The week is the sprint: a cut line on Next up shows what this week's capacity can ship, and the forge writes its own retrospective #8809

Next up now shows a cost range on each issue, a cut line and a one-sentence projection. The forge also writes a retrospective at each weekly reset. I ran bun x tsc --noEmit, bun scripts/gate.ts, bun scripts/design-lint.ts, the new test/week-sprint.test.ts and the neighbouring MCP, next-up, schema and timing tests, and they pass. I did not run the full suite; the forge's check does that on push.

What changed

• Estimates (src/flow/estimate.ts): each row shows the 25th to 75th percentile of what past closed issues of the same size and kind cost. Size comes from title and body length; kind is the first label. Cost is the sum of all agent runs on the issue. With fewer than 3 comparable issues it widens to size alone, then reads "too little history". No agent run is spent.
• Cut line and projection (src/flow/week.ts):

- Capacity is what the crew may still spend in the weekly window, summed over the accounts whose cap is known. - Cut line: walking from the top, the running median estimate is held against that capacity, and the line stops at the first issue that doesn't fit. Rows below it read as next week. - Projection: one sentence says how many should ship by the reset, as a range, with a confidence. - No line: with no capacity or no history there is no line, and the sentence says why. - Live moves: a drag already reloads the page, so the line moves with it. - Determinism: the same list, history and capacity give the same line, and a test checks it.

• Retrospective (src/flow/retro.ts, migration 20261103110000-week-retro with its pin):

- When it is written: an hourly timer in main.ts keeps the projection for the open window. After the window ends it writes the retrospective once and never edits it. - What it holds: shipped against the projection, how many shipped issues cost inside their range, the open work that cost most, and what is stuck (two or more failed runs, with the last failure's words). It ends in at most three proposals. - Where to read it: /:owner/:name/week, linked from Next up, and the MCP week tool. - Sealed issues: the retrospective leaves sealed issues out.

• One tap per proposal: POST /:owner/:name/week or week { app, take }, by a writer who is a person. A proposal puts its issue behind the Later gate, the same move as Not now, and can be taken once. Nothing changes without that tap.
• Reads are indexed, and a test checks their query plans for scans and sorts. The how-to is in docs/features/next-up.md.

What I chose not to do

• The line falls between issues, not whole bets or scopes, because #8713 isn't in the forge yet. The per-group health marks wait for the same reason.
• Proposals are only "put behind the Later gate". Other kinds can be added later.
• I added no terminal screen.
• Rollover: nothing rolls over by itself.

How to try it Open /<owner>/<app>/next. Without a read weekly capacity it shows ranges and "No projection yet: …". Once a subscription's weekly limit has been read, it also shows the line. /<owner>/<app>/week shows the latest retrospective.

Closes #8704

Next up now shows a cost range on each issue, a cut line and a one-sentence projection. The forge also writes a retrospective at each weekly reset. I ran bun x tsc --noEmit, bun scripts/gate.ts, bun scripts/design-lint.ts, the new test/week-sprint.test.ts and the neighbouring MCP, next-up, schema and timing tests, and they pass. I did not run the full suite; the forge's check does that on push.

What changed

• Estimates (src/flow/estimate.ts): each row shows the 25th to 75th percentile of what past closed issues of the same size and kind cost. Size comes from title and body length; kind is the first label. Cost is the sum of all agent runs on the issue. With fewer than 3 comparable issues it widens to size alone, then reads "too little history". No agent run is spent.
• Cut line and projection (src/flow/week.ts):

- Capacity is what the crew may still spend in the weekly window, summed over the accounts whose cap is known. - Cut line: walking from the top, the running median estimate is held against that capacity, and the line stops at the first issue that doesn't fit. Rows below it read as next week. - Projection: one sentence says how many should ship by the reset, as a range, with a confidence. - No line: with no capacity or no history there is no line, and the sentence says why. - Live moves: a drag already reloads the page, so the line moves with it. - Determinism: the same list, history and capacity give the same line, and a test checks it.

• Retrospective (src/flow/retro.ts, migration 20261103110000-week-retro with its pin):

- When it is written: an hourly timer in main.ts keeps the projection for the open window. After the window ends it writes the retrospective once and never edits it. - What it holds: shipped against the projection, how many shipped issues cost inside their range, the open work that cost most, and what is stuck (two or more failed runs, with the last failure's words). It ends in at most three proposals. - Where to read it: /:owner/:name/week, linked from Next up, and the MCP week tool. - Sealed issues: the retrospective leaves sealed issues out.

• One tap per proposal: POST /:owner/:name/week or week { app, take }, by a writer who is a person. A proposal puts its issue behind the Later gate, the same move as Not now, and can be taken once. Nothing changes without that tap.
• Reads are indexed, and a test checks their query plans for scans and sorts. The how-to is in docs/features/next-up.md.

What I chose not to do

• The line falls between issues, not whole bets or scopes, because #8713 isn't in the forge yet. The per-group health marks wait for the same reason.
• Proposals are only "put behind the Later gate". Other kinds can be added later.
• I added no terminal screen.
• Rollover: nothing rolls over by itself.

How to try it Open /<owner>/<app>/next. Without a read weekly capacity it shows ranges and "No projection yet: …". Once a subscription's weekly limit has been read, it also shows the line. /<owner>/<app>/week shows the latest retrospective.

I've addressed all six points from @sentinel's review and pushed to the same branch. I ran bun x tsc --noEmit, bun scripts/gate.ts, bun scripts/design-lint.ts and the affected suites; all passed. I did not run the full suite; the forge's check does that on push.

1. Terminal door.

- Next up screen: it now shows each row's range, a muted "this week ends here (N fit)" line under the last issue that fits, and the projection sentence in the header. - New screen: "The week in a repository" is in the palette. It shows the retrospective, and j/k choose a proposal while enter or a takes it through takeProposal. - Access and live updates: a reader or an agent sees the screen and the key does nothing. A new week.retro event is published after the hourly tick writes a retrospective, and issue.ranked already covers a proposal taken elsewhere, so the open screen updates. - Docs and MCP: docs/features/next-up.md names the keys. MCP next_up rows now carry estimate and thisWeek, plus a projection sentence.

1. Person-only proposals. takeProposal now refuses an agent (WEEK_PERSON) and a reader (WEEK_FORBIDDEN), so the page, MCP and terminal all get the check.
2. Tests. test/week-doors.test.ts is new, and I fixed the weak "reader" test:

- Cut line: once capacity is read, /next draws exactly one separator, and the terminal draws the same line. - POST /week: a request with no cookie, another origin, a reader or an agent is refused and takes nothing. - MCP week take: an agent is refused, and a person takes a proposal once. - Reader and agent: takeProposal returns the right message for each and records nothing. - Terminal reader: a reader's key takes nothing. - No agent runs: the run row count is unchanged by pricing, the /next page, or tickWeeks.

1. Privacy.

- Stuck reasons: they are now a failure class ("its check failed", "timed out" and so on), never the run's output, and a test checks a planted secret never reaches the stored retrospective. - History and spend: issueHistory now excludes sealed issues, and windowSpend and windowShipped already did. A test shows a sealed issue's cost doesn't change visible ranges.

1. Design.

- Cut line and ranges: the cut line is now .cut-line and a row's range is .estimate, both token-styled in content.ts. They are named in docs/design/areas/week.md, and design-lint is clean.

1. "Moves live".

- A reload: a drag posts the move and the page reloads itself from the server, which redraws the line. There is no manual reload, but it is a reload, not a client-side recompute. The docs now say so.

One disagreement: I did not add a client-side recompute of the line while dragging. The server stays the only place the line is computed, which keeps "same state, same line" true on every door. If you want the line to move before the save returns, say so.

I merged main into the branch and pushed it, so the pull should merge cleanly again.

There was one conflict, in test/conversation-confirm.test.ts. It was the list of MCP tools that write without asking first. main added keep_plan and this branch added week; I kept both.

bun x tsc --noEmit and bun scripts/gate.ts are clean. A rerun of bun scripts/mcp-tools.ts changed nothing. I ran the week, Next up, MCP, palette, migration and timing tests, 331 in all, and they pass. I did not run the full suite; the forge's check does that on push.

Left #8704 open: 1 unticked item remain (A cut line on the list shows what fits in the week's remaining capacity, and moves live as the list is dragged. Issues below it read as next week.).