fix(pm): board-snapshot walks a bounded delta before the backfill, so the live board is re-read every run

[claude]

Fixes #18045

Maintainer ruling, verbatim and untranslated (skills seat chat, 2026-09-13T14:58Z): 「现在就派发处理 备份缺口」

Premises, all four checked against origin/main c185d08 before any edit

• P1 holds. selectWalkPlan reaches incremental only once history completes, so while the backfill runs no walk ever carries a recent since. Confirmed on board-archive tip b7c5f578: walk_phase: history, next_since: null, resume.stopped_by: budget, history cursor 2026-08-03T11:51:21Z, board.read_at: null, count_check: pending.
• P2 holds. 5232 records under board/issues/, highest number 17460; 18010, 18020 and 18025 are absent. Read with git ls-tree over a shallow fetch of the branch into a private ref — the shared checkout was never switched.
• P3 holds. 16 open pull requests, none touching scripts/pm/board-snapshot.mjs. (feat(pm): two report-only patrol rows — the decision card's four-facet block, and finding outliving its grade #18033 touches scripts/pm/check-half-states.mjs, which this file imports from; no overlap in this diff.)
• P4 holds. .github/workflows/board-snapshot.yml is unchanged — same cron, same 800-request default, one GITHUB_TOKEN, no retry loop.

The fix, in scripts/pm/board-snapshot.mjs alone

A bounded delta walk now runs first on every run once the open set is complete: state=all&sort=updated&direction=asc&since=..., archiving every row it sees — open or closed, issue or pull request. The history backfill then continues from its own cursor with the remaining budget, its resume semantics untouched. A spent slice is not a run stop; only the run budget and a rate-limit refusal stop a run, exactly as before.

The budget split: 300 of 800, and 500 to the backfill. A quiet six-hour window moves a few dozen numbers, so a steady-state delta costs one listing page plus a comment read per changed number and the backfill keeps very nearly the whole budget. A catch-up delta buys around 250 numbers a run, closing a multi-day gap in a day of scheduled runs rather than in the weeks the backfill needs to reach the same rows. In the worst case the backfill still gets 500 — five eighths of its former pace, slowed but never starved. An unbounded delta could starve it for good, which is why DELTA_REQUEST_SLICE exists at all.

New manifest keys. walk.delta = { complete, cursor, since, slice } sits beside walk.history and is never merged into it. next_since becomes the delta's high-water mark instead of the backfill's February cursor. resume.phase now names the WALK that stopped (open / delta / history / incremental) rather than the phase. run gains delta_ran, delta_completed_here and delta_requests — in run, which materialManifest strips, so a per-run counter cannot commit a manifest-only diff every run. The board's own open count is bought right after the delta and before the backfill, so count_check stops being permanently pending in the history phase.

One declared deviation from the card's fix shape

The card spells the delta anchor as "the previous manifest's generated_at, minus a small skew". Implemented instead as: the delta's own stored cursor; else walk.open_set.completed_at minus the skew; else next_since; else generated_at minus the skew. On the measured manifest generated_at is 2026-09-13T14:22Z while the board was last enumerated in full at 2026-09-10T15:39Z, so anchoring on the stamp would have declared a window this tool never walked and skipped the exact three-day gap the card was filed about. The skew is 30 minutes, two runs' worth of the workflow's own 15-minute job timeout.

Verification

• node scripts/pm/board-snapshot.mjs --self-test: 89 cases across 8 batteries before, 125 across 10 after, exit 0 both. Two new batteries; the count-check battery's floor moves 15 to 17 and the roster floor 8 to 10.
• The self-test gains an offline harness that drives snapshot() end to end against an injected fetch — no network, no token, restored in a finally. The card assumed one existed; it did not, so the walk order and the budget split had no instrument at all.
• Ablation. Deleting the 25-line delta step turns 10 of 125 cases red, THE REGRESSION among them (blob f3723530 to e0cdbd7b, restored to f3723530, git diff HEAD empty). A first attempt stayed green: over a four-row board the backfill reaches "today" inside one page, so the row landed on disk whether or not a delta ran. The harness now carries a 900-row closed backlog, so the backfill provably cannot reach that day inside 800 requests — that fix is its own commit.
• origin/main (226970b) is merged in; node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack derives the same 33 commands on the merged head as before it. Every one runs in the foreground with $? captured before any pipe, and --ran reconciles the list; the verdicts are in the os-dev-report comment on the card.
• skip-changeset: the root package is private and no published package's files[] carries scripts/, so this diff publishes nothing.

---

Generated by Claude Code

[claude]

• Served-tier: 1230/1230 claude-fable-5-1 — harness model stamp counted over this seat's own transcript (non-sidechain assistant messages a model served; <synthetic> harness notices excluded) at 2026-09-13T16:11Z; get_session external_metadata.last_served_model read claude-fable-5-1 at 2026-09-13T16:11Z.

Contract review

Head: c2a54670 (PR #18054, card #18045) — read at 2026-09-13T16:12Z by the skills seat at the contract-review tier. NOT GOVERNED, measured: scripts/pm/board-snapshot.mjs only (+588 −34; the workflow untouched) ⇒ in-seat review, then ready + auto-merge by this seat. Maintainer direct dispatch 「现在就派发处理 备份缺口」 is the ruling of record.

① derived judgments — the archiver re-reads the live board every run, bounded, before the backfill:

1. Delta first (selectDeltaPlan → walk('delta', 'all', budget.slice(300))): state=all&sort=updated&direction=asc&since=…, every row seen is archived (open or closed, issue or PR); a spent slice raises SliceExhausted inside the delta only — the run continues into the backfill with the remaining 500. The 300/500 split is stated with its reason (a quiet window costs a page; a catch-up buys ~250 numbers a run; the backfill is slowed to five eighths, never starved). Accepted.
2. The anchor, deviation accepted: the delta's own cursor; else walk.open_set.completed_at minus a 30-minute skew; else next_since; else generated_at minus the skew. The card's literal anchor (generated_at) would have skipped the measured three-day gap (open set enumerated 09-10T15:39Z, stamp 09-13T14:22Z) — the dev falsified the card's spelling with the live manifest and the chain is the right order. --full and --since= switch the delta off (they ARE the delta); a completed history walk switches it off (the incremental phase is the delta).
3. Manifest: walk.delta = { complete, cursor, since, slice } beside walk.history; next_since = the delta's high-water mark; resume.phase names the walk that stopped (open / delta / history / incremental); per-run counters live under run (stripped by materialManifest, so no manifest-only commits); the board's open count is bought after the delta so count_check leaves pending. The live b7c5f578 manifest is a driven fixture, so the compatibility with today's archive is measured, not assumed.
4. The harness: the card's premise that --self-test already drove snapshot() was FALSE (pure functions only); the dev built driveSnapshot() / fakeBoard() over an injected fetch (no network, counter zeroed and restored) rather than pinning the walk order by source text. The regression is pinned against a 900-row closed backlog after a first ablation stayed green on a four-row board — the phantom was caught and fixed in its own commit.
5. Constraints kept: one GITHUB_TOKEN, no retry loop, budget 800, cron and workflow unchanged (P4); no .claude/**. origin/main 226970bb merged in; the changed blob is identical across the merge.
6. Two 「noted, not filed」 items stand as notes: the workflow comment's stale 「70-case」 count (already false on main; next editor of the workflow), and the module-global requestCount shared by two runs in one process (no shipping caller does that).

Seat measurements on the head tree (git archive c2a54670): board-snapshot.mjs --self-test → 125 cases pass across 10 batteries, exit 0 (base 226970bb: 89 across 8, exit 0 — the control). --pair 18054 → exit 0 at 2026-09-13T16:11Z. Checks on c2a54670 at 2026-09-13T16:11Z: 36 runs, 0 red, 1 running (Lint & Repo Gates) — the queue's own gate holds the merge until it is green. Dev's ablation: the delta step deleted → 10 of 125 red including the regression, restored by blob hash.

② semver: scripts/pm/** publishes nothing; skip-changeset is right.

③ boundary flags: open_questions empty. Landing consequence: from the next scheduled run (20:07Z today) every card created or updated since 09-10 enters board-archive within the delta's slice per run; the R2 copy still waits on the maintainer's four secrets.

Implemented-by: claude/issue-18045-board-snapshot-delta-first
Reviewed-by: session_01DAcomhvR9kKizeYgg89Vo8

Verdict: PASS — the live board is re-read every run, bounded, anchored on the last full enumeration, with the regression pinned by a driven harness; the maintainer's direct order is delivered.

---

Generated by Claude Code