agenthud 0.12.4 → 0.13.1

package/README.md

@@ -10,9 +10,11 @@ An observability layer for [Claude Code](https://github.com/anthropics/claude-co
 
 AgentHUD reads Claude Code's session files from `~/.claude/projects/` and gives you three things:
 
-- **Live monitor** ([`agenthud`](#live-monitor)) — a split-view TUI showing every project, session, sub-agent, and activity as it happens.
-- **Structured export** ([`agenthud report`](#report)) — print activity for any date as Markdown or JSON for piping to scripts, dashboards, or other LLMs.
-- **LLM digest** ([`agenthud summary`](#summary)) — synthesize a day or a date range into an engineering summary via the `claude` CLI, with caching so weekly digests are cheap to regenerate.
+- **Live monitor** (`agenthud`) — a split-view TUI showing every project, session, sub-agent, and activity as it happens
+- **Structured export** (`agenthud report`) — Markdown or JSON for piping to scripts, dashboards, or other LLMs
+- **LLM digest** (`agenthud summary`) — a day or a date range synthesized into an engineering summary via the `claude` CLI
+
+→ **See [FEATURES.md](./FEATURES.md) for the full surface** — every flag, keybinding, config key, file path, and env var.
 
 ## Install
 
@@ -26,17 +28,28 @@ Run this in a separate terminal while using Claude Code. Press `?` inside the TU
 
 > **Platform support.** Primary development is on macOS and Linux; the full test suite runs on all three platforms in CI (including Windows). Windows runtime behavior is exercised by a manual smoke job but isn't daily-driven — issues there are valued bug reports.
 
-Pass `--cwd` to scope the view to the Claude project that contains your current directory — useful when you have many projects but only care about the one you're in right now. Exits 1 with a stderr message if no such project is registered.
+## Quickstart
 
 ```bash
-agenthud           # all Claude projects on the machine
-agenthud --cwd     # only the project containing the current dir
-agenthud --once    # print one frame and exit (no alt-screen)
+# Live monitor
+agenthud                                  # all Claude projects
+agenthud --cwd                            # scope to the project containing $PWD
+agenthud --once                           # snapshot mode, no alt-screen
+
+# Activity report
+agenthud report --date today              # today's activity as markdown
+agenthud report --format json             # script-readable
+agenthud report --with-git                # merge git commits into the timeline
+
+# LLM summary
+agenthud summary --date today             # daily summary via `claude -p`
+agenthud summary --last 7d                # cross-day synthesis of last 7 days
+agenthud summary -oI                      # open the summary + the summaries index
 ```
 
-## Live monitor
+## What you see
 
-AgentHUD's TUI splits the screen into a project tree and an activity viewer:
+The TUI splits into a project tree (top) and an activity viewer (bottom):
 
 ```
 ┌─ Projects ─────────────────────────────────────────────────┐
@@ -57,257 +70,24 @@ AgentHUD's TUI splits the screen into a project tree and an activity viewer:
 └────────────────────────────────────────────────────────────┘
 ```
 
-**Project tree (top pane)**
-- Sessions grouped under their project (project name + path at the top).
-- Session rows show short ID + first user prompt (the session's "topic"). Long titles truncate with a `…` suffix.
-- Right edge of each row shows how long ago it was last touched: `42m`, `17h`, `3d`, `2w`, `1mo`, `1y`. Project rows use the most recent session's mtime.
-- Non-interactive sessions (from `claude -p`, SDK, `agenthud summary`) appear in parens and dimmed.
-- Sub-agents nest one level deeper under their parent session.
-- Cold projects collapse under `... N cold projects` at the bottom (press Enter on the line to expand).
-- Press `h` to hide a project, session, or sub-agent (saved to `~/.agenthud/state.yaml`).
-
-**Activity viewer (bottom pane)**
-- Real-time feed for the selected session: file reads, edits, bash, responses, thinking, git commits. Newest at the bottom, like `tail -f`.
-- The **newest visible row is rendered "alive"** while in LIVE mode: its icon is replaced with a spinning glyph and the text turns bold. When a new activity arrives, the spinner moves to the new bottom row. Hidden when paused or empty.
-- Press `f` to cycle through filter presets (configurable).
-- Press `↵` on any row to open a scrollable detail view; on a commit row this shows `git show --stat --patch`.
-
-### Session status
-
-Each session row carries a colored badge derived from when its JSONL file was last touched:
-
-| Badge | Color | Meaning |
-|-------|-------|---------|
-| `[hot]` | green | Updated in the last 30 minutes — actively running |
-| `[warm]` | yellow | Updated in the last hour |
-| `[cool]` | cyan | Updated earlier today |
-| `[cold]` | gray | Last updated yesterday or earlier — collapsed under `... N cold projects` at the bottom |
-
-Sub-agents use the same scheme. Projects inherit the hottest status of their sessions; a project is treated as "cold" only when all its sessions are cold.
-
-### Activity types
-
-| Icon | Type | Description |
-|------|------|-------------|
-| `○` | Read | File being read |
-| `~` | Edit / Write | File being modified |
-| `$` | Bash | Shell command |
-| `*` | Glob / Grep | File search |
-| `@` | WebFetch / WebSearch | Web request |
-| `»` | Task | Sub-agent task |
-| `<` | Response | Claude's text response |
-| `>` | User | Your message |
-| `…` | Thinking | Claude's thinking (requires `showThinkingSummaries: true`) |
-| `◆` | Commit | Git commit in the project (when `--with-git` or in viewer) |
-
-### Keyboard shortcuts
-
-Full reference is also available inside the app — press `?`.
-
-#### Project tree focus
-
-| Key | Action |
-|-----|--------|
-| `↑` / `k` | Move up |
-| `↓` / `j` | Move down |
-| `←` | Jump to parent (sub-agent → session, session → project) |
-| `↵` | Expand/collapse project, session, or sub-agent summary |
-| `h` | Hide selected (project / session / sub-agent) |
-| `Tab` | Switch focus to activity viewer |
-| `PgUp` / `Ctrl+B` | Page up |
-| `PgDn` / `Ctrl+F` | Page down |
-| `t` | Track — auto-follow the newest live sub-agent (any nav key turns it off) |
-| `r` | Refresh now |
-| `?` | Help |
-| `q` | Quit |
-
-When tracking is on, the tree panel's title shows `[LIVE ⠧]` and the status bar replaces `t: track` with `TRK ●`. Any explicit selection-changing key (`↑/k`, `↓/j`, `PgUp/PgDn`, `↵`, `h`, or `t` again) turns tracking off.
-
-#### Activity viewer focus
-
-| Key | Action |
-|-----|--------|
-| `↑` / `k` | Scroll one line up |
-| `↓` / `j` | Scroll one line down |
-| `PgUp` / `Ctrl+B` | Page up |
-| `PgDn` / `Ctrl+F` | Page down |
-| `Ctrl+U` / `Ctrl+D` | Half page up / down |
-| `g` | Jump to top (oldest) |
-| `G` | Jump to live (newest, bottom) |
-| `↵` | Open detail view |
-| `f` | Cycle filter preset |
-| `r` | Refresh now |
-| `Tab` | Switch focus to project tree |
-| `?` | Help |
-| `q` | Quit |
-
-#### Detail view
-
-| Key | Action |
-|-----|--------|
-| `↑` / `k` / `↓` / `j` | Scroll one line |
-| `Ctrl+U` / `Ctrl+D` | Half page up / down |
-| `↵` / `Esc` / `q` | Close |
-
-Detail view colors the content based on activity type:
-
-- **Git commit detail** (`git show --stat --patch`): added lines green (`+`), removed lines red (`-`), hunk headers cyan (`@@ ... @@`), `commit/Author/Date/diff` metadata dimmed.
-- **Response / thinking / prompt**: text inside triple-backtick code fences renders in cyan so the boundary between prose and code is obvious. No language-specific syntax highlighting — just code-vs-prose separation.
-
-### Behavior
-
-- **Alternate screen buffer.** Watch mode uses the alt-screen (like `vim`, `htop`, `btop`), so quitting (`q`) restores the pre-launch shell completely. No TUI residue, no "is it still running?" confusion.
-- **Minimum terminal size.** 80 cols × 20 rows. Smaller terminals show a one-line hint and redraw automatically when you resize.
-- **Help overlay scrolls.** Press `?` for an in-app reference. The overlay scrolls (`j/k`, `PgUp/PgDn`, `Ctrl+B/F`, `Space`, `g/G`) so the full content is reachable on shorter terminals.
-
-## Report
-
-Print activity for a date in Markdown or JSON — suitable for piping to scripts or LLMs:
-
-```bash
-agenthud report                                      # today (Markdown)
-agenthud report --date 2026-05-14                    # specific date
-agenthud report --date today --include all           # all activity types
-agenthud report --format json                        # JSON output
-agenthud report --detail-limit 0                     # no truncation (full text)
-agenthud report --with-git                           # merge git commits into timeline
-```
-
-Output:
-
-```
-# AgentHUD Report: 2026-05-14
-
-## myproject (10:23 – 14:45)
-
-[10:23] $ npm test
-[10:35] ~ src/ui/App.tsx
-[11:15] < Added spinner hook to make the UI feel alive.
-[14:30] ◆ abc1234 feat: fix auth callback
-```
-
-**Options:**
-
-| Flag | Default | Description |
-|------|---------|-------------|
-| `--date` | today | `YYYY-MM-DD`, `today`, `yesterday`, or `-Nd` (N days ago, local date) |
-| `--include` | `user,response,bash,edit,thinking` | Comma-separated types or `all` |
-| `--format` | `markdown` | `markdown` or `json` |
-| `--detail-limit` | `120` | Max chars per detail field; `0` = unlimited |
-| `--with-git` | off | Merge git commits from each session's project into the timeline |
-
-`--include` types: `response`, `bash`, `edit`, `thinking`, `read`, `glob`, `user`
-
-## Summary
-
-Generate an LLM-based summary of one day or a date range using the `claude` CLI:
-
-```bash
-# Single day
-agenthud summary                            # today (always regenerated)
-agenthud summary --date yesterday           # natural-language date
-agenthud summary --date 2026-05-14          # past date (cached on second run)
-agenthud summary --date 2026-05-14 --force  # ignore cache
-agenthud summary --prompt "Only commits"    # override prompt
-
-# Date range — daily summaries are re-summarized into a meta-summary
-agenthud summary --last 7d                          # last 7 days, ending today
-agenthud summary --from 2026-05-10 --to 2026-05-16  # explicit range
-agenthud summary --last 7d -y                       # skip per-day confirmations
-
-# Tune what gets sent to the LLM — same option shape as `report`
-agenthud summary --include all                      # include every activity type
-agenthud summary --detail-limit 200                 # truncate long tool outputs
-agenthud summary --with-git                         # include git commits (default in config)
-# Defaults are CLI flag → `summary:` config → `report:` config → built-in.
-
-# Cheaper model — summarization doesn't need Opus-tier reasoning
-agenthud summary --date today --model sonnet        # ~40% cheaper than Opus
-agenthud summary --last 7d --model haiku            # ~80% cheaper, 200K context
-
-# Open the resulting summary in your OS default app (browser, VS Code, etc.)
-agenthud summary --last 7d --open                   # -o is the short form
-
-# Open the index (~/.agenthud/summaries/index.md) instead — a hub
-# listing every daily and range summary, grouped by year/month.
-agenthud summary --open-index                       # -I is the short form
-agenthud summary -oI                                # open today + the index
-```
-
-The index is auto-regenerated whenever a summary writes, and each summary file gets a one-line backlink footer at the top (`← all summaries · ← prev · next →`) so any markdown viewer is enough to navigate the whole corpus without leaving the file.
-
-**Daily summaries** are saved to `~/.agenthud/summaries/YYYY-MM-DD.md`. Past dates are cached and returned instantly; today is always regenerated (activity still growing).
+Sessions get colored badges — `[hot]` / `[warm]` / `[cool]` / `[cold]` — based on how recently their JSONL file was touched. Cold projects collapse under a `... N cold projects` sentinel. Press `↵` on any activity to open a scrollable detail view.
 
-**Range summaries** generate any missing daily summaries first, then feed those into a second `claude` call that produces a cross-day synthesis (themes, multi-day workstreams, recurring patterns). Output is cached to `~/.agenthud/summaries/range-FROM_TO.md`. Cached dailies cost nothing to reuse, so weekly summaries are cheap after the first run.
-
-Each missing daily prompts for confirmation just before generation, so you see concrete context (session/activity/commit counts and report size) before deciding. Pass `-y` / `--yes` to skip all prompts. Press Enter to accept the default (`[Y/n]`).
-
-**Prompt customization:** The daily template lives at `~/.agenthud/summary-prompt.md` and the range template at `~/.agenthud/summary-range-prompt.md`. Both are auto-created from built-in templates on first run. Edit them freely.
-
-**`--date` formats:** `YYYY-MM-DD`, `today`, `yesterday`, or `-Nd` (N days ago).
-
-**Model selection:** Summarization is a low-reasoning task (structured input → structured markdown) — Sonnet or Haiku usually beats Opus on cost-per-summary with no quality loss. Pass `--model sonnet`, `--model haiku`, or a full model id (`--model claude-sonnet-4-6`). With no flag, `claude` uses its default model.
-
-**Cost warning:** If the day's activity log is large (~300K tokens or more), AgentHUD prints a warning before sending and asks for one more confirmation in interactive mode. `-y` skips the prompt but still prints the warning.
-
-**Requires:** [`@anthropic-ai/claude-code`](https://www.npmjs.com/package/@anthropic-ai/claude-code) installed and authenticated.
+Full keybinding and badge reference: [FEATURES.md](./FEATURES.md#keybindings).
 
 ## Configuration
 
-`~/.agenthud/config.yaml` is auto-created on first run with sensible defaults. Edit it freely:
-
-```yaml
-# How often to poll for activity updates (Linux fallback when fs.watch isn't used)
-refreshInterval: 2s
-
-# Activity filter presets (cycle with 'f' key in viewer)
-# Each list is one preset. Use "all" (or "*") to show everything.
-# Types: response, user, bash, edit, thinking, read, glob, commit
-filterPresets:
-  - ["all"]
-  - ["response", "user"]
-  - ["commit"]
-
-# Defaults for `agenthud report` (CLI flags still win per-invocation).
-report:
-  include: [user, response, bash, edit, thinking]
-  detailLimit: 120
-  withGit: false
-  format: markdown
-
-# Defaults for `agenthud summary`. Any field omitted here is inherited
-# from `report` above. `model` is summary-specific.
-summary:
-  withGit: true
-  detailLimit: 0
-  # model: sonnet
-```
-
-`report` / `summary` resolve each option as **CLI flag → `summary:` key → `report:` key → built-in default**. The effective set is printed to stderr at the start of each run (e.g. `agenthud: report → include=[user,response,bash,edit,thinking] detail-limit=120 with-git=off format=markdown`), so the actual values are always visible.
-
-App-managed state (hidden items) lives separately in `~/.agenthud/state.yaml` so your config file stays clean. You shouldn't need to edit it manually — use `h` in the TUI to hide things.
-
-## Files
-
-| Path | Purpose |
-|------|---------|
-| `~/.agenthud/config.yaml` | User settings (edit freely) |
-| `~/.agenthud/state.yaml` | Hidden projects/sessions/sub-agents (app-managed) |
-| `~/.agenthud/summary-prompt.md` | LLM prompt template for daily `summary` |
-| `~/.agenthud/summary-range-prompt.md` | LLM prompt template for range `summary --last/--from/--to` |
-| `~/.agenthud/summaries/YYYY-MM-DD.md` | Cached daily summaries |
-| `~/.agenthud/summaries/range-FROM_TO.md` | Cached range summaries |
-| `~/.agenthud/summaries/index.md` | Auto-regenerated hub linking every summary; each summary also gets a backlink footer at the top |
+`~/.agenthud/config.yaml` is auto-created on first run with sensible defaults. CLI flags override config values per-invocation. Resolution order is `CLI flag → summary.<key> → report.<key> → built-in default`, and the effective values print to stderr at the start of every `report` / `summary` run.
 
-## Environment Variables
+App-managed UI state (hidden projects/sessions/sub-agents toggled by `h`) lives separately in `~/.agenthud/state.yaml`.
 
-| Variable | Default | Description |
-|----------|---------|-------------|
-| `CLAUDE_PROJECTS_DIR` | `~/.claude/projects` | Path to Claude Code projects directory. Useful for backups or mounted volumes. |
+Full schema, file paths, and env vars: [FEATURES.md → Config](./FEATURES.md#config).
 
-## Feedback
+## More
 
-Issues and PRs welcome at [GitHub](https://github.com/neochoon/agenthud).
+- **Reference:** [FEATURES.md](./FEATURES.md) — every flag, keybinding, config key, file path, env var
+- **Release history:** [CHANGELOG.md](./CHANGELOG.md)
+- **Deferred items:** [BACKLOG.md](./BACKLOG.md)
+- **Issues / PRs:** [GitHub](https://github.com/neochoon/agenthud)
 
 ## License
 

package/dist/index.js

@@ -15,4 +15,4 @@ Error: Node.js ${MIN_NODE_VERSION}+ is required (current: ${process.version})
   process.exit(1);
 }
 if (!process.env.NODE_ENV) process.env.NODE_ENV = "production";
-import("./main-PNDDQ5WC.js");
+import("./main-GMNV6O4E.js");

package/dist/{main-PNDDQ5WC.js → main-GMNV6O4E.js}

@@ -23,7 +23,8 @@ var DEFAULT_INCLUDE_TYPES = [
   "response",
   "bash",
   "edit",
-  "thinking"
+  "thinking",
+  "task"
 ];
 var ALLOWED_INCLUDE_TYPES = /* @__PURE__ */ new Set([
   "user",
@@ -33,7 +34,8 @@ var ALLOWED_INCLUDE_TYPES = /* @__PURE__ */ new Set([
   "thinking",
   "read",
   "glob",
-  "commit"
+  "commit",
+  "task"
 ]);
 var DEFAULT_GLOBAL_CONFIG = {
   refreshIntervalMs: 2e3,
@@ -944,11 +946,34 @@ function numberLines(content, start) {
   const width = String(start + lines.length - 1).length;
   return lines.map((line, i) => `${String(start + i).padStart(width)}: ${line}`).join("\n");
 }
+function formatBashBody(stdout, stderr, interrupted) {
+  const out = stdout?.replace(/\n+$/, "");
+  const err = stderr?.replace(/\n+$/, "");
+  const sections = [];
+  if (out) sections.push(out);
+  if (err) sections.push(out ? `--- stderr ---
+${err}` : err);
+  if (interrupted) sections.push("[interrupted]");
+  if (sections.length === 0) return null;
+  return sections.join("\n\n");
+}
 function buildToolDetailBody(name, input, result) {
   if (name === "Write") {
     const content = result?.content ?? input?.content;
     if (content) return { text: content, kind: "code" };
   }
+  if (name === "Task") {
+    const content = result?.content;
+    if (content) return { text: content, kind: "code" };
+  }
+  if (name === "Bash") {
+    const text = formatBashBody(
+      result?.stdout,
+      result?.stderr,
+      result?.interrupted
+    );
+    if (text) return { text, kind: "code" };
+  }
   if (name === "Read") {
     const content = result?.file?.content;
     if (content) {
@@ -1123,6 +1148,7 @@ function activityMatchesInclude(activity, include) {
     return true;
   if (include.includes("glob") && (label === "glob" || label === "grep"))
     return true;
+  if (include.includes("task") && label === "task") return true;
   return false;
 }
 function isSameLocalDay(a, b) {
@@ -1137,6 +1163,14 @@ function formatActivity(activity, limit) {
   const suffix = detail ? `: ${detail}` : "";
   return `[${time}] ${activity.icon} ${activity.label}${suffix}`;
 }
+function formatTaskBody(activity, limit) {
+  if (activity.label !== "Task") return null;
+  if (!activity.detailBody) return null;
+  const body = truncateRaw(activity.detailBody, limit);
+  return `<task-result>
+${body}
+</task-result>`;
+}
 function formatDateString2(date) {
   return `${date.getFullYear()}-${String(date.getMonth() + 1).padStart(2, "0")}-${String(date.getDate()).padStart(2, "0")}`;
 }
@@ -1239,6 +1273,8 @@ function generateReport(sessions, options2) {
     lines.push("");
     for (const activity of activities) {
       lines.push(formatActivity(activity, detailLimit));
+      const taskBody = formatTaskBody(activity, detailLimit);
+      if (taskBody) lines.push(taskBody);
     }
     lines.push("");
   }
@@ -2131,6 +2167,16 @@ function resolvePrompt(kind, override) {
     return "Summarize the input below.";
   }
 }
+function buildRangeMetaInput(dailyMarkdowns) {
+  if (dailyMarkdowns.length === 0) return "";
+  return dailyMarkdowns.map(
+    ({ date, markdown }) => `<day date="${dateKey(date)}">
+
+${markdown}
+
+</day>`
+  ).join("\n\n");
+}
 function shouldUseRangeCache(force, dates, today, cacheExists) {
   if (force) return false;
   if (!cacheExists) return false;
@@ -2217,7 +2263,12 @@ function spawnClaude(opts) {
         resolve2({ code: 1, text: "", usage: null });
       }
     });
+    let firstChunkFired = false;
     const writeText = (text) => {
+      if (!firstChunkFired) {
+        firstChunkFired = true;
+        opts.onFirstChunk?.();
+      }
       assembledText += text;
       if (opts.streamToStdout) process.stdout.write(text);
       cacheStream?.write(text);
@@ -2410,22 +2461,17 @@ async function generateDailySummary(opts) {
       }
     }
   }
-  const showTicker = opts.announce && !opts.streamToStdout;
-  if (opts.announce && !showTicker) {
-    process.stderr.write(
-      `sending to claude (this may take a minute)...
-
-`
-    );
-  }
-  const stopTicker = showTicker ? startStderrTicker("sending to claude") : null;
+  const stopTicker = opts.announce ? startStderrTicker("sending to claude") : null;
   const prompt = resolvePrompt("daily", opts.promptOverride);
   const result = await spawnClaude({
     prompt,
     stdin: reportMarkdown,
     cachePath: cached2,
     streamToStdout: opts.streamToStdout,
-    model: opts.model
+    model: opts.model,
+    onFirstChunk: () => {
+      if (stopTicker) stopTicker();
+    }
   });
   if (stopTicker) stopTicker();
   if (opts.announce && result.code === 0) {
@@ -2583,33 +2629,24 @@ async function runRangeSummary(options2) {
     }
     return 0;
   }
-  const metaInput = dailyMarkdowns.map(({ date, markdown }) => `# ${dateKey(date)}
-
-${markdown}`).join("\n\n---\n\n");
+  const metaInput = buildRangeMetaInput(dailyMarkdowns);
   process.stderr.write(
     `
 combining ${dailyMarkdowns.length} daily summaries into range summary...
 `
   );
   const metaStreams = !options2.open;
-  if (!metaStreams) {
-  } else {
-    process.stderr.write(
-      `sending to claude (this may take a minute)...
-
-`
-    );
-  }
-  const stopMetaTicker = metaStreams ? null : startStderrTicker("sending to claude");
+  const stopMetaTicker = startStderrTicker("sending to claude");
   const metaPrompt = resolvePrompt("range");
   const metaResult = await spawnClaude({
     prompt: metaPrompt,
     stdin: metaInput,
     cachePath: rangeCache,
     streamToStdout: metaStreams,
-    model: options2.model
+    model: options2.model,
+    onFirstChunk: () => stopMetaTicker()
   });
-  if (stopMetaTicker) stopMetaTicker();
+  stopMetaTicker();
   if (metaResult.code !== 0) {
     return metaResult.code;
   }
@@ -3829,6 +3866,41 @@ function SessionTreePanel({
   ] });
 }
 
+// src/ui/viewerCursor.ts
+function adjustViewerCursorOnNewActivities(args) {
+  const {
+    prevCursorLine,
+    prevActivityCount,
+    newActivityCount,
+    isLive,
+    viewerRows
+  } = args;
+  const noOp = {
+    cursorLine: prevCursorLine,
+    autoPause: false,
+    scrollDelta: 0
+  };
+  if (!isLive) return noOp;
+  if (prevCursorLine === 0)
+    return { cursorLine: 0, autoPause: false, scrollDelta: 0 };
+  if (newActivityCount <= prevActivityCount) return noOp;
+  const delta = newActivityCount - prevActivityCount;
+  const maxCursor = Math.max(0, viewerRows - 1);
+  const upwardRoom = maxCursor - prevCursorLine;
+  if (delta <= upwardRoom) {
+    return {
+      cursorLine: prevCursorLine + delta,
+      autoPause: false,
+      scrollDelta: 0
+    };
+  }
+  return {
+    cursorLine: maxCursor,
+    autoPause: true,
+    scrollDelta: delta - upwardRoom
+  };
+}
+
 // src/ui/App.tsx
 import { Fragment as Fragment4, jsx as jsx5, jsxs as jsxs5 } from "react/jsx-runtime";
 var VIEWER_HEIGHT_FRACTION = 0.55;
@@ -4063,12 +4135,12 @@ function App({
   useEffect3(() => {
     allFlatRef.current = allFlat;
   }, [allFlat]);
-  const activitiesLengthRef = useRef(0);
   const activitiesRef = useRef(activities);
   useEffect3(() => {
-    activitiesLengthRef.current = activities.length;
     activitiesRef.current = activities;
   }, [activities]);
+  const prevMergedCountRef = useRef(0);
+  const prevPausedCountRef = useRef(0);
   const lastLoadedFileRef = useRef(null);
   useEffect3(() => {
     let node = allFlatRef.current.find((s) => s.id === selectedId);
@@ -4163,13 +4235,8 @@ function App({
     setSessionTree(tree);
     if (!node || !node.filePath) return;
     const newActivities = parseSessionHistory(node.filePath);
-    const delta = newActivities.length - activitiesLengthRef.current;
     setActivities(newActivities);
-    if (!isLive && delta > 0) {
-      setScrollOffset((o) => o + delta);
-      setNewCount((n) => n + delta);
-    }
-  }, [selectedId, isLive, expandedIds, tracking, discoverOptions]);
+  }, [selectedId, expandedIds, tracking, discoverOptions]);
   const refreshRef = useRef(refresh);
   useEffect3(() => {
     refreshRef.current = refresh;
@@ -4234,6 +4301,34 @@ function App({
   const naturalTreeRows = allFlat.length;
   const treeRows = Math.max(1, Math.min(naturalTreeRows, maxTreeRows));
   const viewerRows = Math.max(5, height - 7 - treeRows);
+  useEffect3(() => {
+    const prev = prevMergedCountRef.current;
+    prevMergedCountRef.current = mergedActivities.length;
+    const result = adjustViewerCursorOnNewActivities({
+      prevCursorLine: viewerCursorLine,
+      prevActivityCount: prev,
+      newActivityCount: mergedActivities.length,
+      isLive,
+      viewerRows
+    });
+    if (result.cursorLine !== viewerCursorLine) {
+      setViewerCursorLine(result.cursorLine);
+    }
+    if (result.autoPause) {
+      setIsLive(false);
+      setScrollOffset((o) => o + result.scrollDelta);
+      setNewCount((n) => n + result.scrollDelta);
+    }
+  }, [mergedActivities.length, isLive, viewerRows, viewerCursorLine]);
+  useEffect3(() => {
+    const prev = prevPausedCountRef.current;
+    prevPausedCountRef.current = mergedActivities.length;
+    if (!isLive && mergedActivities.length > prev) {
+      const delta = mergedActivities.length - prev;
+      setScrollOffset((o) => o + delta);
+      setNewCount((n) => n + delta);
+    }
+  }, [mergedActivities.length, isLive]);
   const spinner = useSpinner(isWatchMode, 150);
   const tickActive = isWatchMode && isLive && !helpMode && !detailMode && activities.length > 0;
   const liveTick = useTick(tickActive, 150);

package/dist/templates/summary-prompt.md

@@ -1,8 +1,17 @@
-The following is an activity log from a single day of Claude Code sessions.
-Each entry: `[HH:MM] <icon> <label>: <detail>`. Icons indicate activity type (◆ = git commit).
+The following is an activity log from a single day of Claude Code sessions,
+possibly spanning multiple projects.
 
-Write a concise but high-signal engineering summary in English.
-Aim for roughly 300-500 words total — favor synthesis over completeness.
+Each entry: `[HH:MM] <icon> <label>: <detail>`. Icons:
+  ○ Read           ~ Edit/Write       $ Bash
+  * Glob/Grep      @ WebFetch         » Sub-agent task
+  < Response       > User message     … Thinking
+  ◆ Git commit
+
+Write a concise, high-signal engineering retro in first-person voice
+("I implemented X", "I discovered Y") — treat the developer as the author.
+
+Target ~300-500 words total; individual sections may be 50-100 words.
+Favor synthesis over completeness.
 
 Focus on:
 - meaningful accomplishments
@@ -10,28 +19,38 @@ Focus on:
 - architectural decisions
 - implementation progress
 - important tradeoffs or realizations
-- unfinished work and next steps
+- unfinished work and explicit next steps
+
+Do NOT:
+- mechanically summarize every edit, response, or commit
+- list every modified file (cite paths only for important changes)
+- invent follow-ups not present in the log (no "consider refactoring X")
+- include placeholders like "None" or "N/A" — omit empty sections entirely
 
-Do NOT mechanically summarize every edit, response, or commit.
-Do NOT list every modified file.
-Omit any section that has no substantive content — do not emit placeholders like "None" or "N/A".
-Prefer high-level synthesis grouped by theme or workstream.
+If activity spans multiple projects, group findings by project where clarity
+improves. If they share a theme, synthesize.
 
-Use the following format:
+If the day was light (few entries, no meaningful work), output a single short
+paragraph instead of the full template. Do not pad.
+
+Format (omit any section without substance):
 
 ## Context
-What was the primary goal or investigation today?
+The primary goal(s) of the day. If multi-project, briefly note each.
 
 ## Key Accomplishments
 Meaningful progress, completed implementations, resolved issues.
 
 ## Technical Insights
-Important debugging findings, architectural decisions, tradeoffs, or discoveries.
+Important debugging findings, architectural decisions, tradeoffs, discoveries.
 
 ## Major Code Changes
-Significant codebase or subsystem changes, grouped by theme. Reference relevant commits inline where useful (e.g. `abc1234`). Do not enumerate every commit.
+Significant codebase changes grouped by theme. Reference commits inline
+(e.g., `abc1234`).
 
 ## Open Questions / Next Steps
-Only include items explicitly stated in the log or strongly implied by an unfinished workflow. Do not invent generic follow-ups ("consider refactoring X").
+Only items explicitly stated in the log or directly tied to unfinished workflows.
+
+After reading the log below, write the summary.
 
 Activity log:

package/dist/templates/summary-range-prompt.md

@@ -1,5 +1,18 @@
-The following are daily engineering summaries from a date range, separated by `---` lines.
-Each daily summary uses the structure: Context / Key Accomplishments / Technical Insights / Major Code Changes / Open Questions.
+The input contains daily engineering summaries from a date range.
+
+Each day is wrapped in an XML tag with its date as an attribute:
+
+<day date="YYYY-MM-DD">
+(that day's summary — markdown using sections like
+Context / Key Accomplishments / Technical Insights /
+Major Code Changes / Open Questions)
+</day>
+
+Multiple `<day>` blocks follow in chronological order. Use the
+`date` attribute as the authoritative date for that block — do not
+infer dates from headings or text inside the block. Any `# YYYY-MM-DD`
+heading or `---` line you see inside a `<day>` block is part of the
+day's content, not a structural divider.
 
 Write a range-level synthesis in English.
 Aim for roughly 400-700 words total — the value of a range summary is cross-day pattern extraction, not re-summarization.
@@ -15,6 +28,8 @@ The point of a range summary is to surface what only becomes visible by looking
 - outcomes that matter at the range level, not at any single day
 - work that started in one day and continued, completed, or was abandoned later
 
+When you reference a specific day, use its `date` attribute value verbatim (`2026-06-07`), not a relative term like "Monday" or "yesterday".
+
 Use the following format:
 
 ## Range Overview
@@ -32,4 +47,4 @@ Themes that surfaced on more than one day: repeated debugging classes, recurring
 ## Carried-Over / Unfinished Work
 Work still in progress at the end of the range, or open questions that persisted across multiple days.
 
-Daily summaries:
+Daily summaries follow:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agenthud",
-  "version": "0.12.4",
+  "version": "0.13.1",
   "description": "CLI tool to monitor agent status in real-time. Works with Claude Code, multi-agent workflows, and any AI agent system.",
   "type": "module",
   "main": "dist/index.js",