feat: Token cost visibility — full attribution, budgets, anomaly detection, and efficiency scoring

[vivekchand]

Context

A Reddit comment called out cron job token cost visibility as a game-changer for stopping AI spending bleed. After checking the actual code and UI, this feature does not currently exist. The Crons panel shows timing (runtime_ms as "Took: Xs"), last run status, and errors -- but zero token cost data. This issue proposes building it from scratch.

What the cron_fired event currently stores: job_id, name, status, runtime_ms. No token counts, no cost.
What _build_cron_jobs() sends in the snapshot: id, name, enabled, schedule, task, lastRun=None, lastStatus=None. No cost fields.

This is the full roadmap for token cost visibility in ClawMetry.

---

Proposed improvements

1. Per-cron token cost tracking (foundation -- must ship first)

The cron_fired event needs to capture token usage from the session it spawned. The sync daemon should correlate cron job_id with the session that ran it and pull token counts + cost_usd. Without this, nothing else below is possible.

Fields to add to cron_fired events: tokens_in, tokens_out, cost_usd, model, session_id

2. Per-cron cost display in the Crons panel

Once cost data flows, show it inline in the cron card: "Took: 74s -- 12,400 tokens -- $0.04". Simple text, no redesign needed.

3. Per-cron cost history graph

Show cost per run over time for each cron job. A sparkline in the cron list + an expanded chart in the detail view. Pattern changes are immediately visible: "this heartbeat doubled in cost 3 days ago."

4. Budget thresholds and alerts

Let users set a per-cron or total daily/monthly budget. Alert (Telegram, push, dashboard banner) when a single run exceeds a threshold, daily total crosses a limit, or 7-day rolling average is rising fast.

5. Spending anomaly detection

Automatically flag runs that cost 2x or more than the rolling average for that cron. Show a warning badge. No configuration needed -- just pattern-based detection.

6. Top cost driver leaderboard

A simple ranked list: "Your 5 most expensive crons this week." One glance = instant prioritization.

7. Sub-agent cost attribution tree

When a cron spawns sub-agents, show a cost tree: parent + each spawned child with its share of total spend. Currently sub-agent costs are completely invisible from the parent cron view.

8. Idle cost metric

Show how much a heartbeat or cron spent and produced zero external output (replied HEARTBEAT_OK, no tools called, no messages sent). This is pure waste cost -- surfacing it is the fastest way to help users cut spend.

9. Cost breakdown by tool type

Within a cron run, show what percentage of tokens went to exec, web_search, browser, memory_search, sessions_spawn, etc. Answers "why is this cron expensive" without log diving.

10. Model cost comparison (what-if)

Show estimated monthly savings if the user switches a cron from claude-sonnet to claude-haiku. Use real token counts from history. "Switching to Haiku saves ~$8/month on this cron."

11. Rolling 30-day cost projection

Based on the last 7 days of spend, project the current month total. "On track for $X this month."

12. Efficiency score

Useful output actions (tool calls with side effects, messages sent, files written) divided by total tokens. Show as A/B/C badge on each cron card.

---

Priority order

1. Cron-session cost correlation (foundation -- unblocks everything)
2. Cost display in cron card (immediate user value)
3. Anomaly detection (zero-config, high value)
4. Top cost leaderboard (fast to build)
5. Budget alerts (uses existing notification plumbing)
6. Idle cost metric (directly addresses heartbeat-waste use case)
7. Cost history graph
8. Sub-agent cost tree
9. Tool type breakdown
10. Model comparison, rolling projection, efficiency score

---

Notes

• cron_fired event schema change is backward compatible (new fields are additive)
• Anomaly detection can be client-side using existing Turso run history
• Budget alerts can reuse existing Resend + Telegram notification plumbing
• This is the most direct path to ClawMetry owning the "how much is my AI actually costing me" question