v0.15.0

[0.15.0] - 2026-09-09

Fixed

• Release workflow pins companion WebView 0.1.5 (published assets + Homebrew)
  and crates.io a3s-tui 0.1.15 (with_canvas_rgb,
  DiffView::emphasize_inline_changes). Cargo.lock records the registry
  source/checksum so --locked CI succeeds without the local .cargo path patch.
• Release sandbox proves native policy on macos-latest (arm64 zvec prebuilt).
  Intel (x86_64-apple-darwin) binary jobs provision libzvec_c_api via CMake
  when the zvec-rust prebuilt is absent.
• Pin a3s-code-core to git rev e35e7086 so DeepResearch multi-source Flow
  step identity admits ≈96 KiB selector shards via Core digest-fold (64 KiB
  soft inline / 512 KiB hard ceiling). Capability gate:
  ./scripts/verify-capability-regression.sh --require-published.

Changed

• Executable Tool hermetic no longer depends on unpublished
  PreparedCapability::into_value; prepare readiness + direct spawn instead.
• Code TUI first-principles prune: drop dead wrappers (App.anim, unused
  config dir helpers, orphan open_window/humanize/write_asset_acl, unused
  DeepResearch spawn shim, unused progressive HTTP entrypoints, unused
  OsService/MessageTone/RuntimePolicy variants). Keep live
  execute/journal/RemoteUI/open_window_with and Effect hermetics.
• Code TUI PromptBar idle glyph is a muted non-bold → (Cursor-style), with
  empty-state placeholder Add a follow-up. Shell ! / research ? stay
  bold colored mode marks.
• Sticky Reviewer path: Core Gate admission + independent AuxiliaryExecutor
  (protocol hermetic rubric) on ReplyVerifierLane. Incomplete evidence is
  fail-closed (no published clean). Explicit /review stays on
  GitReviewLane with AgentStyle::CodeReview. Sticky no longer uses a
  CodeReview side-session.
• First-principles full TUI test plan (refuse overfit; Effect vs Contract
  honesty; prioritized backlog beyond Z/M/R/B):
  docs/tui-first-principles-test-plan.md. Phases A–C Exit gates proven
  against existing Effect hermetics (dogfood, Plan/sandbox/HITL, DR1/G1);
  Phase D remains regression-only.
• Use Flow honesty: non-resident catalog/run APIs are hermetic-only;
  production FullCompatibility keeps projection-only staging/preflight.
• Capability regression one-shot + digest-fold pin scripts under scripts/.

Added

• First-principles review of Reviewer + capability matrix work:
  docs/capability-first-principles-review.md. Fixes: split sticky vs git
  host postures; git /review no longer clears sticky open findings;
  main-turn open-findings prefix is unit-tested (skips /loop continuations).
• Capability first-principles test matrix for zvec/BM25, ReMe-like memory,
  Reviewer claim-vs-record, and default Moli web_search (effectiveness +
  efficiency cases + focused runbook): docs/capability-test-matrix.md.
• Sticky Reviewer (/reviewer / Shift+Tab) is a claim-vs-record reply
  verifier: after each main turn it reviews the assistant message against
  the user request and bounded turn tool evidence on an isolated priority
  lane (does not block the main stream). Explicit /review [working-tree|…]
  remains workspace git code review. Open reply findings inject into the next
  main-turn prompt until addressed or waived (w in the checklist). Issue
  checklists deferred while the composer or queue is busy open automatically
  once both are idle.
• Markdown mermaid sequenceDiagram fences render as Cursor-style terminal
  architecture art (participant boxes, lifelines, arrows, alt/opt/loop frames)
  via a3s-tui; other Mermaid dialects stay as code. Open mermaid fences stay
  in the streaming mutable tail until closed so diagrams can reflow.
• Code TUI tool approvals: Cursor-style cyan countdown bar above the permission
  prompt; remaining time shrinks left→right, selection uses -> + hotkeys, and
  the bar auto Skip & tell (approval timed out) when it runs out. Default
  15s; override with A3S_CODE_APPROVAL_TIMEOUT_MS (0 disables countdown).
• Code TUI Ctrl+G Diff review: reopen the latest turn's successful file
  edits as a full DiffView overlay (←/→ switch files, j/k scroll, i seeds a
  follow-up draft about the selected path, Esc closes). Ctrl+R remains prompt
  history. Tracked in docs/tui-cursor-alignment-roadmap.md Phase 1.
• DiffView word-level peek: adjacent delete/insert pairs emphasize changed
  tokens via emphasize_inline_changes (used by Code file-change cards and
  Diff review).
• /ask and /plan composer aliases arm read-only Plan mode (Cursor Ask maps
  to plan without adding a sixth Shift+Tab state). a3s code exec --mode ask
  is accepted as an alias of --mode plan. Help/banner document the
  agent/plan/ask/reviewer/auto/yolo legend.
• /terminal now appends copy-paste repair snippets for weak Shift+Enter /
  enhanced-key / multiplexer passthrough (Cursor /setup-terminal spirit).
  Opt-in turn-complete notify via A3S_CODE_NOTIFY=1 (BEL + OSC 9) and one
  dim idle tip via A3S_CODE_SUGGEST=1 (heuristic; Ctrl+G when the turn
  edited files). Tracked in docs/tui-cursor-alignment-roadmap.md Phase 3.
• a3s code --worktree [NAME] creates an isolated Git worktree under
  .a3s-worktrees and starts the interactive TUI there (cold-start counterpart
  to /fork worktree). Binds managed lifecycle so /worktree cleanup prints
  non-forcing remove guidance. Phase 4.
• Sticky skill mode: Alt/Option+Enter on a $skill menu pick keeps that skill
  attached across turns (footer sticky:<name>); Esc on an empty composer or
  /unstick clears it. Enter alone remains one-shot mention completion.
  Phase 5.

Fixed

• Empty-transcript viewport rebuild preserves the welcome banner (deferred
  startup no longer flashes the logo then blanks it).
• Busy consecutive submits: defer user transcript until queue admission so
  interrupt markers stay under the interrupted turn; after Thought finalize,
  move interrupted after the sealed Thought block.
• Session-store open uses Core new_recovering_corrupt_wal so a duplicate-WAL
  sequence conflict quarantines the corrupt WAL and continues from snapshots
  (requires a3s-code-core = 8.5.1).
• Sticky Alt/Option+Enter clears the composer after attach so Esc can detach
  immediately (leaving $skill previously blocked empty-composer Esc clear).
• A3S_CODE_NOTIFY / A3S_CODE_SUGGEST still fire when sticky Reviewer
  side-sessions spawn after the main turn (primary queue empty).
• Ctrl+G empty notice distinguishes in-flight file edits from a turn with no
  completed edits.
• /worktree and unmanaged-session errors mention a3s code --worktree beside
  /fork worktree.
• Sticky skill injection outranks the disabled-skill list for the current
  session, so a footer sticky:<name> chip cannot silently no-op after a
  plugin disable.
• Diff review Home/End jump to scroll extremes (with Ctrl+G refresh path
  retained).
• Sticky skill footer chip (sticky:<name>) remains visible in Zen display
  profile (mode-like session state was previously dropped by the Zen early
  return).
• Ctrl+G while Diff review is open refreshes from the latest turn and keeps the
  previously selected path when it still exists; empty refresh closes the
  overlay.
• Large bracketed text pastes (≥15 lines or ≥400 characters) collapse into
  PromptBar pills so the six-row composer budget is preserved; submit and
  history recall still carry the full pasted body to the agent (Cursor CLI
  large-paste grammar). Backspace on an empty draft removes the newest pill
  before image chips; click expands a pill into the draft, × removes it.
• Code TUI empty PromptBar now shows a focused dim placeholder
  (Type a message, / for commands…), matching Cursor CLI empty-state
  grammar instead of a blank focused box. Superseded in Unreleased by
  Add a follow-up + muted → glyph.
• PromptBar auto-grow caps at six visible rows (Cursor CLI visual-line
  budget) and scrolls overflow inside the bar; ↑ on the first multiline
  row recalls session history while mid-draft ↑/↓ still move the caret.
  Interaction grammar vs Cursor CLI is tracked in
  docs/composer-input-grammar.md.
• Code TUI pins each turn's reply language from the user's message script
  (zh-CN / ja / ko / en) via Core set_output_language, and the default
  system prompt forbids mixing languages in user-facing prose while keeping
  code, paths, and quoted source unchanged.
• Delegated task fan-out renders a checklist of every task title instead of
  collapsing to a; b +N more on one header line, so parallel work stays
  visible while running and after completion.
• Deferred workspace retrieval and ordinary catalog enable no longer open the
  durable zvec projection during the pre-TUI Loading workspace… window.
  Durable FTS attaches on first Grep/BM25 demand; embeddings still wait for the
  first-frame gate.
• macOS PTY first-frame regression launches a3s through a SIP-safe wrapper so
  expect-driven probes can load native zvec without DYLD_LIBRARY_PATH, and
  asserts the in-process terminal_handoff / first_frame_flushed budget
  instead of a stderr-TTY Loading string that redirect-based probes cannot see.

Added

• Default-mode agents can call Core update_plan to maintain the TUI pinned
  checklist mid-turn (TaskUpdated + existing Codex-compatible plan args).
• Code TUI explore / web tool cards surface pagination windowing (offset,
  limit, max_chars, …) on compact Exploring / Explored and fetch rows;
  read also shows a 1-based L… range when both offset and limit are set.
• Code TUI message-stream image previews: submitted Ctrl+V / @ images keep
  half-block Unicode thumbnails in the user bubble (click opens RemoteUI). Session
  resume rebuilds previews from persisted ContentBlock::Image data.
• Sticky /reviewer composer mode (also on Shift+Tab): arms an isolated
  reviewer lane (own a3s_lane priority queue + ReviewerMsg bus) that runs
  Core AgentStyle::CodeReview side-sessions to critique the latest assistant
  message reply after each main turn (read + verify, no auto-edits) without
  changing or blocking the main agent stream. One-shot /review [working-tree|…]
  enqueues git-scoped code review on the same lane.

Changed

• Code TUI surfaces mid-turn plan progress and explore density from first
  principles: the live viewport appends the plan checklist (progress header
  plan · done/total · N active, shared Checklist glyphs), and adjacent
  explore calls stay one transcript block with count summary, one detail row
  per call via the shared explore_detail formatter, and … N more windowing
  for older rows.

• Advanced the monorepo crates/code gitlink to Core 8.4.0 tip and hardened
  Code TUI integration from first principles:

  • Shift+Tab Plan now calls Core AgentSession::set_agent_style so PROMPT-ALIGN1
    Plan system prompts track the host read-only boundary without a full rebuild.
  • code exec --mode plan installs AgentStyle::Plan on session prompt slots.
  • Lexical catalog / zvec FTS attach lazily on first search/bm25 access
    instead of during session construction.
  • Plugin Manager host construction is deferred past the first TUI frame
    (authorization still loads on the critical path; /packages stays
    fail-closed until ready).

• Code TUI/code exec always attach the lexical chunk catalog and best-effort
  persistent zvec FTS for search bm25, even when semantic
  workspace_retrieval is disabled. Cold queries keep the portable catalog
  scorer until a durable generation is ready.

• Documented Moli as the default headless web_search browser (auto-download
  on first use); Chrome and Lightpanda remain explicit ACL backends.

• Integrated managed Code runtime a3s-code-core 8.4.0 into the Code TUI
  (submodule tip v8.4.0, release pin, and release-workflow contract). Plan
  mode now sets explicit Core AgentStyle::Plan on session prompt slots
  (PROMPT-ALIGN1) so the Plan system prompt and read-only repository-tool
  contract apply at launch and session rebuild.

• Code TUI tool history is brief by default: successful shell/generic
  rows keep the verb line plus at most one result sentence (or a Ctrl+T
  expand hint); successful MCP JSON stays out of the main stream; compact file
  diffs use a short borderless peek. Failures still show a bounded error body.
  Ctrl+T remains the full semantic transcript.

Removed

• Removed the Code TUI/CLI five-pack asset surfaces (/agent, /mcp, /skill,
  /flow, /okf) including pickers, local *_dev modes, clone/scaffold
  workflows, OS publish/activity panels, and matching a3s code subcommands.
  Skill discovery, /plugin, /kb, Use MCP registry, workspace /review, and
  /evolution asset ACL materialization remain. Follow-up honesty scrub removed
  leftover five-pack docs from cli-reference.md, deleted the unused
  os/progressive client module, and pruned orphan chrome helpers that only
  served those surfaces.

• Removed Agent Island, its TUI commands and heartbeat/control bridge, and the
  native floating window mode. RemoteUI and Workspace Host remain available.

• Removed the Anthropic sandbox-runtime package, Node.js bootstrap, managed
  support tree, integrity patcher, legacy archive bridge, installer and
  self-update support-payload transactions, and related release dependencies.

Added

• Added an A3S-owned native local command sandbox across macOS Seatbelt,
  Linux namespaces and seccomp, and Windows AppContainer and Job Objects, with
  a real boundary probe in each platform release gate.

Changed

• Code TUI terminal takeover no longer waits on crossterm's Kitty keyboard
  enhancement device probe (up to two seconds when a host ignores the query).
  Progressive-key flags are pushed unconditionally so the welcome logo and
  first frame appear immediately after session construction.

• Code TUI deferred first-frame services no longer paint a "Loading background
  services" / "starting…" chrome once the editor is ready. /auto and /yolo
  remain typed aliases for Shift+Tab autonomy modes but leave the empty /
  browse list. Session debug writes to editor-local debug-*.log paths were removed from
  the TUI path.

• Code TUI slash IA: /ctx is the context hub (memory / kb / sleep /
  evolution plus existing search/attach/save). /use is the integrations hub
  (plugin / packages / reload plus status/repair). Skin commands, OS asset
  clones, /ide, /goal, and /loop stay typed but leave the empty / browse
  list; legacy slash forms still work and tip the preferred hub path.

• Code TUI /research <query> starts DeepResearch; status / explain /
  replay / diff remain diagnostics. Leading ? is a shortcut that tips
  /research and no longer enters sticky composer research mode.

• Unified Code TUI DeepResearch onto the typed CodeDeepResearchRunner path:
  /research diagnostics read journal-v2 first (Inquiry fallback remains),
  typed fail/cancel can salvage published or recovery artifacts, smoke uses
  the same runner as a3s code research, and startup recovery settles
  interrupted journal-v2 runs.

• WebView health checks now accept published helpers that only expose RemoteUI
  via --help --url, and honor A3S_WEBVIEW_BIN during component discovery so
  first-use setup no longer marks a working companion as unhealthy.

• Homebrew now installs the bundled a3s-webview companion from the CLI
  release archive next to a3s, and no longer depends on a separate
  a3s-webview formula download.

• Updated the managed Code runtime to a3s-code-core 8.4.0 from the monorepo
  crates/code checkout, including a3s-sandbox 0.1.1 for the native local
  command boundary used by a3s code TUI and exec.

• Bash now fails closed in every execution mode when the native boundary is
  unavailable; only an explicit approved require_escalated request can cross
  to the host when a verified sandbox is attached.