Plan: Impact stats, Phase A

Design: docs/design/impact.md. This file is the build plan; tick items as they land so a later session can resume.

Status: done. Phase A merged 2026-10-27; what comes next is in the design note's checklist.

Scope

A daily snapshot of GitHub and the transcripts, pure metrics over it, an /api/impact route and an Impact page with delivery against the baseline, leverage, reviews and cost. No deploy data (Phase B), no new learning events (Phase C).

Files

File	What
lib/impact.py	pure: weeks, baseline and since split, every Phase A metric, PRICES, impact_model(snapshot, now, tz) for the console
lib/impact_collect.py	the only I/O: gh calls, transcript reading through lib/claude_state.py, the per-transcript cache, writing the snapshot atomically; main(argv) for the command
lib/claude_state.py	transcript_activity(path): per message the time, kind (assistant, tool result, prompt), message id, model and usage; the one reader of Claude Code's files
libexec/bosun-impact	bosun impact collect [--since DATE] [--dry-run], Usage header, passes the config
bin/bosun	a usage line
lib/common.sh, examples/generic/bosun.env	set_impact_defaults: BOSUN_IMPACT_REPOS, BOSUN_IMPACT_START, BOSUN_IMPACT_BASELINE_FROM, BOSUN_IMPACT_PROMPT_MINUTES, BOSUN_IMPACT_COLLECT
libexec/bosun-morning	bosun impact collect after the session start when BOSUN_IMPACT_COLLECT=1
lib/console_state.py	read_impact(state_dir): the snapshot, or None
lib/console_server.py	/api/impact
console/index.html, console/app.js, console/app.css	the Impact nav entry and page, a weekly chart with the baseline band and the start line
tests/unit/test_impact.py	every metric on acme fixtures
tests/unit/test_impact_collect.py	the collector with a fake gh and fixture transcripts; cache reuse; dry run writes nothing
tests/js/app_check.js, tests/js/api_models.py	the page renders the model, early-data note, drill-down links
tests/smoke.sh	bosun impact collect --dry-run and a real collect with the fake gh
README.md	Commands table row, a short Impact section under Console

Steps

• [x] 1. Design note, roadmap, this plan.
• [x] 2. lib/impact.py and its tests.
• [x] 3. claude_state.transcript_activity, lib/impact_collect.py, the command, config keys, tests.
• [x] 4. /api/impact and the page, JS checks.
• [x] 5. Smoke checks, README, pre-commit, unit, JS and smoke suites.
• [x] 6. A real collect on this machine and screenshots of #/impact; the numbers sanity-checked against gh by hand.
• [x] 7. Pull request, merged 2026-10-27 (before this repo's public history began, so its number is not one of this repo's).

Status (merged 2026-10-27)

All steps were done and Phase A merged on 2026-10-27. The first real collect took a few minutes, as it reads the GitHub review history of every reviewed PR; later runs reuse the per-transcript cache and the reviews of PRs that did not change. Collecting daily is off by default (BOSUN_IMPACT_COLLECT=0); BOSUN_IMPACT_COLLECT=1 makes the morning run collect. Prices in PRICES are assumed list prices: check them against the current price list before trusting the cost tiles.

Notes for the next session

• The console reads the snapshot only; it never calls gh or reads transcripts itself.
• The snapshot holds no other person's login: reviews store the PR URL, its created time and my review times; authorship is a boolean.