diff --git a/.beans/claudbg-1e1c--tui-quit-confirmation-dialog.md b/.beans/claudbg-1e1c--tui-quit-confirmation-dialog.md
new file mode 100644
index 0000000..9061838
--- /dev/null
+++ b/.beans/claudbg-1e1c--tui-quit-confirmation-dialog.md
@@ -0,0 +1,48 @@
+---
+# claudbg-1e1c
+title: 'TUI: quit confirmation dialog'
+status: todo
+type: task
+priority: normal
+created_at: 2026-03-30T04:47:46Z
+updated_at: 2026-03-30T04:49:03Z
+parent: claudbg-i6l2
+blocked_by:
+    - claudbg-ut9q
+---
+
+Implement a modal quit confirmation dialog.
+
+## Trigger
+
+Press `q` or `Q` from any screen (except when another modal is open).
+
+## Appearance
+
+```
+┌──────────────────────┐
+│   Quit claudbg?      │
+│                      │
+│  q = yes  Esc = no   │
+└──────────────────────┘
+```
+
+Centered overlay on top of the current screen (using `ratatui::widgets::Clear` to clear the background area).
+
+## State
+
+`AppState.show_quit_dialog: bool`
+
+- `q`/`Q` → `show_quit_dialog = true`
+- While dialog is shown: `q` → `should_quit = true` (exit loop), `Escape` → `show_quit_dialog = false`
+- Dialog intercepts all other key input while open.
+
+## Implementation notes
+
+- Use `ratatui::layout::Rect` centered calculation to position the dialog.
+- Render with `Block::bordered().title("Quit?")` and a `Paragraph` inside.
+- Use `Clear` widget behind the dialog block to prevent bleed-through.
+
+## Blocked by
+
+- event loop (claudbg-ut9q)
diff --git a/.beans/claudbg-1tlk--tui-help-modal-listing-all-keyboard-shortcuts.md b/.beans/claudbg-1tlk--tui-help-modal-listing-all-keyboard-shortcuts.md
new file mode 100644
index 0000000..3b32507
--- /dev/null
+++ b/.beans/claudbg-1tlk--tui-help-modal-listing-all-keyboard-shortcuts.md
@@ -0,0 +1,55 @@
+---
+# claudbg-1tlk
+title: 'TUI: help modal listing all keyboard shortcuts'
+status: todo
+type: task
+priority: normal
+created_at: 2026-03-30T04:48:38Z
+updated_at: 2026-03-30T04:49:03Z
+parent: claudbg-i6l2
+blocked_by:
+    - claudbg-ut9q
+---
+
+Implement a `?` help overlay that lists all keyboard shortcuts.
+
+## Trigger
+
+Press `?` (or `Shift+/`) from any screen (except when another modal is open).
+
+## Content
+
+```
+┌─────────── Keyboard Shortcuts ────────────┐
+│ Navigation                                 │
+│   ↑/↓ or k/j    Scroll up/down            │
+│   ←/→ or h/l    Scroll left/right         │
+│   Tab            Cycle panes              │
+│   Enter          Open / select            │
+│   Escape         Go back                  │
+│                                           │
+│ Global                                    │
+│   q / Q          Quit (with confirmation) │
+│   ?              Show this help           │
+└───────────────────────────────────────────┘
+```
+
+Press `Escape` to close.
+
+## State
+
+`AppState.show_help: bool`
+
+- `?` → `show_help = true`
+- `Escape` → `show_help = false`
+- Help modal intercepts all key input while open.
+
+## Implementation notes
+
+- Same centered overlay pattern as quit dialog (Clear + Block + Paragraph).
+- Content is static — hardcoded list of shortcuts.
+- Render on top of whatever screen is active.
+
+## Blocked by
+
+- event loop (claudbg-ut9q)
diff --git a/.beans/claudbg-33n0--agent-discovery-looks-in-wrong-subagents-directory.md b/.beans/claudbg-33n0--agent-discovery-looks-in-wrong-subagents-directory.md
new file mode 100644
index 0000000..cbaa9ac
--- /dev/null
+++ b/.beans/claudbg-33n0--agent-discovery-looks-in-wrong-subagents-directory.md
@@ -0,0 +1,63 @@
+---
+# claudbg-33n0
+title: Agent discovery looks in wrong subagents directory — sessions show 0 agents
+status: todo
+type: bug
+priority: high
+created_at: 2026-03-30T04:43:21Z
+updated_at: 2026-03-30T04:43:21Z
+---
+
+## Problem
+
+`agents list <session-id>` always returns empty results. The subagent directory path used in discovery is wrong.
+
+## Root cause
+
+`discover_agents_for_session(session_file)` in `src/parser/discovery.rs:198-205` does:
+
+```rust
+let parent = session_file.parent();          // ~/.claude/projects/<project-dir>/
+let subagents_dir = parent.join("subagents"); // ~/.claude/projects/<project-dir>/subagents/
+```
+
+That directory never exists.
+
+`discover_all_agents()` has the same bug: it iterates project dirs and looks at `proj_path.join("subagents")` directly — also wrong.
+
+## Actual disk structure
+
+Confirmed across multiple sessions in the live `~/.claude/projects/` dir:
+
+```
+~/.claude/projects/<project-dir>/
+  <session-uuid>.jsonl            ← session file
+  <session-uuid>/                 ← directory with same name as UUID (no .jsonl)
+    subagents/
+      agent-<id>.jsonl
+      agent-<id>.meta.json
+```
+
+Example path that should work but currently returns empty:
+`~/.claude/projects/.../def4776b-25a6-4eca-99e7-f222926c297a/subagents/agent-a8e741de240089eef.jsonl`
+
+## Required fix
+
+`discover_agents_for_session(session_file)` should construct:
+
+```rust
+let stem = session_file.file_stem().unwrap(); // "def4776b-...-f222926c297a"
+let session_dir = session_file.parent().unwrap().join(stem);
+let subagents_dir = session_dir.join("subagents");
+```
+
+`discover_all_agents()` must also be updated to walk UUID subdirectories inside each project dir instead of looking at `<project-dir>/subagents/` directly.
+
+## Secondary issue
+
+The existing unit tests in `src/commands/agents.rs` (`dump_with_real_agent_returns_ok`, `transcribe_with_real_agent_returns_ok`) create subagents at the old wrong path (`projects/subagents/`) and must be updated alongside the fix.
+
+## Relevant files
+
+- `src/parser/discovery.rs` — `discover_agents_for_session` (lines 198–205), `discover_all_agents` (lines 211–266), `collect_agents_in_dir`
+- `src/commands/agents.rs` — integration tests
diff --git a/.beans/claudbg-6dgc--table-overflows-terminal-truncate-long-columns-and.md b/.beans/claudbg-6dgc--table-overflows-terminal-truncate-long-columns-and.md
new file mode 100644
index 0000000..55ae6b2
--- /dev/null
+++ b/.beans/claudbg-6dgc--table-overflows-terminal-truncate-long-columns-and.md
@@ -0,0 +1,30 @@
+---
+# claudbg-6dgc
+title: Table overflows terminal — truncate long columns and respect terminal width
+status: todo
+type: bug
+priority: normal
+created_at: 2026-03-30T04:36:31Z
+updated_at: 2026-03-30T04:41:14Z
+parent: claudbg-tci9
+---
+
+## Problem
+
+`sessions list` table overflows the terminal width. The project path column is the main culprit — full paths like `/run/media/pop/4e1ca4f1-.../Projects/src/github.com/pop/claudbg` make rows far too wide.
+
+## Current behavior
+
+`src/output/table.rs` hardcodes `FALLBACK_WIDTH = 200` and applies `.max(FALLBACK_WIDTH)`, so the table is *always at least 200 chars wide* regardless of the actual terminal. No `ColumnConstraint` is applied to any column.
+
+## Expected behavior
+
+- The table should never exceed the actual terminal width (use `terminal_size` or `comfy_table`'s built-in terminal detection).
+- The project path column should be truncated to show the **last** N characters (tail of the path), since the trailing project name is more meaningful than the leading `/run/media/...` prefix. E.g. display `…/pop/claudbg` instead of `/run/media/pop/4e1ca4f1-…`.
+- Other columns (ID, date, model, message count) should not be truncated.
+- comfy-table supports `ColumnConstraint::MaxWidth` and `ColumnConstraint::LowerBoundary` — use these to cap the project column.
+
+## Relevant files
+
+- `src/output/table.rs` — `render_table`, `FALLBACK_WIDTH` constant
+- `src/commands/sessions.rs` — `list()` function that calls `render_table`
diff --git a/.beans/claudbg-78xt--tui-add-ratatui-crossterm-dependencies.md b/.beans/claudbg-78xt--tui-add-ratatui-crossterm-dependencies.md
new file mode 100644
index 0000000..2473134
--- /dev/null
+++ b/.beans/claudbg-78xt--tui-add-ratatui-crossterm-dependencies.md
@@ -0,0 +1,34 @@
+---
+# claudbg-78xt
+title: 'TUI: add ratatui + crossterm dependencies'
+status: todo
+type: task
+created_at: 2026-03-30T04:45:06Z
+updated_at: 2026-03-30T04:45:06Z
+parent: claudbg-i6l2
+---
+
+Add `ratatui` and `crossterm` to `Cargo.toml`.
+
+## Changes
+
+```toml
+ratatui   = "0.29"   # TUI framework
+crossterm = "0.28"   # terminal backend (raw mode, event reading)
+```
+
+Ratatui ships its own `crossterm` backend feature; use `ratatui = { version = "0.29", features = ["crossterm"] }` to pull both together.
+
+## Verify
+
+Run `cargo check` to confirm the dependency resolves and there are no conflicts with existing deps (`tokio`, `libsql`).
+
+## Notes
+
+- Do NOT add `termion` — crossterm is cross-platform and the right choice.
+- The `ratatui` crate bundles the backend integration; no need for a separate `tui-crossterm` crate.
+- Check that `ratatui 0.29` is available; if not use the latest stable.
+
+## Relevant files
+
+- `Cargo.toml`
diff --git a/.beans/claudbg-8vpb--transcription-output-improvements.md b/.beans/claudbg-8vpb--transcription-output-improvements.md
new file mode 100644
index 0000000..b3bc531
--- /dev/null
+++ b/.beans/claudbg-8vpb--transcription-output-improvements.md
@@ -0,0 +1,10 @@
+---
+# claudbg-8vpb
+title: Transcription output improvements
+status: todo
+type: epic
+created_at: 2026-03-30T04:43:30Z
+updated_at: 2026-03-30T04:43:30Z
+---
+
+Groups bugs in `sessions transcribe` and `agents transcribe` output: tools display format, duration miscalculation, tool result visibility defaults, and tool input truncation.
diff --git a/.beans/claudbg-9c8r--tui-sub-agents-panel-in-transcript-view.md b/.beans/claudbg-9c8r--tui-sub-agents-panel-in-transcript-view.md
new file mode 100644
index 0000000..6a0156d
--- /dev/null
+++ b/.beans/claudbg-9c8r--tui-sub-agents-panel-in-transcript-view.md
@@ -0,0 +1,44 @@
+---
+# claudbg-9c8r
+title: 'TUI: sub-agents panel in transcript view'
+status: todo
+type: feature
+priority: normal
+created_at: 2026-03-30T04:47:15Z
+updated_at: 2026-03-30T04:49:03Z
+parent: claudbg-i6l2
+blocked_by:
+    - claudbg-rudq
+---
+
+Add a sub-agents panel to the right side of the transcript screen, showing all sub-agent runs for the current session.
+
+## Layout
+
+```
+┌── Transcript ──────────────────────┬─── Sub-agents ──┐
+│ [user]: …                          │ a8e741de  agent  │
+│ [assistant]: …                     │ > c3f920aa agent  │◄ selected
+│ [tool: Bash] …                     │                  │
+│ [tool_result]: …                   │                  │
+└────────────────────────────────────┴──────────────────┘
+```
+
+The panel is ~25% of the frame width (or a fixed ~30 chars).
+
+## Implementation
+
+- Use `ratatui::layout::Layout` with `Constraint::Min` for the chat area and `Constraint::Length(30)` for the panel.
+- Panel shows: short agent ID + agent type. If the session has no sub-agents, the panel can be hidden or show "No sub-agents".
+- Panel is focusable via Tab (Tab cycles: chat log → sub-agents panel → chat log).
+- When panel is focused, j/k moves selection. Enter navigates to that sub-agent's transcript (sets `screen` to `SubagentTranscript`).
+- Highlight the focused pane's border (e.g. with a colored `Block` border).
+
+## Data source
+
+`AppState.subagents: Vec<AgentRef>` populated via `discover_agents_for_session` when entering transcript screen.
+
+## Blocked by
+
+- transcript screen (claudbg-rudq)
+- agent discovery fix (claudbg-33n0)
diff --git a/.beans/claudbg-a532--include-help-text-should-enumerate-valid-values.md b/.beans/claudbg-a532--include-help-text-should-enumerate-valid-values.md
new file mode 100644
index 0000000..6be9b6e
--- /dev/null
+++ b/.beans/claudbg-a532--include-help-text-should-enumerate-valid-values.md
@@ -0,0 +1,37 @@
+---
+# claudbg-a532
+title: '`--include` help text should enumerate valid values'
+status: todo
+type: bug
+priority: low
+created_at: 2026-03-30T04:38:03Z
+updated_at: 2026-03-30T04:41:14Z
+parent: claudbg-tci9
+---
+
+## Problem
+
+The `--include` flag help text says "Comma-separated content to include: thinking, output." but clap does not enumerate the valid tokens in `--help` output. Users cannot discover valid values from `--help` alone.
+
+## Current declaration (`src/cli.rs`)
+
+```rust
+/// Comma-separated content to include: thinking, output.
+#[arg(long, global = true, default_value = "")]
+pub include: IncludeList,
+```
+
+The valid tokens are `thinking` and `output` (parsed in `IncludeList::from_str`).
+
+## Expected behavior
+
+The help output should clearly show: `[possible values: thinking, output]` or equivalent. Options:
+
+1. Update the `#[arg(help = ...)]` string to explicitly state the valid values and format.
+2. Or use clap's `value_parser` with a custom validator that also shows valid choices.
+
+Since `IncludeList` is a comma-separated composite (not a single enum value), option 1 (explicit help string) is simplest.
+
+## Relevant files
+
+- `src/cli.rs` — `IncludeList`, `GlobalOpts`, the `include` field annotation
diff --git a/.beans/claudbg-d9ev--sessions-dump-should-collapse-consecutive-identica.md b/.beans/claudbg-d9ev--sessions-dump-should-collapse-consecutive-identica.md
new file mode 100644
index 0000000..512d3b6
--- /dev/null
+++ b/.beans/claudbg-d9ev--sessions-dump-should-collapse-consecutive-identica.md
@@ -0,0 +1,43 @@
+---
+# claudbg-d9ev
+title: '`sessions dump` should collapse consecutive identical rows'
+status: todo
+type: bug
+priority: normal
+created_at: 2026-03-30T04:38:42Z
+updated_at: 2026-03-30T04:41:14Z
+parent: claudbg-tci9
+---
+
+## Problem
+
+`sessions dump` produces hundreds of identical consecutive rows (e.g. repeated `progress` entries) that differ only in timestamp, making the output very noisy.
+
+## Example
+
+Rows 555–661 in session `21fae0a8` are all `progress` entries with the same type, role, and content preview. They currently appear as 107 separate table rows.
+
+## Expected behavior
+
+Consecutive rows that are identical in all fields *except* the timestamp should be collapsed into a single row with a range indicator:
+
+```
+555-661 | progress | <role> | <content preview>
+```
+
+Only collapse truly identical rows (same `type`, `role`, and `content` preview). Do not collapse if any of these differ.
+
+## Scope
+
+- Applies to `sessions dump` table output only.
+- Does **not** apply to `sessions transcribe`.
+- Does **not** apply to `--follow` mode (streaming).
+- JSON/XML output should also collapse (the range can be represented as `{"seq_start": 555, "seq_end": 661, ...}`).
+
+## Implementation hint
+
+After building the `rows` vector in `dump()`, run a post-processing pass: iterate rows and merge consecutive runs where `row[2]` (type), `row[3]` (role), and `row[4]` (content) are all equal. Replace the merged run with a single row whose `#` column shows `N-M` and whose timestamp is dropped or shows the first timestamp.
+
+## Relevant files
+
+- `src/commands/sessions.rs` — `dump()` function, row-building loop
diff --git a/.beans/claudbg-f4ot--transcription-header-shows-tools-as-debug-hashmap.md b/.beans/claudbg-f4ot--transcription-header-shows-tools-as-debug-hashmap.md
new file mode 100644
index 0000000..e5ff5e7
--- /dev/null
+++ b/.beans/claudbg-f4ot--transcription-header-shows-tools-as-debug-hashmap.md
@@ -0,0 +1,48 @@
+---
+# claudbg-f4ot
+title: Transcription header shows tools as debug HashMap, not readable list
+status: todo
+type: bug
+created_at: 2026-03-30T04:43:50Z
+updated_at: 2026-03-30T04:43:50Z
+parent: claudbg-8vpb
+---
+
+## Problem
+
+The transcription header line `Tools:` prints a raw Rust debug-format `HashMap`, e.g.:
+
+```
+Tools:   {"Bash": 3, "Read": 5, "Write": 2}
+```
+
+## Expected behavior
+
+A human-readable comma-separated summary, e.g.:
+
+```
+Tools:   Bash×3, Read×5, Write×2
+```
+
+(or similar format — `Name(n)`, `Name: n`, etc. — sorted by count descending)
+
+## Root cause
+
+`src/commands/sessions.rs:427`:
+```rust
+println!("Tools:   {:?}", stats.tool_calls);
+```
+
+`stats.tool_calls` is `HashMap<String, u64>`. The `:?` debug format is not user-friendly.
+
+Same issue exists in `src/commands/agents.rs` transcribe header (prints same stats).
+
+## Fix
+
+Format `stats.tool_calls` by iterating entries, sorting by count descending, and joining as `"Name×N"`. The `×` (multiplication sign) or `(N)` notation both work.
+
+## Relevant files
+
+- `src/commands/sessions.rs` — transcribe function, `Tools:` println
+- `src/commands/agents.rs` — transcribe function (same issue)
+- `src/models/stats.rs` — `tool_calls: HashMap<String, u64>`
diff --git a/.beans/claudbg-i6l2--tui-interactive-session-browser.md b/.beans/claudbg-i6l2--tui-interactive-session-browser.md
new file mode 100644
index 0000000..fa2c739
--- /dev/null
+++ b/.beans/claudbg-i6l2--tui-interactive-session-browser.md
@@ -0,0 +1,41 @@
+---
+# claudbg-i6l2
+title: 'TUI: interactive session browser'
+status: todo
+type: epic
+created_at: 2026-03-30T04:44:52Z
+updated_at: 2026-03-30T04:44:52Z
+---
+
+Implement a full terminal UI using Ratatui for browsing Claude Code sessions.
+
+## Design
+
+**Navigation model:** drill-down (not split-pane)
+- Screen 1: full-screen session list → Enter opens transcript
+- Screen 2: full-screen transcript with embedded sub-agents panel → Tab cycles panes, Escape returns to list
+- Screen 3: sub-agent transcript (same layout as Screen 2, Escape returns to parent transcript)
+
+**Session list screen**
+Columns: short ID, datetime, project (truncated tail), model, message count, sub-agent count.
+Sorted most-recent-first.
+
+**Transcript screen**
+- Header: session ID, model, tokens (in/out/cache), tool call summary, duration
+- Scrollable chat log: user messages, assistant text, tool use blocks (truncated), tool results (truncated)
+- Long lines truncated with horizontal scroll
+- Sub-agents panel (Tab to focus): list of sub-agent runs for this session; Enter navigates to sub-agent transcript
+- Thinking blocks hidden by default
+
+**Keyboard bindings**
+- `Up`/`Down`/`k`/`j` — navigate within pane
+- `Left`/`Right`/`h`/`l` — horizontal scroll in transcript
+- `Tab` — cycle focus between panes
+- `Enter` — drill into selection
+- `Escape` — go back
+- `q`/`Q` — quit confirmation dialog (q=confirm, Esc=cancel)
+- `?` — help modal listing all shortcuts (Esc to close)
+
+**Non-goals for this milestone**
+- Live refresh / follow mode (future feature)
+- Color themes / configuration
diff --git a/.beans/claudbg-kg0v--tool-use-input-truncated-at-120-chars-regardless-o.md b/.beans/claudbg-kg0v--tool-use-input-truncated-at-120-chars-regardless-o.md
new file mode 100644
index 0000000..dfc6f7b
--- /dev/null
+++ b/.beans/claudbg-kg0v--tool-use-input-truncated-at-120-chars-regardless-o.md
@@ -0,0 +1,49 @@
+---
+# claudbg-kg0v
+title: Tool use input truncated at 120 chars regardless of --verbose
+status: todo
+type: bug
+created_at: 2026-03-30T04:44:31Z
+updated_at: 2026-03-30T04:44:31Z
+parent: claudbg-8vpb
+---
+
+## Problem
+
+Tool use input (arguments passed to a tool) is hard-capped at 120 characters in transcriptions, cutting off commands mid-content:
+
+```
+[tool: Bash] {"command":"nix develop --command bash -c \"./target/debug/claudbg --help && echo '===' && ./target/debug/claudbg sessio
+```
+
+## Expected behavior
+
+- Default: truncate at 120 chars (or a slightly higher value like 200) with a `…` indicator.
+- `--verbose`: show full input, no truncation.
+
+## Current behavior
+
+Both `render_entry_text` implementations use a hardcoded 120-char cap regardless of `opts.verbose`:
+
+```rust
+// src/commands/sessions.rs:108-110
+let input_preview = serde_json::to_string(input).unwrap_or_default();
+let boundary = input_preview.floor_char_boundary(120);
+let input_short = &input_preview[..boundary];
+```
+
+The `opts.verbose` flag is checked nowhere in the tool use branch.
+
+## Fix
+
+```rust
+let cap = if opts.verbose { usize::MAX } else { 120 };
+let boundary = input_preview.floor_char_boundary(cap);
+```
+
+Same change needed in `agents.rs`.
+
+## Relevant files
+
+- `src/commands/sessions.rs` — `render_entry_text`, `ToolUse` branch (~lines 107–111)
+- `src/commands/agents.rs` — `render_entry_text`, `ToolUse` branch (~lines 68–72)
diff --git a/.beans/claudbg-kxii--tui-interactive-session-browser.md b/.beans/claudbg-kxii--tui-interactive-session-browser.md
new file mode 100644
index 0000000..ec13f67
--- /dev/null
+++ b/.beans/claudbg-kxii--tui-interactive-session-browser.md
@@ -0,0 +1,10 @@
+---
+# claudbg-kxii
+title: 'TUI: interactive session browser'
+status: todo
+type: epic
+created_at: 2026-03-30T04:34:21Z
+updated_at: 2026-03-30T04:34:21Z
+---
+
+Implement a full terminal UI for claudbg using ratatui + crossterm. The TUI provides drill-down navigation: a full-screen session list, a transcript view with a sub-agents side panel, and nested sub-agent transcript views. Data is loaded once on screen entry (no live refresh in this iteration). Navigation is keyboard-driven (vim-style + arrow keys). Modals for quit confirmation and a help screen are included.
diff --git a/.beans/claudbg-nq36--tui-app-state-model-and-screen-enum.md b/.beans/claudbg-nq36--tui-app-state-model-and-screen-enum.md
new file mode 100644
index 0000000..53259ac
--- /dev/null
+++ b/.beans/claudbg-nq36--tui-app-state-model-and-screen-enum.md
@@ -0,0 +1,56 @@
+---
+# claudbg-nq36
+title: 'TUI: app state model and screen enum'
+status: todo
+type: task
+priority: normal
+created_at: 2026-03-30T04:45:19Z
+updated_at: 2026-03-30T04:49:03Z
+parent: claudbg-i6l2
+blocked_by:
+    - claudbg-78xt
+---
+
+Define the central `AppState` struct and `Screen` enum that drive the entire TUI.
+
+## Types to define (`src/tui/state.rs` or `src/commands/tui/state.rs`)
+
+```rust
+pub enum Screen {
+    SessionList,
+    Transcript { session_id: String },
+    SubagentTranscript { parent_session_id: String, agent_id: String },
+}
+
+pub enum Focus {
+    ChatLog,
+    SubagentsPanel,
+}
+
+pub struct AppState {
+    pub screen: Screen,
+    // Session list
+    pub sessions: Vec<SessionListItem>,   // loaded once
+    pub list_selected: usize,
+    // Transcript
+    pub transcript_entries: Vec<RawEntry>,
+    pub transcript_scroll: usize,         // vertical scroll offset
+    pub transcript_h_scroll: usize,       // horizontal scroll offset
+    pub subagents: Vec<AgentRef>,
+    pub subagent_selected: usize,
+    pub focus: Focus,
+    // Modals
+    pub show_quit_dialog: bool,
+    pub show_help: bool,
+}
+```
+
+## Notes
+
+- `SessionListItem` is a lightweight struct with the display fields (short_id, date, project, model, msg_count, agent_count).
+- State transitions: `AppState::enter_transcript(session_id)` loads entries + subagents. `AppState::go_back()` pops the screen stack (or just switches back to `SessionList`).
+- Keep all data loading in state transitions so the render function is pure.
+
+## Blocked by
+
+- ratatui dependency task (claudbg-78xt)
diff --git a/.beans/claudbg-nqxz--duration-shows-0ms-rawentryduration-ms-field-name.md b/.beans/claudbg-nqxz--duration-shows-0ms-rawentryduration-ms-field-name.md
new file mode 100644
index 0000000..74ee6bd
--- /dev/null
+++ b/.beans/claudbg-nqxz--duration-shows-0ms-rawentryduration-ms-field-name.md
@@ -0,0 +1,51 @@
+---
+# claudbg-nqxz
+title: Duration shows 0ms — RawEntry.duration_ms field name mismatch (durationMs vs duration_ms)
+status: todo
+type: bug
+priority: high
+created_at: 2026-03-30T04:44:06Z
+updated_at: 2026-03-30T04:44:06Z
+parent: claudbg-8vpb
+---
+
+## Problem
+
+`sessions transcribe` shows `Duration: 0ms` even for long sessions.
+
+## Root cause
+
+The JSONL file uses camelCase `durationMs` at the entry level:
+
+```json
+{"type": "system", "durationMs": 3446227, ...}
+```
+
+But `RawEntry` declares the field as `duration_ms: Option<u64>` with no serde rename:
+
+```rust
+// src/models/session.rs:33
+pub duration_ms: Option<u64>,
+```
+
+Serde looks for `duration_ms` (snake_case), finds nothing, and the value falls into the `extra` catch-all map instead. `compute_stats` then sums zeros.
+
+## Fix
+
+Add `#[serde(rename = "durationMs")]` to the field:
+
+```rust
+#[serde(rename = "durationMs")]
+pub duration_ms: Option<u64>,
+```
+
+Alternatively use `#[serde(alias = "durationMs")]` to accept both forms.
+
+## Confirmed by
+
+Running against session `21fae0a8`: `grep '"durationMs":[0-9]'` finds entries like `{"type":"system","durationMs":3446227,...}` confirming the field name is camelCase.
+
+## Relevant files
+
+- `src/models/session.rs` — `RawEntry`, `duration_ms` field (line 33)
+- `src/models/stats.rs` — `compute_stats`, accumulates `entry.duration_ms`
diff --git a/.beans/claudbg-pfa5--json-output-is-array-of-arrays-instead-of-array-of.md b/.beans/claudbg-pfa5--json-output-is-array-of-arrays-instead-of-array-of.md
new file mode 100644
index 0000000..10bfdb8
--- /dev/null
+++ b/.beans/claudbg-pfa5--json-output-is-array-of-arrays-instead-of-array-of.md
@@ -0,0 +1,40 @@
+---
+# claudbg-pfa5
+title: JSON output is array-of-arrays instead of array-of-objects
+status: todo
+type: bug
+priority: normal
+created_at: 2026-03-30T04:37:05Z
+updated_at: 2026-03-30T04:41:14Z
+parent: claudbg-tci9
+---
+
+## Problem
+
+`sessions list --output json` and `sessions dump --output json` serialize rows as `Vec<Vec<String>>`, producing an array of arrays with no field names:
+
+```json
+[["21fae0a8", "2024-01-01", "/path", "claude-3", "42"], ...]
+```
+
+## Expected behavior
+
+Output should be an array of objects with named keys:
+
+```json
+[{"id": "21fae0a8", "date": "2024-01-01", "project": "/path", "model": "claude-3", "messages": 42, "subagents": 0}, ...]
+```
+
+## Affected commands
+
+- `sessions list --output json`: keys should be `id`, `date`, `project`, `model`, `messages`, `subagents`
+- `sessions dump --output json`: keys should be `seq`, `timestamp`, `type`, `role`, `content`
+
+## Root cause
+
+`src/commands/sessions.rs` passes `&rows: &Vec<Vec<String>>` directly to `render_json`. The fix is to build a typed struct or `serde_json::Value` map per row before serializing.
+
+## Relevant files
+
+- `src/commands/sessions.rs` — `list()` and `dump()` JSON output branches
+- `src/output/json.rs` — generic `render_json`
diff --git a/.beans/claudbg-pta8--tui-session-list-screen-widget.md b/.beans/claudbg-pta8--tui-session-list-screen-widget.md
new file mode 100644
index 0000000..df390ef
--- /dev/null
+++ b/.beans/claudbg-pta8--tui-session-list-screen-widget.md
@@ -0,0 +1,40 @@
+---
+# claudbg-pta8
+title: 'TUI: session list screen widget'
+status: todo
+type: feature
+priority: normal
+created_at: 2026-03-30T04:45:53Z
+updated_at: 2026-03-30T04:49:03Z
+parent: claudbg-i6l2
+blocked_by:
+    - claudbg-nq36
+---
+
+Render the full-screen session list as the TUI's home screen.
+
+## Layout
+
+```
+┌─ Sessions ──────────────────────────────────────────────────────┐
+│ ID       Date                 Project              Model    Msgs Agents │
+│ ────────────────────────────────────────────────────────────────────── │
+│ 21fae0a8 2026-03-29 14:32:01 …/pop/claudbg         claude-3   42    3  │◄ selected (highlighted)
+│ def4776b 2026-03-28 10:15:44 …/pop/claudbg         claude-3  128    1  │
+│ ...                                                                     │
+└─────────────────────────────────────────────────────────────────┘
+```
+
+## Implementation
+
+- Use `ratatui::widgets::Table` with `TableState` for selection tracking.
+- Highlight the selected row with a `Style` that inverts colors or uses a distinct background.
+- Columns: ID (8), Date (20), Project (dynamic, truncated to tail), Model (20), Msgs (6), Agents (7).
+- Project path: display only the last 30 or so characters (prefix with `…`).
+- Populate from `AppState.sessions` (loaded on TUI start from `discover_sessions` + DB).
+- Scroll with arrow keys / j/k; selected index tracked in `AppState.list_selected`.
+
+## Blocked by
+
+- app state model (claudbg-nq36)
+- agent discovery fix (claudbg-33n0) — for accurate agent counts
diff --git a/.beans/claudbg-rudq--tui-transcript-screen-chat-log-and-stats-header.md b/.beans/claudbg-rudq--tui-transcript-screen-chat-log-and-stats-header.md
new file mode 100644
index 0000000..66b8a66
--- /dev/null
+++ b/.beans/claudbg-rudq--tui-transcript-screen-chat-log-and-stats-header.md
@@ -0,0 +1,46 @@
+---
+# claudbg-rudq
+title: 'TUI: transcript screen — chat log and stats header'
+status: todo
+type: feature
+priority: normal
+created_at: 2026-03-30T04:46:56Z
+updated_at: 2026-03-30T04:49:03Z
+parent: claudbg-i6l2
+blocked_by:
+    - claudbg-ut9q
+---
+
+Render the full-screen transcript view for a session or sub-agent.
+
+## Layout
+
+```
+Session: 21fae0a8  Model: claude-3-opus  Tokens: in=12k out=4k  Duration: 57min
+Tools: Bash×12, Read×8, Write×3
+────────────────────────────────────────────────────────────────────────────
+[user]: Please implement the sessions list command…
+[assistant]: I'll start by reading the existing code structure…
+[tool: Read] {"file_path": "/path/to/file.rs"}
+[tool_result]: pub struct SessionRef {\n    pub session_id: String,\n    pub project…
+[assistant]: Here is the implementation:
+…
+```
+
+## Implementation
+
+- Split the frame: top block = stats header (fixed ~3 lines), middle = scrollable chat log, right panel = sub-agents (see sub-agents task).
+- Use `ratatui::widgets::Paragraph` with `Wrap { trim: false }` for the header.
+- Chat log: build as a `Vec<Line>` (ratatui spans/lines). Each entry becomes one or more lines. Use `Paragraph::new(text).scroll((v_offset, h_offset))` for scroll.
+- Horizontal scroll for long lines: track `transcript_h_scroll` in state, increment/decrement with left/right keys.
+- Vertical scroll: `transcript_scroll` in state, up/down keys.
+- Thinking blocks: hidden by default (filter from `entries` before rendering).
+- Tool results: truncated to ~200 chars by default.
+
+## Data source
+
+Load `RawEntry` list from raw JSONL via `read_session_file` when entering transcript screen. Compute `SessionStats` for the header.
+
+## Blocked by
+
+- app state (claudbg-nq36), event loop (claudbg-ut9q)
diff --git a/.beans/claudbg-tci9--session-list-dump-improvements.md b/.beans/claudbg-tci9--session-list-dump-improvements.md
new file mode 100644
index 0000000..03e6f7b
--- /dev/null
+++ b/.beans/claudbg-tci9--session-list-dump-improvements.md
@@ -0,0 +1,10 @@
+---
+# claudbg-tci9
+title: Session list & dump improvements
+status: todo
+type: epic
+created_at: 2026-03-30T04:34:16Z
+updated_at: 2026-03-30T04:34:16Z
+---
+
+Groups bugs and improvements to the `sessions list` and `sessions dump` commands: table overflow, JSON output format, --include help text, duplicate row collapsing, and sub-agent count.
diff --git a/.beans/claudbg-ut9q--tui-terminal-setup-event-loop-and-teardown.md b/.beans/claudbg-ut9q--tui-terminal-setup-event-loop-and-teardown.md
new file mode 100644
index 0000000..9fd6431
--- /dev/null
+++ b/.beans/claudbg-ut9q--tui-terminal-setup-event-loop-and-teardown.md
@@ -0,0 +1,38 @@
+---
+# claudbg-ut9q
+title: 'TUI: terminal setup, event loop, and teardown'
+status: todo
+type: task
+priority: normal
+created_at: 2026-03-30T04:45:32Z
+updated_at: 2026-03-30T04:49:03Z
+parent: claudbg-i6l2
+blocked_by:
+    - claudbg-nq36
+---
+
+Implement terminal initialization, the main event loop, and graceful teardown.
+
+## Responsibilities
+
+1. **Setup:** enable raw mode (`crossterm::terminal::enable_raw_mode`), switch to alternate screen (`EnterAlternateScreen`), create `CrosstermBackend`, instantiate `ratatui::Terminal`.
+
+2. **Panic hook:** install a panic hook that restores the terminal before printing the panic message (prevents garbled terminal on crash).
+
+3. **Event loop:** `loop { terminal.draw(|f| render(f, &state)); if event::poll(Duration::from_millis(50))? { handle_event(event::read()?, &mut state); } if state.should_quit { break; } }`
+
+4. **Teardown:** disable raw mode, leave alternate screen.
+
+## Entry point
+
+Replaces the stub in `src/commands/stubs.rs` (the `Tui` command handler). The `tui()` async function should call into a synchronous `run_tui()` function since ratatui is sync.
+
+## Notes
+
+- Use `crossterm::event::Event` for keyboard input — no mouse needed for this milestone.
+- The 50 ms poll timeout gives ~20 fps without busy-spinning.
+- Teardown must run even on error: use a RAII guard or wrap in a closure.
+
+## Blocked by
+
+- ratatui deps (claudbg-78xt), app state (claudbg-nq36)
diff --git a/.beans/claudbg-xpzp--sessions-list-should-show-sub-agent-run-count-per.md b/.beans/claudbg-xpzp--sessions-list-should-show-sub-agent-run-count-per.md
new file mode 100644
index 0000000..af39bdf
--- /dev/null
+++ b/.beans/claudbg-xpzp--sessions-list-should-show-sub-agent-run-count-per.md
@@ -0,0 +1,42 @@
+---
+# claudbg-xpzp
+title: '`sessions list` should show sub-agent run count per session'
+status: todo
+type: bug
+priority: normal
+created_at: 2026-03-30T04:40:53Z
+updated_at: 2026-03-30T04:43:26Z
+parent: claudbg-tci9
+blocked_by:
+    - claudbg-33n0
+---
+
+## Problem
+
+`sessions list` has no column showing how many sub-agent runs a session spawned, making it impossible to identify sessions that used the Agent tool.
+
+## Expected behavior
+
+Add a `Sub-agents` column to `sessions list` showing the count of sub-agent JSONL files associated with each session. Zero for sessions with no sub-agents.
+
+## Disk structure (important)
+
+The actual subagent files live at:
+```
+~/.claude/projects/<project-dir>/<session-uuid>/subagents/agent-<id>.jsonl
+```
+Note: there is a *directory* named `<session-uuid>` (without `.jsonl`) alongside the session file that contains the `subagents/` folder.
+
+The existing `discover_agents_for_session(session_file)` function currently looks in the *wrong place* (see separate bug). Once that is fixed, the agent count can be obtained by calling `discover_agents_for_session` for each session and taking `.len()`.
+
+## Implementation notes
+
+- In `sessions list`, after syncing sessions, call `discover_agents_for_session(&session_ref.file_path)` for each session to get the count.
+- Add the count as the last column in the table row.
+- Update JSON output to include `subagents` field.
+- This is blocked on the subagent path discovery bug being fixed first.
+
+## Relevant files
+
+- `src/commands/sessions.rs` — `list()` function
+- `src/parser/discovery.rs` — `discover_agents_for_session`
diff --git a/.beans/claudbg-zy1p--tool-results-hidden-by-default-in-transcriptions-s.md b/.beans/claudbg-zy1p--tool-results-hidden-by-default-in-transcriptions-s.md
new file mode 100644
index 0000000..f13e2f3
--- /dev/null
+++ b/.beans/claudbg-zy1p--tool-results-hidden-by-default-in-transcriptions-s.md
@@ -0,0 +1,43 @@
+---
+# claudbg-zy1p
+title: Tool results hidden by default in transcriptions — should show with truncation
+status: todo
+type: bug
+created_at: 2026-03-30T04:44:18Z
+updated_at: 2026-03-30T04:44:18Z
+parent: claudbg-8vpb
+---
+
+## Problem
+
+Tool results (the output of tool calls) are hidden unless the user passes `--include output`. Since tool outputs are a core part of understanding what happened in a session, they should be visible by default.
+
+## Expected behavior
+
+- Tool results shown by default (no flag needed).
+- Truncated to ~200 chars by default (current limit, keep it).
+- `--verbose` removes the truncation cap and shows full output.
+- The `--include output` flag / `opts.include.output` gate should be removed (or inverted).
+
+## Current behavior
+
+`render_entry_text` in both `sessions.rs` and `agents.rs` gates `ToolResult` on `opts.include.output`:
+
+```rust
+if opts.include.output {
+    // show tool result
+}
+```
+
+With the default `--include ""`, `opts.include.output` is `false`, so results are never shown.
+
+## Fix
+
+Remove the `if opts.include.output` guard (or replace it with the inverse: hide only when an explicit `--include no-output` flag is passed, but that's over-engineering — just always show with truncation).
+
+Also apply the same change to the `--verbose` path: when `opts.verbose` is true, show the full content uncapped.
+
+## Relevant files
+
+- `src/commands/sessions.rs` — `render_entry_text`, `ToolResult` branch (~line 113–131)
+- `src/commands/agents.rs` — `render_entry_text`, same branch (~line 74–92)