Weekly usage pulse

A 1-minute self-check. Deterministic only; no prompt scoring, no classification, no content.

1. Window = the last 7 full days ending yesterday (local time). If the user passed a date range, use that instead.
2. Run ${CLAUDE_PLUGIN_ROOT}/bin/ccur-collect --from <start> --to <end> --out <scratchpad>/ccur-pulse (the collector is not on PATH; on Windows prefix it with py -3 or python). Read only metrics.json.
3. Print a pulse of at most 15 lines:

Claude Code pulse <start>..<end>
Active days: n/7 | Prompts: N | Sessions: N | Projects: N
Peak: <weekday> <hour> | Heavy days: n | Interruptions/100 prompts: x
Plan mode: n sessions | Delegation: n sessions | Tool errors: n
Context: median Xk per step | N prompts into 200K+ sessions | M cold resumes
Top projects: a (x%), b (y%)
Top tools: ...
vs last week: prompts +x%, active days +n, interruptions/100 [up|down]
One-line nudge: <single most useful habit change, from the numbers only>

The Context line comes from context_hygiene: X = context_per_call_tokens.median / 1000, N = long_context_prompts.count, M = cold_resume_prompts.count. Each step re-sends the whole conversation, so a falling median is the good direction; never call a long session or a high cache-read share efficient. When the numbers are high, the nudge is: start a new session (/clear or /new) when you switch task. 4. Append the headline numbers (including the three Context numbers) to ~/.claude/usage-reports/weekly/<ISO-year>-W<week>.json (create dirs as needed) so the monthly report and next week's "vs last week" line can use them. 5. No file is meant to be shared; say so if asked. Do not read sessions.jsonl or prompt_sample.jsonl; the pulse is numbers-only.