claude-finops 0.9.0 → 0.10.0

package/README.md

@@ -63,6 +63,13 @@ npm install -g claude-finops    # then `claude-finops` from anywhere
 
 ## What you actually get
 
+**It acts, not just reports.** The overview opens with **⚡ Act now**: one-click
+actions that apply to you right now. Compact a running session that is carrying too
+much context (it types `/compact` into that session's terminal), hand it over to a
+fresh session, turn a shell command you keep re-running into a skill, or move an
+instruction you keep re-typing into `CLAUDE.md`. Every session row has **Resume**
+(copies `claude --resume <id>`), and running ones have **Compact**.
+
 **Where the money goes, per project.** Every project ranked by cost, drilling down
 Project → Session → Prompt.
 
@@ -140,17 +147,21 @@ labelled as your own configured figures.
 
 ## What is in it
 
-**Command center** — Executive overview (spend, tokens, usage %, remaining, forecast),
-an AI FinOps Advisor that answers "what should I do today?" from live data, and a
+**Command center** — Executive overview (spend, tokens, usage %, remaining, forecast,
+today vs yesterday) opening with the ⚡ Act now strip, running sessions you can
+interrupt, compact, hand over or close, an AI FinOps Advisor that answers "what should I do today?" from live data, and a
 0–100 FinOps scorecard with per-dimension reasoning.
 
 **Usage** — Interactive timeline across 10 metrics and 6 time ranges with day
-drill-down; burn rate and limits with a gauge, days-until-limit and projected overage;
+drill-down, plus a weekday × hour peak-hours heatmap in your local time; burn rate and
+limits with a gauge, days-until-limit and projected overage, and your 5-hour and weekly
+plan-limit history from the Claude desktop app;
 per-model FinOps table with superlatives (most expensive, most used, most
 token-efficient, best cost-per-output); context-size distribution and cache
 with-vs-without analysis.
 
-**Drill-down** — Project → Session → Prompt, everywhere. A prompt explorer over every
+**Drill-down** — Project → Session → Prompt, everywhere, with Resume (and Compact for
+running sessions) on every session row. A prompt explorer over every
 prompt with full text, category, tokens, cache split, tool calls, files touched,
 latency and efficiency; five leaderboards; prompt intelligence (spend by activity);
 and a Claude Code view (tools, files, branches, cost per repository).
@@ -212,12 +223,14 @@ depends entirely on how agentic your workload is.
 
 ```
 ~/.claude/projects/**/*.jsonl
+<Claude desktop app>/local-agent-mode-sessions/**/.claude/projects/**/*.jsonl   (Cowork)
         │
         ▼  finops/etl.py     stream-parse, normalize, roll up
    ~/.claude-finops/data/finops.db   SQLite warehouse
         │
         ▼  finops/analytics.py   KPIs · burn · forecast · waste · anomalies · scorecard
    finops/api.py             stdlib HTTP: JSON API + static files (127.0.0.1 only)
+                             (+ <Claude desktop app>/plan-usage-history.json, read on request)
         │
         ▼  web/              vanilla JS, inline-SVG charts, no build step
 ```
@@ -262,6 +275,8 @@ claude-finops --stop          stop it
 claude-finops --where         where your data, settings and keys live
 claude-finops --set-key       store a provider API key (hidden prompt, 0600)
 claude-finops --keys          which provider keys are configured
+claude-finops --install-statusline    model, context % and 5-hour limit under every prompt
+claude-finops --uninstall-statusline  remove it again
 claude-finops --help          everything
 ```
 
@@ -293,6 +308,12 @@ place for you to delete once you are happy.
 
 ## Privacy
 
+**What it reads, all read-only:** Claude Code transcripts in `~/.claude/projects`, and,
+if the Claude desktop app is installed, its Cowork transcripts and
+`plan-usage-history.json` (`~/Library/Application Support/Claude` on macOS,
+`%APPDATA%\Claude` on Windows, `~/.config/Claude` on Linux), plus the other coding
+agents listed under Multi-agent.
+
 `~/.claude-finops/data/finops.db` and the prompt/CSV exports contain **your full prompt text**. The
 server binds to `127.0.0.1` only, but treat the database and any export you
 generate as sensitive.
@@ -310,7 +331,8 @@ configure a key with `--set-key`.
 
 The app has nothing tied to one person. Whoever runs it sees **their own** Claude usage:
 
-- It reads `~/.claude/projects` on the machine it runs on (override with `CLAUDE_PROJECTS=/path`).
+- It reads `~/.claude/projects` on the machine it runs on (override with `CLAUDE_PROJECTS=/path`),
+  plus the Claude desktop app's Cowork sessions when the default location is used.
 - The account label comes from that machine's `~/.claude.json`.
 - Budgets and limits you set in the UI are saved to `~/.claude-finops/settings.local.json`. The committed `config/settings.json` holds only shared defaults.
 
@@ -344,6 +366,19 @@ files on Windows.
   `config/free_models.json`.
 - **Skills & MCP** suggests MCP servers and skills from work you repeat, with the evidence.
   "Add" runs `claude mcp add -s user …`; "Create skill" writes `~/.claude/skills/<name>/SKILL.md`.
+  The strongest skill suggestions also appear in **⚡ Act now** on the overview.
+- **Compact** (Running sessions, Act now, session rows) types `/compact` into that session's
+  terminal: tmux, Terminal.app, iTerm2 or the Windows console. Anywhere else it copies the
+  command for you to paste. It asks you to click twice.
+- **Hand over / split** (Running sessions) writes a brief of the session to
+  `~/.claude/finops-handoffs/` and opens fresh session(s) in new terminals, one per sub-task if
+  you list them; it can close the old session (resumable).
+- **Interrupt / Close / Force kill** (Running sessions) send the session's process a signal.
+  Each asks you to click twice. Interrupt is macOS/Linux only: on Windows there is no way to
+  interrupt one console process without hitting the others in its window.
+- **Install statusline** (Act now, or `--install-statusline`) adds a `statusLine` entry to
+  `~/.claude/settings.json`, keeping a `.finops-backup` copy. It won't replace a statusline you
+  already have.
 
 ---
 

package/finops/act_now.py

@@ -0,0 +1,92 @@
+"""What to do right now: the few one-click actions that apply to this machine today.
+
+Everything here already exists deeper in the dashboard (Running sessions, Skills &
+MCP, Why so many tokens?). This picks the high-signal items only, so the overview
+can act instead of just reporting:
+
+  session     a running session carrying heavy context  -> Compact / Hand over
+  skill       a shell command you keep re-running       -> Create skill
+  memory      an instruction you keep re-typing          -> Add to CLAUDE.md
+  statusline  no statusline configured                   -> Install
+
+Prompt-based skill suggestions and one-off memory lines are left to their own pages:
+they are noisy enough that a one-click card would mostly be wrong.
+"""
+
+MAX_SESSIONS = 3
+MAX_SKILLS = 2
+MAX_MEMORY = 2
+
+
+def _k(n):
+    return f"{n / 1000:.0f}K" if n < 1_000_000 else f"{n / 1_000_000:.1f}M"
+
+
+def build(live, skills, memory, statusline):
+    """live: procs.list_sessions rows; skills: actions.suggestions()["skills"];
+    memory: Diagnoser.memory_suggestions(); statusline: integrate.statusline_state()."""
+    items = []
+    heavy = sorted((x for x in live if x.get("severity") in ("medium", "high") and x.get("context")),
+                   key=lambda x: -x["context"])[:MAX_SESSIONS]
+    for x in heavy:
+        items.append({
+            "kind": "session", "severity": x["severity"], "pid": x.get("pid"),
+            "session_id": x.get("session_id"), "hosts_dashboard": bool(x.get("hosts_dashboard")),
+            "title": f"{x.get('name') or 'A session'} is carrying {_k(x['context'])} tokens of context",
+            "detail": f"{x.get('project') or ''} · every message re-reads all of it. "
+                      + ("Hand it over to a fresh session." if x["severity"] == "high"
+                         else "Compact it at the next break."),
+        })
+    cmds = sorted((s for s in skills if s.get("kind") == "command" and not s.get("installed")),
+                  key=lambda s: -s.get("runs", 0))[:MAX_SKILLS]
+    for s in cmds:
+        items.append({
+            "kind": "skill", "skill": s,
+            "title": f"You ran `{s['trigger']}` {s['runs']} times in {s['sessions']} sessions",
+            "detail": "Make it a skill so Claude runs it the way this project expects, "
+                      "without re-explaining each time.",
+        })
+    themes = sorted((m for m in memory if m.get("kind") == "theme" and not m.get("already_saved")),
+                    key=lambda m: -m.get("sessions", 0))[:MAX_MEMORY]
+    for m in themes:
+        items.append({
+            "kind": "memory",
+            "title": f"You keep re-typing instructions about: {m['text']}",
+            "detail": f"Seen in {m['sessions']} sessions. Put it in {m['target']} once instead.",
+        })
+    if statusline is None:
+        items.append({
+            "kind": "statusline",
+            "title": "See model, context % and 5-hour limit under every Claude Code prompt",
+            "detail": "Adds a statusline to ~/.claude/settings.json (a backup is kept). "
+                      "Undo any time with claude-finops --uninstall-statusline.",
+        })
+    return {"items": items}
+
+
+def act_now(a):
+    """Gather the live inputs and build the strip. Each source fails soft."""
+    from .procs import list_sessions
+    from .actions import suggestions
+    from .diagnose import Diagnoser
+    from .integrate import statusline_state
+
+    def safe(fn, default):
+        try:
+            return fn()
+        except Exception:
+            return default
+
+    live = safe(lambda: list_sessions(a.pricing), [])
+    # the process registry names sessions after their folder; the transcript title reads better
+    ids = [x["session_id"] for x in live if x.get("session_id")]
+    titles = safe(lambda: {r["id"]: r["title"] for r in a.q(
+        "SELECT id, title FROM sessions WHERE title IS NOT NULL AND id IN (%s)" % ",".join("?" * len(ids)),
+        ids)} if ids else {}, {})
+    for x in live:
+        if titles.get(x.get("session_id")):
+            x["name"] = titles[x["session_id"]]
+    return build(live,
+                 safe(lambda: suggestions(a)["skills"], []),
+                 safe(lambda: Diagnoser(a).memory_suggestions({}), []),
+                 safe(statusline_state, "other"))

package/finops/analytics.py

@@ -168,6 +168,16 @@ def _cumsum(values):
         yield total
 
 
+def resume_command(sid, agent, source_file):
+    """`claude --resume <id>` for a Claude Code session; None where it would not work.
+
+    Cowork transcripts live under the desktop app's own config dir, where
+    `claude --resume` would not find them.
+    """
+    cowork = "local-agent-mode-sessions" in (source_file or "")
+    return f"claude --resume {sid}" if (agent or "claude") == "claude" and not cowork else None
+
+
 class Analytics:
     def __init__(self, db_path=DB_PATH):
         # One connection per thread. The HTTP server is threaded, and a single sqlite
@@ -561,6 +571,7 @@ class Analytics:
         rows = self.q(f"""
           SELECT s.id session_id, s.title, s.git_branch, s.cli_version, s.started_at,
                  s.ended_at, s.duration_s, s.files_touched, pr.name project, pr.id project_id,
+                 s.agent, s.source_file,
                  COUNT(DISTINCT r.prompt_id) prompts, COUNT(*) requests,
                  SUM(r.tool_call_count) tool_calls,
                  SUM(r.input_tokens) input_tokens, SUM(r.output_tokens) output_tokens,
@@ -579,6 +590,7 @@ class Analytics:
             r["output_ratio"] = (r["output_tokens"] or 0) / r["tokens"] if r["tokens"] else 0
             cr, cw = r["cache_read_tokens"] or 0, r["cache_write_tokens"] or 0
             r["cache_hit_ratio"] = (cr / (cr + cw)) if (cr + cw) else None
+            r["resume"] = resume_command(r["session_id"], r.pop("agent"), r.pop("source_file"))
         return rows
 
     def sessions_total(self, f=None):
@@ -699,11 +711,7 @@ class Analytics:
         s["files"] = self.q(
             "SELECT path, GROUP_CONCAT(DISTINCT op) ops, COUNT(*) n FROM files_touched"
             " WHERE session_id=? GROUP BY path ORDER BY n DESC LIMIT 100", (sid,))
-        # Cowork transcripts live under the desktop app's own config dir, where
-        # `claude --resume` would not find them.
-        cowork = "local-agent-mode-sessions" in (s.get("source_file") or "")
-        s["resume"] = (f"claude --resume {sid}"
-                       if (s.get("agent") or "claude") == "claude" and not cowork else None)
+        s["resume"] = resume_command(sid, s.get("agent"), s.get("source_file"))
         return s
 
     # ---------------- categories ----------------
@@ -725,7 +733,7 @@ class Analytics:
 
     # ---------------- leaderboards ----------------
     def leaderboards(self, f=None, n=20):
-        return {
+        out = {
             "most_expensive": self.prompts(f, limit=n, order="cost"),
             "most_token_heavy": self.prompts(f, limit=n, order="tokens"),
             "cheapest": self.prompts(f, limit=n, order="cheapest"),
@@ -733,6 +741,17 @@ class Analytics:
             "longest_sessions": self.sessions(f, limit=n, order="duration"),
             "basis": "estimated",
         }
+        prompt_rows = [r for k in ("most_expensive", "most_token_heavy", "cheapest", "most_efficient")
+                       for r in out[k]]
+        sids = sorted({r["session_id"] for r in prompt_rows if r.get("session_id")})
+        src = {r["id"]: r for r in self.q(
+            "SELECT id, agent, source_file FROM sessions WHERE id IN (%s)" % ",".join("?" * len(sids)),
+            sids)} if sids else {}
+        for r in prompt_rows:
+            s = src.get(r.get("session_id")) or {}
+            r["resume"] = (resume_command(r["session_id"], s.get("agent"), s.get("source_file"))
+                           if s else None)
+        return out
 
     # ---------------- efficiency ----------------
     def hygiene(self, f=None, top=12, trajectory_points=80):

package/finops/api.py

@@ -270,6 +270,9 @@ class Handler(BaseHTTPRequestHandler):
             return self.send_json({"job": j["id"]})
         if parts == ["skill"]:
             return self.send_json(X.create_skill(payload))
+        if parts == ["statusline"]:
+            from .integrate import install_statusline
+            return self.send_json(install_statusline())
         if len(parts) == 2 and parts[0] == "mcp":
             return self.send_json(X.add_mcp(parts[1], payload))
         return self.send_json({"error": "unknown action"}, 404)
@@ -390,6 +393,9 @@ class Handler(BaseHTTPRequestHandler):
             return self.send_json(a.timeline(f, g("grain", "day")))
         if route == "heatmap":
             return self.send_json(a.heatmap(f))
+        if route == "act_now":
+            from .act_now import act_now
+            return self.send_json(act_now(a))
         if route == "models":
             return self.send_json(a.models(f))
         if route == "projects":

package/finops/integrate.py

@@ -12,6 +12,7 @@ fail silent, never block, never take long.
 """
 import json
 import os
+import shlex
 import shutil
 import sys
 
@@ -59,14 +60,17 @@ def hook():
 # -------------------------------------------------------------- statusline ----
 
 def statusline():
-    """One line, refreshed constantly: model and context usage."""
+    """One line, refreshed constantly: model, context usage and the 5-hour limit."""
     try:
         d = _stdin_json()
-        pct = _dig(d, "context.percentUsed", "context.percent_used")
+        pct = _dig(d, "context_window.used_percentage", "context.percentUsed", "context.percent_used")
+        five = _dig(d, "rate_limits.five_hour.used_percentage")
         name = _dig(d, "model.display_name", "session.model", "model") or "claude"
         bits = [str(name)]
         if isinstance(pct, (int, float)):
             bits.append(f"{pct:.0f}% ctx")
+        if isinstance(five, (int, float)):
+            bits.append(f"{five:.0f}% 5h")
         print(" · ".join(bits))
     except Exception:
         print("")     # an empty statusline beats a stack trace under the prompt
@@ -81,7 +85,8 @@ def _command():
     if exe:
         return exe
     root = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
-    return f"{sys.executable} {os.path.join(root, 'run.py')}"
+    # settings run this through a shell: an install path with a space would split
+    return f"{shlex.quote(sys.executable)} {shlex.quote(os.path.join(root, 'run.py'))}"
 
 
 def _load_settings():
@@ -113,21 +118,31 @@ def install_hook(remove=False):
     print("The statusline (claude-finops --install-statusline) still shows model and context %.")
 
 
+def statusline_state():
+    """"ours" when our statusline is installed, "other" for someone else's, else None."""
+    cur = _load_settings().get("statusLine")
+    if not cur:
+        return None
+    return "ours" if "finops" in json.dumps(cur) else "other"
+
+
 def install_statusline(remove=False):
-    cmd = f"{_command()} --statusline"
+    """Wire the statusline into ~/.claude/settings.json. Returns {"ok", "message"}.
+
+    Claude Code wants {"type": "command", "command": ...}; an existing statusline that
+    isn't ours is left alone.
+    """
     s = _load_settings()
     if remove:
-        if "finops" in str(s.get("statusLine", "")):
+        if statusline_state() == "ours":
             s.pop("statusLine", None)
-    else:
-        prev = s.get("statusLine")
-        if prev and "finops" not in str(prev):
-            print(f"You already have a statusLine configured:\n  {prev}")
-            print("Leaving it alone. Remove it first if you want ours.")
-            return
-        s["statusLine"] = cmd
+            _save_settings(s)
+        return {"ok": True, "message": f"Removed the statusline from {SETTINGS}"}
+    if statusline_state() == "other":
+        return {"ok": False, "message": "You already have a statusline configured "
+                f"({json.dumps(s['statusLine'])}). Left it alone: remove it first if you want ours."}
+    s["statusLine"] = {"type": "command", "command": f"{_command()} --statusline"}
     _save_settings(s)
-    print(("Removed" if remove else "Installed") + f" the statusline in {SETTINGS}")
-    if not remove:
-        print("  Shows the model and context usage percentage.")
-        print("  Undo: claude-finops --uninstall-statusline")
+    return {"ok": True, "message": f"Installed the statusline in {SETTINGS}. It shows the model, "
+                                   "context % and 5-hour limit % under every prompt. "
+                                   "Undo: claude-finops --uninstall-statusline"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-finops",
-  "version": "0.9.0",
+  "version": "0.10.0",
   "description": "Local FinOps dashboard for Claude Code: what you used, what it cost, why it cost that much, and what to change. Reads your own transcripts, no API key, no data leaves the machine.",
   "bin": {
     "claude-finops": "bin/claude-finops.js"

package/run.py

@@ -153,7 +153,7 @@ HELP = """Claude FinOps Command Center
   claude-finops --version       print the installed version, and whether a newer one is out
   claude-finops --no-update-check   skip the once-a-day npm version check
   claude-finops --install-hook  retired no-op (prints a message, changes nothing)
-  claude-finops --install-statusline   show model and context % in your statusline
+  claude-finops --install-statusline   show model, context % and 5-hour limit in your statusline
   claude-finops --help          this message
 
 Environment:
@@ -296,7 +296,9 @@ def main():
         return install_hook(remove="--uninstall-hook" in args)
     if "--install-statusline" in args or "--uninstall-statusline" in args:
         from finops.integrate import install_statusline
-        return install_statusline(remove="--uninstall-statusline" in args)
+        r = install_statusline(remove="--uninstall-statusline" in args)
+        print(r["message"])
+        return sys.exit(0 if r["ok"] else 1)
     if "--where" in args:
         return where()
     if "--keys" in args:

package/web/app.js

@@ -685,10 +685,110 @@ VIEWS.overview = async (page) => {
   api('diagnose').then(d => {
     if (!page.isConnected) return;
     const el = h(focusStrip(d) || '<div></div>');
-    page.insertBefore(el, page.firstChild);
+    page.insertBefore(el, page.querySelector('.actnow')?.nextSibling || page.firstChild);
+    page._focusData = d;
     wireFocus(page);
   }).catch(() => {});
+  if (hasClaude()) fetch('/api/act_now').then(r => r.json()).then(d => {
+    if (!page.isConnected || !(d.items || []).length) return;
+    const el = h(actNowStrip(d.items));
+    page.insertBefore(el, page.firstChild);
+    wireActNow(el, d.items);
+  }).catch(() => {});
+};
+
+/* ---------- act now: one-click actions, on the page everyone opens first ---------- */
+const ACT_ICON = {session: '🗜', skill: '⚙', memory: '📝', statusline: '▭'};
+const codeTicks = t => esc(t).replace(/`([^`]+)`/g, '<code>$1</code>');
+function actNowStrip(items) {
+  const btn = it => it.kind === 'session'
+    ? '<button class="act" data-do="compact" title="Types /compact into that session\'s terminal">🗜 Compact</button>'
+      + '<button class="act ghost" data-do="handover">⇢ Hand over</button>'
+    : it.kind === 'skill' ? '<button class="act" data-do="skill">＋ Create skill</button>'
+    : it.kind === 'memory' ? '<button class="act ghost" data-do="memory">Show examples</button>'
+    : '<button class="act" data-do="statusline">Install statusline</button>';
+  return `<section class="actnow"><div class="focus-hd">⚡ Act now
+      <span class="note">one click each · nothing changes until you click</span></div>
+    ${items.map((it, i) => `<div class="actnow-it" data-i="${i}">
+      <span class="ai">${ACT_ICON[it.kind] || '•'}</span>
+      <div class="ft"><b>${codeTicks(it.title)}</b><div class="note">${esc(it.detail)}</div></div>
+      <div class="live-actions">${btn(it)}<span class="live-msg"></span></div></div>`).join('')}</section>`;
+}
+function wireActNow(root, items) {
+  root.querySelectorAll('.actnow-it').forEach(row => {
+    const it = items[+row.dataset.i], msg = row.querySelector('.live-msg');
+    const say = (t, ok) => { msg.textContent = t; msg.className = 'live-msg ' + (ok ? 'ok' : 'err'); };
+    row.querySelectorAll('[data-do]').forEach(b => b.onclick = async () => {
+      const what = b.dataset.do;
+      if (what === 'handover') return go('live');
+      if (what === 'memory') return jumpTo('diagnose', 'dx-mem');
+      if (what === 'compact' && !armed(b)) return;
+      b.disabled = true;
+      try {
+        if (what === 'compact') {
+          const r = await sessionAction(it.pid, 'compact', {agent: 'claude'});
+          if (!r.ok && r.copy) { try { await navigator.clipboard.writeText(r.copy); } catch {} }
+          say(r.ok ? r.message : `Failed: ${r.error}`, r.ok);
+          b.disabled = false; return;
+        }
+        const r = what === 'skill' ? await doAction('skill', it.skill) : await doAction('statusline');
+        say(what === 'skill' ? `Created ${r.path}. Use it with ${r.use}.` : r.message, r.ok !== false);
+        if (r.ok !== false) b.remove(); else b.disabled = false;
+      } catch (e) { say('Failed: ' + e.message, false); b.disabled = false; }
+    });
+  });
+}
+// Two-click confirm, inline (no browser dialogs): the first click arms, the second runs.
+function armed(b) {
+  if (b.classList.contains('armed')) { b.classList.remove('armed'); b.textContent = b.dataset.label; return true; }
+  b.dataset.label = b.textContent; b.classList.add('armed'); b.textContent = 'Click again to confirm';
+  setTimeout(() => { if (b.classList.contains('armed')) { b.classList.remove('armed'); b.textContent = b.dataset.label; } }, 4000);
+  return false;
+}
+// Open a view and scroll to an element that appears once the (possibly slow) view renders.
+async function jumpTo(view, anchor) {
+  if (S.view !== view) go(view);
+  for (let i = 0; i < 60; i++) {
+    const el = document.getElementById(anchor);
+    if (el) { el.scrollIntoView({behavior: 'smooth', block: 'start'}); el.classList.add('flash');
+      setTimeout(() => el.classList.remove('flash'), 1600); return; }
+    await new Promise(r => setTimeout(r, 250));
+  }
+}
+
+/* ---------- row actions: Resume (copy) and, for a running session, Compact ---------- */
+async function liveSessions() {
+  try {
+    const d = await fetch('/api/live?agents=claude').then(r => r.json());
+    return new Map((d.sessions || []).filter(x => x.signalable && x.pid).map(x => [x.session_id, x.pid]));
+  } catch { return new Map(); }
+}
+const rowActs = (r, live) => {
+  const pid = live && live.get(r.session_id);
+  return `<span class="row-acts">${r.resume ? `<button class="act ghost" data-resume="${esc(r.resume)}"
+    title="Copy: ${esc(r.resume)}">⧉ Resume</button>` : ''}${pid ? ` <button class="act" data-compact="${pid}"
+    title="Running now: types /compact into its terminal">🗜 Compact</button>` : ''}</span>`;
 };
+function wireRowActs(host) {
+  if (!host) return;
+  host.querySelectorAll('[data-resume]').forEach(b => b.onclick = async e => {
+    e.stopPropagation();
+    try { await navigator.clipboard.writeText(b.dataset.resume); b.textContent = '✓ Copied'; }
+    catch { b.textContent = b.dataset.resume; }
+    setTimeout(() => { b.textContent = '⧉ Resume'; }, 1800);
+  });
+  host.querySelectorAll('[data-compact]').forEach(b => b.onclick = async e => {
+    e.stopPropagation();
+    if (!armed(b)) return;
+    b.disabled = true;
+    try {
+      const r = await sessionAction(+b.dataset.compact, 'compact', {agent: 'claude'});
+      if (!r.ok && r.copy) { try { await navigator.clipboard.writeText(r.copy); } catch {} }
+      b.textContent = r.ok ? '✓ Sent' : 'Failed'; b.title = r.ok ? r.message : r.error;
+    } catch (err) { b.textContent = 'Failed'; b.title = err.message; }
+    setTimeout(() => { b.disabled = false; b.textContent = '🗜 Compact'; }, 2500);
+  });
+}
 
 function renderAdvisorHero(page, advisor) {
   if (!page || !advisor || !Array.isArray(advisor.actions)) return;
@@ -1076,7 +1176,7 @@ VIEWS.sessions = async (page) => {
   const sp = S.sessPage;
   const res = await api('sessions', `&limit=${sp.offset + sp.limit}&order=${order}`);
   const rows = res.rows, total = res.total;
-  const eff = await api('efficiency');
+  const [eff, live] = await Promise.all([api('efficiency'), liveSessions()]);
   const avgTok = rows.length ? rows.reduce((a, r) => a + r.tokens, 0) / rows.length : 0;
   page.innerHTML = `
     <div class="grid g4">
@@ -1121,6 +1221,7 @@ VIEWS.sessions = async (page) => {
     {h: 'Session', trunc: 1, title: r => r.session_id,
      f: r => `${r.tokens > avgTok * 3 ? '🔴 ' : ''}${esc(r.title || shortId(r.session_id))}
        <div class="sub mono">${esc(shortId(r.session_id))}</div>`},
+    {h: '', f: r => rowActs(r, live)},
     {h: 'Project', f: r => esc(r.project)},
     {h: 'Branch', f: r => r.git_branch ? `<span class="mono">${esc(r.git_branch)}</span>` : '—'},
     {h: 'Start', f: r => `<span class="mono">${esc((r.started_at || '').slice(0, 16).replace('T', ' '))}</span>`},
@@ -1141,6 +1242,7 @@ VIEWS.sessions = async (page) => {
     {h: 'Est. cost', num: 1, f: r => fmtUSD(r.cost)},
   ], rows, {onRow: 1});
   wireTable($('#st', page), rows, r => openSession(r.session_id));
+  wireRowActs($('#st', page));
   const effCards = [...page.querySelectorAll('.card')].filter(c =>
     /output yield/.test(c.querySelector('h3')?.textContent || ''));
   wireTable(effCards[0], eff.low_efficiency_sessions, r => openSession(r.session_id));
@@ -1219,7 +1321,7 @@ const RANK_TABS = [
   ['cheapest', 'Top 20 cheapest'], ['most_efficient', 'Top 20 most efficient'],
   ['longest_sessions', 'Top 20 longest sessions']];
 VIEWS.rankings = async (page) => {
-  const lb = await api('leaderboards', '&n=20');
+  const [lb, live] = await Promise.all([api('leaderboards', '&n=20'), liveSessions()]);
   const tab = S.rankTab || 'most_expensive';
   const isSess = tab === 'longest_sessions';
   const rows = lb[tab];
@@ -1233,6 +1335,7 @@ VIEWS.rankings = async (page) => {
   const cols = isSess ? [
     {h: '#', num: 1, f: r => rows.indexOf(r) + 1},
     {h: 'Session', trunc: 1, f: r => esc(r.title || shortId(r.session_id))},
+    {h: '', f: r => rowActs(r, live)},
     {h: 'Project', f: r => esc(r.project)},
     {h: 'Duration', num: 1, f: r => dur(r.duration_s)},
     {h: 'Prompts', num: 1, f: r => fmtInt(r.prompts)},
@@ -1251,10 +1354,12 @@ VIEWS.rankings = async (page) => {
     {h: 'Tok eff', num: 1, f: r => fmtPct(r.efficiency * 100, 2)},
     {h: 'Est. cost', num: 1, f: r => fmtUSD(r.pcost)},
     {h: 'Session', trunc: 1, f: r => esc(r.session_title || shortId(r.session_id))},
+    {h: '', f: r => rowActs(r, live)},
     {h: 'Date', f: r => esc(r.day)},
   ];
   $('#rt', page).innerHTML = table(cols, rows, {onRow: 1});
   wireTable($('#rt', page), rows, r => isSess ? openSession(r.session_id) : openPrompt(r.prompt_id));
+  wireRowActs($('#rt', page));
   addChart(page, 'Top 12 prompts by estimated cost', el => C.barsH(el, {
     rows: (lb.most_expensive || []).slice(0, 12), label: r => clip(r.preview, 46), value: r => r.pcost,
     sub: r => `<div class="row"><span class="k">Project</span><span class="v">${esc(r.project)}</span></div>
@@ -2297,45 +2402,81 @@ VIEWS.exports = async (page) => {
 /* ============================ focus cues ============================ */
 // "What needs me first": one ranked list built from the diagnosis, used for the nav
 // badges and the Focus strip at the top of the overview and diagnosis pages.
+// "What to change": things to fix once and habits to change, each with its reason.
+// Running sessions are left to ⚡ Act now, which can act on them. Items are keyed so a
+// dismissed one comes back only if its evidence changes (e.g. more sessions affected).
+const firstSentence = t => { const m = String(t || '').match(/^.*?[.!?](\s|$)/); return (m ? m[0] : String(t || '')).trim(); };
+// "29.7% of X." + "35.6% of X." -> "29.7–35.6% of X." plus the shared advice; null when they differ.
+function sameFix(fixes) {
+  const pct = /\d+(?:\.\d+)?%/, heads = fixes.map(firstSentence);
+  const shape = heads.map(t => t.replace(pct, '#'));
+  if (!shape.every(t => t === shape[0]) || !pct.test(heads[0])) return null;
+  const nums = heads.map(t => parseFloat(t.match(pct)[0]));
+  const range = `${Math.min(...nums)}–${Math.max(...nums)}%`;
+  const rest = fixes[0].slice(heads[0].length).trim().replace(/\bthis repo\b/g, 'each repo');
+  return `${shape[0].replace('#', range).replace(' here ', ' in these repos ')} ${rest}`.trim();
+}
 function focusItems(d) {
   const out = [];
-  for (const x of d.live_sessions || []) if (x.severity !== 'ok')
-    out.push({lvl: x.severity === 'high' ? 1 : 2, text: `Running session at ${fmtNum(x.context)} context: ${x.title || shortId(x.session_id)}`, view: 'diagnose', anchor: 'dx-live'});
   for (const m of d.memory_suggestions || []) if (m.kind === 'security')
-    out.push({lvl: 1, text: `Credentials pasted in prompts (${m.sessions} sessions): rotate them`, view: 'diagnose', anchor: 'dx-mem'});
+    out.push({lvl: 1, kind: 'fix', key: `sec:${m.sessions}`, view: 'diagnose', anchor: 'dx-mem',
+      text: `Rotate the credentials you pasted into prompts (${m.sessions} session${m.sessions === 1 ? '' : 's'})`,
+      why: 'Closing the sessions does not help: the secrets stay in plain text in those transcripts on disk.'});
+  const byIssue = (sev, lvl) => {
+    const groups = new Map();
+    for (const p of d.projects || []) for (const i of p.issues) if (i.severity === sev) {
+      if (!groups.has(i.title)) groups.set(i.title, []);
+      groups.get(i.title).push({project: p.name, fix: i.fix});
+    }
+    for (const [title, ps] of groups) out.push({lvl, kind: 'fix', view: 'diagnose', anchor: 'dx-issues',
+      key: `proj:${title}:${ps.map(x => x.project).sort().join(',')}`,
+      text: ps.length === 1 ? `${ps[0].project}: ${title}` : `${title} in ${ps.length} projects: ${ps.map(x => x.project).join(', ')}`,
+      why: ps.length === 1 ? ps[0].fix : sameFix(ps.map(x => x.fix))
+        || ps.slice(0, 3).map(x => `${x.project}: ${firstSentence(x.fix)}`).join(' · ')});
+  };
+  byIssue('high', 1);
   for (const r of d.recommendations || []) if (r.priority === 1)
-    out.push({lvl: 1, text: r.title, view: 'diagnose', anchor: 'dx-recs'});
-  for (const p of d.projects || []) for (const i of p.issues) if (i.severity === 'high')
-    out.push({lvl: 1, text: `${p.name}: ${i.title}`, view: 'diagnose', anchor: 'dx-issues'});
-  const s0 = (d.session_health || [])[0];
-  if (s0) out.push({lvl: 2, text: `Heaviest session: ${s0.title || shortId(s0.session_id)} (peak ${fmtNum(s0.peak)})`, view: 'diagnose', anchor: 'dx-past'});
-  for (const p of d.projects || []) for (const i of p.issues) if (i.severity === 'medium')
-    out.push({lvl: 2, text: `${p.name}: ${i.title}`, view: 'diagnose', anchor: 'dx-issues'});
+    out.push({lvl: 2, kind: 'habit', key: `rec:${r.title}`, view: 'diagnose', anchor: 'dx-recs',
+      text: r.title, why: `${firstSentence(r.why)} ${firstSentence(r.how)}`});
+  byIssue('medium', 3);
   return out.sort((a, b) => a.lvl - b.lvl);
 }
+const FOCUS_KEY = 'finops.focus.hidden';
+const focusHidden = () => { try { return new Set(JSON.parse(localStorage.getItem(FOCUS_KEY) || '[]')); } catch { return new Set(); } };
+const setFocusHidden = set => { try { localStorage.setItem(FOCUS_KEY, JSON.stringify([...set])); } catch {} };
 function focusStrip(d, n = 5) {
-  const items = focusItems(d);
-  if (!items.length) return '';
-  return `<section class="focus"><div class="focus-hd">🎯 Focus on these first
-      <span class="note">${items.filter(i => i.lvl === 1).length} urgent · ${items.filter(i => i.lvl === 2).length} next</span></div>
-    ${items.slice(0, n).map((it, k) => `<a class="focus-it l${it.lvl}" data-go="${it.view}" data-anchor="${it.anchor}">
-      <span class="fn">${k + 1}</span><span class="ft">${esc(it.text)}</span>
-      <span class="fl">${it.lvl === 1 ? 'Fix first' : 'Next'}</span></a>`).join('')}</section>`;
+  const all = focusItems(d), hidden = focusHidden();
+  const items = all.filter(i => !hidden.has(i.key));
+  const nHidden = all.length - items.length;
+  if (!all.length) return '';
+  const fixes = items.filter(i => i.kind === 'fix').length, habits = items.filter(i => i.kind === 'habit').length;
+  return `<section class="focus"><div class="focus-hd">🎯 What to change
+      <span class="note">${fixes} to fix once · ${habits} habit${habits === 1 ? '' : 's'} from your past sessions</span>
+      <span class="spacer"></span>${nHidden ? `<button class="act ghost" data-unhide="1">Show ${nHidden} hidden</button>` : ''}</div>
+    ${items.slice(0, n).map((it, k) => `<div class="focus-it l${it.lvl === 1 ? 1 : 2}" data-go="${it.view}" data-anchor="${it.anchor}" data-key="${esc(it.key)}">
+      <span class="fn">${k + 1}</span>
+      <span class="ft"><span class="fx">${esc(it.text)}</span><span class="fw">${esc(it.why || '')}</span></span>
+      <span class="fl">${it.kind === 'habit' ? 'Habit' : 'Fix once'}</span>
+      <button class="fd" title="Hide this until its evidence changes" aria-label="Hide">✕</button></div>`).join('')
+      || '<div class="note">Everything here is hidden. Nice.</div>'}</section>`;
 }
 function wireFocus(root) {
-  root.querySelectorAll('.focus-it').forEach(a => a.onclick = async () => {
-    if (S.view !== a.dataset.go) { go(a.dataset.go); await new Promise(r => setTimeout(r, 400)); }
-    const el = document.getElementById(a.dataset.anchor);
-    if (el) { el.scrollIntoView({behavior: 'smooth', block: 'start'}); el.classList.add('flash');
-      setTimeout(() => el.classList.remove('flash'), 1600); }
+  const redraw = () => { const old = root.querySelector('.focus'); if (!old || !root._focusData) return;
+    const el = h(focusStrip(root._focusData) || '<div></div>'); old.replaceWith(el); wireFocus(root); };
+  root.querySelectorAll('.focus-it').forEach(a => {
+    a.onclick = () => jumpTo(a.dataset.go, a.dataset.anchor);
+    a.querySelector('.fd').onclick = e => { e.stopPropagation();
+      const hid = focusHidden(); hid.add(a.dataset.key); setFocusHidden(hid); redraw(); };
   });
+  const un = root.querySelector('[data-unhide]');
+  if (un) un.onclick = () => { setFocusHidden(new Set()); redraw(); };
 }
 async function updateNavBadges() {
   const set = (id, n, lvl) => { const el = document.querySelector(`[data-nb="${id}"]`);
     if (el) { el.textContent = n ? n : ''; el.className = 'nb' + (n ? ' l' + lvl : ''); } };
   try {
     const [d, w, a] = await Promise.all([hasPriced() ? api('diagnose') : null, api('waste'), api('anomalies')]);
-    if (d) { const f = focusItems(d), urgent = f.filter(i => i.lvl === 1).length;
+    if (d) { const hid = focusHidden(), f = focusItems(d).filter(i => !hid.has(i.key)), urgent = f.filter(i => i.lvl === 1).length;
       set('diagnose', urgent || f.length, urgent ? 1 : 2); }
     const hw = w.findings.filter(x => x.severity === 'high').length;
     set('waste', hw || w.findings.length, hw ? 1 : 2);

package/web/styles.css

@@ -331,6 +331,17 @@ table.tbl .sub { font-size:10.5px; color:var(--muted); }
 .nav a .nb.l1 { background: var(--critical); color: #fff; }
 .nav a .nb.l2 { background: var(--warning); color: #3a2a00; }
 
+.actnow { border: 1px solid color-mix(in srgb, var(--s1) 40%, var(--border));
+  background: color-mix(in srgb, var(--s1) 5%, var(--surface)); border-radius: var(--radius);
+  padding: 10px 12px; margin-bottom: 14px; display: flex; flex-direction: column; gap: 6px; }
+.actnow-it { display: flex; align-items: center; gap: 10px; flex-wrap: wrap; padding: 7px 8px;
+  border: 1px solid var(--border); border-radius: 8px; background: var(--surface); }
+.actnow-it .ai { width: 22px; text-align: center; font-size: 15px; }
+.actnow-it .ft { flex: 1; min-width: 220px; }
+.actnow-it .live-actions { margin-top: 0; }
+.row-acts { white-space: nowrap; }
+.row-acts .act { padding: 2px 8px; font-size: 11px; }
+
 .focus { border: 1px solid color-mix(in srgb, var(--critical) 35%, var(--border));
   background: color-mix(in srgb, var(--critical) 5%, var(--surface)); border-radius: var(--radius);
   padding: 10px 12px; margin-bottom: 14px; display: flex; flex-direction: column; gap: 5px; }
@@ -341,7 +352,12 @@ table.tbl .sub { font-size:10.5px; color:var(--muted); }
 .focus-it .fn { width: 20px; height: 20px; border-radius: 50%; display: grid; place-items: center;
   font-size: 11px; font-weight: 700; color: #fff; flex: none; }
 .focus-it.l1 .fn { background: var(--critical); } .focus-it.l2 .fn { background: var(--warning); }
-.focus-it .ft { flex: 1; min-width: 0; }
+.focus-it .ft { flex: 1; min-width: 0; display: flex; flex-direction: column; gap: 1px; }
+.focus-it .fx { color: var(--s1); }
+.focus-it .fw { font-size: 11.5px; color: var(--muted); }
+.focus-it .fd { border: 0; background: none; color: var(--muted); cursor: pointer; font-size: 12px;
+  padding: 2px 6px; border-radius: 6px; }
+.focus-it .fd:hover { background: var(--surface-2); color: var(--text); }
 .focus-it .fl { font-size: 10.5px; font-weight: 700; text-transform: uppercase; letter-spacing: .05em; }
 .focus-it.l1 .fl { color: var(--critical-ink); } .focus-it.l2 .fl { color: var(--warning-ink); }