Knobs: one panel on the crew page with five plain dials, each showing its live effect, editable in place #10599

I merged main into the branch and pushed it. The merge commit is 8d1771317.

The only conflict was in test/child-quota.test.ts, in the "room for small but not large" test. My side had typed FORGE_CHILD_MEMORY_MIB = "4200". main computes the room from heldByChildren(dir).memoryMiB + SIZES.small.memoryMiB, so a resized small doesn't break the test. I kept main's version. It covers what my line was for.

• child-quota.test.ts: passes, 4 tests.
• bun scripts/gate.ts: reported nothing.
• bun x tsc --noEmit: one error, in test/crew-sandbox-escape.test.ts: it can't find the module @vscode/ripgrep. That file and package aren't part of this merge, and I didn't investigate further. It looks like a missing install in this clone.
• Not run here: the rest of the suite. The forge runs the full check on this push.

Continues #10599

I merged main into the branch and pushed it. The merge commit is 8d1771317.

The only conflict was in test/child-quota.test.ts, in the "room for small but not large" test. My side had typed FORGE_CHILD_MEMORY_MIB = "4200". main computes the room from heldByChildren(dir).memoryMiB + SIZES.small.memoryMiB, so a resized small doesn't break the test. I kept main's version. It covers what my line was for.

• child-quota.test.ts: passes, 4 tests.
• bun scripts/gate.ts: reported nothing.
• bun x tsc --noEmit: one error, in test/crew-sandbox-escape.test.ts: it can't find the module @vscode/ripgrep. That file and package aren't part of this merge, and I didn't investigate further. It looks like a missing install in this clone.
• Not run here: the rest of the suite. The forge runs the full check on this push.

Five plain dials now sit on the crew page, but none of the six removals you listed is in this pull. I held them back because the direction says each removal is its own small pull with its own test, and I can't open pulls.

The gate, tsc, design-lint and 143 tests across the Knobs suites, MCP, palette roster, confirm, child quota, pace and week all passed before I pushed. Those are the Knobs-related tests, not the full suite. I don't have the log from the check that went red twice, and everything I can run here passes. My guess is the releases restarting the pod killed those runs. That is unconfirmed.

What changed

• Panel: one admin-only card on the crew page, src/web/pages/knobs.ts, with one row per dial. Each row shows its value, a one-line meaning, its live effect and the control. The five are Speed, Weekly budget, Polish share, Autonomy per repository (off, ask, go) and Feedback first (shown, not tuned).
• Same state everywhere: the dials are also in the terminal and the MCP, and they write the same settings events the Settings pages write. The logic is in src/crew/knobs.ts and src/actions/knobs.ts.
• Derived values: Speed derives crew.concurrency, the builder slots, the reserve, pace.floor and queue.depth. Weekly budget is a share of the week. Small share and the polish cap are no longer dials, which is part of the first listed removal.
• Merges: I merged origin/main into the branch. I took main's version of the child-quota test, which computes the room from the sizes. The fork's copy of the branch also had a commit I lacked, so I merged that too.

What I left out

• Still in the code: polish.capMinutes and slots.smallShare are not removed. src/crew/polish_cap.ts and smallShare() in src/core/capacity.ts are still there. Only the dials are gone.
• Not touched: the other five removals (the two planner holds, the stalls shun and breaker, the Try: gate, the reviewer's taste block, the design bullets in briefs).
• Next step: the planner should file the six as separate issues, each with a test for the behaviour that replaces it.

Try it Open the crew page as admin, change Speed or Polish share, and check that the live-effect line updates and the feed shows who turned the dial. To run the checks: bun test test/knobs.test.ts test/knobs-page.test.ts.

Knobs puts five dials on the crew page, in the terminal and in the MCP knobs tool: Speed, Weekly budget, Polish share, Autonomy (one row per repository) and Feedback first. The pipeline's own dials are no longer touched. tsc is clean, the design lint passes, and about 190 tests across the files this change reaches pass. The forge's full check hasn't run on this push yet.

Review round

• Pipeline dials restored. queue.depth, pace.floor, slots.smallShare and polish.capMinutes are back as capacity keys with their old behaviour. src/core/capacity.ts and the five tests I had bent around the removals (pace-floor, tui-capacity, week-pump, week-wiring, crew-small-share) now match main. The only new capacity key is week.use, the Weekly budget dial. Speed is plain crew.concurrency, and nothing is derived from it. The smallShare(_db) and polishCapMinutes(_db) stubs are gone with the restore.
• Terminal edit. e on a row opens a prompt and goes through setKnob, for admins who are people. Out-of-range values are refused and change nothing. Non-admins see no knobs to turn. Feedback first is shown and can't be turned. The new test/tui-knobs.test.ts covers a turn, a refusal and a member.
• Web panel. It is now a dl.status-rows list with hairlines, not a boxed section. Each row has the name, one live effect line and the control. The Set button is a quiet ghost button, so the page keeps one accent. The inline width became an .input-num class, named in docs/design/areas/knobs.md. The page has one intro sentence, and the old does line is now a tooltip on the name.
• Page tests added. A guest post, a non-admin post, out-of-range and unknown values, an attempt to set queue.depth, and Autonomy and Polish turns.
• Polish across repositories. The value is checked before any write. A write can't then fail for one repository and not the rest, because the admin check is the same for all of them. I left a comment saying so and didn't add a rollback.
• docs/features/knobs.md. It now says the pipeline dials are untouched and no stored value is dropped. I didn't document stored values for removed keys being ignored, because no key is removed any more.

Not done

• The per-repository dollar ceiling and the reserve are unchanged.
• The slots.smallShare test is the original again, so your point about its weak assertion no longer applies.

To try it Open Settings › Crew as an admin, or open Knobs in the terminal palette and press e.

I raised the statement budgets for the capacity page and pushed. I have not seen the forge's check run on the new push yet.

Cause: this pull adds the week.use setting, and the capacity page reads it. That adds statements to the page, which put it over its counted-work ceilings. Measured here:

The page does not do extra work per change; each setting it lists costs a few statements. I left that as is and kept the page's code unchanged.

Checks run here:

• test/pages-budget.test.ts and test/capacity-timing.test.ts pass.
• bun scripts/gate.ts and bun x tsc --noEmit produce no output.

Still unconfirmed: the check's other failure was test/scaffold-webspec.test.ts ("the scaffold did not boot", after 15 s). It exited with no failure when I ran it here, and I did not change anything it covers. I take it to be load-related, since the check ran 83 shards at once. If it fails again on this push, it needs its own look.

Branch is in sync with the fork and nothing new has been said on the issue, so I made no changes and pushed nothing.

The Knobs work was already pushed in my last run. I haven't re-run the check since then. That run had the gate, tsc, design-lint and the 143 Knobs-related tests passing.

The six removals from the owner's direction are still not done, and I don't plan to do them on this branch. Each is meant as its own small pull with its own test. To start one, the planner would file it as a separate issue.

Nothing has changed since last time, so I made no changes and pushed nothing. The branch matches the fork. The last comment on #10461 is still the owner's removals direction from 03:41 UTC, and the issue is still open.

The Knobs panel is already on the branch from the earlier run. I haven't re-run the gate, tsc or the tests in this run. The six removals from the owner's direction are still not done. They need to be filed as separate issues so each becomes its own small pull with its own test.