The week's subscription is spent by reset: the crew paces to it and pulls the next ranked issues in itself #8701

The crew now paces itself to the week: home shows an on-pace/ahead/behind sentence with efficiency beside it, and the autonomy tick releases gated work when it falls behind. The auto-release itself has no test, because the pure decision is covered but the driveRepo wiring is not. tsc, scripts/gate.ts, design-lint and the covering tests pass: week, pacing, the autonomy suites, hot-reads-indexed and the home tests. scripts/fast.ts named 1321 test files for this diff, so I ran those covering suites by hand instead of the full list.

What changed

• Pace and settings: src/crew/week.ts is a pure decision. From the week's crew capacity, what is spent and the time to the reset, it works out a target burn per day and a status. "On pace" is within 5% of an even burn. Everything is in a plain WEEK object (tolerance, reserve, floor slots) that a person can read.
• Home sentence: src/web/studio/week_line.ts draws it under the figures, with the efficiency line next to it. It is empty while no account's week has a known cap.
• Behind pace: src/crew/autonomy.ts has a new releaseForPace. If builders are idle and nothing is startable, the tick releases the lowest-numbered issue the Later gate alone holds. It does this once per tick, writes a log line, and leaves a comment on the issue saying why. The tick now returns "released" for that.
• Ahead of pace: repoPace shrinks slots in proportion to how far ahead the crew is, never below one while the reserve holds.
• Reserve: 10% of the week's crew capacity is kept for /talk and a person's own requests. At that point slots go to zero until the reset.
• Efficiency: src/crew/week_read.ts reports cost per real landing and runs per real landing since the week opened. Carried pulls (already on the base) and pulls that added nothing are not counted. It uses existing indexes (pull_merged, run_agent_latest), so there is no migration.
• Docs: docs/features/week-pace.md.
• Tests: test/week.test.ts covers the same state giving the same decision, each status, the release conditions, the slot cut ahead of pace, the reserve wall, and the efficiency query.

What I chose not to do

• "Next ranked" is lowest issue number among the gate-only-held issues, not the map's priority order.
• The home sentence and efficiency cover the first autonomous repo, as the idle reason does.
• The home sentence uses zero idle slots and zero startable, so it names the status but never shows a release decision.
• Whether /talk and a person's own runs actually draw from that reserve has not been checked. I did not trace how those runs are counted, so the reserve is a dollar wall on crew slots only.

To try it: open home on a forge whose subscription has a known weekly cap, or run bun test test/week.test.ts.

Closes #8695

The crew now paces itself to the week: home shows an on-pace/ahead/behind sentence with efficiency beside it, and the autonomy tick releases gated work when it falls behind. The auto-release itself has no test, because the pure decision is covered but the driveRepo wiring is not. tsc, scripts/gate.ts, design-lint and the covering tests pass: week, pacing, the autonomy suites, hot-reads-indexed and the home tests. scripts/fast.ts named 1321 test files for this diff, so I ran those covering suites by hand instead of the full list.

What changed

• Pace and settings: src/crew/week.ts is a pure decision. From the week's crew capacity, what is spent and the time to the reset, it works out a target burn per day and a status. "On pace" is within 5% of an even burn. Everything is in a plain WEEK object (tolerance, reserve, floor slots) that a person can read.
• Home sentence: src/web/studio/week_line.ts draws it under the figures, with the efficiency line next to it. It is empty while no account's week has a known cap.
• Behind pace: src/crew/autonomy.ts has a new releaseForPace. If builders are idle and nothing is startable, the tick releases the lowest-numbered issue the Later gate alone holds. It does this once per tick, writes a log line, and leaves a comment on the issue saying why. The tick now returns "released" for that.
• Ahead of pace: repoPace shrinks slots in proportion to how far ahead the crew is, never below one while the reserve holds.
• Reserve: 10% of the week's crew capacity is kept for /talk and a person's own requests. At that point slots go to zero until the reset.
• Efficiency: src/crew/week_read.ts reports cost per real landing and runs per real landing since the week opened. Carried pulls (already on the base) and pulls that added nothing are not counted. It uses existing indexes (pull_merged, run_agent_latest), so there is no migration.
• Docs: docs/features/week-pace.md.
• Tests: test/week.test.ts covers the same state giving the same decision, each status, the release conditions, the slot cut ahead of pace, the reserve wall, and the efficiency query.

What I chose not to do

• "Next ranked" is lowest issue number among the gate-only-held issues, not the map's priority order.
• The home sentence and efficiency cover the first autonomous repo, as the idle reason does.
• The home sentence uses zero idle slots and zero startable, so it names the status but never shows a release decision.
• Whether /talk and a person's own runs actually draw from that reserve has not been checked. I did not trace how those runs are counted, so the reserve is a dollar wall on crew slots only.

To try it: open home on a forge whose subscription has a known weekly cap, or run bun test test/week.test.ts.

I could not make "behind pace" release bet work: bets don't exist in the repo yet (#8713 is still open), so it still releases the lowest-numbered issue only the Later gate holds. This push adds only a doc note; no code changed, so nothing was re-run.

docs/features/week-pace.md now has a short "Bets (#8713)" section. It says the gate-held-issue release is a stand-in. Once bets land, being behind should release the next scope of a backed bet, or offer the next bet to back, in its place. The change is one candidate read in releaseForPace (src/crew/autonomy.ts).

On appetites sitting inside the week's capacity: the pace decision already works out the week's crew capacity (capUsd less the 10% reserve). Bet shares can be taken from that once bets exist, so no bet's share can exceed it.

I did not add a bet interface with no implementation behind it, because it would be dead code. Wiring bets into the release is left for whoever builds #8713, or a follow-up issue once it lands.

Everything from the first push stands: the on-pace/ahead/behind sentence on home with efficiency beside it, the automatic release one issue at a time with a logged reason, slots coming down when ahead, and the reserve for /talk and people. src/crew/week.ts holds the plain-value settings and test/week.test.ts covers the decision.

Two gaps remain from before. The auto-release wiring in the tick has no test, since the decision itself is covered but the driveRepo path is not. I also did not trace how /talk and a person's own runs draw on the reserve, so it is a dollar wall on crew slots only.

To try it, open home on a forge whose subscription has a known weekly cap, or run bun test test/week.test.ts.

🤖 Generated with Claude Code

This round addresses four of juniper's five points and holds the direction question for you. The tsc, scripts/gate.ts and design-lint checks are clean. 91 tests pass across the week, pacing, autonomy, dashboard and home suites. I did not run the full suite; the forge's check does.

What changed

• Both doors:

- Shared sentence: the pace sentence and efficiency now come from one function, weekWords in src/crew/week_read.ts. Home (weekLine) and the terminal dashboard both draw from it. - Terminal gating: in the terminal it sits in the admin-only figures block, next to the idle-reason line. That matches home, where the kpi figures are admin-only too. - No new screen: I used the existing dashboard rather than a new palette entry, because the line belongs with the figures.

• Tests: the new test/week-wiring.test.ts seeds a real subscription with a known weekly cap, and these tests fail if the wiring is removed.

- Release: when behind, with an idle slot and nothing startable, driveRepo returns "released". The Later-gate edge is removed and the "why" comment is on the issue. - No release on pace: the gated issue stays gated when the crew is on pace. - Ahead cut: repoPace goes from 4 to 2 slots when ahead and to 0 at the reserve. - Both doors: the home line and the terminal dashboard print the sentence, and with no known cap there is none.

• Perf: repoPace no longer reads standings on every call. The week's figures are remembered per repository for 5 seconds, keyed on the now passed in.
• Cleanup: weekFor defaults its seen argument, so weekLine no longer passes unused zeros.
• Reserve: I reworded docs/features/week-pace.md so it no longer claims more than the code does. The reserve is a wall on crew slots (a tenth of the week's capacity), and /talk and a person's own runs are not otherwise counted against it. I did not wire it further.

Not done

• Direction (#3): I have not justified why this goes before the pipeline list (#6534 and the rest). That is for you to decide: hold this pull until the pipeline items land, or merge it. The change is self-contained and can wait.
• Preview (#5): the sentence is now shown by the home-line and terminal tests on a seeded cap. I did not re-check the live preview, which has no weekly cap, so home there will still be blank.

Still open from before

• Bets: gate-held release stands in until bets (#8713) exist.
• Remembered week: a subscription's week can lag by up to 5 seconds in production.

🤖 Generated with Claude Code