feat(dor): sweep the whole board, and notify only when findings change

[TaekeK]

Follow-up to #1001. Two changes to dor_reconcile.sh, both aimed at the same complaint: the health report was simultaneously blind and noisy.

1. Membership, not the gate label, decides what is swept

The per-issue walk started from gh issue list --label "$LABEL", so an issue could sit on the board in any phase and be invisible to every check in it — drift, staleness, liveness, zombie sidekicks, the approval count.

Not a corner case. 42 of the Feature board's 74 items carry no enhancement label: a UI cohort parked in Awaiting design (#749–#796), most of the Out of pipeline column, and #758 at Awaiting approval — which is why the value gate reported 2 where the board plainly showed 3.

#758  Awaiting approval  labels=[ui, priority:medium, ready-to-build, state:awaiting-approval]   <- invisible
#665  Awaiting approval  labels=[enhancement, …]
#937  Awaiting approval  labels=[enhancement, …]

The board snapshot now carries each issue's own dates and labels, and the walk runs over board membership. The label query survives for the one thing the board cannot show us: an issue that ought to be on it and is not (the heal path). Board records also move from @tsv to the US separator §2 already used — same reason as #1001: an empty middle field in a tab-joined record shifts everything after it.

2. An item filed on the wrong board is named as such

Driven by membership, a misfiled item would otherwise fall through and be reported as un-routed — "the agent likely never ran" — when the real fault is where it is filed.

#819 sits on the Feature board at Blocked (external) while the Bug board has it at Awaiting functional acceptance. Each sweep only ever reads its own board, so nothing could see the disagreement. It now gets its own line (🧭), and an issue that belongs to the other pipeline entirely is skipped in silence rather than double-reported.

3. The report notifies on change, not on schedule

The sweep commented on every hourly run for as long as anything was open — 87 comments on #886. Comparing bodies cannot fix that: most findings restate a moving age (untouched for 17h), so the text is never identical two sweeps running.

Each exception now also produces a stable identity — marker plus issue (🕳️#941), or marker plus its numbers for the countable lines (🚦3) — and the sorted set rides in the body as an HTML comment:

<!-- dor-fingerprint: 🚦3 🧭#819 -->

The body is refreshed every sweep (editing an issue does not notify); a comment is posted only when that set actually differs, or when a closed health issue has to be reopened. A standing backlog costs one notification instead of one per hour.

Verified against the live boards

Running the real script against the real boards, the Feature report becomes:

- 🧭 #819 sits on this board but its labels route it to the other one …
- 🚦 3 issue(s) waiting in **Awaiting approval** — the Product board's value gate.

Two things worth noting from that run. The 🚦 count corrects itself to 3. And the wave of new findings I expected from the 42 newly-visible items did not materialise — every one of them is correctly routed, label matching Status; they were simply never being checked. The only genuinely new finding is #819.

A second consecutive run produced same exceptions as last sweep — body refreshed silently, which is the notify-on-change behaviour working end to end.

Tests

30 assertions, up from 17. The stub now serves the board as JSON as well and runs the script's own --jq over it, so the board record contract is under test rather than hand-copied into the fixture — the same weakness that hid the #1001 bug. It also strips CR: jq on Windows writes CRLF while gh returns LF, which had been letting a local run disagree with CI about what the script parsed.

Four of the new assertions fail against main:

FAIL  an issue the board carries is swept without the gate label
FAIL  …including in the value-gate backlog
FAIL  a bug on the Feature board is called out as misfiled
FAIL  …but an unchanged exception set posts no comment

jscpd full-repo delta: 123 → 123 clones. The other four ci-scripts suites still pass.

Note

While testing this I ran the script against the live boards with a gh shim that was supposed to block writes. The shim was on PATH as a Windows-style path, which Git Bash ignores for command lookup, so the real gh ran: #886 has one stray comment from me at 08:51:39Z and its body was overwritten with this branch's output. No other issue was touched; #938 is untouched. The next scheduled sweep overwrites that body anyway.

🤖 Generated with Claude Code