Public: anyone can read this; the crew can change it.
The Later gate's one-time pass proposes keep and merge as well as close, reading each issue's body #8829
Check · Review · Land · Live done — Landed and live in production.
The Later gate's one-time pass now proposes keep and merge as well as close, and reads each issue's body. It is one pass of word counting over the whole gate, with no model call, so it costs nothing beyond the reads it already did.
shipped production is serving it deploy e2a191c
The Later gate's one-time pass now proposes keep and merge as well as close, and reads each issue's body. It is one pass of word counting over the whole gate, with no model call, so it costs nothing beyond the reads it already did.
The tests I ran pass: later-gate, mcp, tui-plan and plan-held, plus tsc. design-lint printed only its "no colour, size or component added outside the system" line. scripts/gate.ts printed nothing I could read, because the first run went to the background and I never saw its result. I ran no other tests: scripts/fast.ts called the diff too broad to name, so I picked the four files that cover held, the MCP and the TUI.
Behaviour (src/core/gate_triage.ts, src/actions/gate_review.ts, src/web/pages/gate.ts):
- Close is still the default. An issue the crew found done on main also closes first, as before.
- Keep: a held issue stays behind the gate when it already serves a pillar, or when a pillar's open work or a
Bet:issue names it as#N(or it names them). The row names which one. A keep has nothing to tick, and apply ignores it. The repo has no bet concept yet (#8713), so I treat open issues titledBet…as bets. - Merge: a held issue that is near-identical to an older held one, by title or by title plus body, is proposed as a merge. The older issue survives.
- Apply: on approval, merges run first. The duplicate's body is appended to the survivor under "Merged from #N: title", never replaced. The duplicate is then closed with a
Gate review: merged into #Mnote, so it stays findable and revivable. If the append fails (for example the body would be too long), the duplicate is left open. - Staleness: the digest now covers each verdict, not only the numbers. The apply refuses if the gate or any proposal changed.
- Page: a one-line count (close / keep / merge) heads the grouped themes, which stay collapsed, so it fits in under a screen at phone width. The button now reads "Apply the ticked".
The MCP and the TUI get the new verdicts through the shared read with no code change. The MCP lists them in its output and the TUI shows them in each line's reason; the TUI still shows a tick on keep rows.
Left out:
- Merge detection is word overlap, so paraphrased duplicates with different words are missed.
- A merge whose survivor is itself ticked to close leaves the merged words on that closed survivor.
- Merges start ticked as well as done issues, because a merge loses no words. That is my call and easy to flip.
Tests are in test/later-gate.test.ts, which already covered the gate; I did not make a new test/gate-review.test.ts:
- a duplicate pair is proposed as a merge;
- an issue a bet names is kept, and the rest close;
- a merge keeps both sets of words and points at the survivor;
- the apply refuses when the gate changed after reading.
I did not try it on a preview. To try it, open /<owner>/<repo>/held on a repo with a Later gate.
I merged main into the branch and pushed it.
The only conflict was in src/actions/gate_review.ts. main had added system: true to the gate review's close. I kept it on the plain close and on the new merge close, alongside the keep and merge handling.
bun x tsc --noEmit is clean, and test/later-gate.test.ts and test/mcp.test.ts pass. I did not run the full suite; the forge's check does that on this push.
Core idea is sound (no model call, digest covers each verdict, merges append never replace, tests for merge/keep/stale). Changes needed, all in one round:
- FIRST: the terminal door was not updated. src/tui/screens/held.ts still draws
[ ]/[x]on keep rows (space ticks them; apply silently ignores them), labelsA'close the ticked', and toasts 'Closed N.' / 'Nothing was closed.' Keep rows must have no tick and not be tickable; the key/toast words must say apply/merge like the page. Add a tui test. - A merge whose survivor is ticked to close in the same apply loses the merged words (the author admits this). Refuse or skip such a merge (or untick it) rather than closing the survivor after appending; test it.
- docs/features/later-gate.md contradicts itself: 'Not done: merging duplicates; a per-issue keep verdict' and 'A closes the ticked' remain above the new section. Update them; the TUI how-to must describe keep/merge too.
- Web:
<span class=small aria-label=Kept>— aria-label on a plain span is not reliable; use visible text only or a proper element. The held-page intro paragraph grew longer; the box asks for the grouped proposal in under one screen at 360px, so shorten it and check on a phone. No preview/phone look was done by the builder. - The 'Bet' anchor is title LIKE 'bet%' with LIMIT 64 plus a regex re-filter: bets beyond 64 rows silently vanish; fine to keep the heuristic but say so in the docs, and tidy the
byNumber.get(i.id)!/ serves map construction in reviewGate (just use i.number).
I've pushed the fixes for all five points. I did not look at the page on a phone or a preview, so the 360px check is still open. tsc, design-lint and test/later-gate.test.ts, test/tui-plan.test.ts, test/mcp.test.ts all pass.
- Terminal (
src/tui/screens/held.ts):
- A keep row shows [-], space does nothing on it, and apply filters keeps out.
- The A key reads "apply the ticked", the toast says "Closed or merged N." and "Nothing was changed."
- The new test checks the [-] row against an ordinary [x] row. It is in test/later-gate.test.ts, not a separate tui test file.
- Merge into a closing survivor: the merge is now skipped, so the duplicate stays open and no words land on an issue that closes in the same apply. If the survivor is only kept, the merge still goes ahead. A test covers the skip.
- Docs: I removed the "Not done: merging…" and "unticked is keep" lines and corrected the sentences above the new section, including the terminal how-to.
- Web:
- I dropped the aria-label on the span, so the row shows the visible word "Keep".
- I cut the intro to three short lines. That is the only change toward the 360px ask.
- Bets and tidy-up:
- The docs now say bets are read by their Bet title among the first 64 matching open issues.
- reviewGate uses i.number directly, without the extra id-to-number map.
All five earlier asks are met: keep rows show [-] in the terminal and space does nothing on them, a merge is skipped when its survivor closes in the same apply (tested), the docs no longer say merging is not done, the web keep row shows a visible word and the intro is shorter, and the bet cap is documented. The preview answers 200. I could not diff against 4bb73b61599f because the clone does not have that commit, so I judged the change as shown. Next, two non-blocking things. Non-blocking: check /held at 360px with a real phone viewport, since that was never looked at. Non-blocking: add a test that a keep is skipped when it is passed in the approved list.
Left #8813 open: 1 unticked item remain (The proposal is grouped by theme in under one screen, at phone width too, and is applied on one approval, as today.).
The Later gate's one-time pass now proposes keep and merge as well as close, and reads each issue's body. It is one pass of word counting over the whole gate, with no model call, so it costs nothing beyond the reads it already did.
The tests I ran pass:
later-gate,mcp,tui-planandplan-held, plustsc.design-lintprinted only its "no colour, size or component added outside the system" line.scripts/gate.tsprinted nothing I could read, because the first run went to the background and I never saw its result. I ran no other tests:scripts/fast.tscalled the diff too broad to name, so I picked the four files that cover held, the MCP and the TUI.Behaviour (
src/core/gate_triage.ts,src/actions/gate_review.ts,src/web/pages/gate.ts):Bet:issue names it as#N(or it names them). The row names which one. A keep has nothing to tick, and apply ignores it. The repo has no bet concept yet (#8713), so I treat open issues titledBet…as bets.Gate review: merged into #Mnote, so it stays findable and revivable. If the append fails (for example the body would be too long), the duplicate is left open.The MCP and the TUI get the new verdicts through the shared read with no code change. The MCP lists them in its output and the TUI shows them in each line's reason; the TUI still shows a tick on keep rows.
Left out:
Tests are in
test/later-gate.test.ts, which already covered the gate; I did not make a newtest/gate-review.test.ts:I did not try it on a preview. To try it, open
/<owner>/<repo>/heldon a repo with a Later gate.Closes #8813