The Later gate is in plain sight: what it holds, who set it and when, and one action to lift any of it #8711

The Later gate now shows on Home, on each held issue, and on a new page at /:owner/:name/held. A one-approval close of the unchosen ideas sits on that page, and a closed idea revives with one tap. Before the push, tsc --noEmit, scripts/gate.ts, design-lint and the tests for this change (plus the cross files) passed. The forge's own check on the push has not reported back yet.

What changed

• Home has a line per repository you write whose gate holds issues: N issues are held behind the Later gate in <repo>, set by <who> <when>. It opens the held page. "Who" is the gate issue's author and "when" is the day it was opened. A test checks that a new gate shows up on home.
• A held issue says "Held by the Later gate, set by …" at its top, with Lift for writers and crew. Lift is one action, also available as the MCP lift tool. It drops the edge to the gate, and the Later label goes with it.
• The Later label and the gate are one thing. Giving an issue Later holds it, and removing Later lifts it. This fixes the trap in the issue where removing the label did nothing.
• The pass over the gate (the held page and the MCP held tool) reads titles and bodies only: no model, no run. Following the Shape Up note, it groups held issues by the word their titles share most and proposes to close every one. Issues the crew already found done on main close as done; the rest close as dropped.

- Keeping one: each issue has a checkbox, ticked by default; untick it to leave the issue behind the gate, or press Lift to let it go. One "Close the ticked" button applies exactly what the page showed. If the gate changed since, it refuses and asks you to read again. - Agents: an agent can read the proposal but never applies it.

• Closed ones stay findable. Each closed issue gets a Gate review: note and appears under "Closed by the review" on the held page. Revive at its top, the same lift action, reopens it and lets it out from behind the gate.
• Docs and tests: docs/features/later-gate.md explains it all, and test/later-gate.test.ts covers each of these. I added lift and held to the tool-name list in test/mcp.test.ts.

What I chose not to do

• Bets (#8713): they don't exist yet, so "unless it joins a bet" is only the unticked checkbox for now. The review needs a change once bets land.
• Blockers and enablers: per your override, these are no longer kept. Closing an issue that open work waits for parks that work as "Still holds". The pass does not warn about this, which you may want a look at.
• Merging duplicates: dropped, because every merge target would close too.
• Snooze: not built (until a date or new activity).
• Next up (#8694): it isn't in the code yet, so only Home carries the line.

To try it, open Home on a repo with a Later gate: issue and follow the line to /held. Revive an issue closed by the review from the held page.

Closes #8700

The Later gate now shows on Home, on each held issue, and on a new page at /:owner/:name/held. A one-approval close of the unchosen ideas sits on that page, and a closed idea revives with one tap. Before the push, tsc --noEmit, scripts/gate.ts, design-lint and the tests for this change (plus the cross files) passed. The forge's own check on the push has not reported back yet.

What changed

• Home has a line per repository you write whose gate holds issues: N issues are held behind the Later gate in <repo>, set by <who> <when>. It opens the held page. "Who" is the gate issue's author and "when" is the day it was opened. A test checks that a new gate shows up on home.
• A held issue says "Held by the Later gate, set by …" at its top, with Lift for writers and crew. Lift is one action, also available as the MCP lift tool. It drops the edge to the gate, and the Later label goes with it.
• The Later label and the gate are one thing. Giving an issue Later holds it, and removing Later lifts it. This fixes the trap in the issue where removing the label did nothing.
• The pass over the gate (the held page and the MCP held tool) reads titles and bodies only: no model, no run. Following the Shape Up note, it groups held issues by the word their titles share most and proposes to close every one. Issues the crew already found done on main close as done; the rest close as dropped.

- Keeping one: each issue has a checkbox, ticked by default; untick it to leave the issue behind the gate, or press Lift to let it go. One "Close the ticked" button applies exactly what the page showed. If the gate changed since, it refuses and asks you to read again. - Agents: an agent can read the proposal but never applies it.

• Closed ones stay findable. Each closed issue gets a Gate review: note and appears under "Closed by the review" on the held page. Revive at its top, the same lift action, reopens it and lets it out from behind the gate.
• Docs and tests: docs/features/later-gate.md explains it all, and test/later-gate.test.ts covers each of these. I added lift and held to the tool-name list in test/mcp.test.ts.

What I chose not to do

• Bets (#8713): they don't exist yet, so "unless it joins a bet" is only the unticked checkbox for now. The review needs a change once bets land.
• Blockers and enablers: per your override, these are no longer kept. Closing an issue that open work waits for parks that work as "Still holds". The pass does not warn about this, which you may want a look at.
• Merging duplicates: dropped, because every merge target would close too.
• Snooze: not built (until a date or new activity).
• Next up (#8694): it isn't in the code yet, so only Home carries the line.

To try it, open Home on a repo with a Later gate: issue and follow the line to /held. Revive an issue closed by the review from the held page.

I fixed the three failures from the check and pushed. The forge's check on the new push has not run yet. Before the push, tsc --noEmit was clean and these passed: the server-timing sweep, later-gate, hollow, conversation-confirm, mcp-tools and mcp. I did not run the full suite.

• Server-Timing sweep: the held page (GET /:owner/:name/held) now reports its session and render phases. On a repo with no Later gate it reports render before answering not-found, which is what the sweep hit.
• Tool index: I regenerated src/http/mcp/tools/index.ts with bun scripts/mcp-tools.ts. It was stale after I added the held and lift tools.
• Asks-first list (#8230):

- lift: added to the not-asked list, since a lift reverses by relinking and a revive reverses by lifting again. - held: added to the asks-first list (ASKING), but it asks only when apply is given, because reading the proposal changes nothing. The confirmation sentence says what it does and how to undo it (lift on each issue).

The description from my last message still stands, with these three fixes added.

This round addresses points 1–7 of the review; point 8 is a direction call I did not act on (see the end). bun x tsc --noEmit, bun scripts/gate.ts, bun scripts/design-lint.ts and the covering tests pass: later-gate, tui-held, hot-reads-indexed, mcp, mcp-size, conversation-confirm and neighbours. I did not run the whole suite. scripts/fast.ts named 1384 files, so I picked the covering ones.

What changed

• Terminal door (1).

- src/tui/screens/held.ts is the held page's twin. h on a repository's issues opens it, and so does "Held by the Later gate" in the palette (REPO_VERBS in src/tui/forge.ts). - Keys: space ticks, A closes the ticked, l lifts or revives, and enter opens the issue. - It calls the same reviewGate, reviewedClosed, liftHold and applyGateReview as the page and the MCP tools. Reads go through the palette's canRead check, so a private repository is "not found" to a guest or a stranger. - A write there reaches web viewers live: the test hears issue.closed, issue.reopened and issue.unblocked. - test/tui-held.test.ts covers reader refused, guest changes nothing, private not found, and writer close/revive/lift.

• Safe review (2). The page, the MCP approve list and the terminal start with nothing ticked except issues the crew found done on main. Unticked means keep, and the doc says so. A row shows "N open waits for it: closing parks it" when open issues depend on it.
• Stale check (3). The form, the MCP held read and the terminal carry a digest of the sorted candidate numbers. applyGateReview takes it and refuses with stale, closing nothing, if it is missing or differs. Tests cover a newly held issue and a missing digest.
• Marker (4).

- The marker is now an issue.gate_reviewed event that only the apply path writes. It is written after a close that succeeded, then the comment note. - An issue reads as review-closed only while that event is newer than its last close and reopen. A comment cannot forge it, and a reopen followed by a manual close does not read as the review's. - The new read uses the event index; test/hot-reads-indexed.test.ts checks its query plan.

• Lift (5).

- liftHold checks the hold first (gateOf ignores state) and returns not_held with nothing touched. Otherwise it unlinks, then reopens, and re-links if the reopen fails. - followLabel returns its result. setIssueLabels restores the previous labels and returns forbidden if the link or unlink fails.

• Tests (6). New tests, each of which fails without its check:

- A guest and a reader cannot lift, revive or apply on the page. - A reader gets no Revive form. - The MCP lift and held tools refuse a reader and require the digest. - A private repository's /held is 404 for a guest and a stranger. - A forged note is ignored. - A close that does not happen leaves no marker or note.

• Wording (7). The triage doc and comments now say "titles only", and triage no longer takes body. liftHold no longer reassigns its parameter, and the followLabel import in actions/issues.ts sits with the other action imports.

Not done

• Merge is still not built, and there is no separate keep verdict; keep is an unticked box. The doc says both.
• Point 8: I did not widen the pull, but I also did not drop the feature. If it should wait behind the pipeline list, that is a call for a person.

Try it

• Web: open /<owner>/<repo>/held on the preview, https://lab--8711--forge-trevato-apps.open-platform.sh.
• Terminal: h on a repository's issues, or the palette entry "Held by the Later gate".
• MCP: held { app, apply?, digest? }, lift { app, issue }.

Two of the failing tests were my doing, and I fixed them. The other four pass here without any change, so I could not reproduce them. I have pushed the fix.

• Palette roster (4 tests in test/palette-roster.test.ts): the new "Held by the Later gate" palette entry now appears in the roster. I added it to the expected list in all four places, since the entry is intentional.
• Event audience coverage: the issue.gate_reviewed event I introduced had no audience. It is now listed under reader kinds in src/core/event_audience.ts, which is safe because its payload is only an issue number. I also wrote the kind as a string literal at the place that emits it, because the coverage test only finds literal strings.
• Not reproduced: the image-blob header test, the byte-budget test in studio-words, and the looks test "a public app's previews are taken as nobody". All three pass when I run them alone here, and the image-blob test passes on the base branch too. They share nothing with the held page. I took them to be flakes, so look at the next check run to confirm.

bun x tsc --noEmit, bun scripts/gate.ts and the affected tests pass here. Other looks tests that need a real Chromium fail in this sandbox; the check runs them.

I merged current main into this branch and pushed. I think neither of the two failures was caused by this pull. I haven't confirmed that against a new check run.

• "a row wears a dot, not a word badge, and the page stays under its byte budget" is in test/issues-quiet.test.ts. main has since landed "The issue list's byte budget allows for instant actions' script: 51.7 KB had turned every pull red", which is this failure. The merge brings that fix in. I did not touch the held page or its code for this.
• "the screen shows the share, the budgets that refused and the agent refused" is in test/tui-crawlers.test.ts. It passed when I ran it on its own, before and after the merge. I could not reproduce the failure, and I found no connection to the held-gate work. If it fails again on the next check, send me the log.

After the merge, bun x tsc --noEmit and bun scripts/gate.ts are clean. The tests that cover this change also pass: later-gate, tui-held, palette-roster, event-audience-coverage, issues-quiet and tui-crawlers.

Left #8700 open: 1 unticked item remain (A one-time planner pass over what the gate holds proposes keep, merge or close per issue, grouped by theme in under one screen, with closes applied only on one approval. It stays cheap: titles and bodies only, one run.).