## Summary
This PR implements dual model support (main + turbo) for cost-optimized
operations and adds model sorting functionality.

### Key Changes

#### Dual Model Support
- Added `turbo_model` and `turbo_cost_threshold` config options
- New dual-pane model selector (Tab to switch panes)
- Visual indicators: ⚡ (turbo), 🧠 (reasoning), 🔧 (tools)
- Smart turbo model selection with fallback logic
- Fixed Anthropic auth showing incorrect turbo status

#### Model Sorting
- Added `release_date` and `last_updated` fields to model schema
- Press 'S' to cycle through sort modes: Name, Last Updated, Release
Date
- Stable multi-level sorting with consistent ordering
- Helps discover newest models and track updates

### Impact
Enables cost optimization for PRs anomalyco#269 (status verbs) and anomalyco#275 (window
titles) to use configured turbo models instead of hardcoded ones.

### Demo



https://github.com/user-attachments/assets/d076e840-d790-4e23-8cf7-a179282d1f0b

…ompaction

feat: intelligent compaction with model-initiated trigger, handoff prompts, and persistent memory

- Add VoyageEmbeddingProvider for Voyage AI embeddings
- Add MemoryStore with high-level store/search/list API
- Configure Nebius cloud API for Qwen3-Embedding-8B (#1 on MTEB)
- Fix LocalEmbeddingProvider to support both /v1/embeddings and /embeddings
- Fix QdrantVectorStorage to set collection on existing collections

Model: Qwen3-Embedding-8B (70.58 MTEB score, 99% zero-shot, 4096 dims)
Provider: Nebius API (https://api.tokenfactory.nebius.com/v1)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

…s / smaller nodes

Address user's 6-point list after first visual pass:

1. Trace overlap + session id readability
   - SessionPicker now displays session TITLE (not the cryptic id), pulled
     from sync.data.session. Fallback to "…<last8>" only if no Session
     record loaded yet.
   - SessionSummary type gains `title` field.
   - .rt-pop-title CSS clamps to 240px with ellipsis so long titles never
     overflow the trigger row or dropdown.
   - .rt-pop-menu / .rt-pop-scrim z-index bumped (60/70) so the dropdown
     floats above any module body.

2. Rail icons + hierarchy
   - Replaced @opencode-ai/ui Icon usage with inline SVG ICONS map
     (workflow / knowledge / trace) matching the design template's
     stroke-1.5 rounded-cap shapes.
   - .rune-rail-subs gained a 1px left guide line + 26px margin-left,
     and each sub gets a 4px dot prefix that turns accent-blue when
     active. Hierarchy is now visually obvious instead of the previous
     flush-with-parent layout.

3. Rail badges
   - Knowledge: refiner-page publishes `pendingReviewCount` (count of
     experiences with review_status === "pending") via the bridge as
     `railBadges.knowledge`. Shows the bubble next to "Knowledge".
   - Workflow: session.tsx publishes `runningCount` (root sessions
     updated within last 60s as a freshness proxy for "running") via
     `railBadges.workflow`.

4. Trace session by title — same as anomalyco#1 above.

5. Workflow tasks as Rail sub-list
   - session.tsx now publishes `railSubs` with the same root-sessions
     list it pushes to the Tasks Drawer. Each task renders as a Rail
     sub-row under Workflow. Drawer stays as the richer overlay.

6. Workflow node cards smaller + minimal status
   - workflow-node.tsx rewritten to compact form: type icon (24px),
     single-line title with truncate (was 2-line wrap), tiny status
     dot/spinner, arrow. Removed inline progress percent + summary
     chips (those bloated the card to 80+px tall).
   - Progress bar moved to a 2px full-width strip beneath the row.
   - .wf-node-compact CSS: 6px×10px padding, max-width 240px, 8px radius.
     Cards now fit ~3-up in a 760px lane and scan at a glance.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…yout

Batch fix for issues found in the user's workflow session:

1. **Master not woken on final node.completed (anomalyco#7)**. The wake
   subscriber at workflow/index.ts:1820 dropped ALL wakes when
   workflow was in a terminal state. The user reported the FIRST
   nodes wake master fine, but the LAST node's completion races with
   the master's `workflow_finalize` call — by the time the
   node.completed event hits the subscriber, workflow.status is
   already "completed" and the wake gets silently dropped. Now
   "signal" kinds (node.completed / failed / interrupted /
   attempt_reported / budget_exceeded / stalled / checkpoint.*) ALWAYS
   wake master regardless of workflow status; chrome kinds (wake_*,
   pulled, command_acked, graph.edit.*) still drop on terminal so
   we don't wake forever.

2. **Session compaction silent failure (anomalyco#8)**. compaction.ts only
   published `session.compacted` on success — failure paths only
   set `message.error` on the last assistant message which was
   invisible unless the user scrolled back. The session just
   "froze" with no indication. Added a new `session.compaction_failed`
   bus event, wired the frontend event-reducer to park it on a new
   `sessionError[sessionID]` slot, and the workflow chat panel now
   appends a system message banner (`⚠ 会话压缩失败：…`) with
   actionable guidance (new session, bigger model context window).
   Successful re-compaction clears the banner.

3. **Question dialog deadlock in node sessions (anomalyco#9)**. Frontend
   polling for /question and /permission was gated on
   `status() === "running"` — but a sub-agent calling question()
   transitions to idle BEFORE the master picks it up, so polling
   stopped and the dialog never appeared, locking the session.
   Removed the gate; polling runs unconditionally every 2s now
   (cheap, returns [] when empty).

4. **Per-task node-card state leak (anomalyco#4)**. `workflowOpenTabs` was
   a global array — opening a node card under Task A left it
   visible after switching to Task B. Re-keyed by sessionId (now
   `Record<sessionID, OpenTab[]>`). Substrip render reads
   `workflowOpenTabs[params.id]`, opens/closes write through the
   current sessionId. Switching tasks now shows that task's own
   opened tabs.

5. **Refiner merge button no-op (anomalyco#1)**. Merge button was reading
   `selectedIds()` but the actual list-level multi-select bound to
   `mergeIDs` (a separate signal used by the MergeTray). Header
   button now reads/writes the same `mergeIDs` set the list uses,
   so shift-click selection -> merge button enable -> merge action
   all flow through one state.

6. **Refiner search input only takes one char per keystroke (anomalyco#11)**.
   The two search `<input>`s (line ~3815 and ~6107 of refiner-page)
   were missing `onCompositionStart/End` handlers. Chinese pinyin
   IME fires intermediate input events during composition; the
   controlled signal update reset the input every keystroke,
   breaking multi-char input. Added composition tracking that
   suppresses setQuery() while composing.

7. **Node-session layout (anomalyco#5)**. The outer `<div className="flex
   h-full flex-col">` of NodeSessionView had no horizontal sizing
   — `flex-col` parent's `flex-1` granted vertical grow but not
   width fill. Result: chat collapsed to content width, plugin
   slot didn't reach right edge, splitbar drag moved the whole
   right column left instead of resizing chat. Adding `w-full
   min-w-0 flex-1` to the root + `w-full min-w-0` to the body row
   fixes all three symptoms (whitespace left, narrow right, drag
   misbehavior).

8. **Rail overflow with many tasks (anomalyco#10)**. `.rune-rail-sec` had
   no `overflow-y: auto` or flex sizing — task list pushed beyond
   viewport when >N tasks. Set `flex: 1 1 auto; min-height: 0;
   overflow-y: auto` so the section scrolls independently while
   the rail brand / footer stay anchored.

Not addressed in this batch (deferred for separate discussion):
- Bug 2 / Bug 3 (action-budget architecture): structural change
- Bug 6 (context limit on node create): largely covered by the
  earlier workflow_read trim; ship and observe.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

anomalyco#1 question 跨 task 显示: backend /question 列出所有 session 的 pending
  请求，前端 chatsWithDialogs 现在按 ownedSessionIDs (root + workflow
  nodes' session_id) 过滤；其他 task 的 question/permission 不再串台

anomalyco#6 slave 调用 task 被禁: workflowSlaveAllow 加 `task: {"*":"deny"}`,
  所有 workflow node agent (coding/debug/build-flash/deploy/explore)
  现在禁用 task 工具——子节点不能再 spawn 子子任务，避免 off-canvas
  helper 绕过 workflow graph

anomalyco#9 refiner log 显示 session 名: RefinerLogsModal 加 useSync 反查
  sync.data.session[id].title, 列表行渲染 labelForSessionID(id) 取代
  s.id.slice(0,12); manual entry 仍显示 "manual / 手动新建"

  附带修复 anomalyco#9 配套问题——新建 task session 完全不触发 refiner:
  之前的 NON_REFINABLE_AGENTS=["orchestrator"] gate 太狠把所有
  workflow task 的 user 消息一棒打死。改成 META_COMMAND_PATTERN 只
  跳过纯命令式短消息 (go/run/approve/继续/确认/批准/中止//cmd 等),
  正常 domain query 走 refiner

#2a 创建 workflow 后 inspector 模型空白要切换 task 才显示: newWorkflow
  Task 现在在 navigate 前 sync.session.sync({force:true}), 让新建路由
  挂载时 snapshot 已经在手

#2b 改一个 slave 模型连 master 都跟着变: 之前 fix 是行为层 (clear-
  before-pick), 这次补视觉层——AgentCard 的 inherited 状态用 italic +
  60% opacity + (继承) 后缀 + 显眼的 "继承 master" 黄色 chip。让用户
  一眼看出 "这不是 slave 自己的模型,是临时继承 master session 的",
  改 master 时 inherited slaves 跟着显示也不再误以为 "全部被改了"

anomalyco#3 slave update success 不唤醒 master: classify() 之前只在 result !==
  "success" 时把 node.attempt_reported 算作 wake 事件,理由是
  "node.completed 会跟着来,到时一起唤醒"。但 slave 不总转 completed
  状态——经常只 update state_json + history 就停在 running/waiting
  等 master 决定下一步,master 永远不被唤醒。改为 attempt_reported
  无条件 wake (master 看到 status 已 completed 时可以自行 no-op)

anomalyco#10 timeline 看不到 slave 在干什么: 两个 fix 一起
  - workflow events projection 现在同时接受 payload.summary 和
    payload.summary_short (后者是上一版 trim 后的字段)。之前只读
    summary 导致整个 attempt_reported 摘要丢失,显示空字符串
  - events-panel 给 node.attempt_reported 和 workflow.* 加 data-primary
    + data-scope, CSS 让 attempt summary 用更大字号 + bold + dot 加
    halo;workflow-level 事件用 accent ribbon 左竖条,从 node-level
    chatter 中视觉分组分开。tool/update 等次要事件保留但回归背景

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…rectory tree UI

anomalyco#3 — Model picker "改一个变所有": root-fix
  Root cause: nodes are persisted with `node.model = null` when INSERT_NODE
  doesn't carry an explicit model (e.g., orchestrator-generated nodes).
  Inspector renders via `node.model || fallback`, where fallback is
  master's `local.model.current()`. When pickModel clears local.model
  briefly to detect "same model picked = no change" and reopens with the
  user's selection M, every inheriting slave's fallback display flips to
  M (until forceRouteNodes' .finally restores). Looks like "all change".

  Fix: proactive auto-route — new createEffect in workflow-panel.tsx
  watches `nodes()` and PATCHes every node lacking model with
  `local.model.current()` immediately. After this, every node has its
  own explicit `node.model`; inspector reads it directly; substrip pick
  has zero cascade; inspector pill for slave X touches only X.

  `routedThisSession` Set guards against re-PATCHing the same node on
  every snapshot poll. Rollback markers on failure so a transient route-
  service failure can retry on the next tick.

anomalyco#1 — Nested-git multi-layer code change
  Shadow git was treating nested `.git` dirs as gitlinks (mode 160000)
  and skipping their contents. New mechanism in snapshot/index.ts:

  - findNestedGitDirs walks worktree to depth 4 (skipping node_modules /
    dist / build / target / out / .git / .pnpm / .yarn). Records every
    nested `.git` directory under sub-paths.
  - withHiddenNestedGit wraps the diff/stage pipeline:
    1. Writes the planned rename list to `<gitdir>/snapshot_pending_
       renames.json` BEFORE any rename (crash-recovery marker).
    2. Renames each `<dir>/.git → <dir>/.git__snapshot_hidden__` (stable
       name, not a random suffix).
    3. Runs the shadow git pipeline as `addUnsafe` (the renamed dir is
       now a plain directory, git recurses into it normally).
    4. `Effect.ensuring` guarantees the restore loop runs even on
       exception — every hidden marker gets renamed back to `.git`.
    5. Marker file deleted after clean restore.
  - recoverPendingRenames runs FIRST inside `track()` on every snapshot
    init: if a previous process died between forward-rename and restore,
    the boot scan finds the marker, renames each `.git__snapshot_
    hidden__` back to `.git`, then deletes the marker. User's nested
    repos are guaranteed-usable before snapshot touches the worktree.

anomalyco#1 (cont) — Code change page now shows directory tree
  Flat horizontal tab strip replaced with a vertical indented tree:
  - buildFileTree groups files by their directory path.
  - Single-child dir chains collapse into combined names so deep paths
    like `packages/opencode/src/foo` read as ONE row, not four nested.
  - FileTreeRow recursively renders dirs (collapsible chevron + folder
    icon) and files (FileCode2 icon + +/− deltas right-aligned).
  - State: per-component `collapsedDirs` Set + selected file index.
  - New `.wf-detail-file-tree` + `.wf-detail-file-tree-row` CSS in
    theme.css; supports active highlight, hover, max-height 240px
    with internal scroll.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

anomalyco#3 slave card state: persist `sessionNode` overlay to localStorage
keyed by the workflow's root session id. Switching to Trace /
Knowledge rails no longer wipes the opened slave-card overlay; the
key is scoped per workflow so unrelated runs don't bleed state.

anomalyco#10 image attachments in chat composer: paste / drop image files
into the chat textarea, render thumbnails above the input, send as
`file` parts via promptAsync(). Optimistic update mirrors the
on-wire shape so the user sees their thumbnail in their own bubble
immediately. Backward compatible — old callers that ignore the new
`attachments` callback arg still work.

anomalyco#1+anomalyco#8 push-driven runtime commands: split the slave-session prompt
helper into `kickoffSlave()` (full preamble, fires once when a node
starts) and `pushRuntimeCommand()` (just the command+ack, fires on
every workflow_control). Previous code re-injected the full
"You are executing workflow node X" preamble on every command push,
so the slave saw the same framing 5+ times per run. New
`pushRuntimeCommand` includes the actual command payload inline and
explicitly tells the slave NOT to call workflow_pull — push IS the
authoritative path; pull is fault-tolerance only.

anomalyco#4 reasoning-step re-retrieve + compact re-injection: prompt loop
now re-retrieves experiences every 6 tool-call steps (was: only
step 1 of a turn), to counter slave drift on long chains. After
compaction, the auto-continue path detects workflow sessions via
Workflow.nodeBySession and prepends a brief "[compact-restoration]"
framing that lists the node id, status, and any pending commands —
so the slave doesn't think it's plain chat after history truncation.

anomalyco#5 real pause semantics: pauseNode now flips status BEFORE
cancelling the session, so any tool-call mid-flight sees the
status-machine guard and its workflow_update gets rejected as
illegal transition. Workflow.control refuses to enqueue new
commands when status is `paused` or `cancelled`, returning a
structured `refused: true` so the orchestrator stops retrying.

anomalyco#6 serial-port multiplex: Serial.create checks for an existing
session on the same device path with matching transport params
before opening a fresh SerialPort instance. Returns the existing
SerialID so the Serial Monitor UI and the debug-node serial_create
tool can share a single physical port instead of racing for EBUSY.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…loop retrieve

anomalyco#1 Global LLM refine — one shot, structured plan:
  - New Refiner.globalReRefineLLMPlan() sends the entire active
    library (compact projection: id/kind/title/abstract/scope/
    categories/observation_count) to one LLM call, asks it to
    classify each entry as merge / delete / conflict / keep, and
    returns a structured plan with reasons.
  - Refiner.globalReRefineLLMApply() commits via existing
    mergeExperiences / deleteExperience so refinement_history and
    audit logs stay coherent and the time-windowed undo button can
    roll the whole batch back.
  - Server routes: POST /experimental/refiner/global-rerefine/llm/plan
    and /llm/apply. Cap default 200 exps (newest-refined first);
    user can pass `unbounded: true`.
  - UI: new 🪄 LLM 整理 button in the global re-refine modal;
    confirm dialog shows the LLM's overall_summary + counts before
    anything mutates.

anomalyco#3 Live-push refiner changes via SSE:
  - New Refiner.Event.ExperienceChanged emitted from writeExperience
    and deleteExperience. The instance SSE relay already proxies
    every Bus event, so the frontend just subscribes by event-type
    string.
  - refiner-page subscribes to "refiner.experience.changed" via
    useSDK().event.on and triggers refetch() on any mutation — no
    more waiting 5 minutes for the poll interval to fire.

anomalyco#2 Mid-loop user-message retrieve trigger:
  - Track `lastRetrievedUserID` across the reasoning loop. When a
    new user message arrives mid-loop (lastUser.id changes but step
    keeps marching), force a retrieve on the next step instead of
    waiting up to 5 more steps for the step % 6 cadence to align.
    This unblocks the case where the user types a follow-up while
    the agent is still tool-calling.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>