app** — start a session in one surface, resume it in another. The UI is a three-column shell (session tree | conversation | details) with streaming replies and collapsible reasoning, live tool cards (ANSI-colored terminal transcripts, unified diffs, line-numbered file reads), permission approvals as a composer takeover with once/session/always grants, a prompt queue for follow-ups typed mid-turn, slash-command completion, a context meter with per-category breakdown, a To-dos panel folded live from the transcript, and light/dark/system themes. Two views per session: **Chat** is the conversation; **Trajectory** is the same run as a metered ledger — every model request and tool call on a three-lane timeline with per-step tokens and timing (TTFT, throughput, cache-hit rate), model time and tool time reported separately. A **Settings** page covers what used to need slash commands or hand-editing `config.json`: provider **API keys** (add / replace / disconnect, stored where `clawcodex login` stores them), the **default provider** new sessions start on, the approvals default (with a Full-access confirmation that means it), response language, output style, and the end-of-turn recap toggle. `GET /` inlines the session token so the bare URL works after launch — and because of that, `clawcodex web` binds loopback only unless you explicitly pass `--allow-remote`.
+- **2026-08-09 (later that day):** **ClawCodex Desktop runs and packages on Windows** — `npm run dist:win:nsis` in `ui-desktop/` now produces a working unsigned NSIS installer (`ClawCodex-<version>-win-x64.exe`, crab icon and version info stamped via rcedit, node-pty **conpty** binaries staged), and `clawcodex desktop` launches the dev app natively. The packaged app boots the same `%USERPROFILE%\.clawcodex\clawcodex` backend that `install.ps1` creates — **one shared install, config, and session store across CLI, TUI, and Desktop**. Verified live on Windows 11: silent NSIS install → app boot → backend spawned from the shared venv → healthy loopback gateway → real DeepSeek turns on the shared keys. Also fixed along the way: the in-app updater's relaunch resolver pointed at the upstream `apps/desktop/` layout (never matching `ui-desktop/`), its bash relaunch handoff now honestly reports manual-restart on Windows, and a hashbang broke vitest collection of the packaging tests.
+- **2026-08-09:** **Native Windows support for the CLI** — ClawCodex now runs first-class on Windows 10/11 (PowerShell / cmd / Windows Terminal, no WSL required), installed with one line: `irm https://clawcodex.app/install.ps1 | iex`. The new `install.ps1` mirrors `install.sh` end to end — uv install, Python provisioning, lock-pinned deps, PATH registration, TUI build, plus the same `doctor` / `verify` / `update` / `uninstall` lifecycle. Under the hood the port is structural, not cosmetic: a shell platform layer resolves **Git Bash** for the Bash tool (explicitly refusing the WSL `System32\bash.exe` shim), so shell commands keep their POSIX semantics everywhere; process-tree kills go through `taskkill /T`; the mailbox/transcript/lockfile gain real `msvcrt` locking (the Windows CRT's `O_APPEND` emulation can silently *lose* concurrent writes); the permission layer folds NTFS case-insensitivity and refuses whole-drive grants and drive-relative escapes (deny-side only); and `@`-mentions, `CLAWCODEX.md` `@includes`, and persistent-`cd` tracking all round-trip real `C:\` paths. The full test suite now runs on `windows-latest` alongside Ubuntu in CI.
 - **2026-08-08 (v1.5.0):** **ClawCodex Desktop — the whole agent in a native app (#802–#808)** — ClawCodex now ships a real desktop application (`ui-desktop/`): streaming chat with a live tool trail and reasoning, permission approvals with once/session/always grants, a session sidebar that lists and resumes the **same durable sessions as the TUI**, side-by-side previews, settings, and the official pixel-art crab as the dock icon and in-app brand mark. `clawcodex desktop` launches it from a checkout (first run installs the UI deps; `--no-dev` builds once and launches Electron directly). The architecture is the interesting part: the app spawns **`clawcodex serve`** — one loopback port serving `/api/*` REST plus a JSON-RPC WebSocket gateway at `/api/ws` — and sessions run on the **same in-process agent core the TUI uses**, so both surfaces share one config, one session store, one skills set, and one permission system; the wire contract is the TUI's own gateway vocabulary, adapted server-side. The port itself is one of the largest single features ClawCodex has landed: ~310K lines of TypeScript across ~1,500 files brought over from the reference desktop implementation, rebranded end to end, with every quality gate held to the reference's own baseline (typecheck green, lint identical, unit-test failure set byte-identical) and the whole loop verified live — boot → real chat turn → streamed reply rendered in the window. macOS packaging works today (`npm run dist:mac` → DMG/zip via electron-builder, hardened-runtime config in place). Ship-week fixes landed the same day: the root `.gitignore`'s Python-oriented `lib/` pattern had silently kept 180 renderer source files out of the initial import — fresh clones failed at boot until #806; the default UI scale moved from the reference's dense 90% preset to Chromium's 100% actual size (#807); and a second `clawcodex desktop` now gets a friendly "already running" message instead of a vite stack trace (#808).
 - **2026-08-02 (v1.4.0):** **Fusion models — give a text-only model vision (#771, #787)** — several strong reasoning models cannot see images at all: `deepseek-v4-pro` rejects an image content block outright (`400 unknown variant \`image_url\``), so pasting a screenshot, `@`-mentioning one, or letting `Read` return one ended the turn. A **fusion model** pairs that base model with a second, vision-capable one — every image is described by the vision model first, and the base model reads the description. `/fusion create <name> <base> <vision>` saves one; it then behaves like a normal model in the `/model` picker, as `--model <name>`, in `-p`, and across restarts. Ported from [claude-code-router](https://ccrdesk.top/en/configuration/fusion-models/)'s Fusion Model concept, with one deliberate difference: CCR is a proxy, so it can only offer vision as a *tool* the model may choose to call — which cannot help a pasted image, already on the wire before the model gets a turn. ClawCodex owns the agent loop, so it substitutes images in place, covering paste, `@file.png`, `Read`, and Bash image output at once. Verified end to end on Terminal-Bench 2.1's `code-from-image` task — transcribing handwritten pseudocode from a PNG and reproducing its output — with `deepseek-v4-flash` + `openai:gpt-5.6-luna` (#787); the base model alone returns a 400 on the same image. **Also in v1.4.0:** GPT-5.6 Sol/Terra/Luna (#773); four more OpenAI-compatible providers — groq, cerebras, baseten, xai — taking the registry to 30 (#784); `/mode` becomes `/permissions` with a three-level picker and Full Access by default (#768); `AskUserQuestion` finally renders a real picker instead of returning JSON to the model (#774); the OpenAI provider now picks its wire protocol from the model rather than the auth mode, which is what makes `gpt-5.6-luna` usable on an API key (#783); cached prompt tokens are billed at the cache rate instead of the full input rate, and OpenRouter's streamed reasoning is no longer discarded (#785, #786); and headless runs stop reporting a cut-short run as a success (#777–#782).
 - **2026-07-29 (v1.3.0):** **ClawCodex scores 80.9% on Terminal-Bench 2.1 — a top-tier open-source result on Opus 5 (#720–#725, #747–#754)** — running headless on `claude-opus-5` at `effort=xhigh`, ClawCodex solved **72 of 89** Terminal-Bench 2.1 tasks: **80.9% pass@1** on a single run. On the [public 2.1 leaderboard](https://www.tbench.ai/leaderboard/terminal-bench/2.1) (k=5 averages) that would slot **around third** — behind Claude Code / Fable 5 (83.8%) and Codex / GPT-5.5 (83.1%), statistically level with the 79–80% cluster, and **ahead of Claude Code on Opus 4.8 (78.9%) and Sonnet 5 (74.6%)**. Getting there was open, unglamorous parity work: a Harbor eval adapter (`eval/harbor/`) for three-way ClawCodex-vs-openclaude-vs-Claude-Code runs (#720, #724, #725), then a run of prompt- and reliability-parity fixes — restored task-tool skip conditions and parallel-tool guidance, deferred nonessential initial tools, and recovery of trials lost to empty turns and transport drops (#747–#754). **Also in v1.3.0:** `claude-opus-5` support with an interactive `/effort` fix (#746), bounded persistent memory with a background self-improvement review (#731), a VS Code extension driving the agent-server over stdio (#727), image-paste input with an `[Image #N]` un-attach chip (#761, #762), the `CLAUDE.md → CLAWCODEX.md` context-file rebrand (#732), and transport-retry hardening (#757, #760). Stated plainly: this is a single k=1 pass (binomial 1σ ±4.2pp) against the board's k=5 ± ~1.2pp averages, benchmarked on `main` at #756 (before the v1.3.0 tag), so read it as directional rather than a ranked submission.