diff --git a/CLAUDE.md b/CLAUDE.md
index c257ef15..44372b20 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -84,3 +84,4 @@ dotnet test tests/ClaudeDo.Worker.Tests/ClaudeDo.Worker.Tests.csproj -c Release
 - `docs/improvement-plan.md` — improvement snapshot from 2026-04-13 (historical)
 - `docs/prompts-inventory.md`, `docs/mailbox-proposal.md` — reference material (mailbox integration is parked)
 - `CHANGELOG.md` — Keep a Changelog format, maintained on release
+- `docs/explore-notes/` — distilled maps of complex subsystems from deep exploration (detail too fine for a CLAUDE.md). **Before** deep-exploring a subsystem, check for a matching note first; **after** a deep explore, distill durable findings back and bump its "verified against" commit. Always verify against current code before trusting. See `docs/explore-notes/README.md`.
diff --git a/docs/explore-notes/README.md b/docs/explore-notes/README.md
new file mode 100644
index 00000000..e68e0e79
--- /dev/null
+++ b/docs/explore-notes/README.md
@@ -0,0 +1,38 @@
+# Explore-notes
+
+Distilled, reusable maps of complex subsystems, produced by deep code exploration.
+The goal: stop re-exploring the same subsystem from scratch in every new session.
+
+These sit **between** the CLAUDE.md files and the code:
+
+- **CLAUDE.md** — high-level orientation, hand-maintained, always-loaded.
+- **explore-notes** — deeper subsystem detail (flows, who-calls-whom, invariants) that is
+  too fine-grained for a CLAUDE.md but stable enough to be worth caching. Read on demand.
+- **code** — the only source of truth.
+
+## Rules
+
+- **Only stable structure.** Flows, responsibilities, entry points, invariants, relative
+  file paths. **No line numbers**, no exhaustive symbol dumps — those rot fastest.
+- **Verify before trusting.** A note is a starting map, not authority. Always confirm
+  against current code before acting on it. Each note records the commit it was verified
+  against so you can diff for drift.
+- **Not a substitute for CLAUDE.md.** If a fact belongs in orientation, put it there.
+
+## Header every note must carry
+
+```
+> **Explore-note — verify before trusting.** Distilled map of a subsystem, not authoritative.
+> Last verified against commit `<short-hash>` (<date>).
+> Drift check: `git log --oneline <short-hash>..HEAD -- <paths this note covers>`
+> Stable structure only (no line numbers). See docs/explore-notes/README.md.
+```
+
+## Workflow
+
+1. **Before** deep-exploring a subsystem, check for a matching note here and read it first;
+   explore only to fill gaps or confirm.
+2. **After** a deep explore, distill the durable findings into a new/updated note and bump
+   its "verified against" commit line.
+3. If the drift check shows the covered paths changed a lot since the verified commit, treat
+   the note as suspect and re-verify the parts you rely on.
diff --git a/docs/explore-notes/worker-task-pipeline.md b/docs/explore-notes/worker-task-pipeline.md
new file mode 100644
index 00000000..e5eadd63
--- /dev/null
+++ b/docs/explore-notes/worker-task-pipeline.md
@@ -0,0 +1,124 @@
+> **Explore-note — verify before trusting.** Distilled map of a subsystem, not authoritative.
+> Last verified against commit `896d4b5` (2026-07-23).
+> Drift check: `git log --oneline 896d4b5..HEAD -- src/ClaudeDo.Worker`
+> Stable structure only (no line numbers). See docs/explore-notes/README.md.
+
+# Worker: Task Execution Pipeline
+
+How a task moves Queued → Running → terminal, across `src/ClaudeDo.Worker`
+(Queue, Runner, Lifecycle, State, Agents, Worktrees, Hub).
+
+## End-to-End Flow (Queued → Terminal)
+
+1. **Enqueue** — `ITaskStateService.EnqueueAsync()` (State/TaskStateService.cs)
+   - Idle → Queued, then wakes the dispatcher via `IQueueWaker.Wake()`.
+
+2. **Dispatch** — `QueueService` loop (Queue/QueueService.cs)
+   - `BackgroundService`; waits for a wake signal or a backstop timer.
+   - Reads the max-parallel limit from settings; claims a free slot if under limit.
+
+3. **Atomic Claim** — `IQueuePicker.ClaimNextAsync()` (Queue/QueuePicker.cs)
+   - Raw SQL `UPDATE ... RETURNING` in one transaction: picks an eligible Queued task
+     (unblocked, due or unscheduled; sorted by sort_order/created_at), sets status→Running
+     + started_at, returns the row. Prevents two workers claiming the same task (TOCTOU).
+
+4. **Slot Execution** — `QueueService.RunInSlotAsync()` (Queue/QueueService.cs)
+   - For review feedback: resume the prior session if one exists, else fold feedback into
+     the prompt. Calls `TaskRunner.RunAsync()` / `ContinueAsync()` with `alreadyClaimed=true`.
+
+5. **Run Preparation** — `TaskRunner.RunAsync()` (Runner/TaskRunner.cs)
+   - Loads task, list config, subtasks, attachments from the DB.
+   - `PrepareRunDirectoryAsync()`: worktree (via WorktreeManager) if the list has a WorkingDir,
+     else sandbox. Generates a per-run MCP token, writes MCP config to disk.
+   - `StartRunningAsync()` (only if not pre-claimed): atomic Queued → Running. Broadcasts TaskStarted.
+
+6. **Claude Execution** — `TaskRunner.RunOnceAsync()` (Runner/TaskRunner.cs)
+   - Creates a TaskRunEntity, points the task at the run's log path.
+   - Builds claude CLI args (ClaudeArgsBuilder), spawns the process via `IClaudeProcess.RunAsync()`
+     with prompt + working dir + streaming callback.
+   - Stream lines → NDJSON log + broadcast via TaskMessage. MCP tools (AskUser, SuggestImprovement)
+     are scoped by the per-run token.
+
+7. **Result Handling** — `TaskRunner.HandleSuccess()` / `MarkFailed()` (Runner/TaskRunner.cs)
+   - Success (exit 0 + result markdown): if worktree, commit + broadcast WorktreeUpdated; then
+     transition to Done / WaitingForReview / WaitingForChildren (CompleteAsync / SubmitForReviewAsync
+     / SubmitForChildrenAsync).
+   - Failure: if a session exists, auto-retry once via ContinueAsync; else MarkFailed → FailAsync.
+   - All terminal writes use `CancellationToken.None` so a task is never left Running.
+
+8. **Terminal States** — `ITaskStateService` transitions (State/TaskStateService.cs)
+   - **Done** CompleteAsync (Running → Done) — top-level success.
+   - **WaitingForReview** SubmitForReviewAsync (Running → WaitingForReview) — review gate.
+   - **WaitingForChildren** SubmitForChildrenAsync (Running → WaitingForChildren) — blocks on children.
+   - **Failed** FailAsync (Running/Queued → Failed).
+   - **Cancelled** CancelAsync (Running/Queued/WaitingForReview/WaitingForChildren → Cancelled).
+
+## Component Responsibilities
+
+**Queue/**
+- `QueueService` — main dispatch loop; slot limit; decides when to start tasks.
+- `QueuePicker` — atomic Queued→Running claim via raw SQL.
+- `QueueWaker` — semaphore for non-blocking, idempotent wake signals.
+- `OverrideSlotService` — owns the RunNow / ContinueTask slot (bypasses the queue).
+
+**Runner/**
+- `TaskRunner` — orchestrates the run (prepare, execute, handle result).
+- `WorktreeManager` — creates/manages git worktrees; self-heals stale branches.
+- `ClaudeProcess` — spawns the claude CLI subprocess; manages streams/logs.
+- `TaskRunMcpService` — runtime MCP tools (AskUser, SuggestImprovement).
+- `TaskRunTokenRegistry` — per-run MCP identity for tool-access control.
+- `InteractiveLaunchSpecService` — config for the task's claude run.
+
+**State/**
+- `TaskStateService` — all task status transitions; guards preconditions; signals queue/hub.
+
+**Lifecycle/** (startup recovery)
+- `StaleTaskRecovery` — tasks stuck Running after a crash/restart → Failed.
+- `OrphanRecovery` — dequeues children whose parent is no longer planning (stays attached).
+- `AttachmentOrphanRecovery` — cleans orphaned attachment files.
+- `TaskResetService` — manual reset to Idle.
+- `TaskMergeService` — conflict resolution for worktree merges.
+
+**Hub/**
+- `HubBroadcaster` — single SignalR broadcast point (TaskStarted/TaskUpdated/TaskMessage/RunCreated…).
+- `WorkerHub` — SignalR hub + client methods.
+
+**Agents/**
+- `AgentFileService` — file I/O for custom agents.
+- `DefaultAgentSeeder` — seeds built-in agents on startup.
+
+**Worktrees/**
+- `WorktreeMaintenanceService` — cleanup, state tracking, overview reporting.
+
+## Entry Points & Call Chain
+
+```
+Program.cs (DI setup)
+ ├─ QueueService (BackgroundService) → ExecuteAsync loop
+ │   ├─ waits: IQueueWaker.WaitAsync() or timer
+ │   ├─ claims: IQueuePicker.ClaimNextAsync()
+ │   └─ runs:  TaskRunner.RunAsync() / ContinueAsync()
+ ├─ Hub clients → WorkerHub methods
+ │   ├─ Enqueue     → ITaskStateService.EnqueueAsync() → Wake()
+ │   ├─ RunNow      → OverrideSlotService.RunNow()      → TaskRunner.RunAsync()
+ │   ├─ ContinueTask→ OverrideSlotService.ContinueTask()→ TaskRunner.ContinueAsync()
+ │   └─ CancelTask  → QueueService.CancelTask()
+ ├─ Lifecycle recovery (startup): StaleTaskRecovery / OrphanRecovery / AttachmentOrphanRecovery
+ └─ State transitions → HubBroadcaster.TaskUpdated()
+```
+
+## Invariants & Conventions
+
+- **Atomic claiming** — QueuePicker's `UPDATE ... RETURNING` makes Queued→Running atomic.
+- **Slot limit** — respects MaxParallelExecutions; a backstop timer wakes even if a Wake() is missed.
+- **Pre-claimed tasks** — the dispatcher pre-claims via the picker; the override slot
+  (RunNow/ContinueTask) must call StartRunningAsync if a task is not pre-claimed.
+- **Terminal writes** — use `CancellationToken.None`; a task is never left Running after crash/cancel.
+- **Per-run MCP tokens** — each run gets a unique token scoping tool access; unregistered on end.
+- **Auto-retry** — one automatic retry if a session exists and the first run failed.
+- **Worktree self-heal** — on branch collision, remove phantom worktrees, prune, delete branch, retry add.
+- **Review feedback** — stored on the task; consumed once a run reaches a terminal state; a re-queued
+  task resumes the session or folds feedback into the prompt.
+- **Child tasks** — planning creates draft children; finalization requires no Queued children remain;
+  OrphanRecovery dequeues children if the parent is not planning.
+- **Lifecycle recovery** runs at startup: stale-Running → Failed; orphaned children → dequeued but attached.
diff --git a/docs/verification-handoff.md b/docs/verification-handoff.md
new file mode 100644
index 00000000..77c8e136
--- /dev/null
+++ b/docs/verification-handoff.md
@@ -0,0 +1,89 @@
+# Verifikations-Handoff (2026-07-23)
+
+Konsolidierte manuelle Verifikationen aus `docs/open.md` + Memory-Ständen — gedacht für eine frische Session, die diese Punkte am laufenden System durchspielt. Mika bedient die UI, die Session protokolliert Pass/Fail.
+
+**Vorbedingungen:** Worker + App laufen (SignalR-Port lt. `~/.todo-app/worker.config.json`, aktuell 37821; External MCP 47822). Testliste `ClaudeDoTests` (`C:\TestRepos\ClaudeDoTests`) für zerstörungsfreie Runs nutzen.
+
+**Achtung:** 7 Bugfix-/Chore-Tasks vom 2026-07-23 (createdBy `claude-code-gap-analysis`) liegen in der Queue der Liste „Claude do". Die Merge-bezogenen Checks (Abschnitte 2 und 4) erst NACH Review/Merge dieser Tasks final abhaken — sie ändern ggf. genau dieses Verhalten.
+
+**Nacharbeit:** Erledigtes aus `docs/open.md` austragen, dieses File danach löschen.
+
+---
+
+## 1. Detail-Insel & Diff-Viewer (reines Durchklicken)
+
+- [ ] Detail-Insel komplett: Output/Git/Session-Tabs, Merge-Sektion, Agent-Settings-Overrides (InheritedBadge korrekt), Prep-Panel — nach dem VM-Split (`DetailsIslandViewModel` → Sektions-VMs) alles gebunden, keine leeren Panels.
+- [ ] Diff-Viewer: Dateiliste, Added/Deleted/Renamed/Binary-Erkennung, Commit-Range-Diff nach einem Merge.
+- [ ] DiffModal-Fehler-State: Commit-Range ohne aufgezeichnete Commits → „Diff nicht mehr verfügbar" statt Crash/leer.
+- [ ] „children need attention"-Band auf dem Session-Tab eines Parents mit failed/blocked Kind.
+
+## 2. Worktree-Pipeline (3 falsifizierbare Fälle)
+
+- [ ] Happy-Path: Task mit WorkingDir → `worktrees.state='active'`, `head_commit` gesetzt, `diff_stat` non-empty, Branch `claudedo/<id[:8]>` existiert auf Disk.
+- [ ] No-Changes-Run: → `status='Done'`, `head_commit IS NULL`, `diff_stat IS NULL`.
+- [ ] Kein Git-Repo (WorkingDir = `C:\Temp`): → `status='Failed'`, KEINE `worktrees`-Row, Git-Fehler im Log.
+
+## 3. Planning-Flow-Walkthrough
+
+- [ ] Draft → Finalize → Kette: Finalize queued NICHT automatisch (Kinder bleiben Idle); „Queue plan" setzt alle nicht-terminalen Kinder Queued, Kette läuft sequenziell durch (blocked-by löst sich je Vorgänger).
+- [ ] Parent landet nach letztem terminalen Kind in WaitingForReview; Approve merged die ganze Unit (Parent-Worktree falls Active + jedes Done-Kind in Reihenfolge).
+- [ ] UnfinishedPlanning-Modal: Resume / FinalizeNow / Discard.
+
+## 4. Merge-Editor (Rider-Style 3-Pane) mit echtem Konflikt
+
+Konflikt provozieren: gleiche Datei auf main ändern, während der Task-Branch sie ändert. Beide Wege testen: **(a)** Single-Task-Approve mit Konflikt, **(b)** Planning-Unit-Merge mit Konflikt in einem Subtask (`PlanningMergeConflict` → Editor öffnet pro Subtask).
+
+- [ ] Drei Panes: MAIN read-only | Result editierbar | INCOMING read-only; Konfliktblöcke rot, aufgelöst grün, in allen Panes.
+- [ ] Gutter-Toggle `›`/`‹`: Seite rein/raus, Klickreihenfolge = Reihenfolge im Result; main/incoming/beide/keine möglich.
+- [ ] Nur Konfliktregionen im Result editierbar (Stable read-only); Edits fließen in den Block zurück.
+- [ ] Synchrones vertikales Scrollen; File-Switcher bei mehreren Dateien; `M conflicts · K resolved`-Readout; Conflict-Ruler (Klick springt).
+- [ ] Continue erst aktiv, wenn ALLE Konflikte in ALLEN Dateien gelöst; Binär-Guard greift.
+- [ ] Abort: Tree sauber, Task bleibt WaitingForReview.
+- Bekannte Kanten (nur gegenprüfen, nicht als Fail werten): leere Ours-Seite → null-lange Result-Region (Accept geht, Handtippen fummelig); Gutter-Y-Ausrichtung bei sehr hohen Fenstern/großem Scroll; vertikaler Drift nachfolgender Blöcke nach Konflikt mit ungleicher Zeilenzahl.
+
+## 5. Embedded ConPTY / Mission Control
+
+- [ ] **Kritisch:** Task-basiert (Kontextmenü „Open ConPTY session"): frischer Task → Worktree wird on-demand angelegt, `claude` startet — verifizieren, dass der Task-Prompt wirklich GESENDET wird (Antwort beginnt), nicht nur im Eingabefeld vorbefüllt ist.
+- [ ] Ad-hoc: „New session"-Button → Ordnerwahl → freie Session im gewählten Verzeichnis.
+- [ ] Grid↔Tabs-Toggle; Close killt den Prozess + entfernt die Kachel; mehrere Sessions parallel; Pane-Resize reflowt das TUI.
+- [ ] Resume: im Worktree-Dir per `claude --continue` möglich (ClaudeDo persistiert die ConPTY-Session-Id bewusst nicht).
+
+## 6. Pick up in terminal
+
+- [ ] Sichtbarkeit: Kontextmenü-Eintrag + Terminal-Button (ArrowOut) NUR bei WaitingForReview und Failed; bei Idle/Running/Queued/Done nicht.
+- [ ] Klick → neues Windows-Terminal im Worktree-Verzeichnis, `claude --resume <id>` nimmt die Session mit Kontext wieder auf.
+- [ ] Fehlerfälle surfacen sauber (Footer-Strip bzw. Fehlerdialog): laufende/gequeuete Task, keine persistierte Session-Id, kein aktiver Worktree.
+- Bekannte Kante: parked-Idle (reject-park) hat oft Session+Worktree, zeigt die Aktion aber bewusst NICHT (Idle nicht unterscheidbar). Nervt das in der Praxis → `CanPickUpInTerminal` erweitern.
+
+## 7. AskUser (Frage aus laufendem Task)
+
+- [ ] Task, dessen Prompt eine Rückfrage erzwingt → Frage erscheint im Task-Monitor (Mission Control), Prozess wartet.
+- [ ] Antwort inline absenden → Run läuft mit der Antwort weiter.
+- [ ] Timeout-/Cleanup-Verhalten der PendingQuestionRegistry (Frage unbeantwortet lassen): Task schlägt kontrolliert fehl, UI räumt die Frage auf (MCP_TOOL_TIMEOUT-Gotcha).
+
+## 8. Session Skills (E2E + UI)
+
+- [ ] Settings → Skills: `https://github.com/DietrichGebert/ponytail` installieren → 6 Skills erscheinen (ponytail, -help, -review, -audit, -debt, -gain), auf Commit gepinnt, Dateien unter `~/.todo-app/session-skills/<name>/`.
+- [ ] Skill per-Task (Agent-Settings-Flyout) oder global aktivieren → Task laufen lassen → im Worktree liegt `.claude/skills/<name>/`, Agent kann ihn nutzen, `git status` im Worktree bleibt sauber (info/exclude greift).
+- [ ] Gegenprobe: nicht-aktivierte/andere interaktive Session sieht den Skill NICHT (kein Leak nach `~/.claude`).
+- [ ] UI: Skills-Tab (Install-Zeile, Karten mit Update/Remove), Checkbox-Listen im General-Tab + AgentConfigEditor (Flyout-Höhe!), leerer Zustand (0 Skills — Empty-State fehlt evtl., dann entscheiden), lange Namen/URLs (Trimming).
+
+## 9. Attachments (Drag & Drop + MCP)
+
+- [ ] Drop aufs Detail-Pane: „Drop to attach"-Overlay, Datei erscheint in der Liste, landet unter `~/.todo-app/attachments/<taskId>/`; „Add file…"-Picker; Remove-Button.
+- [ ] `ComposedPreview` enthält die Attachment-Pfade („## Reference files").
+- [ ] MCP: `add_task_attachment` / `list_task_attachments` / `remove_task_attachment`; Running-Task verweigert add/remove.
+
+## 10. Daily Prep (Prime) & Weekly Report
+
+- [ ] Prime-Trigger: Schedule feuert bzw. „Plan day" manuell → Prep-Log streamt live, `daily-prep.log` enthält letzten Run, MyDay-Auswahl respektiert `DailyPrepMaxTasks`.
+- [ ] Weekly Report: Range-Default „seit letztem Standup-Wochentag → heute", Markdown rendert, Cache pro Range.
+
+## 11. Self-Update & Autostart (am Gerät)
+
+- [ ] Update-Banner → Update durchführen → danach „up to date".
+- [ ] Autostart: Logoff/Logon startet den Worker (Startup-`.lnk`); Update-Pfad erhält den Autostart; Uninstall entfernt die `.lnk`.
+
+## 12. Status-Bar / RunNow (Mini-Codecheck, erst messen)
+
+- [ ] Worker trennen/verbinden → prüfen, ob RunNow-Enable pro Task-Row sauber re-evaluiert (Connection-State lebt in `IslandsShellViewModel`). Nur fixen, wenn tatsächlich kaputt.