v0.4.0

Removed

• Workbench tab and Hub-internal chat proxy retired. The Workbench tab
  (SQLite-backed session history browser, timeline/summaries/redacted exports,
  prompt templates, task/policy notes, diagnostics, usage summaries,
  stale-session and worktree helpers) and the associated /api/workbench/*
  routes have been removed. The Hub-internal chat proxy that intercepted
  ANTHROPIC_BASE_URL / OPENAI_BASE_URL to capture request payloads
  (internal/proxy/, internal/hub/chat_proxy*.go) has also been removed;
  wrapped claude / codex sessions now talk directly to the upstream API.
• Side effect: Sonnet 5+ default 1M context restored under the Hub. Because
  Claude Code no longer sees a rewritten ANTHROPIC_BASE_URL pointing at an
  unknown gateway, it stops downgrading to the 200K cap and honours the
  Sonnet 5-era default 1M context window again. No more per-model
  claude-sonnet-5[1m] workaround. Ollama / LM Studio routing (which legitimately
  needs a rewritten base URL) is unaffected.
• Temporary /api/debug/cursor-hide-log diagnostic endpoint (introduced and
  removed within the v0.4.0 cycle).

Added

• Grok Build CLI provider support. many-ai-cli grok (and wrap grok)
  now wraps xAI's official grok terminal coding agent in a PTY, joining
  Claude Code / Codex CLI / GitHub Copilot CLI / Cursor Agent CLI as the fifth
  supported provider. It is handled like Cursor Agent — grok --model <id>
  passthrough with no route or env injection, the shared approval-rules block
  injected into the project-root AGENTS.md, and official approval patterns
  from resources/approval-patterns/grok.md. /api/spawn validation,
  diagnostics, slash-command/approval-pattern sources, and the model picker
  were extended accordingly. Requires a SuperGrok or X Premium+ subscription;
  many-ai-cli never reads, stores, or proxies xAI session tokens.
• opencode CLI wrapper. many-ai-cli opencode (and wrap opencode) wraps
  the community opencode CLI in a PTY. Instead of pattern-scraping approval
  prompts, the Hub writes the provider's own opencode.json (permission
  field) into the session cwd: normal sessions get ask (surface the approval
  to the Hub UI), while orchestration child sessions get allow (bypass the
  provider's own approval so the conductor stays in charge). The original
  opencode.json is restored on session end.
• Light orchestration API. POST /api/sessions/:id/spawn-child lets a
  conductor session spawn child AI sessions with a role, provider, model,
  initial prompt, and optional cwd. The Hub creates
  ~/.many-ai-cli/orchestration/<orchestration_id>/board.md, injects the
  board path into the child prompt, and watches the board for appended
  progress and ## DONE <role> session=<child_id> markers. When the parent
  cwd is a git repository, children run in dedicated worktrees under
  .many-ai-cli/worktrees/<orchestration_id>/<role> by default. The Hub does
  not auto-merge child branches; the conductor / user decides what to merge.
  Also ships companion improvements: send-side duplicate guard, idle-false-
  positive suppression, board notification path isolation, and a
  waitForInputReady staleguard (internal/wrapper/staleguard.go).
• VT-based approval marker detection. In addition to the existing
  xterm.js buffer scan, the Hub now detects [MANY-AI-CLI] markers directly
  from the Hub-side VT stream (internal/hub/approval_marker.go). This
  catches approval prompts that the pure xterm scan used to lose to CUP
  positioning or the cursor-hide filter (a recurring class of v0.3.x bugs).
• Configurable Ollama daemon base URL. ollama.base_url in
  ~/.many-ai-cli/config.yaml can now point Ollama routing at any HTTP(S)
  daemon reachable from the Hub process, not just localhost:11434. This
  supports host/guest setups such as Windows host + Hyper-V guest or Windows
  host + WSL guest: /api/models reads <base_url>/api/tags, Claude Code gets
  ANTHROPIC_BASE_URL=<base_url>, and Codex gets
  OPENAI_BASE_URL=<base_url>/v1.
• Mobile UX overhaul. A dedicated mobile home screen
  (web/src/app/mobile-home.ts + web/src/styles/mobile-home.css), a
  lightweight mobile terminal (mobile-terminal-lite.ts), and a mobile
  transcript view (mobile-transcript.ts) replace the previous
  responsive-only layout. The redesign covers the approval sheet, drawer,
  swipe-based navigation gestures, and PWA polish (manifest, icons, service
  worker updates).
• Claude Code skill-derived slash-command auto-discovery.
  internal/hub/slash_cmd_fetch.go (discoverSkillSlashCmds) walks
  ~/.claude/skills/*/SKILL.md and surfaces registered skills in the Hub's
  slash-command listing.
• Attachment filename sanitization. sanitizeFilename in
  internal/attach/store.go keeps the original filename visible while
  stripping path-traversal and unsafe characters from the on-disk path.
• Consolidated build info. internal/hub/buildinfo.go centralizes the
  version / build metadata exposed by /api/info and the About panel, so the
  ldflags-injected values have a single source of truth on the server side.
• Session log noise filter. A repetitive-progress-line filter is applied
  when writing the clean transcript (internal/sessionstore/store.go +
  noise_output_test.go), so re-drawn progress bars no longer dominate the
  saved .txt.
• Regression tests for input splitting (internal/wrapper/input_split_test.go)
  and the cursor-hide filter (web/src/app/cursor-hide-filter-fixtures.ts).

Changed

• README (English / Japanese) opening hook rewritten around "run multiple AI
  CLIs in parallel and approve from your phone" so the primary value
  proposition is clear from the first paragraph.
• Ollama routing env-var assembly now flows through
  EnvPresetForWithOllamaBase, deriving ANTHROPIC_BASE_URL /
  OPENAI_BASE_URL from ollama.base_url in one place.
• Windows resource metadata (winres/winres.json and the regenerated
  rsrc_windows_*.syso) is bumped to 0.4.0 so the executable properties
  dialog matches the release tag.
• The session-list / spawn-panel / settings / terminal / token-statusbar
  frontends were reorganized to share layout primitives with the new mobile
  UI without regressing the desktop Hub.

Fixed

• Hotfix → develop merges no longer leave the develop-side version constant
  stale (develop-version-stale-v0.3.4).
• Cursor Agent no longer drops chunks of larger inputs that contain a Ctrl+U
  (cursor-agent-ctrl-u-chunk-input-drop).
• Grok history viewer no longer renders blank rows caused by leftover SGR
  state (grok-history-viewer-sgr-blank), and the Grok build display label
  no longer breaks alignment (grok-build-display-label).
• Grok responses and opencode pickers are no longer eaten by the cursor-hide
  filter (grok-response-discarded-by-cursor-hide-filter,
  opencode-picker-discarded-by-cursor-hide-filter).
• The orchestration conductor's initial prompt is now submitted with an
  Enter (orchestration-conductor-initial-prompt-not-submitted), and
  codex child sessions spawn reliably again
  (orchestration-codex-child-spawn-failures).
• Mobile drawer no longer shows a transparent background layer or lingers on
  PC-width viewports (mobile-drawer-bg-layer-transparent,
  mobile-drawer-residue-on-pc-width). The floating hamburger, header height,
  and tab bar dead space were also fixed
  (mobile-header-band-floating-hamburger, mobile-header-height-reduction,
  mobile-header-tab-bar-dead-space).
• Empty input bar no longer renders a stale placeholder height
  (input-bar-empty-placeholder-height), and clicks in the empty wrap area
  no longer lose focus (input-wrap-empty-area-focus).
• The Hub marker filter no longer swallows CUP positioning, so layouts
  driven by cursor positioning stay intact
  (hub-marker-filter-strips-cup-positioning).
• Batch approval Q-tabs no longer overflow the action bar
  (batch-approval-tabs-overflow), and marker blocks no longer collapse
  onto a single line (marker-block-line-collapse).
• The spawn cwd chip no longer collides with a full-path prefix
  (spawn-cwd-chip-fullpath-prefix-collision).
• Approval popups no longer leak across sessions
  (approval-popup-cross-session-leak).
• Local-LLM routing fixes for Qwen via LM Studio / Ollama
  (local-llm-qwen-lmstudio-ollama-unusable).
• WebGL terminal renderer works again (vendored xterm.js generation mismatch).
  The vendored xterm-addon-webgl.min.js had been updated to 0.19.0 (the
  xterm.js 6.0.0 generation, which requires the mainDocument API of the
  terminal core), while xterm.min.js and the fit / unicode11 / web-links
  addons were still the older 5.x-generation files — only the license ledger
  and the About dialog had been bumped to 6.0.0. Every loadAddon call then
  threw Cannot read properties of undefined (reading 'createElement') inside
  the addon, so the UI silently fell back to the DOM renderer on each session
  switch, re-introducing the fullwidth-glyph / selection-highlight drift the
  WebGL renderer exists to fix. The vendored core (xterm.min.js,
  xterm.min.css) and the three addons are now the actual 6.0.0-generation
  artifacts, matching the ledger.
• Orchestration fallback ID no longer collides across Hub restarts. When a
  conductor started a child via spawn-child without going through
  /api/orchestration/create, the orchestration ID was s<parentSessionID>
  (e.g. s3). After a Hub restart the session number was reassigned from #1,
  so the same conductor session ID could reuse the previous run's
  ~/.many-ai-cli/orchestration/s<N>/board.md (header/purpose stayed from the
  old run) and its child-<sessionID>.md history. The fallback is now
  s<parentSessionID>-<UnixNano>, which is unique per Hub start. Existing
  s<N>/ folders are harmless leftovers and can be moved aside or deleted
  manually.
• No more scroll residue when switching between terminals. attachTerminal
  used to append the terminal DOM node first and then flush the buffered
  chunks that had accumulated while the tab was hidden, which made the target
  tab briefly show the pre-switch viewport and then scroll down through the
  buffered output. flushPending now takes an onDrained callback, and
  attachTerminal flushes first — appending the container, releasing hidden
  WebGL renderers, scrolling to the bottom, and fitting the terminal only
  after the buffered chunks are fully written.