Estimates are recorded against an issue, start from the forge's own history,… #11964

The ESTIMATE_RULE doc comment at src/crew/brief.ts:373 is now wrapped to two lines under 100 columns. That was the only thing the check named.

bun scripts/gate.ts and bun x tsc --noEmit both pass locally, and the fix is pushed as 7d7044b98. I ran no tests, since the change is a comment only; the forge's check will run its fast tier on this push.

The gate only judges lines in committed history. My first rewrap was uncommitted and it kept reporting line 373, so I committed it before the gate would pass.

Nothing else in the pull changed: estimates are still recorded against an issue through the estimate tool and the estimate routes, start from the forge's history, and are scored.

Continues #11964

The ESTIMATE_RULE doc comment at src/crew/brief.ts:373 is now wrapped to two lines under 100 columns. That was the only thing the check named.

bun scripts/gate.ts and bun x tsc --noEmit both pass locally, and the fix is pushed as 7d7044b98. I ran no tests, since the change is a comment only; the forge's check will run its fast tier on this push.

The gate only judges lines in committed history. My first rewrap was uncommitted and it kept reporting line 373, so I committed it before the gate would pass.

Nothing else in the pull changed: estimates are still recorded against an issue through the estimate tool and the estimate routes, start from the forge's history, and are scored.

Gate and tsc were clean earlier (no output); branch is already pushed and up to date.

Estimates are now recorded against an issue, start from the forge's own history, and are scored when the work goes live. This is the branch as already pushed; I made no new commits this run.

• Tool: estimate works over MCP, POST and GET /api/v1/repos/:owner/:name/issues/:n/estimates, and the issue page.

- With no times it returns the base rate for that issue. - With event, p50 and p90 it records an estimate. Estimates are never edited; a new one supersedes the old and the revisions are kept.

• Resolution: the forge fills in the actual instant when the event happens. Issues closed as split, not-now or duplicate void their open estimates, which are kept and marked void.
• Scoring: each estimator is scored on median ratio and p90 hit rate. The base rate is scored as history.
• Shown on the page: the issue page and /next rows carry the promise line.
• Crew rule: the "by <time>" rule is in the task prompt.
• Migration: the migration is pinned.

Checks

• bun x tsc --noEmit and bun scripts/gate.ts printed no errors.
• 89 tests pass across test/estimates.test.ts, test/mcp.test.ts, test/route-timing-coverage.test.ts and test/conversation-confirm.test.ts.
• scripts/fast.ts said the diff was too broad to name tests, so I ran those four files, the ones the branch changed. The full check is left to the forge on push.
• A broader bun test I started in the background hit sandbox network denials; I did not wait for it.

Not done: the terminal surface and the /flow scorecard are not verified against a running preview.

To try: POST an estimate to /api/v1/repos/:owner/:name/issues/:n/estimates, then open the issue page and /next.

Estimates are recorded against an issue, start from the forge's own history, and are scored. All of it was already on the pushed branch, and this run changed nothing. The tree is clean and up to date with origin, so there was nothing to push.

What the branch does

• Tool: estimate is available over MCP, as POST and GET /api/v1/repos/:owner/:name/issues/:n/estimates, and on the issue page.

- With no times it returns the base rate for that issue. - With event, p50 and p90 it records an estimate. Estimates are never edited; a new one supersedes the old and the revisions are kept.

• Resolution: the forge fills in the actual instant when the event happens. Closing an issue as split, not-now or duplicate voids its open estimates; they are kept and marked void.
• Scoring: each estimator is scored on median ratio and p90 hit rate, and the base rate is scored as history.
• Shown on pages: the issue page and /next rows carry the promise line.
• Crew rule: the "by <time>" rule is in the task prompt.
• Migration: the migration is pinned.

Checks this run

• bun x tsc --noEmit and bun scripts/gate.ts are clean.
• test/estimates.test.ts passes (14 tests).
• Last run, test/estimates.test.ts, test/mcp.test.ts, test/route-timing-coverage.test.ts and test/conversation-confirm.test.ts passed together (89 tests).
• I did not run the full suite or try the running preview. The forge runs the full check on push.
• origin/main has 10 commits that are not on this branch (my fetch was in the command that was killed, so this count may be stale). I did not rebase or merge, so I haven't checked that the branch merges cleanly with them.

Not done

• The terminal surface and the /flow scorecard are not verified against a running preview.

To try: POST an estimate to /api/v1/repos/:owner/:name/issues/:n/estimates, then open the issue page and /next.

Estimates are recorded against an issue, start from the forge's own history, and are scored. All of this was already on the pushed branch, and this run changed nothing. The tree is clean and up to date with origin, so there was nothing to push.

• Tool: estimate is available over MCP, as POST and GET /api/v1/repos/:owner/:name/issues/:n/estimates, and on the issue page.

- With no times it returns the base rate for that issue. - With event, p50 and p90 it records an estimate. Estimates are never edited; a new one supersedes the old and the revisions are kept.

• Resolution: the forge fills in the actual instant when the event happens. Closing an issue as split, not-now or duplicate voids its open estimates; they are kept and marked void.
• Scoring: each estimator is scored on median ratio and p90 hit rate, and the base rate is scored as history.
• Shown on pages: the issue page and /next rows carry the promise line.
• Crew rule: the "by <time>" rule is in the task prompt.
• Migration: the migration is pinned.

What I checked this run

• A trial merge of origin/main, which is 11 commits ahead of the branch, completes without conflicts.
• I did not rerun bun x tsc --noEmit, bun scripts/gate.ts or the tests this run, because nothing changed. Last run bun x tsc --noEmit and bun scripts/gate.ts were clean and test/estimates.test.ts passed (14 tests). Earlier, those tests plus test/mcp.test.ts, test/route-timing-coverage.test.ts and test/conversation-confirm.test.ts passed together (89 tests).
• The full suite and the running preview were not exercised. The forge runs the full check on push.

Not done: the terminal surface and the /flow scorecard are not verified against a running preview.

To try: POST an estimate to /api/v1/repos/:owner/:name/issues/:n/estimates, then open the issue page and /next.

Estimates are now recorded against an issue: the pages that read them have their statement budgets raised, and the pushed branch type-checks. That budget bump was the only failure I could trace to this change. I did not re-run the full check, so the pull's verdict is still open.

What changed

• test/pages-budget.test.ts now allows one more statement on the issue page, 70 to 71, for the open-estimates read.
• The /next page goes from 64 to 65, for one read of open estimates covering every row.
• The unfixtured-routes cap goes from 147 to 148 for the new GET …/issues/:number/estimates.
• test/pages-budget.test.ts passes locally with these.

Failures I did not fix

• Four of the check's failures also fail on a clean origin/main checkout in this sandbox, in files this pull doesn't touch:

- no src/flow file reads the wall clock outside a default parameter: it flags src/flow/pipeline.ts. - background timers run inside during: it flags src/runs/log.ts:149. - every export under src/ has a caller: it flags mergedPullClosingScan and bufferRunLogs. - the assigned sweep drives from the agent's issues.

• The assigned-sweep timing test and the two "on production's size" timing tests in slow-pages.test.ts I did not run. The check ran them on a busy machine, so they may be load-related.
• The scaffold OpenID Connect test failed only when run alongside other tests, so I took it for a flake.

How to try it

• Run bun test test/estimates.test.ts test/pages-budget.test.ts.

🤖 Generated with Claude Code

All seven of sentinel's points and the terminal door are done. bun x tsc --noEmit and bun scripts/gate.ts are clean. I ran the estimates, page-budget, MCP, map, next, keeper, issue and flow tests, and they pass. I did not run the full suite: scripts/fast.ts named 2,607 of 2,990 files, so I picked these tests myself. The forge's check will run the rest on this push.

What changed

• Terminal door. New src/tui/screens/estimates.ts, listed in the repository screens in src/tui/forge.ts. It lists the open ready issues (the top 100) with their promise line. a asks the base rate for the issue under the cursor and r records one, typed as event p50 p90 basis…. It gates as the page does: canRead and sealedScope to list, mayView on the issue to ask, and recordEstimate itself (canWrite) to record. The how-to moved from docs/FLOW.md to docs/features/estimates.md; FLOW.md keeps a one-line pointer.
• Leak across users. allEstimates is replaced by estimatesVisibleTo(db, viewer, byId?), which uses the same visibility SQL as the repository lists. /flow and both asker scorecards read through it.
• Bounded reads.

- historySamples now reads only the closed issues the merged pulls name, via json_each. - currentByNumber reads only the listed numbers, at most 500. - The scorecard reads the newest 5,000 rows.

• Voiding. The migration comment, the code comments and the docs now say dropped, split, or moved to Not now. CloseReason has no "duplicate", so a duplicate closes as a drop. New tests cover done (does not void), dropped, split and Not now.
• estimateIssue comment fixed: it only checks the issue is not a pull; recordEstimate checks open.
• Flow section is now one sentence.
• Page budget.

- The raised caps (70→71, 64→65, 147→148) are back at main's values. - Issue gains an estimated flag (an EXISTS inside the existing issue read). The issue page and /next skip the estimates read when nothing was ever recorded. - The estimates API route now has a fixture in test/pages-budget.test.ts, so it is walked instead of counted as unfixtured. - test/map.test.ts builds an Issue by hand, so it gained estimated: false.

• Main is merged.
• Layering. core/issues.ts, core/pulls.ts and keeper/record.ts are back to main's versions. src/actions/estimates_settle.ts is the one place that settles, subscribed on the bus in app.ts:

- pull.opened, pull.merged, and issue.closed as dropped or split settle or void their estimates. - release.verified_live resolves the live estimates. - The Not now void stays in actions/next.ts, the action that does it.

Behaviour changes to check

• live now resolves at release.verified_live, not when the release first answers (markLive). A release that rolls back no longer resolves anything; the base rate already excluded rolled-back releases.
• settleLive now takes the list of served releases and picks the first cut after the merge.

Not done

• live still resolves only for the forge's own releases, not an app's production deploy.

To try

• Open a repository in the terminal, pick Estimates in a repository, press a, then r.
• Visit /flow signed out: private repositories' estimators no longer appear.
• Run bun test test/estimates.test.ts.

The check failed on the gate's 100-column comment rule and its "no as never in tests" rule. I fixed both and pushed.

• Comment lines: I rewrapped the four lines the gate named: src/app.ts:262, src/core/issues.ts:69, and two comments in test/estimates.test.ts.
• Key press in the terminal test: the as never cast on the key event now builds a real event with keyFor(key).

With these commits applied, bun scripts/gate.ts and bun x tsc --noEmit pass here, and test/estimates.test.ts passes (18 tests). The gate only recognised the fixes after I committed them.

The forge's check will run the rest of the suite on this push.

The only failure this pull caused was test/palette-roster.test.ts, and it now passes; I pushed the fix as 280feed46. The forge will re-run the full check on that push.

The cause. The terminal's new "Estimates in a repository" entry (estimates, hint owner/name) appears in the palette. The test pins the palette roster for each kind of session, so it failed in four places. I added that entry before "The week in a repository" in all four expectations. The test was out of date, not the code. bun x tsc --noEmit is clean.

Left alone. Four other test files were already red on main at a59cf33, and the check log says they aren't counted against this pull:

• flow-clock
• unused-exports
• timers-named
• stall-scans

I re-ran stall-scans here and its failure is the same one the check reported. I didn't run the other three, and none of the four touch the estimates change.