style: Biome format baseline + check-mode CI gate (outside apps/desktop) (#415)

[UncertaintyDeterminesYou4ndMe]

The formatter follow-up promised in #1019 — one atomic PR, structured for a merge moment of your choosing. Closes the formatter half of #415.

Structure (3 commits)

1. Config + gate — formatter config restored in biome.jsonc (2-space / 100-col / single quotes / semicolons / trailing commas), format + format:check scripts, and npm run format:check in the CI typecheck job (right after lint, before build).
2. Baseline — mechanical biome format --write to fixpoint: 805 files, zero logic changes. This is the commit to point .git-blame-ignore-revs at.
3. .git-blame-ignore-revs — carries the baseline commit's branch hash. ⚠️ If you squash/rebase-merge, the hash changes; the file has the one-liner to regenerate it (git log --oneline --grep 'Biome format baseline' -1) — happy to push the corrected hash post-merge instead if you prefer merge-commit-off.

⚠️ The discovery you should weigh in on: apps/desktop is formatter-excluded

Running the baseline over the whole tree broke 145 assertions across 64 desktop test files. Desktop has a pervasive source-contract test style — tests read .ts/.tsx source and regex-match exact single-line shapes (e.g. account-auth-ui.test.ts matches a full <div className=... aria-label=...> line). Any reformat — Biome's or a future hand edit — snaps them.

Rather than scattering ~145 // biome-ignore format: comments through desktop source (which would exempt much of desktop from the formatter this PR introduces — self-defeating), this PR excludes apps/desktop/** from the formatter only (it stays fully linted), and defers desktop to a follow-up. Options for that follow-up, your call:

• (a) Make the source-contract assertions whitespace-tolerant first, then format desktop. Two shared helpers (main-process-contract-source-helpers.js, 20 files; provider-contract-source-helpers.js, 6 files) could centralize normalization, but many exact-spacing regexes are inline per test — non-trivial.
• (b) Keep desktop formatter-excluded long-term and accept the split.
• (c) Treat the source-contract style itself as the issue (they're fragile against any edit, not just Biome) and migrate them to AST-based checks — bigger conversation.

Outside desktop, the same class of problem existed at small scale and IS handled in the baseline: 5 statements carry a // biome-ignore format: guard — two @ts-expect-error directives that must stay adjacent to the line they suppress (proxy-test.ts, run-trace.ts — splitting the statement orphans the directive: TS2578 + the suppressed error resurfaces) and three single-line shapes asserted by source-contract tests (run-prompt-ab.mjs, run-prompt-optimization.mjs, run-harness-ab.mjs).

Also excluded (byte-sensitive, not hand-authored): apps/desktop/resources/licenses/** (SOURCE.json is whole-file SHA-256'd + size-checked by prepare-cua-driver.mjs) and bundled-tools.json.

Verification

• npm run lint ✅ and npm run format:check ✅ (formatter run to fixpoint — Biome's formatter isn't idempotent on some multi-line generic signatures, so one extra pass was needed)
• npm run build ✅, npm run typecheck ✅
• Full npm test: scripts 109, core 1067, storage 287, runtime 1983, computer-use 135, desktop (untouched) all green; headless 1063/1067 with only the 2 known-flaky opencode stop runner subtests (fail identically on unmodified main, pass in isolation)

Merge timing

Touches ~805 files, so it conflicts with most open PRs — merge whenever suits; I'll rebase + re-run the baseline to fixpoint on request the same day you want it landed (regenerating is cheap and deterministic, ~1 min).

Heads-up per the #1019 thread: biome lint --write --unsafe still hangs on apps/desktop (~100% CPU indefinitely; plain format is fine) — upstream minimization still on my list, tracked separately from this PR.

[UncertaintyDeterminesYou4ndMe]

CI's test failure was real and informative — fixed and force-pushed.

Root cause: desktop's source-contract tests don't only introspect desktop's own source — ~45 contracts read packages/ui source too (composer.tsx, daily-review-panel.tsx, primitives/*.tsx, …), and packages/ui was in the baseline. So the exclusion boundary isn't "apps/desktop", it's the whole UI surface.

Also owning a verification gap on my side: locally npm test short-circuits at the 2 known-flaky headless subtests (the script chains suites with &&), so the desktop/ui suites never actually ran after my rebase — I've re-verified per-suite explicitly this time.

Changes in the force-push:

• Formatter now excludes apps/desktop/**andpackages/ui/** (both still fully linted); baseline shrinks 805 → 660 files (packages/ui's 142 reverted, byte-identical to main).
• .git-blame-ignore-revs updated to the new baseline hash.

Verified: npm run lint ✅, npm run format:check ✅ (fixpoint), build ✅, typecheck ✅, and explicitly: @maka/ui ✅, @maka/desktop ✅ (2,571 tests), @maka/storage ✅.

This also sharpens the follow-up question: the source-contract web spans the whole UI surface (desktop + ui, 190+ single-line assertions), so making those contracts format-tolerant (or AST-based) is the real gate for ever formatting UI code — option (a)/(c) from the PR description.

[Astro-Han]

Approved.

Non-blocking P3 follow-up: this repository squash-merges PRs, so the branch-only baseline hash in .git-blame-ignore-revs will not exist in a fresh clone after merge; git blame configured with the file then exits with "cannot find revision ... to ignore". Please remove the branch hash before merge and add the resulting squash commit hash in a follow-up, or otherwise update the entry immediately after merge.

[UncertaintyDeterminesYou4ndMe]

Done — went with your first option. The branch hash is removed from .git-blame-ignore-revs before merge; the file now ships as documentation/scaffold only, with a note that entries must be hashes that exist on main (you're right that a hash unknown to the clone makes git blame hard-error, not just no-op — stricter than my original in-file note assumed). I'll open the one-line follow-up adding the squash-commit hash immediately after this lands.

One operational note: GitHub hasn't been starting CI runs on my last few pushes to this branch ("no checks reported", no check-suites on the head SHA — not the fork-approval gate). The commits are content-identical to the previously verified tree apart from the .git-blame-ignore-revs text change, and everything is verified locally (lint / format:check fixpoint / build / typecheck / per-suite tests including desktop's 2,571). If checks stay absent on your side, a close/reopen should kick a fresh pull_request event — or happy to try anything else you prefer.

[Astro-Han]

LGTM, but there is a lot of rebase to be done.

[likun666661]

LGTM, but there is a lot of rebase to be done.

[UncertaintyDeterminesYou4ndMe]

已在最新 main(8b9d69a)上重建整个基线(force-push,a385f96b)。全仓机械格式化 PR 的 rebase 策略是重新生成而非解决 660 文件的文本冲突:配置 commit 重放 → 重新跑 biome format(750 文件)→ 补 ignore-revs 脚手架(占位设计不变)。

重建中发现并处理了两个机械格式化的实际危害(已写进基线 commit message):

1. proxy-test.ts 的 @ts-expect-error 因 fetch 调用被拆行而错位(build 失败)→ 指令移到它压制的属性行;
2. Biome 对两个文件的首遍输出不满足自身 format:check(非幂等边界)→ 二遍收敛;
3. 一个源码契约测试(tool-runtime-extraction-contract)把 RunTracePhase 联合钉成单行字面量 → 改为对格式容忍的正则(契约是变体集合,不是排版)。

验证:format:check / lint / build / typecheck 干净,全套件通过——唯一失败是 session-title.test.js 的「aborts title generation when the provider exceeds its deadline」,它在 pristine origin/main 上同样失败(#1199 引入,套件内稳定挂、单文件跑通过,cancelledByParent 挂起类),与本 PR 无关,另行报告。

[UncertaintyDeterminesYou4ndMe]

The formatter follow-up promised in #1019 — one atomic PR, structured for a merge moment of your choosing. Closes the formatter half of #415.

Structure (3 commits)

1. Config + gate — formatter config restored in biome.jsonc (2-space / 100-col / single quotes / semicolons / trailing commas), format + format:check scripts, and npm run format:check in the CI typecheck job (right after lint, before build).
2. Baseline — mechanical biome format --write to fixpoint: 805 files, zero logic changes. This is the commit to point .git-blame-ignore-revs at.
3. .git-blame-ignore-revs — carries the baseline commit's branch hash. ⚠️ If you squash/rebase-merge, the hash changes; the file has the one-liner to regenerate it (git log --oneline --grep 'Biome format baseline' -1) — happy to push the corrected hash post-merge instead if you prefer merge-commit-off.

⚠️ The discovery you should weigh in on: apps/desktop is formatter-excluded

Running the baseline over the whole tree broke 145 assertions across 64 desktop test files. Desktop has a pervasive source-contract test style — tests read .ts/.tsx source and regex-match exact single-line shapes (e.g. account-auth-ui.test.ts matches a full <div className=... aria-label=...> line). Any reformat — Biome's or a future hand edit — snaps them.

Rather than scattering ~145 // biome-ignore format: comments through desktop source (which would exempt much of desktop from the formatter this PR introduces — self-defeating), this PR excludes apps/desktop/** from the formatter only (it stays fully linted), and defers desktop to a follow-up. Options for that follow-up, your call:

• (a) Make the source-contract assertions whitespace-tolerant first, then format desktop. Two shared helpers (main-process-contract-source-helpers.js, 20 files; provider-contract-source-helpers.js, 6 files) could centralize normalization, but many exact-spacing regexes are inline per test — non-trivial.
• (b) Keep desktop formatter-excluded long-term and accept the split.
• (c) Treat the source-contract style itself as the issue (they're fragile against any edit, not just Biome) and migrate them to AST-based checks — bigger conversation.

Outside desktop, the same class of problem existed at small scale and IS handled in the baseline: 5 statements carry a // biome-ignore format: guard — two @ts-expect-error directives that must stay adjacent to the line they suppress (proxy-test.ts, run-trace.ts — splitting the statement orphans the directive: TS2578 + the suppressed error resurfaces) and three single-line shapes asserted by source-contract tests (run-prompt-ab.mjs, run-prompt-optimization.mjs, run-harness-ab.mjs).

Also excluded (byte-sensitive, not hand-authored): apps/desktop/resources/licenses/** (SOURCE.json is whole-file SHA-256'd + size-checked by prepare-cua-driver.mjs) and bundled-tools.json.

Verification

• npm run lint ✅ and npm run format:check ✅ (formatter run to fixpoint — Biome's formatter isn't idempotent on some multi-line generic signatures, so one extra pass was needed)
• npm run build ✅, npm run typecheck ✅
• Full npm test: scripts 109, core 1067, storage 287, runtime 1983, computer-use 135, desktop (untouched) all green; headless 1063/1067 with only the 2 known-flaky opencode stop runner subtests (fail identically on unmodified main, pass in isolation)

Merge timing

Touches ~805 files, so it conflicts with most open PRs — merge whenever suits; I'll rebase + re-run the baseline to fixpoint on request the same day you want it landed (regenerating is cheap and deterministic, ~1 min).

Heads-up per the #1019 thread: biome lint --write --unsafe still hangs on apps/desktop (~100% CPU indefinitely; plain format is fine) — upstream minimization still on my list, tracked separately from this PR.

[UncertaintyDeterminesYou4ndMe]

CI's test failure was real and informative — fixed and force-pushed.

Root cause: desktop's source-contract tests don't only introspect desktop's own source — ~45 contracts read packages/ui source too (composer.tsx, daily-review-panel.tsx, primitives/*.tsx, …), and packages/ui was in the baseline. So the exclusion boundary isn't "apps/desktop", it's the whole UI surface.

Also owning a verification gap on my side: locally npm test short-circuits at the 2 known-flaky headless subtests (the script chains suites with &&), so the desktop/ui suites never actually ran after my rebase — I've re-verified per-suite explicitly this time.

Changes in the force-push:

• Formatter now excludes apps/desktop/**andpackages/ui/** (both still fully linted); baseline shrinks 805 → 660 files (packages/ui's 142 reverted, byte-identical to main).
• .git-blame-ignore-revs updated to the new baseline hash.

Verified: npm run lint ✅, npm run format:check ✅ (fixpoint), build ✅, typecheck ✅, and explicitly: @maka/ui ✅, @maka/desktop ✅ (2,571 tests), @maka/storage ✅.

This also sharpens the follow-up question: the source-contract web spans the whole UI surface (desktop + ui, 190+ single-line assertions), so making those contracts format-tolerant (or AST-based) is the real gate for ever formatting UI code — option (a)/(c) from the PR description.

[Astro-Han]

Approved.

Non-blocking P3 follow-up: this repository squash-merges PRs, so the branch-only baseline hash in .git-blame-ignore-revs will not exist in a fresh clone after merge; git blame configured with the file then exits with "cannot find revision ... to ignore". Please remove the branch hash before merge and add the resulting squash commit hash in a follow-up, or otherwise update the entry immediately after merge.

[UncertaintyDeterminesYou4ndMe]

Done — went with your first option. The branch hash is removed from .git-blame-ignore-revs before merge; the file now ships as documentation/scaffold only, with a note that entries must be hashes that exist on main (you're right that a hash unknown to the clone makes git blame hard-error, not just no-op — stricter than my original in-file note assumed). I'll open the one-line follow-up adding the squash-commit hash immediately after this lands.

One operational note: GitHub hasn't been starting CI runs on my last few pushes to this branch ("no checks reported", no check-suites on the head SHA — not the fork-approval gate). The commits are content-identical to the previously verified tree apart from the .git-blame-ignore-revs text change, and everything is verified locally (lint / format:check fixpoint / build / typecheck / per-suite tests including desktop's 2,571). If checks stay absent on your side, a close/reopen should kick a fresh pull_request event — or happy to try anything else you prefer.

[Astro-Han]

LGTM, but there is a lot of rebase to be done.

[likun666661]

LGTM, but there is a lot of rebase to be done.

[UncertaintyDeterminesYou4ndMe]

已在最新 main(8b9d69a)上重建整个基线(force-push,a385f96b)。全仓机械格式化 PR 的 rebase 策略是重新生成而非解决 660 文件的文本冲突:配置 commit 重放 → 重新跑 biome format(750 文件)→ 补 ignore-revs 脚手架(占位设计不变)。

重建中发现并处理了两个机械格式化的实际危害(已写进基线 commit message):

1. proxy-test.ts 的 @ts-expect-error 因 fetch 调用被拆行而错位(build 失败)→ 指令移到它压制的属性行;
2. Biome 对两个文件的首遍输出不满足自身 format:check(非幂等边界)→ 二遍收敛;
3. 一个源码契约测试(tool-runtime-extraction-contract)把 RunTracePhase 联合钉成单行字面量 → 改为对格式容忍的正则(契约是变体集合,不是排版)。

验证:format:check / lint / build / typecheck 干净,全套件通过——唯一失败是 session-title.test.js 的「aborts title generation when the provider exceeds its deadline」,它在 pristine origin/main 上同样失败(#1199 引入,套件内稳定挂、单文件跑通过,cancelledByParent 挂起类),与本 PR 无关,另行报告。

[UncertaintyDeterminesYou4ndMe]

The formatter follow-up promised in #1019 — one atomic PR, structured for a merge moment of your choosing. Closes the formatter half of #415.

Structure (3 commits)

1. Config + gate — formatter config restored in biome.jsonc (2-space / 100-col / single quotes / semicolons / trailing commas), format + format:check scripts, and npm run format:check in the CI typecheck job (right after lint, before build).
2. Baseline — mechanical biome format --write to fixpoint: 805 files, zero logic changes. This is the commit to point .git-blame-ignore-revs at.
3. .git-blame-ignore-revs — carries the baseline commit's branch hash. ⚠️ If you squash/rebase-merge, the hash changes; the file has the one-liner to regenerate it (git log --oneline --grep 'Biome format baseline' -1) — happy to push the corrected hash post-merge instead if you prefer merge-commit-off.

⚠️ The discovery you should weigh in on: apps/desktop is formatter-excluded

Running the baseline over the whole tree broke 145 assertions across 64 desktop test files. Desktop has a pervasive source-contract test style — tests read .ts/.tsx source and regex-match exact single-line shapes (e.g. account-auth-ui.test.ts matches a full <div className=... aria-label=...> line). Any reformat — Biome's or a future hand edit — snaps them.

Rather than scattering ~145 // biome-ignore format: comments through desktop source (which would exempt much of desktop from the formatter this PR introduces — self-defeating), this PR excludes apps/desktop/** from the formatter only (it stays fully linted), and defers desktop to a follow-up. Options for that follow-up, your call:

• (a) Make the source-contract assertions whitespace-tolerant first, then format desktop. Two shared helpers (main-process-contract-source-helpers.js, 20 files; provider-contract-source-helpers.js, 6 files) could centralize normalization, but many exact-spacing regexes are inline per test — non-trivial.
• (b) Keep desktop formatter-excluded long-term and accept the split.
• (c) Treat the source-contract style itself as the issue (they're fragile against any edit, not just Biome) and migrate them to AST-based checks — bigger conversation.

Outside desktop, the same class of problem existed at small scale and IS handled in the baseline: 5 statements carry a // biome-ignore format: guard — two @ts-expect-error directives that must stay adjacent to the line they suppress (proxy-test.ts, run-trace.ts — splitting the statement orphans the directive: TS2578 + the suppressed error resurfaces) and three single-line shapes asserted by source-contract tests (run-prompt-ab.mjs, run-prompt-optimization.mjs, run-harness-ab.mjs).

Also excluded (byte-sensitive, not hand-authored): apps/desktop/resources/licenses/** (SOURCE.json is whole-file SHA-256'd + size-checked by prepare-cua-driver.mjs) and bundled-tools.json.

Verification

• npm run lint ✅ and npm run format:check ✅ (formatter run to fixpoint — Biome's formatter isn't idempotent on some multi-line generic signatures, so one extra pass was needed)
• npm run build ✅, npm run typecheck ✅
• Full npm test: scripts 109, core 1067, storage 287, runtime 1983, computer-use 135, desktop (untouched) all green; headless 1063/1067 with only the 2 known-flaky opencode stop runner subtests (fail identically on unmodified main, pass in isolation)

Merge timing

Touches ~805 files, so it conflicts with most open PRs — merge whenever suits; I'll rebase + re-run the baseline to fixpoint on request the same day you want it landed (regenerating is cheap and deterministic, ~1 min).

Heads-up per the #1019 thread: biome lint --write --unsafe still hangs on apps/desktop (~100% CPU indefinitely; plain format is fine) — upstream minimization still on my list, tracked separately from this PR.

[UncertaintyDeterminesYou4ndMe]

CI's test failure was real and informative — fixed and force-pushed.

Root cause: desktop's source-contract tests don't only introspect desktop's own source — ~45 contracts read packages/ui source too (composer.tsx, daily-review-panel.tsx, primitives/*.tsx, …), and packages/ui was in the baseline. So the exclusion boundary isn't "apps/desktop", it's the whole UI surface.

Also owning a verification gap on my side: locally npm test short-circuits at the 2 known-flaky headless subtests (the script chains suites with &&), so the desktop/ui suites never actually ran after my rebase — I've re-verified per-suite explicitly this time.

Changes in the force-push:

• Formatter now excludes apps/desktop/**andpackages/ui/** (both still fully linted); baseline shrinks 805 → 660 files (packages/ui's 142 reverted, byte-identical to main).
• .git-blame-ignore-revs updated to the new baseline hash.

Verified: npm run lint ✅, npm run format:check ✅ (fixpoint), build ✅, typecheck ✅, and explicitly: @maka/ui ✅, @maka/desktop ✅ (2,571 tests), @maka/storage ✅.

This also sharpens the follow-up question: the source-contract web spans the whole UI surface (desktop + ui, 190+ single-line assertions), so making those contracts format-tolerant (or AST-based) is the real gate for ever formatting UI code — option (a)/(c) from the PR description.

[Astro-Han]

Approved.

Non-blocking P3 follow-up: this repository squash-merges PRs, so the branch-only baseline hash in .git-blame-ignore-revs will not exist in a fresh clone after merge; git blame configured with the file then exits with "cannot find revision ... to ignore". Please remove the branch hash before merge and add the resulting squash commit hash in a follow-up, or otherwise update the entry immediately after merge.

[UncertaintyDeterminesYou4ndMe]

Done — went with your first option. The branch hash is removed from .git-blame-ignore-revs before merge; the file now ships as documentation/scaffold only, with a note that entries must be hashes that exist on main (you're right that a hash unknown to the clone makes git blame hard-error, not just no-op — stricter than my original in-file note assumed). I'll open the one-line follow-up adding the squash-commit hash immediately after this lands.

One operational note: GitHub hasn't been starting CI runs on my last few pushes to this branch ("no checks reported", no check-suites on the head SHA — not the fork-approval gate). The commits are content-identical to the previously verified tree apart from the .git-blame-ignore-revs text change, and everything is verified locally (lint / format:check fixpoint / build / typecheck / per-suite tests including desktop's 2,571). If checks stay absent on your side, a close/reopen should kick a fresh pull_request event — or happy to try anything else you prefer.

[Astro-Han]

LGTM, but there is a lot of rebase to be done.

[likun666661]

LGTM, but there is a lot of rebase to be done.

[UncertaintyDeterminesYou4ndMe]

已在最新 main(8b9d69a)上重建整个基线(force-push,a385f96b)。全仓机械格式化 PR 的 rebase 策略是重新生成而非解决 660 文件的文本冲突:配置 commit 重放 → 重新跑 biome format(750 文件)→ 补 ignore-revs 脚手架(占位设计不变)。

重建中发现并处理了两个机械格式化的实际危害(已写进基线 commit message):

1. proxy-test.ts 的 @ts-expect-error 因 fetch 调用被拆行而错位(build 失败)→ 指令移到它压制的属性行;
2. Biome 对两个文件的首遍输出不满足自身 format:check(非幂等边界)→ 二遍收敛;
3. 一个源码契约测试(tool-runtime-extraction-contract)把 RunTracePhase 联合钉成单行字面量 → 改为对格式容忍的正则(契约是变体集合,不是排版)。

验证:format:check / lint / build / typecheck 干净,全套件通过——唯一失败是 session-title.test.js 的「aborts title generation when the provider exceeds its deadline」,它在 pristine origin/main 上同样失败(#1199 引入,套件内稳定挂、单文件跑通过,cancelledByParent 挂起类),与本 PR 无关,另行报告。

[UncertaintyDeterminesYou4ndMe]

The formatter follow-up promised in #1019 — one atomic PR, structured for a merge moment of your choosing. Closes the formatter half of #415.

Structure (3 commits)

1. Config + gate — formatter config restored in biome.jsonc (2-space / 100-col / single quotes / semicolons / trailing commas), format + format:check scripts, and npm run format:check in the CI typecheck job (right after lint, before build).
2. Baseline — mechanical biome format --write to fixpoint: 805 files, zero logic changes. This is the commit to point .git-blame-ignore-revs at.
3. .git-blame-ignore-revs — carries the baseline commit's branch hash. ⚠️ If you squash/rebase-merge, the hash changes; the file has the one-liner to regenerate it (git log --oneline --grep 'Biome format baseline' -1) — happy to push the corrected hash post-merge instead if you prefer merge-commit-off.

⚠️ The discovery you should weigh in on: apps/desktop is formatter-excluded

Running the baseline over the whole tree broke 145 assertions across 64 desktop test files. Desktop has a pervasive source-contract test style — tests read .ts/.tsx source and regex-match exact single-line shapes (e.g. account-auth-ui.test.ts matches a full <div className=... aria-label=...> line). Any reformat — Biome's or a future hand edit — snaps them.

Rather than scattering ~145 // biome-ignore format: comments through desktop source (which would exempt much of desktop from the formatter this PR introduces — self-defeating), this PR excludes apps/desktop/** from the formatter only (it stays fully linted), and defers desktop to a follow-up. Options for that follow-up, your call:

• (a) Make the source-contract assertions whitespace-tolerant first, then format desktop. Two shared helpers (main-process-contract-source-helpers.js, 20 files; provider-contract-source-helpers.js, 6 files) could centralize normalization, but many exact-spacing regexes are inline per test — non-trivial.
• (b) Keep desktop formatter-excluded long-term and accept the split.
• (c) Treat the source-contract style itself as the issue (they're fragile against any edit, not just Biome) and migrate them to AST-based checks — bigger conversation.

Outside desktop, the same class of problem existed at small scale and IS handled in the baseline: 5 statements carry a // biome-ignore format: guard — two @ts-expect-error directives that must stay adjacent to the line they suppress (proxy-test.ts, run-trace.ts — splitting the statement orphans the directive: TS2578 + the suppressed error resurfaces) and three single-line shapes asserted by source-contract tests (run-prompt-ab.mjs, run-prompt-optimization.mjs, run-harness-ab.mjs).

Also excluded (byte-sensitive, not hand-authored): apps/desktop/resources/licenses/** (SOURCE.json is whole-file SHA-256'd + size-checked by prepare-cua-driver.mjs) and bundled-tools.json.

Verification

• npm run lint ✅ and npm run format:check ✅ (formatter run to fixpoint — Biome's formatter isn't idempotent on some multi-line generic signatures, so one extra pass was needed)
• npm run build ✅, npm run typecheck ✅
• Full npm test: scripts 109, core 1067, storage 287, runtime 1983, computer-use 135, desktop (untouched) all green; headless 1063/1067 with only the 2 known-flaky opencode stop runner subtests (fail identically on unmodified main, pass in isolation)

Merge timing

Touches ~805 files, so it conflicts with most open PRs — merge whenever suits; I'll rebase + re-run the baseline to fixpoint on request the same day you want it landed (regenerating is cheap and deterministic, ~1 min).

Heads-up per the #1019 thread: biome lint --write --unsafe still hangs on apps/desktop (~100% CPU indefinitely; plain format is fine) — upstream minimization still on my list, tracked separately from this PR.

[UncertaintyDeterminesYou4ndMe]

CI's test failure was real and informative — fixed and force-pushed.

Root cause: desktop's source-contract tests don't only introspect desktop's own source — ~45 contracts read packages/ui source too (composer.tsx, daily-review-panel.tsx, primitives/*.tsx, …), and packages/ui was in the baseline. So the exclusion boundary isn't "apps/desktop", it's the whole UI surface.

Also owning a verification gap on my side: locally npm test short-circuits at the 2 known-flaky headless subtests (the script chains suites with &&), so the desktop/ui suites never actually ran after my rebase — I've re-verified per-suite explicitly this time.

Changes in the force-push:

• Formatter now excludes apps/desktop/**andpackages/ui/** (both still fully linted); baseline shrinks 805 → 660 files (packages/ui's 142 reverted, byte-identical to main).
• .git-blame-ignore-revs updated to the new baseline hash.

Verified: npm run lint ✅, npm run format:check ✅ (fixpoint), build ✅, typecheck ✅, and explicitly: @maka/ui ✅, @maka/desktop ✅ (2,571 tests), @maka/storage ✅.

This also sharpens the follow-up question: the source-contract web spans the whole UI surface (desktop + ui, 190+ single-line assertions), so making those contracts format-tolerant (or AST-based) is the real gate for ever formatting UI code — option (a)/(c) from the PR description.

[Astro-Han]

Approved.

Non-blocking P3 follow-up: this repository squash-merges PRs, so the branch-only baseline hash in .git-blame-ignore-revs will not exist in a fresh clone after merge; git blame configured with the file then exits with "cannot find revision ... to ignore". Please remove the branch hash before merge and add the resulting squash commit hash in a follow-up, or otherwise update the entry immediately after merge.

[UncertaintyDeterminesYou4ndMe]

Done — went with your first option. The branch hash is removed from .git-blame-ignore-revs before merge; the file now ships as documentation/scaffold only, with a note that entries must be hashes that exist on main (you're right that a hash unknown to the clone makes git blame hard-error, not just no-op — stricter than my original in-file note assumed). I'll open the one-line follow-up adding the squash-commit hash immediately after this lands.

One operational note: GitHub hasn't been starting CI runs on my last few pushes to this branch ("no checks reported", no check-suites on the head SHA — not the fork-approval gate). The commits are content-identical to the previously verified tree apart from the .git-blame-ignore-revs text change, and everything is verified locally (lint / format:check fixpoint / build / typecheck / per-suite tests including desktop's 2,571). If checks stay absent on your side, a close/reopen should kick a fresh pull_request event — or happy to try anything else you prefer.

[Astro-Han]

LGTM, but there is a lot of rebase to be done.

[likun666661]

LGTM, but there is a lot of rebase to be done.

[UncertaintyDeterminesYou4ndMe]

已在最新 main(8b9d69a)上重建整个基线(force-push,a385f96b)。全仓机械格式化 PR 的 rebase 策略是重新生成而非解决 660 文件的文本冲突:配置 commit 重放 → 重新跑 biome format(750 文件)→ 补 ignore-revs 脚手架(占位设计不变)。

重建中发现并处理了两个机械格式化的实际危害(已写进基线 commit message):

1. proxy-test.ts 的 @ts-expect-error 因 fetch 调用被拆行而错位(build 失败)→ 指令移到它压制的属性行;
2. Biome 对两个文件的首遍输出不满足自身 format:check(非幂等边界)→ 二遍收敛;
3. 一个源码契约测试(tool-runtime-extraction-contract)把 RunTracePhase 联合钉成单行字面量 → 改为对格式容忍的正则(契约是变体集合,不是排版)。

验证:format:check / lint / build / typecheck 干净,全套件通过——唯一失败是 session-title.test.js 的「aborts title generation when the provider exceeds its deadline」,它在 pristine origin/main 上同样失败(#1199 引入,套件内稳定挂、单文件跑通过,cancelledByParent 挂起类),与本 PR 无关,另行报告。

[UncertaintyDeterminesYou4ndMe]

The formatter follow-up promised in #1019 — one atomic PR, structured for a merge moment of your choosing. Closes the formatter half of #415.

Structure (3 commits)

1. Config + gate — formatter config restored in biome.jsonc (2-space / 100-col / single quotes / semicolons / trailing commas), format + format:check scripts, and npm run format:check in the CI typecheck job (right after lint, before build).
2. Baseline — mechanical biome format --write to fixpoint: 805 files, zero logic changes. This is the commit to point .git-blame-ignore-revs at.
3. .git-blame-ignore-revs — carries the baseline commit's branch hash. ⚠️ If you squash/rebase-merge, the hash changes; the file has the one-liner to regenerate it (git log --oneline --grep 'Biome format baseline' -1) — happy to push the corrected hash post-merge instead if you prefer merge-commit-off.

⚠️ The discovery you should weigh in on: apps/desktop is formatter-excluded

Running the baseline over the whole tree broke 145 assertions across 64 desktop test files. Desktop has a pervasive source-contract test style — tests read .ts/.tsx source and regex-match exact single-line shapes (e.g. account-auth-ui.test.ts matches a full <div className=... aria-label=...> line). Any reformat — Biome's or a future hand edit — snaps them.

Rather than scattering ~145 // biome-ignore format: comments through desktop source (which would exempt much of desktop from the formatter this PR introduces — self-defeating), this PR excludes apps/desktop/** from the formatter only (it stays fully linted), and defers desktop to a follow-up. Options for that follow-up, your call:

• (a) Make the source-contract assertions whitespace-tolerant first, then format desktop. Two shared helpers (main-process-contract-source-helpers.js, 20 files; provider-contract-source-helpers.js, 6 files) could centralize normalization, but many exact-spacing regexes are inline per test — non-trivial.
• (b) Keep desktop formatter-excluded long-term and accept the split.
• (c) Treat the source-contract style itself as the issue (they're fragile against any edit, not just Biome) and migrate them to AST-based checks — bigger conversation.

Outside desktop, the same class of problem existed at small scale and IS handled in the baseline: 5 statements carry a // biome-ignore format: guard — two @ts-expect-error directives that must stay adjacent to the line they suppress (proxy-test.ts, run-trace.ts — splitting the statement orphans the directive: TS2578 + the suppressed error resurfaces) and three single-line shapes asserted by source-contract tests (run-prompt-ab.mjs, run-prompt-optimization.mjs, run-harness-ab.mjs).

Also excluded (byte-sensitive, not hand-authored): apps/desktop/resources/licenses/** (SOURCE.json is whole-file SHA-256'd + size-checked by prepare-cua-driver.mjs) and bundled-tools.json.

Verification

• npm run lint ✅ and npm run format:check ✅ (formatter run to fixpoint — Biome's formatter isn't idempotent on some multi-line generic signatures, so one extra pass was needed)
• npm run build ✅, npm run typecheck ✅
• Full npm test: scripts 109, core 1067, storage 287, runtime 1983, computer-use 135, desktop (untouched) all green; headless 1063/1067 with only the 2 known-flaky opencode stop runner subtests (fail identically on unmodified main, pass in isolation)

Merge timing

Touches ~805 files, so it conflicts with most open PRs — merge whenever suits; I'll rebase + re-run the baseline to fixpoint on request the same day you want it landed (regenerating is cheap and deterministic, ~1 min).

Heads-up per the #1019 thread: biome lint --write --unsafe still hangs on apps/desktop (~100% CPU indefinitely; plain format is fine) — upstream minimization still on my list, tracked separately from this PR.

[UncertaintyDeterminesYou4ndMe]

CI's test failure was real and informative — fixed and force-pushed.

Root cause: desktop's source-contract tests don't only introspect desktop's own source — ~45 contracts read packages/ui source too (composer.tsx, daily-review-panel.tsx, primitives/*.tsx, …), and packages/ui was in the baseline. So the exclusion boundary isn't "apps/desktop", it's the whole UI surface.

Also owning a verification gap on my side: locally npm test short-circuits at the 2 known-flaky headless subtests (the script chains suites with &&), so the desktop/ui suites never actually ran after my rebase — I've re-verified per-suite explicitly this time.

Changes in the force-push:

• Formatter now excludes apps/desktop/**andpackages/ui/** (both still fully linted); baseline shrinks 805 → 660 files (packages/ui's 142 reverted, byte-identical to main).
• .git-blame-ignore-revs updated to the new baseline hash.

Verified: npm run lint ✅, npm run format:check ✅ (fixpoint), build ✅, typecheck ✅, and explicitly: @maka/ui ✅, @maka/desktop ✅ (2,571 tests), @maka/storage ✅.

This also sharpens the follow-up question: the source-contract web spans the whole UI surface (desktop + ui, 190+ single-line assertions), so making those contracts format-tolerant (or AST-based) is the real gate for ever formatting UI code — option (a)/(c) from the PR description.

[Astro-Han]

Approved.

Non-blocking P3 follow-up: this repository squash-merges PRs, so the branch-only baseline hash in .git-blame-ignore-revs will not exist in a fresh clone after merge; git blame configured with the file then exits with "cannot find revision ... to ignore". Please remove the branch hash before merge and add the resulting squash commit hash in a follow-up, or otherwise update the entry immediately after merge.

[UncertaintyDeterminesYou4ndMe]

Done — went with your first option. The branch hash is removed from .git-blame-ignore-revs before merge; the file now ships as documentation/scaffold only, with a note that entries must be hashes that exist on main (you're right that a hash unknown to the clone makes git blame hard-error, not just no-op — stricter than my original in-file note assumed). I'll open the one-line follow-up adding the squash-commit hash immediately after this lands.

One operational note: GitHub hasn't been starting CI runs on my last few pushes to this branch ("no checks reported", no check-suites on the head SHA — not the fork-approval gate). The commits are content-identical to the previously verified tree apart from the .git-blame-ignore-revs text change, and everything is verified locally (lint / format:check fixpoint / build / typecheck / per-suite tests including desktop's 2,571). If checks stay absent on your side, a close/reopen should kick a fresh pull_request event — or happy to try anything else you prefer.

[Astro-Han]

LGTM, but there is a lot of rebase to be done.

[likun666661]

LGTM, but there is a lot of rebase to be done.

[UncertaintyDeterminesYou4ndMe]

已在最新 main(8b9d69a)上重建整个基线(force-push,a385f96b)。全仓机械格式化 PR 的 rebase 策略是重新生成而非解决 660 文件的文本冲突:配置 commit 重放 → 重新跑 biome format(750 文件)→ 补 ignore-revs 脚手架(占位设计不变)。

重建中发现并处理了两个机械格式化的实际危害(已写进基线 commit message):

1. proxy-test.ts 的 @ts-expect-error 因 fetch 调用被拆行而错位(build 失败)→ 指令移到它压制的属性行;
2. Biome 对两个文件的首遍输出不满足自身 format:check(非幂等边界)→ 二遍收敛;
3. 一个源码契约测试(tool-runtime-extraction-contract)把 RunTracePhase 联合钉成单行字面量 → 改为对格式容忍的正则(契约是变体集合,不是排版)。

验证:format:check / lint / build / typecheck 干净,全套件通过——唯一失败是 session-title.test.js 的「aborts title generation when the provider exceeds its deadline」,它在 pristine origin/main 上同样失败(#1199 引入,套件内稳定挂、单文件跑通过,cancelledByParent 挂起类),与本 PR 无关,另行报告。

[UncertaintyDeterminesYou4ndMe]

The formatter follow-up promised in #1019 — one atomic PR, structured for a merge moment of your choosing. Closes the formatter half of #415.

Structure (3 commits)

1. Config + gate — formatter config restored in biome.jsonc (2-space / 100-col / single quotes / semicolons / trailing commas), format + format:check scripts, and npm run format:check in the CI typecheck job (right after lint, before build).
2. Baseline — mechanical biome format --write to fixpoint: 805 files, zero logic changes. This is the commit to point .git-blame-ignore-revs at.
3. .git-blame-ignore-revs — carries the baseline commit's branch hash. ⚠️ If you squash/rebase-merge, the hash changes; the file has the one-liner to regenerate it (git log --oneline --grep 'Biome format baseline' -1) — happy to push the corrected hash post-merge instead if you prefer merge-commit-off.

⚠️ The discovery you should weigh in on: apps/desktop is formatter-excluded

Running the baseline over the whole tree broke 145 assertions across 64 desktop test files. Desktop has a pervasive source-contract test style — tests read .ts/.tsx source and regex-match exact single-line shapes (e.g. account-auth-ui.test.ts matches a full <div className=... aria-label=...> line). Any reformat — Biome's or a future hand edit — snaps them.

Rather than scattering ~145 // biome-ignore format: comments through desktop source (which would exempt much of desktop from the formatter this PR introduces — self-defeating), this PR excludes apps/desktop/** from the formatter only (it stays fully linted), and defers desktop to a follow-up. Options for that follow-up, your call:

• (a) Make the source-contract assertions whitespace-tolerant first, then format desktop. Two shared helpers (main-process-contract-source-helpers.js, 20 files; provider-contract-source-helpers.js, 6 files) could centralize normalization, but many exact-spacing regexes are inline per test — non-trivial.
• (b) Keep desktop formatter-excluded long-term and accept the split.
• (c) Treat the source-contract style itself as the issue (they're fragile against any edit, not just Biome) and migrate them to AST-based checks — bigger conversation.

Outside desktop, the same class of problem existed at small scale and IS handled in the baseline: 5 statements carry a // biome-ignore format: guard — two @ts-expect-error directives that must stay adjacent to the line they suppress (proxy-test.ts, run-trace.ts — splitting the statement orphans the directive: TS2578 + the suppressed error resurfaces) and three single-line shapes asserted by source-contract tests (run-prompt-ab.mjs, run-prompt-optimization.mjs, run-harness-ab.mjs).

Also excluded (byte-sensitive, not hand-authored): apps/desktop/resources/licenses/** (SOURCE.json is whole-file SHA-256'd + size-checked by prepare-cua-driver.mjs) and bundled-tools.json.

Verification

• npm run lint ✅ and npm run format:check ✅ (formatter run to fixpoint — Biome's formatter isn't idempotent on some multi-line generic signatures, so one extra pass was needed)
• npm run build ✅, npm run typecheck ✅
• Full npm test: scripts 109, core 1067, storage 287, runtime 1983, computer-use 135, desktop (untouched) all green; headless 1063/1067 with only the 2 known-flaky opencode stop runner subtests (fail identically on unmodified main, pass in isolation)

Merge timing

Touches ~805 files, so it conflicts with most open PRs — merge whenever suits; I'll rebase + re-run the baseline to fixpoint on request the same day you want it landed (regenerating is cheap and deterministic, ~1 min).

Heads-up per the #1019 thread: biome lint --write --unsafe still hangs on apps/desktop (~100% CPU indefinitely; plain format is fine) — upstream minimization still on my list, tracked separately from this PR.

[UncertaintyDeterminesYou4ndMe]

CI's test failure was real and informative — fixed and force-pushed.

Root cause: desktop's source-contract tests don't only introspect desktop's own source — ~45 contracts read packages/ui source too (composer.tsx, daily-review-panel.tsx, primitives/*.tsx, …), and packages/ui was in the baseline. So the exclusion boundary isn't "apps/desktop", it's the whole UI surface.

Also owning a verification gap on my side: locally npm test short-circuits at the 2 known-flaky headless subtests (the script chains suites with &&), so the desktop/ui suites never actually ran after my rebase — I've re-verified per-suite explicitly this time.

Changes in the force-push:

• Formatter now excludes apps/desktop/**andpackages/ui/** (both still fully linted); baseline shrinks 805 → 660 files (packages/ui's 142 reverted, byte-identical to main).
• .git-blame-ignore-revs updated to the new baseline hash.

Verified: npm run lint ✅, npm run format:check ✅ (fixpoint), build ✅, typecheck ✅, and explicitly: @maka/ui ✅, @maka/desktop ✅ (2,571 tests), @maka/storage ✅.

This also sharpens the follow-up question: the source-contract web spans the whole UI surface (desktop + ui, 190+ single-line assertions), so making those contracts format-tolerant (or AST-based) is the real gate for ever formatting UI code — option (a)/(c) from the PR description.

[Astro-Han]

Approved.

Non-blocking P3 follow-up: this repository squash-merges PRs, so the branch-only baseline hash in .git-blame-ignore-revs will not exist in a fresh clone after merge; git blame configured with the file then exits with "cannot find revision ... to ignore". Please remove the branch hash before merge and add the resulting squash commit hash in a follow-up, or otherwise update the entry immediately after merge.

[UncertaintyDeterminesYou4ndMe]

Done — went with your first option. The branch hash is removed from .git-blame-ignore-revs before merge; the file now ships as documentation/scaffold only, with a note that entries must be hashes that exist on main (you're right that a hash unknown to the clone makes git blame hard-error, not just no-op — stricter than my original in-file note assumed). I'll open the one-line follow-up adding the squash-commit hash immediately after this lands.

One operational note: GitHub hasn't been starting CI runs on my last few pushes to this branch ("no checks reported", no check-suites on the head SHA — not the fork-approval gate). The commits are content-identical to the previously verified tree apart from the .git-blame-ignore-revs text change, and everything is verified locally (lint / format:check fixpoint / build / typecheck / per-suite tests including desktop's 2,571). If checks stay absent on your side, a close/reopen should kick a fresh pull_request event — or happy to try anything else you prefer.

[Astro-Han]

LGTM, but there is a lot of rebase to be done.

[likun666661]

LGTM, but there is a lot of rebase to be done.

[UncertaintyDeterminesYou4ndMe]

已在最新 main(8b9d69a)上重建整个基线(force-push,a385f96b)。全仓机械格式化 PR 的 rebase 策略是重新生成而非解决 660 文件的文本冲突:配置 commit 重放 → 重新跑 biome format(750 文件)→ 补 ignore-revs 脚手架(占位设计不变)。

重建中发现并处理了两个机械格式化的实际危害(已写进基线 commit message):

1. proxy-test.ts 的 @ts-expect-error 因 fetch 调用被拆行而错位(build 失败)→ 指令移到它压制的属性行;
2. Biome 对两个文件的首遍输出不满足自身 format:check(非幂等边界)→ 二遍收敛;
3. 一个源码契约测试(tool-runtime-extraction-contract)把 RunTracePhase 联合钉成单行字面量 → 改为对格式容忍的正则(契约是变体集合,不是排版)。

验证:format:check / lint / build / typecheck 干净,全套件通过——唯一失败是 session-title.test.js 的「aborts title generation when the provider exceeds its deadline」,它在 pristine origin/main 上同样失败(#1199 引入,套件内稳定挂、单文件跑通过,cancelledByParent 挂起类),与本 PR 无关,另行报告。

[UncertaintyDeterminesYou4ndMe]

The formatter follow-up promised in #1019 — one atomic PR, structured for a merge moment of your choosing. Closes the formatter half of #415.

Structure (3 commits)

1. Config + gate — formatter config restored in biome.jsonc (2-space / 100-col / single quotes / semicolons / trailing commas), format + format:check scripts, and npm run format:check in the CI typecheck job (right after lint, before build).
2. Baseline — mechanical biome format --write to fixpoint: 805 files, zero logic changes. This is the commit to point .git-blame-ignore-revs at.
3. .git-blame-ignore-revs — carries the baseline commit's branch hash. ⚠️ If you squash/rebase-merge, the hash changes; the file has the one-liner to regenerate it (git log --oneline --grep 'Biome format baseline' -1) — happy to push the corrected hash post-merge instead if you prefer merge-commit-off.

⚠️ The discovery you should weigh in on: apps/desktop is formatter-excluded

Running the baseline over the whole tree broke 145 assertions across 64 desktop test files. Desktop has a pervasive source-contract test style — tests read .ts/.tsx source and regex-match exact single-line shapes (e.g. account-auth-ui.test.ts matches a full <div className=... aria-label=...> line). Any reformat — Biome's or a future hand edit — snaps them.

Rather than scattering ~145 // biome-ignore format: comments through desktop source (which would exempt much of desktop from the formatter this PR introduces — self-defeating), this PR excludes apps/desktop/** from the formatter only (it stays fully linted), and defers desktop to a follow-up. Options for that follow-up, your call:

• (a) Make the source-contract assertions whitespace-tolerant first, then format desktop. Two shared helpers (main-process-contract-source-helpers.js, 20 files; provider-contract-source-helpers.js, 6 files) could centralize normalization, but many exact-spacing regexes are inline per test — non-trivial.
• (b) Keep desktop formatter-excluded long-term and accept the split.
• (c) Treat the source-contract style itself as the issue (they're fragile against any edit, not just Biome) and migrate them to AST-based checks — bigger conversation.

Outside desktop, the same class of problem existed at small scale and IS handled in the baseline: 5 statements carry a // biome-ignore format: guard — two @ts-expect-error directives that must stay adjacent to the line they suppress (proxy-test.ts, run-trace.ts — splitting the statement orphans the directive: TS2578 + the suppressed error resurfaces) and three single-line shapes asserted by source-contract tests (run-prompt-ab.mjs, run-prompt-optimization.mjs, run-harness-ab.mjs).

Also excluded (byte-sensitive, not hand-authored): apps/desktop/resources/licenses/** (SOURCE.json is whole-file SHA-256'd + size-checked by prepare-cua-driver.mjs) and bundled-tools.json.

Verification

• npm run lint ✅ and npm run format:check ✅ (formatter run to fixpoint — Biome's formatter isn't idempotent on some multi-line generic signatures, so one extra pass was needed)
• npm run build ✅, npm run typecheck ✅
• Full npm test: scripts 109, core 1067, storage 287, runtime 1983, computer-use 135, desktop (untouched) all green; headless 1063/1067 with only the 2 known-flaky opencode stop runner subtests (fail identically on unmodified main, pass in isolation)

Merge timing

Touches ~805 files, so it conflicts with most open PRs — merge whenever suits; I'll rebase + re-run the baseline to fixpoint on request the same day you want it landed (regenerating is cheap and deterministic, ~1 min).

Heads-up per the #1019 thread: biome lint --write --unsafe still hangs on apps/desktop (~100% CPU indefinitely; plain format is fine) — upstream minimization still on my list, tracked separately from this PR.

[UncertaintyDeterminesYou4ndMe]

CI's test failure was real and informative — fixed and force-pushed.

Root cause: desktop's source-contract tests don't only introspect desktop's own source — ~45 contracts read packages/ui source too (composer.tsx, daily-review-panel.tsx, primitives/*.tsx, …), and packages/ui was in the baseline. So the exclusion boundary isn't "apps/desktop", it's the whole UI surface.

Also owning a verification gap on my side: locally npm test short-circuits at the 2 known-flaky headless subtests (the script chains suites with &&), so the desktop/ui suites never actually ran after my rebase — I've re-verified per-suite explicitly this time.

Changes in the force-push:

• Formatter now excludes apps/desktop/**andpackages/ui/** (both still fully linted); baseline shrinks 805 → 660 files (packages/ui's 142 reverted, byte-identical to main).
• .git-blame-ignore-revs updated to the new baseline hash.

Verified: npm run lint ✅, npm run format:check ✅ (fixpoint), build ✅, typecheck ✅, and explicitly: @maka/ui ✅, @maka/desktop ✅ (2,571 tests), @maka/storage ✅.

This also sharpens the follow-up question: the source-contract web spans the whole UI surface (desktop + ui, 190+ single-line assertions), so making those contracts format-tolerant (or AST-based) is the real gate for ever formatting UI code — option (a)/(c) from the PR description.

[Astro-Han]

Approved.

Non-blocking P3 follow-up: this repository squash-merges PRs, so the branch-only baseline hash in .git-blame-ignore-revs will not exist in a fresh clone after merge; git blame configured with the file then exits with "cannot find revision ... to ignore". Please remove the branch hash before merge and add the resulting squash commit hash in a follow-up, or otherwise update the entry immediately after merge.

[UncertaintyDeterminesYou4ndMe]

Done — went with your first option. The branch hash is removed from .git-blame-ignore-revs before merge; the file now ships as documentation/scaffold only, with a note that entries must be hashes that exist on main (you're right that a hash unknown to the clone makes git blame hard-error, not just no-op — stricter than my original in-file note assumed). I'll open the one-line follow-up adding the squash-commit hash immediately after this lands.

One operational note: GitHub hasn't been starting CI runs on my last few pushes to this branch ("no checks reported", no check-suites on the head SHA — not the fork-approval gate). The commits are content-identical to the previously verified tree apart from the .git-blame-ignore-revs text change, and everything is verified locally (lint / format:check fixpoint / build / typecheck / per-suite tests including desktop's 2,571). If checks stay absent on your side, a close/reopen should kick a fresh pull_request event — or happy to try anything else you prefer.

[Astro-Han]

LGTM, but there is a lot of rebase to be done.

[likun666661]

LGTM, but there is a lot of rebase to be done.

[UncertaintyDeterminesYou4ndMe]

已在最新 main(8b9d69a)上重建整个基线(force-push,a385f96b)。全仓机械格式化 PR 的 rebase 策略是重新生成而非解决 660 文件的文本冲突:配置 commit 重放 → 重新跑 biome format(750 文件)→ 补 ignore-revs 脚手架(占位设计不变)。

重建中发现并处理了两个机械格式化的实际危害(已写进基线 commit message):

1. proxy-test.ts 的 @ts-expect-error 因 fetch 调用被拆行而错位(build 失败)→ 指令移到它压制的属性行;
2. Biome 对两个文件的首遍输出不满足自身 format:check(非幂等边界)→ 二遍收敛;
3. 一个源码契约测试(tool-runtime-extraction-contract)把 RunTracePhase 联合钉成单行字面量 → 改为对格式容忍的正则(契约是变体集合,不是排版)。

验证:format:check / lint / build / typecheck 干净,全套件通过——唯一失败是 session-title.test.js 的「aborts title generation when the provider exceeds its deadline」,它在 pristine origin/main 上同样失败(#1199 引入,套件内稳定挂、单文件跑通过,cancelledByParent 挂起类),与本 PR 无关,另行报告。

[UncertaintyDeterminesYou4ndMe]

The formatter follow-up promised in #1019 — one atomic PR, structured for a merge moment of your choosing. Closes the formatter half of #415.

Structure (3 commits)

1. Config + gate — formatter config restored in biome.jsonc (2-space / 100-col / single quotes / semicolons / trailing commas), format + format:check scripts, and npm run format:check in the CI typecheck job (right after lint, before build).
2. Baseline — mechanical biome format --write to fixpoint: 805 files, zero logic changes. This is the commit to point .git-blame-ignore-revs at.
3. .git-blame-ignore-revs — carries the baseline commit's branch hash. ⚠️ If you squash/rebase-merge, the hash changes; the file has the one-liner to regenerate it (git log --oneline --grep 'Biome format baseline' -1) — happy to push the corrected hash post-merge instead if you prefer merge-commit-off.

⚠️ The discovery you should weigh in on: apps/desktop is formatter-excluded

Running the baseline over the whole tree broke 145 assertions across 64 desktop test files. Desktop has a pervasive source-contract test style — tests read .ts/.tsx source and regex-match exact single-line shapes (e.g. account-auth-ui.test.ts matches a full <div className=... aria-label=...> line). Any reformat — Biome's or a future hand edit — snaps them.

Rather than scattering ~145 // biome-ignore format: comments through desktop source (which would exempt much of desktop from the formatter this PR introduces — self-defeating), this PR excludes apps/desktop/** from the formatter only (it stays fully linted), and defers desktop to a follow-up. Options for that follow-up, your call:

• (a) Make the source-contract assertions whitespace-tolerant first, then format desktop. Two shared helpers (main-process-contract-source-helpers.js, 20 files; provider-contract-source-helpers.js, 6 files) could centralize normalization, but many exact-spacing regexes are inline per test — non-trivial.
• (b) Keep desktop formatter-excluded long-term and accept the split.
• (c) Treat the source-contract style itself as the issue (they're fragile against any edit, not just Biome) and migrate them to AST-based checks — bigger conversation.

Outside desktop, the same class of problem existed at small scale and IS handled in the baseline: 5 statements carry a // biome-ignore format: guard — two @ts-expect-error directives that must stay adjacent to the line they suppress (proxy-test.ts, run-trace.ts — splitting the statement orphans the directive: TS2578 + the suppressed error resurfaces) and three single-line shapes asserted by source-contract tests (run-prompt-ab.mjs, run-prompt-optimization.mjs, run-harness-ab.mjs).

Also excluded (byte-sensitive, not hand-authored): apps/desktop/resources/licenses/** (SOURCE.json is whole-file SHA-256'd + size-checked by prepare-cua-driver.mjs) and bundled-tools.json.

Verification

• npm run lint ✅ and npm run format:check ✅ (formatter run to fixpoint — Biome's formatter isn't idempotent on some multi-line generic signatures, so one extra pass was needed)
• npm run build ✅, npm run typecheck ✅
• Full npm test: scripts 109, core 1067, storage 287, runtime 1983, computer-use 135, desktop (untouched) all green; headless 1063/1067 with only the 2 known-flaky opencode stop runner subtests (fail identically on unmodified main, pass in isolation)

Merge timing

Touches ~805 files, so it conflicts with most open PRs — merge whenever suits; I'll rebase + re-run the baseline to fixpoint on request the same day you want it landed (regenerating is cheap and deterministic, ~1 min).

Heads-up per the #1019 thread: biome lint --write --unsafe still hangs on apps/desktop (~100% CPU indefinitely; plain format is fine) — upstream minimization still on my list, tracked separately from this PR.

[UncertaintyDeterminesYou4ndMe]

CI's test failure was real and informative — fixed and force-pushed.

Root cause: desktop's source-contract tests don't only introspect desktop's own source — ~45 contracts read packages/ui source too (composer.tsx, daily-review-panel.tsx, primitives/*.tsx, …), and packages/ui was in the baseline. So the exclusion boundary isn't "apps/desktop", it's the whole UI surface.

Also owning a verification gap on my side: locally npm test short-circuits at the 2 known-flaky headless subtests (the script chains suites with &&), so the desktop/ui suites never actually ran after my rebase — I've re-verified per-suite explicitly this time.

Changes in the force-push:

• Formatter now excludes apps/desktop/**andpackages/ui/** (both still fully linted); baseline shrinks 805 → 660 files (packages/ui's 142 reverted, byte-identical to main).
• .git-blame-ignore-revs updated to the new baseline hash.

Verified: npm run lint ✅, npm run format:check ✅ (fixpoint), build ✅, typecheck ✅, and explicitly: @maka/ui ✅, @maka/desktop ✅ (2,571 tests), @maka/storage ✅.

This also sharpens the follow-up question: the source-contract web spans the whole UI surface (desktop + ui, 190+ single-line assertions), so making those contracts format-tolerant (or AST-based) is the real gate for ever formatting UI code — option (a)/(c) from the PR description.

[Astro-Han]

Approved.

Non-blocking P3 follow-up: this repository squash-merges PRs, so the branch-only baseline hash in .git-blame-ignore-revs will not exist in a fresh clone after merge; git blame configured with the file then exits with "cannot find revision ... to ignore". Please remove the branch hash before merge and add the resulting squash commit hash in a follow-up, or otherwise update the entry immediately after merge.

[UncertaintyDeterminesYou4ndMe]

Done — went with your first option. The branch hash is removed from .git-blame-ignore-revs before merge; the file now ships as documentation/scaffold only, with a note that entries must be hashes that exist on main (you're right that a hash unknown to the clone makes git blame hard-error, not just no-op — stricter than my original in-file note assumed). I'll open the one-line follow-up adding the squash-commit hash immediately after this lands.

One operational note: GitHub hasn't been starting CI runs on my last few pushes to this branch ("no checks reported", no check-suites on the head SHA — not the fork-approval gate). The commits are content-identical to the previously verified tree apart from the .git-blame-ignore-revs text change, and everything is verified locally (lint / format:check fixpoint / build / typecheck / per-suite tests including desktop's 2,571). If checks stay absent on your side, a close/reopen should kick a fresh pull_request event — or happy to try anything else you prefer.

[Astro-Han]

LGTM, but there is a lot of rebase to be done.

[likun666661]

LGTM, but there is a lot of rebase to be done.

[UncertaintyDeterminesYou4ndMe]

已在最新 main(8b9d69a)上重建整个基线(force-push,a385f96b)。全仓机械格式化 PR 的 rebase 策略是重新生成而非解决 660 文件的文本冲突:配置 commit 重放 → 重新跑 biome format(750 文件)→ 补 ignore-revs 脚手架(占位设计不变)。

重建中发现并处理了两个机械格式化的实际危害(已写进基线 commit message):

1. proxy-test.ts 的 @ts-expect-error 因 fetch 调用被拆行而错位(build 失败)→ 指令移到它压制的属性行;
2. Biome 对两个文件的首遍输出不满足自身 format:check(非幂等边界)→ 二遍收敛;
3. 一个源码契约测试(tool-runtime-extraction-contract)把 RunTracePhase 联合钉成单行字面量 → 改为对格式容忍的正则(契约是变体集合,不是排版)。

验证:format:check / lint / build / typecheck 干净,全套件通过——唯一失败是 session-title.test.js 的「aborts title generation when the provider exceeds its deadline」,它在 pristine origin/main 上同样失败(#1199 引入,套件内稳定挂、单文件跑通过,cancelledByParent 挂起类),与本 PR 无关,另行报告。