feat: distinguish a quiet arXiv inbox from a broken one

[Jammy2211]

Overview

The knowledge board renders an empty arXiv inbox as nothing waiting — the nightly arXiv digest fills this. — the same line for two opposite states:
arXiv genuinely announced no strong-lensing papers, or the digest ran and the
filing failed so the papers were lost. On 2026-08-25 it was the second, and
nothing on the page said so.

arxiv_papers.yml already killed this exact ambiguity on the Slack side — empty
days still post a heartbeat, so silence in #papers always means a broken run.
The board has no equivalent. This task gives it one.

Split out of complete/2026/08/arxiv-inbox-pat-scope.md, which fixed the filing
bug (PR PyAutoMind#317 makes a denied push redden the run) but deliberately left
this open: that half only helps someone reading CI, and this half is for the
person reading the board.

Plan

• Give arxiv-inbox.md a single last digest: <YYYY-MM-DD> line, owned by
  scripts/inbox_actions.py like every other part of the inbox format. No new
  state file.
• One line covers both options the prompt listed: a stamp dated today over an
  empty inbox is the quiet-day marker (absence becomes positive evidence);
  a stamp four days old is the last filled: signal. Because it is replaced
  rather than accumulated, the sweep never has to age it out.
• Make the digest stamp on quiet days too. Today the inbox step is skipped
  whenever the fetch returns zero papers, so a stamp would go stale on exactly
  the days it exists to prove were quiet. This is the PyAutoMind leg.
• Render the date on the board, and compute the staleness warning in the
  reader's browser, not at render time (see the trap below).
• Warn after 2 weekdays with no digest.

Detailed implementation plan

Affected Repositories

• PyAutoMemory (primary) — the stamp and the board
• PyAutoMind — the workflow leg that writes the stamp on quiet days

The prompt scopes this to PyAutoMemory alone. The PyAutoMind leg is not
scope creep: a stamp that only updates on days papers arrive cannot
distinguish a quiet day from a broken one, which is the entire point.

Branch Survey

Repository	Current Branch	Dirty?
./PyAutoMemory	main	clean
./PyAutoMind	claude/inbox-board-staleness-signal-onzg34	clean

Suggested branch: claude/inbox-board-staleness-signal-onzg34 (both repos —
the session's designated branch; no worktree, this is a web-github run)

The load-bearing trap: a render-time warning can never fire

knowledge_board.yml re-renders on pushes to arxiv-inbox.md, plus a weekly
Monday cron. When filing breaks there is no push — so the published page freezes
at its last good render, and a staleness warning computed server-side would be
frozen with it, showing "0 days since last digest" forever. It would be absent
in precisely the failure it exists for.

So the split is:

• the absolute date is rendered into the page — honest even when frozen;
• the warning is computed from it client-side, at view time.

The weekly cron re-renders the page anyway; the JS makes it honest in between.

Implementation Steps

1. PyAutoMemory/scripts/inbox_actions.py
   • INBOX_STALE_WEEKDAYS = 2, beside INBOX_WINDOW_DAYS, with the reasoning:
     the digest is nominally 02:00 UTC Mon–Fri and GitHub cron only ever jitters
     later (0–3 h), so a single missed weekday is reachable without anything
     being wrong; two is a pattern.
   • STAMP_RE = ^last digest:\s*(\d{4}-\d{2}-\d{2})\s*$.
   • last_digest(text) -> str | None.
   • set_last_digest(text, date) -> str — replaces the stamp in place, else
     inserts directly after the --- separator.
   • weekdays_since(stamp, today) -> int.
   • stamp subcommand, printing stamped:<date> like the other one-token
     subcommands.
   • append's insertion point is unchanged: STAMP_RE cannot match
     INBOX_LINE_RE, and the stamp sits above the papers under every ordering
     (append inserts after the last paper line, or at EOF when there are none).
2. PyAutoMemory/scripts/board.py
   • collect() gains inbox_last_digest; carried through --json.
   • Markdown: a freshness clause in both branches — populated
     (2 waiting · last digest 2026-08-25 · lapse after 7 days) and empty
     (nothing waiting — last digest ran 2026-08-25 and found nothing).
   • HTML: the same, wrapped in <time datetime=… data-stale-weekdays=2>.
   • _EXTRA_JS: compute the weekday age at view time and swap in
     ⚠ no digest for N weekdays — filing may be broken.
   • A missing stamp reads neutrally (no digest run recorded yet), not as
     a warning: spawn.py maps arxiv-inbox.md to EMPTY, so a freshly
     spawned template legitimately has no stamp and has never run a digest.
3. PyAutoMemory/arxiv-inbox.md — document the stamp in the format block.
4. PyAutoMind/.github/workflows/arxiv_papers.yml
   • The count == '0' heartbeat step also writes arxiv_survivors.json as
     {"papers": []} — exactly what the Claude step already writes on its
     empty day, so every path produces the file.
   • Drop if: count != '0' from the inbox step and add the stamp call, so
     every weekday run stamps. The existing warn-and-exit-0 guard on a
     missing survivors file stays as the real-failure signal.
   • Unchanged: the step still runs after the Slack POST, so "a cross-repo
     failure must never cost the morning's post" holds.
   • Deliberate side effect: one commit per weekday to PyAutoMemory, which
     re-renders the board daily. That is the heartbeat.
5. Tests — tests/test_inbox_actions.py (stamp round-trip, insertion point,
   replacement not accumulation, weekday arithmetic across a weekend, stamp is
   not parsed as a paper) and tests/test_board.py (both render branches,
   missing stamp, the date reaching the HTML).

Not provable in CI

arxiv_papers.yml cannot be fired from a branch that edits it —
claude-code-action refuses to run when the workflow file differs from the
default branch's copy, and then exits success (recorded trap, cost a wasted
dispatch in arxiv-inbox-pat-scope). That leg gets bash -n, a YAML parse and
reasoned review only; the first genuine exercise is the scheduled run on main.

Pre-existing failure to clear first

PyAutoMemory@main is currently red on both make validate and make test:
commit 498e1a8 added a tracked .claude/ directory without adding it to
ALLOWED_TOP_DIRS in scripts/validate_structure.py. Unrelated to this task,
but any PR is red on arrival until it is fixed. One line, in its own commit.

Key Files

• PyAutoMemory/scripts/inbox_actions.py — single owner of the inbox line
  format, the window and every transition; the stamp joins them
• PyAutoMemory/scripts/board.py — renders the inbox above the reading queue
• PyAutoMemory/arxiv-inbox.md — the inbox itself, and its format block
• PyAutoMind/.github/workflows/arxiv_papers.yml — the nightly digest
• PyAutoMemory/scripts/validate_structure.py — the pre-existing red

Original Prompt

Click to expand starting prompt

The knowledge board's empty inbox should distinguish quiet from broken

Type: feature
Target: PyAutoMemory
Repos:

• PyAutoMemory
  Difficulty: small
  Autonomy: supervised
  Priority: normal
  Status: formalised
  Filed: 2026-08-25

Split out of complete/2026/08/arxiv-inbox-pat-scope.md, which fixed the filing
bug but deliberately left this open.

The gap

The board renders an empty arXiv inbox as:

arXiv inbox (suggested overnight — un-acted papers lapse after 7 days)
nothing waiting — the nightly arXiv digest fills this.

That line is shown for two completely different states:

• arXiv genuinely announced no strong-lensing papers; or
• the digest ran and the filing failed, so papers were lost.

On 2026-08-25 it was the second, and nothing on the page said so. The reader has
no way to tell without opening GitHub Actions.

Why this is a known-solved problem here

arxiv_papers.yml already fixed the identical ambiguity on the Slack side. From
its header comment (user, 2026-07-13):

Empty days STILL post a short "no new strong-lensing papers today" heartbeat —
so silence in #papers always means a broken run, never a genuinely empty one.

The board has no equivalent. The digest's own invariant stops at Slack and does
not reach the surface the paper-management feature actually asks the human to
use.

Task

Give the inbox a freshness signal so an empty board is readable. Options, in
rough order of cost:

• last filled: line — the board renders the date of the most recent
  successful append/sweep. An empty inbox dated today is quiet; one dated
  four days ago is suspect. Cheapest, no new state if it can be derived from
  arxiv-inbox.md's git history or a marker the digest writes.
• Explicit quiet-day marker — the digest writes a dated "no papers" line on
  genuinely empty days, exactly mirroring the Slack heartbeat. Makes absence
  positive evidence, at the cost of a line the sweep must then age out.
• Freshness warning — the board flags an inbox with no write in N weekdays.
  Strongest signal, needs a threshold nobody has tuned yet.

Prefer whichever needs no new state file. Note that scripts/inbox_actions.py
is the single owner of the line format, the window and every transition — a
marker line must go through it, not through board.py or a workflow's shell.

Note on the sibling failure mode

PR PyAutoMind#317 made the workflow fail loudly when the push is denied, so a
future filing failure reddens the run. This task covers the other half: the
board should be honest even when nobody is reading CI.

[Jammy2211]

Shipped

• PyAutoMemory — #60 merged, main 963388a
• PyAutoMind — #320 merged, main a82b96f

Merged library-first: arxiv_papers.yml calls inbox_actions.py stamp, so the Memory side had to land before the Mind side or the nightly run would have failed on a subcommand that did not exist yet.

What landed

inbox_actions.py owns a last digest: <date> line in arxiv-inbox.md, rewritten on every digest run whether or not papers arrive. One line does both options this issue scoped: dated today over an empty inbox it is the quiet-day marker, making absence positive evidence; days old it is the "last filled" signal. Replaced rather than accumulated, so the sweep has nothing to age out.

The board reads it back as:

nothing waiting — the last digest ran 2026-08-25 and found nothing
nothing waiting — ⚠ no digest since 2026-08-19 (4 weekdays) — the nightly filing may be broken
nothing waiting — the nightly arXiv digest fills this — no run recorded yet

Two decisions worth keeping

The warning is computed in the reader's browser, not at render time. knowledge_board.yml republishes on pushes to arxiv-inbox.md — exactly what stops happening when filing breaks. The page then freezes at its last good render, so a server-side warning would sit at "0 weekdays" forever and never fire in the one failure it exists for. The date is a fact and survives the freeze; the verdict is recomputed against the reader's clock. A page already stale when rendered still warns without JS.

Threshold 2 weekdays, not 1 day. Cron only ever jitters later, so a board rendered early can see yesterday's stamp with nothing wrong. Weekdays because the digest is Mon–Fri: a Monday board reading Friday's stamp is three calendar days but zero missed runs.

A missing stamp reads neutrally, never as a warning — spawn.py empties arxiv-inbox.md for the template repos, so a freshly spawned Memory has never run a digest and has nothing to be late for.

Scope note

The issue scoped this to PyAutoMemory. The PyAutoMind leg was added deliberately: a stamp that only updates on days papers arrive cannot distinguish a quiet day from a broken one, which is the whole point. The digest's inbox step lost its count != 0 guard and the empty-day heartbeat now writes arxiv_survivors.json as {"papers": []}, so every path still produces the file and a missing one stays a real failure. Side effects, both intended: a quiet day is now a one-line commit that re-renders the board daily, and the sweep runs every weekday instead of waiting for the next paper day.

Validation

92 PyAutoMemory tests (23 new, was 69) · make validate green · 197 PyAutoMind tests · lifecycle.py check OK · YAML parses · every run: block passes bash -n · both PRs green on their final head shas.

Beyond CI: the shipped JS (extracted from board.py, not retyped) was cross-checked against the Python weekday arithmetic over 30 consecutive dates including weekends, and all four digest paths were rehearsed against a real checkout with the shell blocks extracted verbatim from the workflow — quiet day/same date (no commit), quiet day/new date (heartbeat commit), paper day (paper lands below the stamp), quiet day sweeping lapsed papers.

Not proven live: claude-code-action refuses to run when arxiv_papers.yml differs from the default branch's copy and then exits success, so no branch dispatch can reach those steps. The scheduled 02:00 UTC run on main is the first genuine end-to-end exercise.

Incidental

main was red on make validate and make test from 498e1a8 (tracked .claude/ missing from ALLOWED_TOP_DIRS). #59 landed the same one-line fix while this branch was in flight; the conflict was resolved in main's favour and the redundant commit contributes nothing.

---

Generated by Claude Code