[bhavsarpratik]

What this adds

A new Work page (⌥6, macOS only) shows what your coding-agent spend went into, whether it worked, and where a cheaper model or lower reasoning effort would likely have done. Google's Gemma 4 E4B, running in Ollama on your Mac, labels each request with its kind of work, area, and complexity, and flags follow-ups where you pushed back. Each request carries its share of the session's cost. Prompts never leave the machine; only salted keys, labels, and counts are stored.

How it works, rule by rule: specs/WORK_INSIGHTS.md. Evaluation history: specs/2026-09-30-work-insights-design.md (iteration 13 covers Gemma).

How to use it

1. Check out this branch and run ./scripts/install.
2. Turn on Settings → Work insights. Setup runs in the background:
   • It reuses your Ollama 0.34+, or installs a pinned, signature-checked Ollama 0.34.4 under Token Meter's folder (no sudo).
   • It downloads Google's official 4-bit model, gemma-4-E4B_q4_0-it.gguf (Apache-2.0, about 5 GB, pinned by commit and SHA-256), adds it to Ollama as token-meter-gemma, and deletes the download.
   • It needs about 11 GB free while it runs. The model uses about 5.5 GB of memory while it labels.
3. Labels fill in at 5 model calls a minute by default. Labeling pauses on battery and when the Mac is busy, and the Work page shows progress.

What the page shows

• Right-sizing: three cards (Model, Reasoning, Pushback) with estimated savings, plus spend by model tier and reasoning effort.
• Where the spend went: spend by area, on the day it happened, with a trend line per area.
• How sessions ended: accepted, recovered after pushback, or ended on pushback.
• Session tags: such as Marathon, Long thread, Big spender, Subagent team, Rescued, and One-shot.
• When you work, Pushback over time, Cost per resolved task, and Model choices.

Every bar, tag, and suggestion opens the sessions behind it. The page uses the shared design system from main: tokens, type, weight, radius, and spacing scales.

Memory safety

Labeling never pushes a Mac into swapping on its own account:

• The model is loaded only when free memory covers it plus a margin of 2 GB (or a tenth of the Mac's memory, if larger).
• Before every model call, Token Meter checks memory. On critical pressure, or when free memory drops below the margin, it unloads the model and waits ("Waiting: memory is low").
• If memory stays short, the wait doubles from 1 to 15 minutes, and labeling resumes on its own once memory frees up.
• Macs with less than about 10 GB of memory never download or load the model, and the page says why.

Also in this branch

• Models page: a sortable Pushback column. It uses the same rule as Work: each request is credited to the model that answered it.
• Current sessions: a session counts as live from its last model reply, not from its trace file changing. Opening an old session no longer makes it live or triggers a notification.
• Sessions summary: "vs yesterday" always shows a number on partial days, and the estimate suffix is gone from the spend value.
• Performance page: Work appears in its sidebar, and its ⌥ shortcuts match the dashboard.
• Migration: saved settings that name Winnow or Jet move to token-meter-gemma. The Work page asks before the new download, and Token Meter's own Ollama removes the retired models.

Label accuracy

Real local requests from three session-disjoint sets. Gold labels were set blind, and the data stays private on the developer's machine. Every setting was fitted on dev only.

Test / held-out	Winnow-E4B (previous)	Gemma 4 E4B QAT (this PR)
Work type	76% / 85%	76% / 82%
Area	71% / 75%	77% / 80%
Complexity, exact	64% / 75%	69% / 74%
Pushback F1	0.83 / 0.86	0.80 / 0.69

Pushback on the held-out set is the remaining gap. The pushback decision point keeps reported rates close to the truth: 22% flagged against 18.5% real on test.

Testing

• Full suite: 1583 tests OK. The design-system contract passes, the embedded JS parses, and py_compile and git diff --check are clean.
• High-risk route on 20f7758: an independent tester and two reviewers (correctness, and resource safety/privacy) pass. The reviews caught and closed a reload loop in the memory guard.
• The tester ran the real classifier against Ollama with a simulated memory budget:
  • Labeling completes when the loaded model leaves only modest free memory.
  • Critical pressure unloads the model.
  • The same script reproduces the reload loop on the earlier commit.
• The prompts the app sends are byte-identical to the evaluated ones: 1,040 request prompts and 780 pushback prompts.
• Google's pinned file gives exactly the same log-probabilities as the Ollama build used for tuning.
• Installed and running from 20f7758: /health and /menubar OK, and the runtime matches the source. 1bb98e8 after it only removes a stale diagram from the spec.

Not verified

• Browser checks of the Work page after the design-system port, and of the new memory messages, at 1440 and 1024 px.
• A fresh-machine setup end to end (download, signature check, LaunchAgent) is covered only by tests with faked downloads.
• Real behaviour under memory pressure is covered by simulation and unit tests, not an actual low-memory Mac.
• Relabeling a long history with the new model takes hours at the default pace.
• Linux and Windows: Work insights stay disabled.

🤖 Generated with Claude Code