perf(coil): halve the desktop release bundle (#53)

[radroid]

Closes#53.

app.asar goes from 189.66 MiB / 14,765 files to 99.02 MiB / 3,429 files (−47.8% bytes, −76.8% files). Measured, not estimated, and re-measured on 136b3514f after the #71 Coil rename.

	Baseline	After	Δ
app.asar	189.66 MiB	99.02 MiB	−90.64 MiB (−47.8%)
Files in the asar	14,765	3,429	−76.8%
.zip (what the updater downloads)	144,592,118 B	124,467,727 B	−20.1 MB
.dmg	150,392,957 B	129,226,158 B	−21.2 MB
electron-builder phase	~126 s	~48 s	−62%
Staged vp install --prod	12.6 s	4.4 s	−65%

The compressed artifacts fall 14% rather than 48% because what was removed is text, which compressed well already. The .zip number is the one users pay, on every release.

Rebuilt on the renamed fork

This branch was originally cut before #71. It has been rebuilt from 136b3514f, not merged forward, so nothing carries an old path:

• scripts/t3x/ → scripts/coil/ (7 new files), docs/t3x/ → docs/coil/.
• .github/workflows/t3x-release.yml → coil-release.yml; every step anchor re-verified against the current file.
• refactor(coil): retire the local desktop autobuild #92 retired the local autobuild, so the edit this branch previously made to auto-build-desktop.sh is dropped. The release workflow is now the only path that builds a shipping artifact, and the test that used to assert the autobuild set the variables now asserts that no second build path has reappeared without them.
• T3X_* env names are kept, matching what Rename the fork's identity from t3x to coil (everything except the landing site) #71 deliberately did with T3X_DESKTOP_APP_ID and the T3X_MAC_CSC_* secrets. The new variable sits next to that one and reads as a pair.
• Doc references to T3 Code (Alpha).app → T3 Coil (Alpha).app. artifactName is still upstream's T3-Code-…, which Rename the fork's identity from t3x to coil (everything except the landing site) #71 deliberately left alone, so the artifact filenames in the doc are unchanged and correct.

The rename moved none of the measurements. Both builds were re-run on the new base: the packed asar is the same size on both sides (189.66 → 99.02 MiB), only the .dmg/.zip moved by ~3 KB from the renamed product string. All 66 non-wildcard globs still match a package that is actually shipped, so none has gone stale. The 385-chunk renderer byte-identity proof and the Clerk execution probe were both re-run and both reproduce.

Sync cost: one upstream file, strictly additive

Upstream-owned files touched	1 (scripts/build-desktop-artifact.ts)
This issue's delta on it	+21 / −0
That file's total fork delta	+32 / −1 — the −1 is #70's pre-existing displacement
New seam rows	0
New fork-owned files	8, under scripts/coil/ and docs/coil/ — paths upstream has never had

The additive-seam invariant holds: no deletion count moved and the file count did not change. The only realistic conflict is textual — upstream adding its own entry to DESKTOP_FILE_EXCLUSIONS collides with the ...coilExtraFileExclusions, line at the end of that array, a keep-both resolution.

The justification for the env hook has been corrected. The earlier version of this PR said an inline list would "cost a second seam row on a second upstream file." Since #71 that is no longer true — build-desktop-artifact.test.ts already carries a fork row. The real reason is stronger and survives the rename: the list is 67 globs and grows, so inline it makes every future size fix an edit to an upstream test assertion, in a file resolved by hand at every sync, where the conflict is decided by re-reading 67 lines instead of one. Through the environment, an unset environment packages exactly what upstream packages and upstream's deepStrictEqual keeps passing untouched.

What was cut, and the evidence for each

Web renderer source maps — 37.4 MiB. The largest line item in the bundle, unmentioned in #53. Turned off with T3CODE_WEB_SOURCEMAP=0, an existing upstream variable, so the biggest win costs zero seam rows. Provably behaviour-preserving: all 385 renderer chunks are byte-identical after stripping the trailing //# sourceMappingURL= comment, and their content-hashed filenames do not move.

Third-party source maps — 18.2 MiB. Node only parses a .map under --enable-source-maps, which the desktop backend passes through rather than sets. First-party maps (~12 MiB) are kept so an operator who does export it still gets legible frames for our own code.

Clerk's browser SDK — 23.2 MiB, 16 packages. The main process requires only two of @clerk/electron's six entry points (root and /storage), which between them require only electron and electron-store. The entries pulling the browser SDK are /react and /passkeys, whose only importer is apps/web/src/main.tsx — compiled before staging, and run in a renderer created with sandbox: true, nodeIntegration: false, which has no module resolution at all.

Verified by execution: loading @clerk/electron and @clerk/electron/storage from inside the optimized asar under ELECTRON_RUN_AS_NODE returns createClerkBridge, setupPasskeysMain and storage, and the transitive module list touches none of the excluded packages. @clerk/electron, @clerk/electron-passkeys and its native binaries are not excluded, and a test asserts they never will be.

The shiki / @pierre/diffs tree — 21.0 MiB, 50 packages. This answers #53 §3.3. The server imports one subpath (utils/parsePatchFiles) and the server build bundles it — bin.mjs has no @pierre/* import left. The highlighting users see is the web client's own per-language chunks, so @shikijs/langs was shipped twice and this was the copy nobody could load.

Two tools, and two bugs testing found in them

desktop-bundle-reachability.mjs (analysis) follows literal import edges and string-mention edges. The second exists because PlaywrightInjectedRuntime.ts resolves playwright-core through a const — an import-only scan calls 10.2 MiB of live code dead.

verify-desktop-bundle.mjs gates every release on both platforms. Testing the gate rather than trusting it found two real defects:

• It originally read only the archive. On Windows asarUnpack puts the bundles and all of node_modules on disk, so it would have failed every Windows release.
• Its severity originally depended on attribution. A real build with a deliberately over-broad !**/node_modules/effect/**/* produced a bundle that cannot start and the gate went green. Severity is now unconditional; attribution only makes the message actionable. That case is now a fixture test.

Five corrections to issue #53

In docs/coil/desktop-bundle-size.md:

1. The size table measured du block allocation, not content.core-js "15 MB (3,684 files)" is 1.25 MiB — 3,684 files × 4 KiB blocks. Every many-small-files package is overstated the same way, which is why the Clerk tree was not the biggest win.
2. Source maps were 35.6% of the asar and go unmentioned.
3. DESKTOP_FILE_EXCLUSIONS is not fork-owned. The recommendation to prefer it is right; the reason is not.
4. The mac leg is not the release critical path. The legs run in parallel; Windows is 17.5 min to mac's 9.6. And --concurrency-limit 1 costs only 48–58 s, so raising it buys ~20 s while re-opening Windows desktop build intermittently fails with STATUS_DLL_INIT_FAILED (0xC0000142) in parallel vp tasks #47's crashes — recommend leaving it alone.
5. The pnpm store is already cached.setup-vp's cache: true restores a 1,943 MB cache with downloaded 0.

The real remaining build-time target is the 442 s Spectre MSVC install — 42% of the Windows leg. Not touched: it cannot be verified without a Windows release run, and this workflow publishes on success.

Testing

• 86 tests across the affected files pass; upstream's build-desktop-artifact.test.ts passes unmodified.
• Full before/after builds on the post-rename base; the verifier passes on both and fails the deliberately-broken one.
• Real-asar fixture tests pin the header offset arithmetic.
• tsgo --noEmit clean, vp lint clean, vp fmt applied. Formatter churn outside this change's scope was reverted; the 50 deletions in the diff are all table-padding in the fork-owned ledger.

Not done: a full GUI launch with an interactive sign-in. A live app instance was running on port 3773 and a second instance risked disturbing that session. The renderer half is covered by the byte-identity proof and the main-process half by the execution probe — installing this build and signing in once is the natural last check before it ships.

🤖 Generated with Claude Code

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: ecdf249e-1114-4881-912b-c683a53efde6

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[radroid]

Closes#53.

app.asar goes from 189.66 MiB / 14,765 files to 99.02 MiB / 3,429 files (−47.8% bytes, −76.8% files). Measured, not estimated, and re-measured on 136b3514f after the #71 Coil rename.

	Baseline	After	Δ
app.asar	189.66 MiB	99.02 MiB	−90.64 MiB (−47.8%)
Files in the asar	14,765	3,429	−76.8%
.zip (what the updater downloads)	144,592,118 B	124,467,727 B	−20.1 MB
.dmg	150,392,957 B	129,226,158 B	−21.2 MB
electron-builder phase	~126 s	~48 s	−62%
Staged vp install --prod	12.6 s	4.4 s	−65%

The compressed artifacts fall 14% rather than 48% because what was removed is text, which compressed well already. The .zip number is the one users pay, on every release.

Rebuilt on the renamed fork

This branch was originally cut before #71. It has been rebuilt from 136b3514f, not merged forward, so nothing carries an old path:

• scripts/t3x/ → scripts/coil/ (7 new files), docs/t3x/ → docs/coil/.
• .github/workflows/t3x-release.yml → coil-release.yml; every step anchor re-verified against the current file.
• refactor(coil): retire the local desktop autobuild #92 retired the local autobuild, so the edit this branch previously made to auto-build-desktop.sh is dropped. The release workflow is now the only path that builds a shipping artifact, and the test that used to assert the autobuild set the variables now asserts that no second build path has reappeared without them.
• T3X_* env names are kept, matching what Rename the fork's identity from t3x to coil (everything except the landing site) #71 deliberately did with T3X_DESKTOP_APP_ID and the T3X_MAC_CSC_* secrets. The new variable sits next to that one and reads as a pair.
• Doc references to T3 Code (Alpha).app → T3 Coil (Alpha).app. artifactName is still upstream's T3-Code-…, which Rename the fork's identity from t3x to coil (everything except the landing site) #71 deliberately left alone, so the artifact filenames in the doc are unchanged and correct.

The rename moved none of the measurements. Both builds were re-run on the new base: the packed asar is the same size on both sides (189.66 → 99.02 MiB), only the .dmg/.zip moved by ~3 KB from the renamed product string. All 66 non-wildcard globs still match a package that is actually shipped, so none has gone stale. The 385-chunk renderer byte-identity proof and the Clerk execution probe were both re-run and both reproduce.

Sync cost: one upstream file, strictly additive

Upstream-owned files touched	1 (scripts/build-desktop-artifact.ts)
This issue's delta on it	+21 / −0
That file's total fork delta	+32 / −1 — the −1 is #70's pre-existing displacement
New seam rows	0
New fork-owned files	8, under scripts/coil/ and docs/coil/ — paths upstream has never had

The additive-seam invariant holds: no deletion count moved and the file count did not change. The only realistic conflict is textual — upstream adding its own entry to DESKTOP_FILE_EXCLUSIONS collides with the ...coilExtraFileExclusions, line at the end of that array, a keep-both resolution.

The justification for the env hook has been corrected. The earlier version of this PR said an inline list would "cost a second seam row on a second upstream file." Since #71 that is no longer true — build-desktop-artifact.test.ts already carries a fork row. The real reason is stronger and survives the rename: the list is 67 globs and grows, so inline it makes every future size fix an edit to an upstream test assertion, in a file resolved by hand at every sync, where the conflict is decided by re-reading 67 lines instead of one. Through the environment, an unset environment packages exactly what upstream packages and upstream's deepStrictEqual keeps passing untouched.

What was cut, and the evidence for each

Web renderer source maps — 37.4 MiB. The largest line item in the bundle, unmentioned in #53. Turned off with T3CODE_WEB_SOURCEMAP=0, an existing upstream variable, so the biggest win costs zero seam rows. Provably behaviour-preserving: all 385 renderer chunks are byte-identical after stripping the trailing //# sourceMappingURL= comment, and their content-hashed filenames do not move.

Third-party source maps — 18.2 MiB. Node only parses a .map under --enable-source-maps, which the desktop backend passes through rather than sets. First-party maps (~12 MiB) are kept so an operator who does export it still gets legible frames for our own code.

Clerk's browser SDK — 23.2 MiB, 16 packages. The main process requires only two of @clerk/electron's six entry points (root and /storage), which between them require only electron and electron-store. The entries pulling the browser SDK are /react and /passkeys, whose only importer is apps/web/src/main.tsx — compiled before staging, and run in a renderer created with sandbox: true, nodeIntegration: false, which has no module resolution at all.

Verified by execution: loading @clerk/electron and @clerk/electron/storage from inside the optimized asar under ELECTRON_RUN_AS_NODE returns createClerkBridge, setupPasskeysMain and storage, and the transitive module list touches none of the excluded packages. @clerk/electron, @clerk/electron-passkeys and its native binaries are not excluded, and a test asserts they never will be.

The shiki / @pierre/diffs tree — 21.0 MiB, 50 packages. This answers #53 §3.3. The server imports one subpath (utils/parsePatchFiles) and the server build bundles it — bin.mjs has no @pierre/* import left. The highlighting users see is the web client's own per-language chunks, so @shikijs/langs was shipped twice and this was the copy nobody could load.

Two tools, and two bugs testing found in them

desktop-bundle-reachability.mjs (analysis) follows literal import edges and string-mention edges. The second exists because PlaywrightInjectedRuntime.ts resolves playwright-core through a const — an import-only scan calls 10.2 MiB of live code dead.

verify-desktop-bundle.mjs gates every release on both platforms. Testing the gate rather than trusting it found two real defects:

• It originally read only the archive. On Windows asarUnpack puts the bundles and all of node_modules on disk, so it would have failed every Windows release.
• Its severity originally depended on attribution. A real build with a deliberately over-broad !**/node_modules/effect/**/* produced a bundle that cannot start and the gate went green. Severity is now unconditional; attribution only makes the message actionable. That case is now a fixture test.

Five corrections to issue #53

In docs/coil/desktop-bundle-size.md:

1. The size table measured du block allocation, not content.core-js "15 MB (3,684 files)" is 1.25 MiB — 3,684 files × 4 KiB blocks. Every many-small-files package is overstated the same way, which is why the Clerk tree was not the biggest win.
2. Source maps were 35.6% of the asar and go unmentioned.
3. DESKTOP_FILE_EXCLUSIONS is not fork-owned. The recommendation to prefer it is right; the reason is not.
4. The mac leg is not the release critical path. The legs run in parallel; Windows is 17.5 min to mac's 9.6. And --concurrency-limit 1 costs only 48–58 s, so raising it buys ~20 s while re-opening Windows desktop build intermittently fails with STATUS_DLL_INIT_FAILED (0xC0000142) in parallel vp tasks #47's crashes — recommend leaving it alone.
5. The pnpm store is already cached.setup-vp's cache: true restores a 1,943 MB cache with downloaded 0.

The real remaining build-time target is the 442 s Spectre MSVC install — 42% of the Windows leg. Not touched: it cannot be verified without a Windows release run, and this workflow publishes on success.

Testing

• 86 tests across the affected files pass; upstream's build-desktop-artifact.test.ts passes unmodified.
• Full before/after builds on the post-rename base; the verifier passes on both and fails the deliberately-broken one.
• Real-asar fixture tests pin the header offset arithmetic.
• tsgo --noEmit clean, vp lint clean, vp fmt applied. Formatter churn outside this change's scope was reverted; the 50 deletions in the diff are all table-padding in the fork-owned ledger.

Not done: a full GUI launch with an interactive sign-in. A live app instance was running on port 3773 and a second instance risked disturbing that session. The renderer half is covered by the byte-identity proof and the main-process half by the execution probe — installing this build and signing in once is the natural last check before it ships.

🤖 Generated with Claude Code

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: ecdf249e-1114-4881-912b-c683a53efde6

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[radroid]

Closes#53.

app.asar goes from 189.66 MiB / 14,765 files to 99.02 MiB / 3,429 files (−47.8% bytes, −76.8% files). Measured, not estimated, and re-measured on 136b3514f after the #71 Coil rename.

	Baseline	After	Δ
app.asar	189.66 MiB	99.02 MiB	−90.64 MiB (−47.8%)
Files in the asar	14,765	3,429	−76.8%
.zip (what the updater downloads)	144,592,118 B	124,467,727 B	−20.1 MB
.dmg	150,392,957 B	129,226,158 B	−21.2 MB
electron-builder phase	~126 s	~48 s	−62%
Staged vp install --prod	12.6 s	4.4 s	−65%

The compressed artifacts fall 14% rather than 48% because what was removed is text, which compressed well already. The .zip number is the one users pay, on every release.

Rebuilt on the renamed fork

This branch was originally cut before #71. It has been rebuilt from 136b3514f, not merged forward, so nothing carries an old path:

• scripts/t3x/ → scripts/coil/ (7 new files), docs/t3x/ → docs/coil/.
• .github/workflows/t3x-release.yml → coil-release.yml; every step anchor re-verified against the current file.
• refactor(coil): retire the local desktop autobuild #92 retired the local autobuild, so the edit this branch previously made to auto-build-desktop.sh is dropped. The release workflow is now the only path that builds a shipping artifact, and the test that used to assert the autobuild set the variables now asserts that no second build path has reappeared without them.
• T3X_* env names are kept, matching what Rename the fork's identity from t3x to coil (everything except the landing site) #71 deliberately did with T3X_DESKTOP_APP_ID and the T3X_MAC_CSC_* secrets. The new variable sits next to that one and reads as a pair.
• Doc references to T3 Code (Alpha).app → T3 Coil (Alpha).app. artifactName is still upstream's T3-Code-…, which Rename the fork's identity from t3x to coil (everything except the landing site) #71 deliberately left alone, so the artifact filenames in the doc are unchanged and correct.

The rename moved none of the measurements. Both builds were re-run on the new base: the packed asar is the same size on both sides (189.66 → 99.02 MiB), only the .dmg/.zip moved by ~3 KB from the renamed product string. All 66 non-wildcard globs still match a package that is actually shipped, so none has gone stale. The 385-chunk renderer byte-identity proof and the Clerk execution probe were both re-run and both reproduce.

Sync cost: one upstream file, strictly additive

Upstream-owned files touched	1 (scripts/build-desktop-artifact.ts)
This issue's delta on it	+21 / −0
That file's total fork delta	+32 / −1 — the −1 is #70's pre-existing displacement
New seam rows	0
New fork-owned files	8, under scripts/coil/ and docs/coil/ — paths upstream has never had

The additive-seam invariant holds: no deletion count moved and the file count did not change. The only realistic conflict is textual — upstream adding its own entry to DESKTOP_FILE_EXCLUSIONS collides with the ...coilExtraFileExclusions, line at the end of that array, a keep-both resolution.

The justification for the env hook has been corrected. The earlier version of this PR said an inline list would "cost a second seam row on a second upstream file." Since #71 that is no longer true — build-desktop-artifact.test.ts already carries a fork row. The real reason is stronger and survives the rename: the list is 67 globs and grows, so inline it makes every future size fix an edit to an upstream test assertion, in a file resolved by hand at every sync, where the conflict is decided by re-reading 67 lines instead of one. Through the environment, an unset environment packages exactly what upstream packages and upstream's deepStrictEqual keeps passing untouched.

What was cut, and the evidence for each

Web renderer source maps — 37.4 MiB. The largest line item in the bundle, unmentioned in #53. Turned off with T3CODE_WEB_SOURCEMAP=0, an existing upstream variable, so the biggest win costs zero seam rows. Provably behaviour-preserving: all 385 renderer chunks are byte-identical after stripping the trailing //# sourceMappingURL= comment, and their content-hashed filenames do not move.

Third-party source maps — 18.2 MiB. Node only parses a .map under --enable-source-maps, which the desktop backend passes through rather than sets. First-party maps (~12 MiB) are kept so an operator who does export it still gets legible frames for our own code.

Clerk's browser SDK — 23.2 MiB, 16 packages. The main process requires only two of @clerk/electron's six entry points (root and /storage), which between them require only electron and electron-store. The entries pulling the browser SDK are /react and /passkeys, whose only importer is apps/web/src/main.tsx — compiled before staging, and run in a renderer created with sandbox: true, nodeIntegration: false, which has no module resolution at all.

Verified by execution: loading @clerk/electron and @clerk/electron/storage from inside the optimized asar under ELECTRON_RUN_AS_NODE returns createClerkBridge, setupPasskeysMain and storage, and the transitive module list touches none of the excluded packages. @clerk/electron, @clerk/electron-passkeys and its native binaries are not excluded, and a test asserts they never will be.

The shiki / @pierre/diffs tree — 21.0 MiB, 50 packages. This answers #53 §3.3. The server imports one subpath (utils/parsePatchFiles) and the server build bundles it — bin.mjs has no @pierre/* import left. The highlighting users see is the web client's own per-language chunks, so @shikijs/langs was shipped twice and this was the copy nobody could load.

Two tools, and two bugs testing found in them

desktop-bundle-reachability.mjs (analysis) follows literal import edges and string-mention edges. The second exists because PlaywrightInjectedRuntime.ts resolves playwright-core through a const — an import-only scan calls 10.2 MiB of live code dead.

verify-desktop-bundle.mjs gates every release on both platforms. Testing the gate rather than trusting it found two real defects:

• It originally read only the archive. On Windows asarUnpack puts the bundles and all of node_modules on disk, so it would have failed every Windows release.
• Its severity originally depended on attribution. A real build with a deliberately over-broad !**/node_modules/effect/**/* produced a bundle that cannot start and the gate went green. Severity is now unconditional; attribution only makes the message actionable. That case is now a fixture test.

Five corrections to issue #53

In docs/coil/desktop-bundle-size.md:

1. The size table measured du block allocation, not content.core-js "15 MB (3,684 files)" is 1.25 MiB — 3,684 files × 4 KiB blocks. Every many-small-files package is overstated the same way, which is why the Clerk tree was not the biggest win.
2. Source maps were 35.6% of the asar and go unmentioned.
3. DESKTOP_FILE_EXCLUSIONS is not fork-owned. The recommendation to prefer it is right; the reason is not.
4. The mac leg is not the release critical path. The legs run in parallel; Windows is 17.5 min to mac's 9.6. And --concurrency-limit 1 costs only 48–58 s, so raising it buys ~20 s while re-opening Windows desktop build intermittently fails with STATUS_DLL_INIT_FAILED (0xC0000142) in parallel vp tasks #47's crashes — recommend leaving it alone.
5. The pnpm store is already cached.setup-vp's cache: true restores a 1,943 MB cache with downloaded 0.

The real remaining build-time target is the 442 s Spectre MSVC install — 42% of the Windows leg. Not touched: it cannot be verified without a Windows release run, and this workflow publishes on success.

Testing

• 86 tests across the affected files pass; upstream's build-desktop-artifact.test.ts passes unmodified.
• Full before/after builds on the post-rename base; the verifier passes on both and fails the deliberately-broken one.
• Real-asar fixture tests pin the header offset arithmetic.
• tsgo --noEmit clean, vp lint clean, vp fmt applied. Formatter churn outside this change's scope was reverted; the 50 deletions in the diff are all table-padding in the fork-owned ledger.

Not done: a full GUI launch with an interactive sign-in. A live app instance was running on port 3773 and a second instance risked disturbing that session. The renderer half is covered by the byte-identity proof and the main-process half by the execution probe — installing this build and signing in once is the natural last check before it ships.

🤖 Generated with Claude Code

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: ecdf249e-1114-4881-912b-c683a53efde6

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[radroid]

Closes#53.

app.asar goes from 189.66 MiB / 14,765 files to 99.02 MiB / 3,429 files (−47.8% bytes, −76.8% files). Measured, not estimated, and re-measured on 136b3514f after the #71 Coil rename.

	Baseline	After	Δ
app.asar	189.66 MiB	99.02 MiB	−90.64 MiB (−47.8%)
Files in the asar	14,765	3,429	−76.8%
.zip (what the updater downloads)	144,592,118 B	124,467,727 B	−20.1 MB
.dmg	150,392,957 B	129,226,158 B	−21.2 MB
electron-builder phase	~126 s	~48 s	−62%
Staged vp install --prod	12.6 s	4.4 s	−65%

The compressed artifacts fall 14% rather than 48% because what was removed is text, which compressed well already. The .zip number is the one users pay, on every release.

Rebuilt on the renamed fork

This branch was originally cut before #71. It has been rebuilt from 136b3514f, not merged forward, so nothing carries an old path:

• scripts/t3x/ → scripts/coil/ (7 new files), docs/t3x/ → docs/coil/.
• .github/workflows/t3x-release.yml → coil-release.yml; every step anchor re-verified against the current file.
• refactor(coil): retire the local desktop autobuild #92 retired the local autobuild, so the edit this branch previously made to auto-build-desktop.sh is dropped. The release workflow is now the only path that builds a shipping artifact, and the test that used to assert the autobuild set the variables now asserts that no second build path has reappeared without them.
• T3X_* env names are kept, matching what Rename the fork's identity from t3x to coil (everything except the landing site) #71 deliberately did with T3X_DESKTOP_APP_ID and the T3X_MAC_CSC_* secrets. The new variable sits next to that one and reads as a pair.
• Doc references to T3 Code (Alpha).app → T3 Coil (Alpha).app. artifactName is still upstream's T3-Code-…, which Rename the fork's identity from t3x to coil (everything except the landing site) #71 deliberately left alone, so the artifact filenames in the doc are unchanged and correct.

The rename moved none of the measurements. Both builds were re-run on the new base: the packed asar is the same size on both sides (189.66 → 99.02 MiB), only the .dmg/.zip moved by ~3 KB from the renamed product string. All 66 non-wildcard globs still match a package that is actually shipped, so none has gone stale. The 385-chunk renderer byte-identity proof and the Clerk execution probe were both re-run and both reproduce.

Sync cost: one upstream file, strictly additive

Upstream-owned files touched	1 (scripts/build-desktop-artifact.ts)
This issue's delta on it	+21 / −0
That file's total fork delta	+32 / −1 — the −1 is #70's pre-existing displacement
New seam rows	0
New fork-owned files	8, under scripts/coil/ and docs/coil/ — paths upstream has never had

The additive-seam invariant holds: no deletion count moved and the file count did not change. The only realistic conflict is textual — upstream adding its own entry to DESKTOP_FILE_EXCLUSIONS collides with the ...coilExtraFileExclusions, line at the end of that array, a keep-both resolution.

The justification for the env hook has been corrected. The earlier version of this PR said an inline list would "cost a second seam row on a second upstream file." Since #71 that is no longer true — build-desktop-artifact.test.ts already carries a fork row. The real reason is stronger and survives the rename: the list is 67 globs and grows, so inline it makes every future size fix an edit to an upstream test assertion, in a file resolved by hand at every sync, where the conflict is decided by re-reading 67 lines instead of one. Through the environment, an unset environment packages exactly what upstream packages and upstream's deepStrictEqual keeps passing untouched.

What was cut, and the evidence for each

Web renderer source maps — 37.4 MiB. The largest line item in the bundle, unmentioned in #53. Turned off with T3CODE_WEB_SOURCEMAP=0, an existing upstream variable, so the biggest win costs zero seam rows. Provably behaviour-preserving: all 385 renderer chunks are byte-identical after stripping the trailing //# sourceMappingURL= comment, and their content-hashed filenames do not move.

Third-party source maps — 18.2 MiB. Node only parses a .map under --enable-source-maps, which the desktop backend passes through rather than sets. First-party maps (~12 MiB) are kept so an operator who does export it still gets legible frames for our own code.

Clerk's browser SDK — 23.2 MiB, 16 packages. The main process requires only two of @clerk/electron's six entry points (root and /storage), which between them require only electron and electron-store. The entries pulling the browser SDK are /react and /passkeys, whose only importer is apps/web/src/main.tsx — compiled before staging, and run in a renderer created with sandbox: true, nodeIntegration: false, which has no module resolution at all.

Verified by execution: loading @clerk/electron and @clerk/electron/storage from inside the optimized asar under ELECTRON_RUN_AS_NODE returns createClerkBridge, setupPasskeysMain and storage, and the transitive module list touches none of the excluded packages. @clerk/electron, @clerk/electron-passkeys and its native binaries are not excluded, and a test asserts they never will be.

The shiki / @pierre/diffs tree — 21.0 MiB, 50 packages. This answers #53 §3.3. The server imports one subpath (utils/parsePatchFiles) and the server build bundles it — bin.mjs has no @pierre/* import left. The highlighting users see is the web client's own per-language chunks, so @shikijs/langs was shipped twice and this was the copy nobody could load.

Two tools, and two bugs testing found in them

desktop-bundle-reachability.mjs (analysis) follows literal import edges and string-mention edges. The second exists because PlaywrightInjectedRuntime.ts resolves playwright-core through a const — an import-only scan calls 10.2 MiB of live code dead.

verify-desktop-bundle.mjs gates every release on both platforms. Testing the gate rather than trusting it found two real defects:

• It originally read only the archive. On Windows asarUnpack puts the bundles and all of node_modules on disk, so it would have failed every Windows release.
• Its severity originally depended on attribution. A real build with a deliberately over-broad !**/node_modules/effect/**/* produced a bundle that cannot start and the gate went green. Severity is now unconditional; attribution only makes the message actionable. That case is now a fixture test.

Five corrections to issue #53

In docs/coil/desktop-bundle-size.md:

1. The size table measured du block allocation, not content.core-js "15 MB (3,684 files)" is 1.25 MiB — 3,684 files × 4 KiB blocks. Every many-small-files package is overstated the same way, which is why the Clerk tree was not the biggest win.
2. Source maps were 35.6% of the asar and go unmentioned.
3. DESKTOP_FILE_EXCLUSIONS is not fork-owned. The recommendation to prefer it is right; the reason is not.
4. The mac leg is not the release critical path. The legs run in parallel; Windows is 17.5 min to mac's 9.6. And --concurrency-limit 1 costs only 48–58 s, so raising it buys ~20 s while re-opening Windows desktop build intermittently fails with STATUS_DLL_INIT_FAILED (0xC0000142) in parallel vp tasks #47's crashes — recommend leaving it alone.
5. The pnpm store is already cached.setup-vp's cache: true restores a 1,943 MB cache with downloaded 0.

The real remaining build-time target is the 442 s Spectre MSVC install — 42% of the Windows leg. Not touched: it cannot be verified without a Windows release run, and this workflow publishes on success.

Testing

• 86 tests across the affected files pass; upstream's build-desktop-artifact.test.ts passes unmodified.
• Full before/after builds on the post-rename base; the verifier passes on both and fails the deliberately-broken one.
• Real-asar fixture tests pin the header offset arithmetic.
• tsgo --noEmit clean, vp lint clean, vp fmt applied. Formatter churn outside this change's scope was reverted; the 50 deletions in the diff are all table-padding in the fork-owned ledger.

Not done: a full GUI launch with an interactive sign-in. A live app instance was running on port 3773 and a second instance risked disturbing that session. The renderer half is covered by the byte-identity proof and the main-process half by the execution probe — installing this build and signing in once is the natural last check before it ships.

🤖 Generated with Claude Code

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: ecdf249e-1114-4881-912b-c683a53efde6

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[radroid]

Closes#53.

app.asar goes from 189.66 MiB / 14,765 files to 99.02 MiB / 3,429 files (−47.8% bytes, −76.8% files). Measured, not estimated, and re-measured on 136b3514f after the #71 Coil rename.

	Baseline	After	Δ
app.asar	189.66 MiB	99.02 MiB	−90.64 MiB (−47.8%)
Files in the asar	14,765	3,429	−76.8%
.zip (what the updater downloads)	144,592,118 B	124,467,727 B	−20.1 MB
.dmg	150,392,957 B	129,226,158 B	−21.2 MB
electron-builder phase	~126 s	~48 s	−62%
Staged vp install --prod	12.6 s	4.4 s	−65%

The compressed artifacts fall 14% rather than 48% because what was removed is text, which compressed well already. The .zip number is the one users pay, on every release.

Rebuilt on the renamed fork

This branch was originally cut before #71. It has been rebuilt from 136b3514f, not merged forward, so nothing carries an old path:

• scripts/t3x/ → scripts/coil/ (7 new files), docs/t3x/ → docs/coil/.
• .github/workflows/t3x-release.yml → coil-release.yml; every step anchor re-verified against the current file.
• refactor(coil): retire the local desktop autobuild #92 retired the local autobuild, so the edit this branch previously made to auto-build-desktop.sh is dropped. The release workflow is now the only path that builds a shipping artifact, and the test that used to assert the autobuild set the variables now asserts that no second build path has reappeared without them.
• T3X_* env names are kept, matching what Rename the fork's identity from t3x to coil (everything except the landing site) #71 deliberately did with T3X_DESKTOP_APP_ID and the T3X_MAC_CSC_* secrets. The new variable sits next to that one and reads as a pair.
• Doc references to T3 Code (Alpha).app → T3 Coil (Alpha).app. artifactName is still upstream's T3-Code-…, which Rename the fork's identity from t3x to coil (everything except the landing site) #71 deliberately left alone, so the artifact filenames in the doc are unchanged and correct.

The rename moved none of the measurements. Both builds were re-run on the new base: the packed asar is the same size on both sides (189.66 → 99.02 MiB), only the .dmg/.zip moved by ~3 KB from the renamed product string. All 66 non-wildcard globs still match a package that is actually shipped, so none has gone stale. The 385-chunk renderer byte-identity proof and the Clerk execution probe were both re-run and both reproduce.

Sync cost: one upstream file, strictly additive

Upstream-owned files touched	1 (scripts/build-desktop-artifact.ts)
This issue's delta on it	+21 / −0
That file's total fork delta	+32 / −1 — the −1 is #70's pre-existing displacement
New seam rows	0
New fork-owned files	8, under scripts/coil/ and docs/coil/ — paths upstream has never had

The additive-seam invariant holds: no deletion count moved and the file count did not change. The only realistic conflict is textual — upstream adding its own entry to DESKTOP_FILE_EXCLUSIONS collides with the ...coilExtraFileExclusions, line at the end of that array, a keep-both resolution.

The justification for the env hook has been corrected. The earlier version of this PR said an inline list would "cost a second seam row on a second upstream file." Since #71 that is no longer true — build-desktop-artifact.test.ts already carries a fork row. The real reason is stronger and survives the rename: the list is 67 globs and grows, so inline it makes every future size fix an edit to an upstream test assertion, in a file resolved by hand at every sync, where the conflict is decided by re-reading 67 lines instead of one. Through the environment, an unset environment packages exactly what upstream packages and upstream's deepStrictEqual keeps passing untouched.

What was cut, and the evidence for each

Web renderer source maps — 37.4 MiB. The largest line item in the bundle, unmentioned in #53. Turned off with T3CODE_WEB_SOURCEMAP=0, an existing upstream variable, so the biggest win costs zero seam rows. Provably behaviour-preserving: all 385 renderer chunks are byte-identical after stripping the trailing //# sourceMappingURL= comment, and their content-hashed filenames do not move.

Third-party source maps — 18.2 MiB. Node only parses a .map under --enable-source-maps, which the desktop backend passes through rather than sets. First-party maps (~12 MiB) are kept so an operator who does export it still gets legible frames for our own code.

Clerk's browser SDK — 23.2 MiB, 16 packages. The main process requires only two of @clerk/electron's six entry points (root and /storage), which between them require only electron and electron-store. The entries pulling the browser SDK are /react and /passkeys, whose only importer is apps/web/src/main.tsx — compiled before staging, and run in a renderer created with sandbox: true, nodeIntegration: false, which has no module resolution at all.

Verified by execution: loading @clerk/electron and @clerk/electron/storage from inside the optimized asar under ELECTRON_RUN_AS_NODE returns createClerkBridge, setupPasskeysMain and storage, and the transitive module list touches none of the excluded packages. @clerk/electron, @clerk/electron-passkeys and its native binaries are not excluded, and a test asserts they never will be.

The shiki / @pierre/diffs tree — 21.0 MiB, 50 packages. This answers #53 §3.3. The server imports one subpath (utils/parsePatchFiles) and the server build bundles it — bin.mjs has no @pierre/* import left. The highlighting users see is the web client's own per-language chunks, so @shikijs/langs was shipped twice and this was the copy nobody could load.

Two tools, and two bugs testing found in them

desktop-bundle-reachability.mjs (analysis) follows literal import edges and string-mention edges. The second exists because PlaywrightInjectedRuntime.ts resolves playwright-core through a const — an import-only scan calls 10.2 MiB of live code dead.

verify-desktop-bundle.mjs gates every release on both platforms. Testing the gate rather than trusting it found two real defects:

• It originally read only the archive. On Windows asarUnpack puts the bundles and all of node_modules on disk, so it would have failed every Windows release.
• Its severity originally depended on attribution. A real build with a deliberately over-broad !**/node_modules/effect/**/* produced a bundle that cannot start and the gate went green. Severity is now unconditional; attribution only makes the message actionable. That case is now a fixture test.

Five corrections to issue #53

In docs/coil/desktop-bundle-size.md:

1. The size table measured du block allocation, not content.core-js "15 MB (3,684 files)" is 1.25 MiB — 3,684 files × 4 KiB blocks. Every many-small-files package is overstated the same way, which is why the Clerk tree was not the biggest win.
2. Source maps were 35.6% of the asar and go unmentioned.
3. DESKTOP_FILE_EXCLUSIONS is not fork-owned. The recommendation to prefer it is right; the reason is not.
4. The mac leg is not the release critical path. The legs run in parallel; Windows is 17.5 min to mac's 9.6. And --concurrency-limit 1 costs only 48–58 s, so raising it buys ~20 s while re-opening Windows desktop build intermittently fails with STATUS_DLL_INIT_FAILED (0xC0000142) in parallel vp tasks #47's crashes — recommend leaving it alone.
5. The pnpm store is already cached.setup-vp's cache: true restores a 1,943 MB cache with downloaded 0.

The real remaining build-time target is the 442 s Spectre MSVC install — 42% of the Windows leg. Not touched: it cannot be verified without a Windows release run, and this workflow publishes on success.

Testing

• 86 tests across the affected files pass; upstream's build-desktop-artifact.test.ts passes unmodified.
• Full before/after builds on the post-rename base; the verifier passes on both and fails the deliberately-broken one.
• Real-asar fixture tests pin the header offset arithmetic.
• tsgo --noEmit clean, vp lint clean, vp fmt applied. Formatter churn outside this change's scope was reverted; the 50 deletions in the diff are all table-padding in the fork-owned ledger.

Not done: a full GUI launch with an interactive sign-in. A live app instance was running on port 3773 and a second instance risked disturbing that session. The renderer half is covered by the byte-identity proof and the main-process half by the execution probe — installing this build and signing in once is the natural last check before it ships.

🤖 Generated with Claude Code

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: ecdf249e-1114-4881-912b-c683a53efde6

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[radroid]

Closes#53.

app.asar goes from 189.66 MiB / 14,765 files to 99.02 MiB / 3,429 files (−47.8% bytes, −76.8% files). Measured, not estimated, and re-measured on 136b3514f after the #71 Coil rename.

	Baseline	After	Δ
app.asar	189.66 MiB	99.02 MiB	−90.64 MiB (−47.8%)
Files in the asar	14,765	3,429	−76.8%
.zip (what the updater downloads)	144,592,118 B	124,467,727 B	−20.1 MB
.dmg	150,392,957 B	129,226,158 B	−21.2 MB
electron-builder phase	~126 s	~48 s	−62%
Staged vp install --prod	12.6 s	4.4 s	−65%

The compressed artifacts fall 14% rather than 48% because what was removed is text, which compressed well already. The .zip number is the one users pay, on every release.

Rebuilt on the renamed fork

This branch was originally cut before #71. It has been rebuilt from 136b3514f, not merged forward, so nothing carries an old path:

• scripts/t3x/ → scripts/coil/ (7 new files), docs/t3x/ → docs/coil/.
• .github/workflows/t3x-release.yml → coil-release.yml; every step anchor re-verified against the current file.
• refactor(coil): retire the local desktop autobuild #92 retired the local autobuild, so the edit this branch previously made to auto-build-desktop.sh is dropped. The release workflow is now the only path that builds a shipping artifact, and the test that used to assert the autobuild set the variables now asserts that no second build path has reappeared without them.
• T3X_* env names are kept, matching what Rename the fork's identity from t3x to coil (everything except the landing site) #71 deliberately did with T3X_DESKTOP_APP_ID and the T3X_MAC_CSC_* secrets. The new variable sits next to that one and reads as a pair.
• Doc references to T3 Code (Alpha).app → T3 Coil (Alpha).app. artifactName is still upstream's T3-Code-…, which Rename the fork's identity from t3x to coil (everything except the landing site) #71 deliberately left alone, so the artifact filenames in the doc are unchanged and correct.

The rename moved none of the measurements. Both builds were re-run on the new base: the packed asar is the same size on both sides (189.66 → 99.02 MiB), only the .dmg/.zip moved by ~3 KB from the renamed product string. All 66 non-wildcard globs still match a package that is actually shipped, so none has gone stale. The 385-chunk renderer byte-identity proof and the Clerk execution probe were both re-run and both reproduce.

Sync cost: one upstream file, strictly additive

Upstream-owned files touched	1 (scripts/build-desktop-artifact.ts)
This issue's delta on it	+21 / −0
That file's total fork delta	+32 / −1 — the −1 is #70's pre-existing displacement
New seam rows	0
New fork-owned files	8, under scripts/coil/ and docs/coil/ — paths upstream has never had

The additive-seam invariant holds: no deletion count moved and the file count did not change. The only realistic conflict is textual — upstream adding its own entry to DESKTOP_FILE_EXCLUSIONS collides with the ...coilExtraFileExclusions, line at the end of that array, a keep-both resolution.

The justification for the env hook has been corrected. The earlier version of this PR said an inline list would "cost a second seam row on a second upstream file." Since #71 that is no longer true — build-desktop-artifact.test.ts already carries a fork row. The real reason is stronger and survives the rename: the list is 67 globs and grows, so inline it makes every future size fix an edit to an upstream test assertion, in a file resolved by hand at every sync, where the conflict is decided by re-reading 67 lines instead of one. Through the environment, an unset environment packages exactly what upstream packages and upstream's deepStrictEqual keeps passing untouched.

What was cut, and the evidence for each

Web renderer source maps — 37.4 MiB. The largest line item in the bundle, unmentioned in #53. Turned off with T3CODE_WEB_SOURCEMAP=0, an existing upstream variable, so the biggest win costs zero seam rows. Provably behaviour-preserving: all 385 renderer chunks are byte-identical after stripping the trailing //# sourceMappingURL= comment, and their content-hashed filenames do not move.

Third-party source maps — 18.2 MiB. Node only parses a .map under --enable-source-maps, which the desktop backend passes through rather than sets. First-party maps (~12 MiB) are kept so an operator who does export it still gets legible frames for our own code.

Clerk's browser SDK — 23.2 MiB, 16 packages. The main process requires only two of @clerk/electron's six entry points (root and /storage), which between them require only electron and electron-store. The entries pulling the browser SDK are /react and /passkeys, whose only importer is apps/web/src/main.tsx — compiled before staging, and run in a renderer created with sandbox: true, nodeIntegration: false, which has no module resolution at all.

Verified by execution: loading @clerk/electron and @clerk/electron/storage from inside the optimized asar under ELECTRON_RUN_AS_NODE returns createClerkBridge, setupPasskeysMain and storage, and the transitive module list touches none of the excluded packages. @clerk/electron, @clerk/electron-passkeys and its native binaries are not excluded, and a test asserts they never will be.

The shiki / @pierre/diffs tree — 21.0 MiB, 50 packages. This answers #53 §3.3. The server imports one subpath (utils/parsePatchFiles) and the server build bundles it — bin.mjs has no @pierre/* import left. The highlighting users see is the web client's own per-language chunks, so @shikijs/langs was shipped twice and this was the copy nobody could load.

Two tools, and two bugs testing found in them

desktop-bundle-reachability.mjs (analysis) follows literal import edges and string-mention edges. The second exists because PlaywrightInjectedRuntime.ts resolves playwright-core through a const — an import-only scan calls 10.2 MiB of live code dead.

verify-desktop-bundle.mjs gates every release on both platforms. Testing the gate rather than trusting it found two real defects:

• It originally read only the archive. On Windows asarUnpack puts the bundles and all of node_modules on disk, so it would have failed every Windows release.
• Its severity originally depended on attribution. A real build with a deliberately over-broad !**/node_modules/effect/**/* produced a bundle that cannot start and the gate went green. Severity is now unconditional; attribution only makes the message actionable. That case is now a fixture test.

Five corrections to issue #53

In docs/coil/desktop-bundle-size.md:

1. The size table measured du block allocation, not content.core-js "15 MB (3,684 files)" is 1.25 MiB — 3,684 files × 4 KiB blocks. Every many-small-files package is overstated the same way, which is why the Clerk tree was not the biggest win.
2. Source maps were 35.6% of the asar and go unmentioned.
3. DESKTOP_FILE_EXCLUSIONS is not fork-owned. The recommendation to prefer it is right; the reason is not.
4. The mac leg is not the release critical path. The legs run in parallel; Windows is 17.5 min to mac's 9.6. And --concurrency-limit 1 costs only 48–58 s, so raising it buys ~20 s while re-opening Windows desktop build intermittently fails with STATUS_DLL_INIT_FAILED (0xC0000142) in parallel vp tasks #47's crashes — recommend leaving it alone.
5. The pnpm store is already cached.setup-vp's cache: true restores a 1,943 MB cache with downloaded 0.

The real remaining build-time target is the 442 s Spectre MSVC install — 42% of the Windows leg. Not touched: it cannot be verified without a Windows release run, and this workflow publishes on success.

Testing

• 86 tests across the affected files pass; upstream's build-desktop-artifact.test.ts passes unmodified.
• Full before/after builds on the post-rename base; the verifier passes on both and fails the deliberately-broken one.
• Real-asar fixture tests pin the header offset arithmetic.
• tsgo --noEmit clean, vp lint clean, vp fmt applied. Formatter churn outside this change's scope was reverted; the 50 deletions in the diff are all table-padding in the fork-owned ledger.

Not done: a full GUI launch with an interactive sign-in. A live app instance was running on port 3773 and a second instance risked disturbing that session. The renderer half is covered by the byte-identity proof and the main-process half by the execution probe — installing this build and signing in once is the natural last check before it ships.

🤖 Generated with Claude Code

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: ecdf249e-1114-4881-912b-c683a53efde6

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[radroid]

Closes#53.

app.asar goes from 189.66 MiB / 14,765 files to 99.02 MiB / 3,429 files (−47.8% bytes, −76.8% files). Measured, not estimated, and re-measured on 136b3514f after the #71 Coil rename.

	Baseline	After	Δ
app.asar	189.66 MiB	99.02 MiB	−90.64 MiB (−47.8%)
Files in the asar	14,765	3,429	−76.8%
.zip (what the updater downloads)	144,592,118 B	124,467,727 B	−20.1 MB
.dmg	150,392,957 B	129,226,158 B	−21.2 MB
electron-builder phase	~126 s	~48 s	−62%
Staged vp install --prod	12.6 s	4.4 s	−65%

The compressed artifacts fall 14% rather than 48% because what was removed is text, which compressed well already. The .zip number is the one users pay, on every release.

Rebuilt on the renamed fork

This branch was originally cut before #71. It has been rebuilt from 136b3514f, not merged forward, so nothing carries an old path:

• scripts/t3x/ → scripts/coil/ (7 new files), docs/t3x/ → docs/coil/.
• .github/workflows/t3x-release.yml → coil-release.yml; every step anchor re-verified against the current file.
• refactor(coil): retire the local desktop autobuild #92 retired the local autobuild, so the edit this branch previously made to auto-build-desktop.sh is dropped. The release workflow is now the only path that builds a shipping artifact, and the test that used to assert the autobuild set the variables now asserts that no second build path has reappeared without them.
• T3X_* env names are kept, matching what Rename the fork's identity from t3x to coil (everything except the landing site) #71 deliberately did with T3X_DESKTOP_APP_ID and the T3X_MAC_CSC_* secrets. The new variable sits next to that one and reads as a pair.
• Doc references to T3 Code (Alpha).app → T3 Coil (Alpha).app. artifactName is still upstream's T3-Code-…, which Rename the fork's identity from t3x to coil (everything except the landing site) #71 deliberately left alone, so the artifact filenames in the doc are unchanged and correct.

The rename moved none of the measurements. Both builds were re-run on the new base: the packed asar is the same size on both sides (189.66 → 99.02 MiB), only the .dmg/.zip moved by ~3 KB from the renamed product string. All 66 non-wildcard globs still match a package that is actually shipped, so none has gone stale. The 385-chunk renderer byte-identity proof and the Clerk execution probe were both re-run and both reproduce.

Sync cost: one upstream file, strictly additive

Upstream-owned files touched	1 (scripts/build-desktop-artifact.ts)
This issue's delta on it	+21 / −0
That file's total fork delta	+32 / −1 — the −1 is #70's pre-existing displacement
New seam rows	0
New fork-owned files	8, under scripts/coil/ and docs/coil/ — paths upstream has never had

The additive-seam invariant holds: no deletion count moved and the file count did not change. The only realistic conflict is textual — upstream adding its own entry to DESKTOP_FILE_EXCLUSIONS collides with the ...coilExtraFileExclusions, line at the end of that array, a keep-both resolution.

The justification for the env hook has been corrected. The earlier version of this PR said an inline list would "cost a second seam row on a second upstream file." Since #71 that is no longer true — build-desktop-artifact.test.ts already carries a fork row. The real reason is stronger and survives the rename: the list is 67 globs and grows, so inline it makes every future size fix an edit to an upstream test assertion, in a file resolved by hand at every sync, where the conflict is decided by re-reading 67 lines instead of one. Through the environment, an unset environment packages exactly what upstream packages and upstream's deepStrictEqual keeps passing untouched.

What was cut, and the evidence for each

Web renderer source maps — 37.4 MiB. The largest line item in the bundle, unmentioned in #53. Turned off with T3CODE_WEB_SOURCEMAP=0, an existing upstream variable, so the biggest win costs zero seam rows. Provably behaviour-preserving: all 385 renderer chunks are byte-identical after stripping the trailing //# sourceMappingURL= comment, and their content-hashed filenames do not move.

Third-party source maps — 18.2 MiB. Node only parses a .map under --enable-source-maps, which the desktop backend passes through rather than sets. First-party maps (~12 MiB) are kept so an operator who does export it still gets legible frames for our own code.

Clerk's browser SDK — 23.2 MiB, 16 packages. The main process requires only two of @clerk/electron's six entry points (root and /storage), which between them require only electron and electron-store. The entries pulling the browser SDK are /react and /passkeys, whose only importer is apps/web/src/main.tsx — compiled before staging, and run in a renderer created with sandbox: true, nodeIntegration: false, which has no module resolution at all.

Verified by execution: loading @clerk/electron and @clerk/electron/storage from inside the optimized asar under ELECTRON_RUN_AS_NODE returns createClerkBridge, setupPasskeysMain and storage, and the transitive module list touches none of the excluded packages. @clerk/electron, @clerk/electron-passkeys and its native binaries are not excluded, and a test asserts they never will be.

The shiki / @pierre/diffs tree — 21.0 MiB, 50 packages. This answers #53 §3.3. The server imports one subpath (utils/parsePatchFiles) and the server build bundles it — bin.mjs has no @pierre/* import left. The highlighting users see is the web client's own per-language chunks, so @shikijs/langs was shipped twice and this was the copy nobody could load.

Two tools, and two bugs testing found in them

desktop-bundle-reachability.mjs (analysis) follows literal import edges and string-mention edges. The second exists because PlaywrightInjectedRuntime.ts resolves playwright-core through a const — an import-only scan calls 10.2 MiB of live code dead.

verify-desktop-bundle.mjs gates every release on both platforms. Testing the gate rather than trusting it found two real defects:

• It originally read only the archive. On Windows asarUnpack puts the bundles and all of node_modules on disk, so it would have failed every Windows release.
• Its severity originally depended on attribution. A real build with a deliberately over-broad !**/node_modules/effect/**/* produced a bundle that cannot start and the gate went green. Severity is now unconditional; attribution only makes the message actionable. That case is now a fixture test.

Five corrections to issue #53

In docs/coil/desktop-bundle-size.md:

1. The size table measured du block allocation, not content.core-js "15 MB (3,684 files)" is 1.25 MiB — 3,684 files × 4 KiB blocks. Every many-small-files package is overstated the same way, which is why the Clerk tree was not the biggest win.
2. Source maps were 35.6% of the asar and go unmentioned.
3. DESKTOP_FILE_EXCLUSIONS is not fork-owned. The recommendation to prefer it is right; the reason is not.
4. The mac leg is not the release critical path. The legs run in parallel; Windows is 17.5 min to mac's 9.6. And --concurrency-limit 1 costs only 48–58 s, so raising it buys ~20 s while re-opening Windows desktop build intermittently fails with STATUS_DLL_INIT_FAILED (0xC0000142) in parallel vp tasks #47's crashes — recommend leaving it alone.
5. The pnpm store is already cached.setup-vp's cache: true restores a 1,943 MB cache with downloaded 0.

The real remaining build-time target is the 442 s Spectre MSVC install — 42% of the Windows leg. Not touched: it cannot be verified without a Windows release run, and this workflow publishes on success.

Testing

• 86 tests across the affected files pass; upstream's build-desktop-artifact.test.ts passes unmodified.
• Full before/after builds on the post-rename base; the verifier passes on both and fails the deliberately-broken one.
• Real-asar fixture tests pin the header offset arithmetic.
• tsgo --noEmit clean, vp lint clean, vp fmt applied. Formatter churn outside this change's scope was reverted; the 50 deletions in the diff are all table-padding in the fork-owned ledger.

Not done: a full GUI launch with an interactive sign-in. A live app instance was running on port 3773 and a second instance risked disturbing that session. The renderer half is covered by the byte-identity proof and the main-process half by the execution probe — installing this build and signing in once is the natural last check before it ships.

🤖 Generated with Claude Code

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: ecdf249e-1114-4881-912b-c683a53efde6

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[radroid]

Closes#53.

app.asar goes from 189.66 MiB / 14,765 files to 99.02 MiB / 3,429 files (−47.8% bytes, −76.8% files). Measured, not estimated, and re-measured on 136b3514f after the #71 Coil rename.

	Baseline	After	Δ
app.asar	189.66 MiB	99.02 MiB	−90.64 MiB (−47.8%)
Files in the asar	14,765	3,429	−76.8%
.zip (what the updater downloads)	144,592,118 B	124,467,727 B	−20.1 MB
.dmg	150,392,957 B	129,226,158 B	−21.2 MB
electron-builder phase	~126 s	~48 s	−62%
Staged vp install --prod	12.6 s	4.4 s	−65%

The compressed artifacts fall 14% rather than 48% because what was removed is text, which compressed well already. The .zip number is the one users pay, on every release.

Rebuilt on the renamed fork

This branch was originally cut before #71. It has been rebuilt from 136b3514f, not merged forward, so nothing carries an old path:

• scripts/t3x/ → scripts/coil/ (7 new files), docs/t3x/ → docs/coil/.
• .github/workflows/t3x-release.yml → coil-release.yml; every step anchor re-verified against the current file.
• refactor(coil): retire the local desktop autobuild #92 retired the local autobuild, so the edit this branch previously made to auto-build-desktop.sh is dropped. The release workflow is now the only path that builds a shipping artifact, and the test that used to assert the autobuild set the variables now asserts that no second build path has reappeared without them.
• T3X_* env names are kept, matching what Rename the fork's identity from t3x to coil (everything except the landing site) #71 deliberately did with T3X_DESKTOP_APP_ID and the T3X_MAC_CSC_* secrets. The new variable sits next to that one and reads as a pair.
• Doc references to T3 Code (Alpha).app → T3 Coil (Alpha).app. artifactName is still upstream's T3-Code-…, which Rename the fork's identity from t3x to coil (everything except the landing site) #71 deliberately left alone, so the artifact filenames in the doc are unchanged and correct.

The rename moved none of the measurements. Both builds were re-run on the new base: the packed asar is the same size on both sides (189.66 → 99.02 MiB), only the .dmg/.zip moved by ~3 KB from the renamed product string. All 66 non-wildcard globs still match a package that is actually shipped, so none has gone stale. The 385-chunk renderer byte-identity proof and the Clerk execution probe were both re-run and both reproduce.

Sync cost: one upstream file, strictly additive

Upstream-owned files touched	1 (scripts/build-desktop-artifact.ts)
This issue's delta on it	+21 / −0
That file's total fork delta	+32 / −1 — the −1 is #70's pre-existing displacement
New seam rows	0
New fork-owned files	8, under scripts/coil/ and docs/coil/ — paths upstream has never had

The additive-seam invariant holds: no deletion count moved and the file count did not change. The only realistic conflict is textual — upstream adding its own entry to DESKTOP_FILE_EXCLUSIONS collides with the ...coilExtraFileExclusions, line at the end of that array, a keep-both resolution.

The justification for the env hook has been corrected. The earlier version of this PR said an inline list would "cost a second seam row on a second upstream file." Since #71 that is no longer true — build-desktop-artifact.test.ts already carries a fork row. The real reason is stronger and survives the rename: the list is 67 globs and grows, so inline it makes every future size fix an edit to an upstream test assertion, in a file resolved by hand at every sync, where the conflict is decided by re-reading 67 lines instead of one. Through the environment, an unset environment packages exactly what upstream packages and upstream's deepStrictEqual keeps passing untouched.

What was cut, and the evidence for each

Web renderer source maps — 37.4 MiB. The largest line item in the bundle, unmentioned in #53. Turned off with T3CODE_WEB_SOURCEMAP=0, an existing upstream variable, so the biggest win costs zero seam rows. Provably behaviour-preserving: all 385 renderer chunks are byte-identical after stripping the trailing //# sourceMappingURL= comment, and their content-hashed filenames do not move.

Third-party source maps — 18.2 MiB. Node only parses a .map under --enable-source-maps, which the desktop backend passes through rather than sets. First-party maps (~12 MiB) are kept so an operator who does export it still gets legible frames for our own code.

Clerk's browser SDK — 23.2 MiB, 16 packages. The main process requires only two of @clerk/electron's six entry points (root and /storage), which between them require only electron and electron-store. The entries pulling the browser SDK are /react and /passkeys, whose only importer is apps/web/src/main.tsx — compiled before staging, and run in a renderer created with sandbox: true, nodeIntegration: false, which has no module resolution at all.

Verified by execution: loading @clerk/electron and @clerk/electron/storage from inside the optimized asar under ELECTRON_RUN_AS_NODE returns createClerkBridge, setupPasskeysMain and storage, and the transitive module list touches none of the excluded packages. @clerk/electron, @clerk/electron-passkeys and its native binaries are not excluded, and a test asserts they never will be.

The shiki / @pierre/diffs tree — 21.0 MiB, 50 packages. This answers #53 §3.3. The server imports one subpath (utils/parsePatchFiles) and the server build bundles it — bin.mjs has no @pierre/* import left. The highlighting users see is the web client's own per-language chunks, so @shikijs/langs was shipped twice and this was the copy nobody could load.

Two tools, and two bugs testing found in them

desktop-bundle-reachability.mjs (analysis) follows literal import edges and string-mention edges. The second exists because PlaywrightInjectedRuntime.ts resolves playwright-core through a const — an import-only scan calls 10.2 MiB of live code dead.

verify-desktop-bundle.mjs gates every release on both platforms. Testing the gate rather than trusting it found two real defects:

• It originally read only the archive. On Windows asarUnpack puts the bundles and all of node_modules on disk, so it would have failed every Windows release.
• Its severity originally depended on attribution. A real build with a deliberately over-broad !**/node_modules/effect/**/* produced a bundle that cannot start and the gate went green. Severity is now unconditional; attribution only makes the message actionable. That case is now a fixture test.

Five corrections to issue #53

In docs/coil/desktop-bundle-size.md:

1. The size table measured du block allocation, not content.core-js "15 MB (3,684 files)" is 1.25 MiB — 3,684 files × 4 KiB blocks. Every many-small-files package is overstated the same way, which is why the Clerk tree was not the biggest win.
2. Source maps were 35.6% of the asar and go unmentioned.
3. DESKTOP_FILE_EXCLUSIONS is not fork-owned. The recommendation to prefer it is right; the reason is not.
4. The mac leg is not the release critical path. The legs run in parallel; Windows is 17.5 min to mac's 9.6. And --concurrency-limit 1 costs only 48–58 s, so raising it buys ~20 s while re-opening Windows desktop build intermittently fails with STATUS_DLL_INIT_FAILED (0xC0000142) in parallel vp tasks #47's crashes — recommend leaving it alone.
5. The pnpm store is already cached.setup-vp's cache: true restores a 1,943 MB cache with downloaded 0.

The real remaining build-time target is the 442 s Spectre MSVC install — 42% of the Windows leg. Not touched: it cannot be verified without a Windows release run, and this workflow publishes on success.

Testing

• 86 tests across the affected files pass; upstream's build-desktop-artifact.test.ts passes unmodified.
• Full before/after builds on the post-rename base; the verifier passes on both and fails the deliberately-broken one.
• Real-asar fixture tests pin the header offset arithmetic.
• tsgo --noEmit clean, vp lint clean, vp fmt applied. Formatter churn outside this change's scope was reverted; the 50 deletions in the diff are all table-padding in the fork-owned ledger.

Not done: a full GUI launch with an interactive sign-in. A live app instance was running on port 3773 and a second instance risked disturbing that session. The renderer half is covered by the byte-identity proof and the main-process half by the execution probe — installing this build and signing in once is the natural last check before it ships.

🤖 Generated with Claude Code

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: ecdf249e-1114-4881-912b-c683a53efde6

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}