First study, so no previous report to follow up. Written with AI assistance.
Window
2026-10-01 to 2026-10-15 (14 days). The local clone's history is partial (the git fetch --unshallow origin main did not finish in this session), so per-day commit counts from git log are unreliable and are not used. PR counts below come from the GitHub search API and are complete.
Evidence (commands re-takeable)
| Fact | Number | Source |
|---|---|---|
| PRs merged since 2026-10-01 | 814 | search is:pr is:merged merged:>=2026-10-01 |
| PRs closed unmerged since 2026-10-01 | 42 | search is:pr is:closed is:unmerged closed:>=2026-10-01 |
Of those, titled ... new venture ... | 33 | same list, counted by title |
| Of those, same-slug duplicate venture pairs | 3 (#6698/#6699, #6337/#6338, #6343/#6347); plus one duplicate fix pair (#6265/#6266) | titles in the list |
| "Fix page so Pages publish unblocks" PRs (merged or closed) | 5 seen: #6265, #6266, #6558, #6727, #6645 | titles |
launched.json ordering / date repair commits on 2026-10-14 | 4 (#6936, #6962, #6963, #6975) | git log origin/main --grep=launched.json |
| Open PRs now | 8 (3 are 1-9 days old drafts or parked: #6504, #5967, #6854/#6906) | list open search |
| Run records | 175 total, 170 research, 1 process, 2 ventures | python3 prompts/promptctl.py measure |
unlanded_work.py | exits 1 only for this session's own unpushed branch | run in this session |
"Never measured" items: CI failures before green per PR (the run records carry null), run_workflow_locally.py --tree on main, and guard_state.py were not run in this session. Ventures have 2 run records against ~hundreds of venture PRs, so the run records cannot describe the venture track at all.
Sample audit
Not done. I did not re-run five PRs' tests or read their diffs this session, so this study makes no claim about the quality of what merged. The next study should do it first.
Findings by root cause
- Shared sorted files are edited by parallel branches (4 repairs in one day).
platform/pages/launched.jsonhas a byte-for-byte sort test (launched.test.mjs:124). Four merges on 2026-10-14 repaired order or missing dates after sibling venture PRs landed. The commit messages say the entry was "appended out of order", which fits a hand edit instead ofnode platform/pages/launched.mjs --write(the commandPRODUCT_PLAYBOOK.mdnames). I did not confirm the hand-edit cause from the diffs. - Duplicate venture PRs (at least 3 venture pairs and 1 fix pair, 8 closed unmerged). Same slug opened twice within hours. This is the failure
CLAUDE.md"Duplicate PRs" already describes; the check there reads open PRs, which cannot see a sibling that opened minutes earlier. - Venture page defects that block the Pages publish (5 PRs). Titles name a
[hidden]panel that renders anyway (#6558, #6645) and a dead inline script (#6265, #6266).platform/webcheck/page-integrityandinline-script-braces.test.mjsexist for these, so the open question is whether they run on every new venture before merge or only afterwards. - High venture volume, little measurement. at least 77 "new venture" subjects on
mainin the window (lower bound: partial local history) versus no recorded visitor numbers (portfolio/usage/,live_url_visitors: null). Throughput is not the problem the evidence shows; whether any of it is used is unmeasured.
Evidence is thin on causes for 1 and 3: I read titles and two commit bodies, not diffs.
Recommendations (ranked)
| # | Problem (count) | Change | Touches | Metric | Who |
|---|---|---|---|---|---|
| 1 | launched.json repairs (4/day) | Make the sort a non-conflict: have check/merge run launched.mjs --write automatically, or let the test normalise instead of failing | platform/pages/launched.mjs, test | repair commits per week → 0 | agent |
| 2 | Duplicate venture PRs (4 pairs) | Before opening a venture PR, fail on an existing ventures/<slug> in main or an open/closed-in-last-24h PR with that slug in its title | new check in platform/ci/ | closed-unmerged venture PRs | agent |
| 3 | Page defects blocking publish (5) | Confirm page-integrity runs for every ventures/**/*.html in the PR check; add if not | platform/webcheck, workflow | "unblock Pages" PRs per week | agent |
| 4 | No usage measurement | Owner decision on a privacy-respecting counter; until then, weight effort toward deploying, not new calculators | portfolio/usage/ | share of ventures with a reading | owner |
| 5 | Study could not see history | Document that the cloud clone is shallow and git fetch --unshallow origin main takes over 2 min; run it in the background first | prompts/process/prompt-0.md | study setup minutes | agent |
No recommendation was implemented in this PR; none was both small and verified enough to ship without more evidence.
Follow-up on the last report
None exists; this is the first study.