sync: absorb the 113-commit upstream range three daily syncs could not (#117) by radroid · Pull Request #119 · radroid/t3code

radroid · 2026-08-17T18:01:51Z

Closes#117. Absorbs the 113-commit upstream range that three consecutive daily syncs (08-15, 08-16, 08-17) could not, and replays all 166 fork patches on top of cd096b9ad.

Important

Do not use the GitHub merge button. This is the fork's patch series rebased onto new upstream, so main is not an ancestor of it and GitHub reports it CONFLICTING. That is expected and is not a defect in this branch. Land it by force-updating main:

git fetch origin
OLD_MAIN=$(git rev-parse origin/main)# save: needed to rebase in-flight branches
git push --force-with-lease origin origin/coil/sync-20260817:main

Rollback point is coil/last-good-20260817-083055 (verified == the pre-sync origin/main). If you'd rather have a genuinely button-mergeable PR, say so and I'll re-shape this as a merge commit instead — the tree would be identical.

Why the automation kept failing

Each daily run stops at the first conflict and escalates. Two different ones were in the way, so fixing either alone would not have unblocked it:

Day	Conflicted file	Upstream culprit
08-15	`apps/desktop/src/ipc/DesktopIpcHandlers.ts`	`57a299a7` open remote environments over SSH (pingdotgg#6572)
08-16, 08-17	`apps/server/src/sourceControl/SourceControlProviderDiscovery.ts`	`d715c2e56` raise the discovery probe budget (pingdotgg#6223)

Six conflicts surfaced in total across the replay. The three interesting ones:

1. Upstream independently fixed half of fork issue #4 — and did it better.d715c2e56 added a per-spec probeTimeoutMs with az at 20s; the fork carried a global 15s CLI_PROBE_TIMEOUT_MS. The fork's constant is deleted and the timeout ceded entirely to upstream. What upstream still does not do is the behavioural half — classifying only VcsProcessSpawnError as "missing", so a slow-but-present CLI stays available and the auth probe still runs — so that is all the fork keeps. Seam: +20/-8 → +12/-5.

2. The stale-hoist hazard on ComposerPrimaryActions.tsx is retired. Upstream 7afa184a9 (pingdotgg#4781) hoisted its inline send button to a const — the exact hoist the fork had been carrying since #35, and the one the ledger has flagged since 08-08 as able to silently revert an upstream restyle, because a copy cannot conflict. The two copies turned out byte-identical apart from one aria-label branch, so the fork's copy is deleted and upstream's kept. The fork's queue/steer branch was then folded into upstream's new bottom dispatch rather than left as the early return it used to be — as an early return it would have pre-empted upstream's whole mobile path and silently reverted pingdotgg#4781. Seam: +128/-55 → +74/-5; the -55 → -5 is the real signal, the fork now extends upstream's dispatch instead of displacing it.

3. manifest.webmanifest became an add/add seam. Upstream now ships its own. Upstream's has no name/short_name/description and no maskable icon — an installable PWA needs all four, which is what Web Push (#23) rides on — so the fork's fields are unioned onto upstream's, dropping the fork's duplicate apple-touch-icon.

Adapting to upstream, beyond the conflicts

Four commits sit on top of the replayed series:

no-native-title-tooltip — upstream added this oxlint rule in fee10def1 (feat(lint): ban native title tooltips and migrate to styled Tooltip pingdotgg/t3code#7209) and migrated its own call sites. The fork's UpdateToast.tsx was not in that sweep, so a new upstream rule turned fork lint red on a file upstream has never seen. Migrated to the same Tooltip/TooltipTrigger/TooltipPopup idiom.
DMG title assertion — ad117235b (feat(desktop): add signal macOS DMG installer background pingdotgg/t3code#6201) added a deepStrictEqual over the whole mac.dmg config, including a title built from resolveDesktopProductName, which Rename the fork's identity from t3x to coil (everything except the landing site) #71 renamed. Same one-word substitution the fork already makes ~300 lines above in that file.
Running-send aria-label — fix(web): keep send reachable while a turn is running on mobile pingdotgg/t3code#4781's new test expects aria-label="Send message"; the fork lengthens exactly that label to "Send message to the running turn" (Thread outbox: UI polish, queue images, send-and-interrupt — and decide whether the queue should survive #35), because on a steer-capable driver the submit folds into the work in progress and nothing else says so. The label stays the fork's; upstream's actual subject (a submit button beside Stop, at the larger size) is still pinned by the untouched assertions around it.
Seam ledger re-baseline — both halves against cd096b9ad, including the churn/risk columns, which are measured over the 60 days before the merge-base and so go stale every time it moves.

The checks CI cannot do

Logic mirrors re-checked. The steer allowlist has no capability flag behind it — if upstream changed an adapter to stop steering, nothing fails, and the fork would keep sending mid-turn into a provider that drops it. All four allowlisted drivers still steer (three still covered by upstream's own "steers a running turn..." tests, Grok still reuses ctx.activeTurnId), and grep -rn 'turn/steer' apps/server is still empty, so the deliberate Codex exclusion holds. The mobile send-label mirror also still matches — upstream's only mobile composer commit this range was theming.
Parallel path reviewed, nothing ported.apps/coil-home/ is a copy of apps/marketing/, so it drifts without ever conflicting. Both upstream marketing commits were checked and deliberately skipped: the Mac-chip detection (fix(marketing): detect Mac chip on homepage download button pingdotgg/t3code#4197) has nothing to choose between, since the release matrix builds only darwin-arm64 and win32-x64; and the Grok-mark mobile fix (fix(marketing): keep Grok mark clear of mobile hero copy pingdotgg/t3code#4542) repositions absolutely-positioned marks the fork already replaced with an in-flow row, so porting it would regress the fork's own fix.
Lockfile untouched. Upstream changed no dependency in this range — its only manifest edit adds two exports entries to packages/shared/package.json. pnpm-lock.yaml is byte-identical to pre-sync main, so the dependency sweep is intact and no regeneration was needed. (This is the step that destroyed the sweep on 08-08.)

Verification

vp run typecheck — exit 0, 17/17 packages.
vp run lint — exit 0, no error-level findings.
vp run test --testTimeout=120000 --hookTimeout=120000 — one fully green run (2708 passed | 7 skipped, exit 0).
Patch integrity: 166/166 fork patches replayed, subject-for-subject identical to pre-sync main. Nothing dropped, nothing absorbed.

Two caveats, both honest:

Later full-suite runs each hit one intermittently failing test — composerInlineTokens (a < 1000ms wall-clock assertion) and nativeReviewDiffHighlighter. Both are upstream-owned, untouched by the fork, and pass 3/3 in isolation; they fail only under full-suite parallel load on this machine. CI is the arbiter.
build-desktop-artifact's cross-arch Windows probe test fails locally for an environment reason, not a code one: the probe does {...process.env}, and this machine's shell has ELECTRON_RUN_AS_NODE=1 because the agent runs inside Electron-as-Node. Verified by control run — pristine upstream code fails it identically here — and it passes with the variable unset. It will pass in CI.

Not included

#118 (auto-resume never armed) is open and unrelated to the sync; it is untouched here.

radroid · 2026-08-17T18:12:26Z

CI is green — run 32052894658: Typecheck ✅ Lint ✅ Test ✅.

That settles both caveats in the description. Neither of the intermittently-failing tests (composerInlineTokens, nativeReviewDiffHighlighter) failed on CI, and neither did the build-desktop-artifact cross-arch probe test — confirming all three were local environment artifacts (parallel-load contention, and this machine's ambient ELECTRON_RUN_AS_NODE=1) rather than anything this branch changed.

coderabbitai · 2026-08-17T20:02:01Z

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: 1a471a64-016a-4afb-a024-f7b5c6541362

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

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

radroid · 2026-08-17T18:01:51Z

Closes#117. Absorbs the 113-commit upstream range that three consecutive daily syncs (08-15, 08-16, 08-17) could not, and replays all 166 fork patches on top of cd096b9ad.

Important

Do not use the GitHub merge button. This is the fork's patch series rebased onto new upstream, so main is not an ancestor of it and GitHub reports it CONFLICTING. That is expected and is not a defect in this branch. Land it by force-updating main:

git fetch origin
OLD_MAIN=$(git rev-parse origin/main)# save: needed to rebase in-flight branches
git push --force-with-lease origin origin/coil/sync-20260817:main

Rollback point is coil/last-good-20260817-083055 (verified == the pre-sync origin/main). If you'd rather have a genuinely button-mergeable PR, say so and I'll re-shape this as a merge commit instead — the tree would be identical.

Why the automation kept failing

Each daily run stops at the first conflict and escalates. Two different ones were in the way, so fixing either alone would not have unblocked it:

Day	Conflicted file	Upstream culprit
08-15	`apps/desktop/src/ipc/DesktopIpcHandlers.ts`	`57a299a7` open remote environments over SSH (pingdotgg#6572)
08-16, 08-17	`apps/server/src/sourceControl/SourceControlProviderDiscovery.ts`	`d715c2e56` raise the discovery probe budget (pingdotgg#6223)

Six conflicts surfaced in total across the replay. The three interesting ones:

1. Upstream independently fixed half of fork issue #4 — and did it better.d715c2e56 added a per-spec probeTimeoutMs with az at 20s; the fork carried a global 15s CLI_PROBE_TIMEOUT_MS. The fork's constant is deleted and the timeout ceded entirely to upstream. What upstream still does not do is the behavioural half — classifying only VcsProcessSpawnError as "missing", so a slow-but-present CLI stays available and the auth probe still runs — so that is all the fork keeps. Seam: +20/-8 → +12/-5.

2. The stale-hoist hazard on ComposerPrimaryActions.tsx is retired. Upstream 7afa184a9 (pingdotgg#4781) hoisted its inline send button to a const — the exact hoist the fork had been carrying since #35, and the one the ledger has flagged since 08-08 as able to silently revert an upstream restyle, because a copy cannot conflict. The two copies turned out byte-identical apart from one aria-label branch, so the fork's copy is deleted and upstream's kept. The fork's queue/steer branch was then folded into upstream's new bottom dispatch rather than left as the early return it used to be — as an early return it would have pre-empted upstream's whole mobile path and silently reverted pingdotgg#4781. Seam: +128/-55 → +74/-5; the -55 → -5 is the real signal, the fork now extends upstream's dispatch instead of displacing it.

3. manifest.webmanifest became an add/add seam. Upstream now ships its own. Upstream's has no name/short_name/description and no maskable icon — an installable PWA needs all four, which is what Web Push (#23) rides on — so the fork's fields are unioned onto upstream's, dropping the fork's duplicate apple-touch-icon.

Adapting to upstream, beyond the conflicts

Four commits sit on top of the replayed series:

no-native-title-tooltip — upstream added this oxlint rule in fee10def1 (feat(lint): ban native title tooltips and migrate to styled Tooltip pingdotgg/t3code#7209) and migrated its own call sites. The fork's UpdateToast.tsx was not in that sweep, so a new upstream rule turned fork lint red on a file upstream has never seen. Migrated to the same Tooltip/TooltipTrigger/TooltipPopup idiom.
DMG title assertion — ad117235b (feat(desktop): add signal macOS DMG installer background pingdotgg/t3code#6201) added a deepStrictEqual over the whole mac.dmg config, including a title built from resolveDesktopProductName, which Rename the fork's identity from t3x to coil (everything except the landing site) #71 renamed. Same one-word substitution the fork already makes ~300 lines above in that file.
Running-send aria-label — fix(web): keep send reachable while a turn is running on mobile pingdotgg/t3code#4781's new test expects aria-label="Send message"; the fork lengthens exactly that label to "Send message to the running turn" (Thread outbox: UI polish, queue images, send-and-interrupt — and decide whether the queue should survive #35), because on a steer-capable driver the submit folds into the work in progress and nothing else says so. The label stays the fork's; upstream's actual subject (a submit button beside Stop, at the larger size) is still pinned by the untouched assertions around it.
Seam ledger re-baseline — both halves against cd096b9ad, including the churn/risk columns, which are measured over the 60 days before the merge-base and so go stale every time it moves.

The checks CI cannot do

Logic mirrors re-checked. The steer allowlist has no capability flag behind it — if upstream changed an adapter to stop steering, nothing fails, and the fork would keep sending mid-turn into a provider that drops it. All four allowlisted drivers still steer (three still covered by upstream's own "steers a running turn..." tests, Grok still reuses ctx.activeTurnId), and grep -rn 'turn/steer' apps/server is still empty, so the deliberate Codex exclusion holds. The mobile send-label mirror also still matches — upstream's only mobile composer commit this range was theming.
Parallel path reviewed, nothing ported.apps/coil-home/ is a copy of apps/marketing/, so it drifts without ever conflicting. Both upstream marketing commits were checked and deliberately skipped: the Mac-chip detection (fix(marketing): detect Mac chip on homepage download button pingdotgg/t3code#4197) has nothing to choose between, since the release matrix builds only darwin-arm64 and win32-x64; and the Grok-mark mobile fix (fix(marketing): keep Grok mark clear of mobile hero copy pingdotgg/t3code#4542) repositions absolutely-positioned marks the fork already replaced with an in-flow row, so porting it would regress the fork's own fix.
Lockfile untouched. Upstream changed no dependency in this range — its only manifest edit adds two exports entries to packages/shared/package.json. pnpm-lock.yaml is byte-identical to pre-sync main, so the dependency sweep is intact and no regeneration was needed. (This is the step that destroyed the sweep on 08-08.)

Verification

vp run typecheck — exit 0, 17/17 packages.
vp run lint — exit 0, no error-level findings.
vp run test --testTimeout=120000 --hookTimeout=120000 — one fully green run (2708 passed | 7 skipped, exit 0).
Patch integrity: 166/166 fork patches replayed, subject-for-subject identical to pre-sync main. Nothing dropped, nothing absorbed.

Two caveats, both honest:

Later full-suite runs each hit one intermittently failing test — composerInlineTokens (a < 1000ms wall-clock assertion) and nativeReviewDiffHighlighter. Both are upstream-owned, untouched by the fork, and pass 3/3 in isolation; they fail only under full-suite parallel load on this machine. CI is the arbiter.
build-desktop-artifact's cross-arch Windows probe test fails locally for an environment reason, not a code one: the probe does {...process.env}, and this machine's shell has ELECTRON_RUN_AS_NODE=1 because the agent runs inside Electron-as-Node. Verified by control run — pristine upstream code fails it identically here — and it passes with the variable unset. It will pass in CI.

Not included

#118 (auto-resume never armed) is open and unrelated to the sync; it is untouched here.

radroid · 2026-08-17T18:12:26Z

CI is green — run 32052894658: Typecheck ✅ Lint ✅ Test ✅.

That settles both caveats in the description. Neither of the intermittently-failing tests (composerInlineTokens, nativeReviewDiffHighlighter) failed on CI, and neither did the build-desktop-artifact cross-arch probe test — confirming all three were local environment artifacts (parallel-load contention, and this machine's ambient ELECTRON_RUN_AS_NODE=1) rather than anything this branch changed.

coderabbitai · 2026-08-17T20:02:01Z

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: 1a471a64-016a-4afb-a024-f7b5c6541362

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

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

radroid · 2026-08-17T18:01:51Z

Closes#117. Absorbs the 113-commit upstream range that three consecutive daily syncs (08-15, 08-16, 08-17) could not, and replays all 166 fork patches on top of cd096b9ad.

Important

Do not use the GitHub merge button. This is the fork's patch series rebased onto new upstream, so main is not an ancestor of it and GitHub reports it CONFLICTING. That is expected and is not a defect in this branch. Land it by force-updating main:

git fetch origin
OLD_MAIN=$(git rev-parse origin/main)# save: needed to rebase in-flight branches
git push --force-with-lease origin origin/coil/sync-20260817:main

Rollback point is coil/last-good-20260817-083055 (verified == the pre-sync origin/main). If you'd rather have a genuinely button-mergeable PR, say so and I'll re-shape this as a merge commit instead — the tree would be identical.

Why the automation kept failing

Each daily run stops at the first conflict and escalates. Two different ones were in the way, so fixing either alone would not have unblocked it:

Day	Conflicted file	Upstream culprit
08-15	`apps/desktop/src/ipc/DesktopIpcHandlers.ts`	`57a299a7` open remote environments over SSH (pingdotgg#6572)
08-16, 08-17	`apps/server/src/sourceControl/SourceControlProviderDiscovery.ts`	`d715c2e56` raise the discovery probe budget (pingdotgg#6223)

Six conflicts surfaced in total across the replay. The three interesting ones:

1. Upstream independently fixed half of fork issue #4 — and did it better.d715c2e56 added a per-spec probeTimeoutMs with az at 20s; the fork carried a global 15s CLI_PROBE_TIMEOUT_MS. The fork's constant is deleted and the timeout ceded entirely to upstream. What upstream still does not do is the behavioural half — classifying only VcsProcessSpawnError as "missing", so a slow-but-present CLI stays available and the auth probe still runs — so that is all the fork keeps. Seam: +20/-8 → +12/-5.

2. The stale-hoist hazard on ComposerPrimaryActions.tsx is retired. Upstream 7afa184a9 (pingdotgg#4781) hoisted its inline send button to a const — the exact hoist the fork had been carrying since #35, and the one the ledger has flagged since 08-08 as able to silently revert an upstream restyle, because a copy cannot conflict. The two copies turned out byte-identical apart from one aria-label branch, so the fork's copy is deleted and upstream's kept. The fork's queue/steer branch was then folded into upstream's new bottom dispatch rather than left as the early return it used to be — as an early return it would have pre-empted upstream's whole mobile path and silently reverted pingdotgg#4781. Seam: +128/-55 → +74/-5; the -55 → -5 is the real signal, the fork now extends upstream's dispatch instead of displacing it.

3. manifest.webmanifest became an add/add seam. Upstream now ships its own. Upstream's has no name/short_name/description and no maskable icon — an installable PWA needs all four, which is what Web Push (#23) rides on — so the fork's fields are unioned onto upstream's, dropping the fork's duplicate apple-touch-icon.

Adapting to upstream, beyond the conflicts

Four commits sit on top of the replayed series:

no-native-title-tooltip — upstream added this oxlint rule in fee10def1 (feat(lint): ban native title tooltips and migrate to styled Tooltip pingdotgg/t3code#7209) and migrated its own call sites. The fork's UpdateToast.tsx was not in that sweep, so a new upstream rule turned fork lint red on a file upstream has never seen. Migrated to the same Tooltip/TooltipTrigger/TooltipPopup idiom.
DMG title assertion — ad117235b (feat(desktop): add signal macOS DMG installer background pingdotgg/t3code#6201) added a deepStrictEqual over the whole mac.dmg config, including a title built from resolveDesktopProductName, which Rename the fork's identity from t3x to coil (everything except the landing site) #71 renamed. Same one-word substitution the fork already makes ~300 lines above in that file.
Running-send aria-label — fix(web): keep send reachable while a turn is running on mobile pingdotgg/t3code#4781's new test expects aria-label="Send message"; the fork lengthens exactly that label to "Send message to the running turn" (Thread outbox: UI polish, queue images, send-and-interrupt — and decide whether the queue should survive #35), because on a steer-capable driver the submit folds into the work in progress and nothing else says so. The label stays the fork's; upstream's actual subject (a submit button beside Stop, at the larger size) is still pinned by the untouched assertions around it.
Seam ledger re-baseline — both halves against cd096b9ad, including the churn/risk columns, which are measured over the 60 days before the merge-base and so go stale every time it moves.

The checks CI cannot do

Logic mirrors re-checked. The steer allowlist has no capability flag behind it — if upstream changed an adapter to stop steering, nothing fails, and the fork would keep sending mid-turn into a provider that drops it. All four allowlisted drivers still steer (three still covered by upstream's own "steers a running turn..." tests, Grok still reuses ctx.activeTurnId), and grep -rn 'turn/steer' apps/server is still empty, so the deliberate Codex exclusion holds. The mobile send-label mirror also still matches — upstream's only mobile composer commit this range was theming.
Parallel path reviewed, nothing ported.apps/coil-home/ is a copy of apps/marketing/, so it drifts without ever conflicting. Both upstream marketing commits were checked and deliberately skipped: the Mac-chip detection (fix(marketing): detect Mac chip on homepage download button pingdotgg/t3code#4197) has nothing to choose between, since the release matrix builds only darwin-arm64 and win32-x64; and the Grok-mark mobile fix (fix(marketing): keep Grok mark clear of mobile hero copy pingdotgg/t3code#4542) repositions absolutely-positioned marks the fork already replaced with an in-flow row, so porting it would regress the fork's own fix.
Lockfile untouched. Upstream changed no dependency in this range — its only manifest edit adds two exports entries to packages/shared/package.json. pnpm-lock.yaml is byte-identical to pre-sync main, so the dependency sweep is intact and no regeneration was needed. (This is the step that destroyed the sweep on 08-08.)

Verification

vp run typecheck — exit 0, 17/17 packages.
vp run lint — exit 0, no error-level findings.
vp run test --testTimeout=120000 --hookTimeout=120000 — one fully green run (2708 passed | 7 skipped, exit 0).
Patch integrity: 166/166 fork patches replayed, subject-for-subject identical to pre-sync main. Nothing dropped, nothing absorbed.

Two caveats, both honest:

Later full-suite runs each hit one intermittently failing test — composerInlineTokens (a < 1000ms wall-clock assertion) and nativeReviewDiffHighlighter. Both are upstream-owned, untouched by the fork, and pass 3/3 in isolation; they fail only under full-suite parallel load on this machine. CI is the arbiter.
build-desktop-artifact's cross-arch Windows probe test fails locally for an environment reason, not a code one: the probe does {...process.env}, and this machine's shell has ELECTRON_RUN_AS_NODE=1 because the agent runs inside Electron-as-Node. Verified by control run — pristine upstream code fails it identically here — and it passes with the variable unset. It will pass in CI.

Not included

#118 (auto-resume never armed) is open and unrelated to the sync; it is untouched here.

radroid · 2026-08-17T18:12:26Z

CI is green — run 32052894658: Typecheck ✅ Lint ✅ Test ✅.

That settles both caveats in the description. Neither of the intermittently-failing tests (composerInlineTokens, nativeReviewDiffHighlighter) failed on CI, and neither did the build-desktop-artifact cross-arch probe test — confirming all three were local environment artifacts (parallel-load contention, and this machine's ambient ELECTRON_RUN_AS_NODE=1) rather than anything this branch changed.

coderabbitai · 2026-08-17T20:02:01Z

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: 1a471a64-016a-4afb-a024-f7b5c6541362

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

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

radroid · 2026-08-17T18:01:51Z

Closes#117. Absorbs the 113-commit upstream range that three consecutive daily syncs (08-15, 08-16, 08-17) could not, and replays all 166 fork patches on top of cd096b9ad.

Important

Do not use the GitHub merge button. This is the fork's patch series rebased onto new upstream, so main is not an ancestor of it and GitHub reports it CONFLICTING. That is expected and is not a defect in this branch. Land it by force-updating main:

git fetch origin
OLD_MAIN=$(git rev-parse origin/main)# save: needed to rebase in-flight branches
git push --force-with-lease origin origin/coil/sync-20260817:main

Rollback point is coil/last-good-20260817-083055 (verified == the pre-sync origin/main). If you'd rather have a genuinely button-mergeable PR, say so and I'll re-shape this as a merge commit instead — the tree would be identical.

Why the automation kept failing

Each daily run stops at the first conflict and escalates. Two different ones were in the way, so fixing either alone would not have unblocked it:

Day	Conflicted file	Upstream culprit
08-15	`apps/desktop/src/ipc/DesktopIpcHandlers.ts`	`57a299a7` open remote environments over SSH (pingdotgg#6572)
08-16, 08-17	`apps/server/src/sourceControl/SourceControlProviderDiscovery.ts`	`d715c2e56` raise the discovery probe budget (pingdotgg#6223)

Six conflicts surfaced in total across the replay. The three interesting ones:

1. Upstream independently fixed half of fork issue #4 — and did it better.d715c2e56 added a per-spec probeTimeoutMs with az at 20s; the fork carried a global 15s CLI_PROBE_TIMEOUT_MS. The fork's constant is deleted and the timeout ceded entirely to upstream. What upstream still does not do is the behavioural half — classifying only VcsProcessSpawnError as "missing", so a slow-but-present CLI stays available and the auth probe still runs — so that is all the fork keeps. Seam: +20/-8 → +12/-5.

2. The stale-hoist hazard on ComposerPrimaryActions.tsx is retired. Upstream 7afa184a9 (pingdotgg#4781) hoisted its inline send button to a const — the exact hoist the fork had been carrying since #35, and the one the ledger has flagged since 08-08 as able to silently revert an upstream restyle, because a copy cannot conflict. The two copies turned out byte-identical apart from one aria-label branch, so the fork's copy is deleted and upstream's kept. The fork's queue/steer branch was then folded into upstream's new bottom dispatch rather than left as the early return it used to be — as an early return it would have pre-empted upstream's whole mobile path and silently reverted pingdotgg#4781. Seam: +128/-55 → +74/-5; the -55 → -5 is the real signal, the fork now extends upstream's dispatch instead of displacing it.

3. manifest.webmanifest became an add/add seam. Upstream now ships its own. Upstream's has no name/short_name/description and no maskable icon — an installable PWA needs all four, which is what Web Push (#23) rides on — so the fork's fields are unioned onto upstream's, dropping the fork's duplicate apple-touch-icon.

Adapting to upstream, beyond the conflicts

Four commits sit on top of the replayed series:

no-native-title-tooltip — upstream added this oxlint rule in fee10def1 (feat(lint): ban native title tooltips and migrate to styled Tooltip pingdotgg/t3code#7209) and migrated its own call sites. The fork's UpdateToast.tsx was not in that sweep, so a new upstream rule turned fork lint red on a file upstream has never seen. Migrated to the same Tooltip/TooltipTrigger/TooltipPopup idiom.
DMG title assertion — ad117235b (feat(desktop): add signal macOS DMG installer background pingdotgg/t3code#6201) added a deepStrictEqual over the whole mac.dmg config, including a title built from resolveDesktopProductName, which Rename the fork's identity from t3x to coil (everything except the landing site) #71 renamed. Same one-word substitution the fork already makes ~300 lines above in that file.
Running-send aria-label — fix(web): keep send reachable while a turn is running on mobile pingdotgg/t3code#4781's new test expects aria-label="Send message"; the fork lengthens exactly that label to "Send message to the running turn" (Thread outbox: UI polish, queue images, send-and-interrupt — and decide whether the queue should survive #35), because on a steer-capable driver the submit folds into the work in progress and nothing else says so. The label stays the fork's; upstream's actual subject (a submit button beside Stop, at the larger size) is still pinned by the untouched assertions around it.
Seam ledger re-baseline — both halves against cd096b9ad, including the churn/risk columns, which are measured over the 60 days before the merge-base and so go stale every time it moves.

The checks CI cannot do

Logic mirrors re-checked. The steer allowlist has no capability flag behind it — if upstream changed an adapter to stop steering, nothing fails, and the fork would keep sending mid-turn into a provider that drops it. All four allowlisted drivers still steer (three still covered by upstream's own "steers a running turn..." tests, Grok still reuses ctx.activeTurnId), and grep -rn 'turn/steer' apps/server is still empty, so the deliberate Codex exclusion holds. The mobile send-label mirror also still matches — upstream's only mobile composer commit this range was theming.
Parallel path reviewed, nothing ported.apps/coil-home/ is a copy of apps/marketing/, so it drifts without ever conflicting. Both upstream marketing commits were checked and deliberately skipped: the Mac-chip detection (fix(marketing): detect Mac chip on homepage download button pingdotgg/t3code#4197) has nothing to choose between, since the release matrix builds only darwin-arm64 and win32-x64; and the Grok-mark mobile fix (fix(marketing): keep Grok mark clear of mobile hero copy pingdotgg/t3code#4542) repositions absolutely-positioned marks the fork already replaced with an in-flow row, so porting it would regress the fork's own fix.
Lockfile untouched. Upstream changed no dependency in this range — its only manifest edit adds two exports entries to packages/shared/package.json. pnpm-lock.yaml is byte-identical to pre-sync main, so the dependency sweep is intact and no regeneration was needed. (This is the step that destroyed the sweep on 08-08.)

Verification

vp run typecheck — exit 0, 17/17 packages.
vp run lint — exit 0, no error-level findings.
vp run test --testTimeout=120000 --hookTimeout=120000 — one fully green run (2708 passed | 7 skipped, exit 0).
Patch integrity: 166/166 fork patches replayed, subject-for-subject identical to pre-sync main. Nothing dropped, nothing absorbed.

Two caveats, both honest:

Later full-suite runs each hit one intermittently failing test — composerInlineTokens (a < 1000ms wall-clock assertion) and nativeReviewDiffHighlighter. Both are upstream-owned, untouched by the fork, and pass 3/3 in isolation; they fail only under full-suite parallel load on this machine. CI is the arbiter.
build-desktop-artifact's cross-arch Windows probe test fails locally for an environment reason, not a code one: the probe does {...process.env}, and this machine's shell has ELECTRON_RUN_AS_NODE=1 because the agent runs inside Electron-as-Node. Verified by control run — pristine upstream code fails it identically here — and it passes with the variable unset. It will pass in CI.

Not included

#118 (auto-resume never armed) is open and unrelated to the sync; it is untouched here.

radroid · 2026-08-17T18:12:26Z

CI is green — run 32052894658: Typecheck ✅ Lint ✅ Test ✅.

That settles both caveats in the description. Neither of the intermittently-failing tests (composerInlineTokens, nativeReviewDiffHighlighter) failed on CI, and neither did the build-desktop-artifact cross-arch probe test — confirming all three were local environment artifacts (parallel-load contention, and this machine's ambient ELECTRON_RUN_AS_NODE=1) rather than anything this branch changed.

coderabbitai · 2026-08-17T20:02:01Z

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: 1a471a64-016a-4afb-a024-f7b5c6541362

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

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

radroid · 2026-08-17T18:01:51Z

Closes#117. Absorbs the 113-commit upstream range that three consecutive daily syncs (08-15, 08-16, 08-17) could not, and replays all 166 fork patches on top of cd096b9ad.

Important

Do not use the GitHub merge button. This is the fork's patch series rebased onto new upstream, so main is not an ancestor of it and GitHub reports it CONFLICTING. That is expected and is not a defect in this branch. Land it by force-updating main:

git fetch origin
OLD_MAIN=$(git rev-parse origin/main)# save: needed to rebase in-flight branches
git push --force-with-lease origin origin/coil/sync-20260817:main

Rollback point is coil/last-good-20260817-083055 (verified == the pre-sync origin/main). If you'd rather have a genuinely button-mergeable PR, say so and I'll re-shape this as a merge commit instead — the tree would be identical.

Why the automation kept failing

Each daily run stops at the first conflict and escalates. Two different ones were in the way, so fixing either alone would not have unblocked it:

Day	Conflicted file	Upstream culprit
08-15	`apps/desktop/src/ipc/DesktopIpcHandlers.ts`	`57a299a7` open remote environments over SSH (pingdotgg#6572)
08-16, 08-17	`apps/server/src/sourceControl/SourceControlProviderDiscovery.ts`	`d715c2e56` raise the discovery probe budget (pingdotgg#6223)

Six conflicts surfaced in total across the replay. The three interesting ones:

1. Upstream independently fixed half of fork issue #4 — and did it better.d715c2e56 added a per-spec probeTimeoutMs with az at 20s; the fork carried a global 15s CLI_PROBE_TIMEOUT_MS. The fork's constant is deleted and the timeout ceded entirely to upstream. What upstream still does not do is the behavioural half — classifying only VcsProcessSpawnError as "missing", so a slow-but-present CLI stays available and the auth probe still runs — so that is all the fork keeps. Seam: +20/-8 → +12/-5.

2. The stale-hoist hazard on ComposerPrimaryActions.tsx is retired. Upstream 7afa184a9 (pingdotgg#4781) hoisted its inline send button to a const — the exact hoist the fork had been carrying since #35, and the one the ledger has flagged since 08-08 as able to silently revert an upstream restyle, because a copy cannot conflict. The two copies turned out byte-identical apart from one aria-label branch, so the fork's copy is deleted and upstream's kept. The fork's queue/steer branch was then folded into upstream's new bottom dispatch rather than left as the early return it used to be — as an early return it would have pre-empted upstream's whole mobile path and silently reverted pingdotgg#4781. Seam: +128/-55 → +74/-5; the -55 → -5 is the real signal, the fork now extends upstream's dispatch instead of displacing it.

3. manifest.webmanifest became an add/add seam. Upstream now ships its own. Upstream's has no name/short_name/description and no maskable icon — an installable PWA needs all four, which is what Web Push (#23) rides on — so the fork's fields are unioned onto upstream's, dropping the fork's duplicate apple-touch-icon.

Adapting to upstream, beyond the conflicts

Four commits sit on top of the replayed series:

no-native-title-tooltip — upstream added this oxlint rule in fee10def1 (feat(lint): ban native title tooltips and migrate to styled Tooltip pingdotgg/t3code#7209) and migrated its own call sites. The fork's UpdateToast.tsx was not in that sweep, so a new upstream rule turned fork lint red on a file upstream has never seen. Migrated to the same Tooltip/TooltipTrigger/TooltipPopup idiom.
DMG title assertion — ad117235b (feat(desktop): add signal macOS DMG installer background pingdotgg/t3code#6201) added a deepStrictEqual over the whole mac.dmg config, including a title built from resolveDesktopProductName, which Rename the fork's identity from t3x to coil (everything except the landing site) #71 renamed. Same one-word substitution the fork already makes ~300 lines above in that file.
Running-send aria-label — fix(web): keep send reachable while a turn is running on mobile pingdotgg/t3code#4781's new test expects aria-label="Send message"; the fork lengthens exactly that label to "Send message to the running turn" (Thread outbox: UI polish, queue images, send-and-interrupt — and decide whether the queue should survive #35), because on a steer-capable driver the submit folds into the work in progress and nothing else says so. The label stays the fork's; upstream's actual subject (a submit button beside Stop, at the larger size) is still pinned by the untouched assertions around it.
Seam ledger re-baseline — both halves against cd096b9ad, including the churn/risk columns, which are measured over the 60 days before the merge-base and so go stale every time it moves.

The checks CI cannot do

Logic mirrors re-checked. The steer allowlist has no capability flag behind it — if upstream changed an adapter to stop steering, nothing fails, and the fork would keep sending mid-turn into a provider that drops it. All four allowlisted drivers still steer (three still covered by upstream's own "steers a running turn..." tests, Grok still reuses ctx.activeTurnId), and grep -rn 'turn/steer' apps/server is still empty, so the deliberate Codex exclusion holds. The mobile send-label mirror also still matches — upstream's only mobile composer commit this range was theming.
Parallel path reviewed, nothing ported.apps/coil-home/ is a copy of apps/marketing/, so it drifts without ever conflicting. Both upstream marketing commits were checked and deliberately skipped: the Mac-chip detection (fix(marketing): detect Mac chip on homepage download button pingdotgg/t3code#4197) has nothing to choose between, since the release matrix builds only darwin-arm64 and win32-x64; and the Grok-mark mobile fix (fix(marketing): keep Grok mark clear of mobile hero copy pingdotgg/t3code#4542) repositions absolutely-positioned marks the fork already replaced with an in-flow row, so porting it would regress the fork's own fix.
Lockfile untouched. Upstream changed no dependency in this range — its only manifest edit adds two exports entries to packages/shared/package.json. pnpm-lock.yaml is byte-identical to pre-sync main, so the dependency sweep is intact and no regeneration was needed. (This is the step that destroyed the sweep on 08-08.)

Verification

vp run typecheck — exit 0, 17/17 packages.
vp run lint — exit 0, no error-level findings.
vp run test --testTimeout=120000 --hookTimeout=120000 — one fully green run (2708 passed | 7 skipped, exit 0).
Patch integrity: 166/166 fork patches replayed, subject-for-subject identical to pre-sync main. Nothing dropped, nothing absorbed.

Two caveats, both honest:

Later full-suite runs each hit one intermittently failing test — composerInlineTokens (a < 1000ms wall-clock assertion) and nativeReviewDiffHighlighter. Both are upstream-owned, untouched by the fork, and pass 3/3 in isolation; they fail only under full-suite parallel load on this machine. CI is the arbiter.
build-desktop-artifact's cross-arch Windows probe test fails locally for an environment reason, not a code one: the probe does {...process.env}, and this machine's shell has ELECTRON_RUN_AS_NODE=1 because the agent runs inside Electron-as-Node. Verified by control run — pristine upstream code fails it identically here — and it passes with the variable unset. It will pass in CI.

Not included

#118 (auto-resume never armed) is open and unrelated to the sync; it is untouched here.

radroid · 2026-08-17T18:12:26Z

CI is green — run 32052894658: Typecheck ✅ Lint ✅ Test ✅.

That settles both caveats in the description. Neither of the intermittently-failing tests (composerInlineTokens, nativeReviewDiffHighlighter) failed on CI, and neither did the build-desktop-artifact cross-arch probe test — confirming all three were local environment artifacts (parallel-load contention, and this machine's ambient ELECTRON_RUN_AS_NODE=1) rather than anything this branch changed.

coderabbitai · 2026-08-17T20:02:01Z

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: 1a471a64-016a-4afb-a024-f7b5c6541362

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

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

radroid · 2026-08-17T18:01:51Z

Closes#117. Absorbs the 113-commit upstream range that three consecutive daily syncs (08-15, 08-16, 08-17) could not, and replays all 166 fork patches on top of cd096b9ad.

Important

Do not use the GitHub merge button. This is the fork's patch series rebased onto new upstream, so main is not an ancestor of it and GitHub reports it CONFLICTING. That is expected and is not a defect in this branch. Land it by force-updating main:

git fetch origin
OLD_MAIN=$(git rev-parse origin/main)# save: needed to rebase in-flight branches
git push --force-with-lease origin origin/coil/sync-20260817:main

Rollback point is coil/last-good-20260817-083055 (verified == the pre-sync origin/main). If you'd rather have a genuinely button-mergeable PR, say so and I'll re-shape this as a merge commit instead — the tree would be identical.

Why the automation kept failing

Each daily run stops at the first conflict and escalates. Two different ones were in the way, so fixing either alone would not have unblocked it:

Day	Conflicted file	Upstream culprit
08-15	`apps/desktop/src/ipc/DesktopIpcHandlers.ts`	`57a299a7` open remote environments over SSH (pingdotgg#6572)
08-16, 08-17	`apps/server/src/sourceControl/SourceControlProviderDiscovery.ts`	`d715c2e56` raise the discovery probe budget (pingdotgg#6223)

Six conflicts surfaced in total across the replay. The three interesting ones:

1. Upstream independently fixed half of fork issue #4 — and did it better.d715c2e56 added a per-spec probeTimeoutMs with az at 20s; the fork carried a global 15s CLI_PROBE_TIMEOUT_MS. The fork's constant is deleted and the timeout ceded entirely to upstream. What upstream still does not do is the behavioural half — classifying only VcsProcessSpawnError as "missing", so a slow-but-present CLI stays available and the auth probe still runs — so that is all the fork keeps. Seam: +20/-8 → +12/-5.

2. The stale-hoist hazard on ComposerPrimaryActions.tsx is retired. Upstream 7afa184a9 (pingdotgg#4781) hoisted its inline send button to a const — the exact hoist the fork had been carrying since #35, and the one the ledger has flagged since 08-08 as able to silently revert an upstream restyle, because a copy cannot conflict. The two copies turned out byte-identical apart from one aria-label branch, so the fork's copy is deleted and upstream's kept. The fork's queue/steer branch was then folded into upstream's new bottom dispatch rather than left as the early return it used to be — as an early return it would have pre-empted upstream's whole mobile path and silently reverted pingdotgg#4781. Seam: +128/-55 → +74/-5; the -55 → -5 is the real signal, the fork now extends upstream's dispatch instead of displacing it.

3. manifest.webmanifest became an add/add seam. Upstream now ships its own. Upstream's has no name/short_name/description and no maskable icon — an installable PWA needs all four, which is what Web Push (#23) rides on — so the fork's fields are unioned onto upstream's, dropping the fork's duplicate apple-touch-icon.

Adapting to upstream, beyond the conflicts

Four commits sit on top of the replayed series:

no-native-title-tooltip — upstream added this oxlint rule in fee10def1 (feat(lint): ban native title tooltips and migrate to styled Tooltip pingdotgg/t3code#7209) and migrated its own call sites. The fork's UpdateToast.tsx was not in that sweep, so a new upstream rule turned fork lint red on a file upstream has never seen. Migrated to the same Tooltip/TooltipTrigger/TooltipPopup idiom.
DMG title assertion — ad117235b (feat(desktop): add signal macOS DMG installer background pingdotgg/t3code#6201) added a deepStrictEqual over the whole mac.dmg config, including a title built from resolveDesktopProductName, which Rename the fork's identity from t3x to coil (everything except the landing site) #71 renamed. Same one-word substitution the fork already makes ~300 lines above in that file.
Running-send aria-label — fix(web): keep send reachable while a turn is running on mobile pingdotgg/t3code#4781's new test expects aria-label="Send message"; the fork lengthens exactly that label to "Send message to the running turn" (Thread outbox: UI polish, queue images, send-and-interrupt — and decide whether the queue should survive #35), because on a steer-capable driver the submit folds into the work in progress and nothing else says so. The label stays the fork's; upstream's actual subject (a submit button beside Stop, at the larger size) is still pinned by the untouched assertions around it.
Seam ledger re-baseline — both halves against cd096b9ad, including the churn/risk columns, which are measured over the 60 days before the merge-base and so go stale every time it moves.

The checks CI cannot do

Logic mirrors re-checked. The steer allowlist has no capability flag behind it — if upstream changed an adapter to stop steering, nothing fails, and the fork would keep sending mid-turn into a provider that drops it. All four allowlisted drivers still steer (three still covered by upstream's own "steers a running turn..." tests, Grok still reuses ctx.activeTurnId), and grep -rn 'turn/steer' apps/server is still empty, so the deliberate Codex exclusion holds. The mobile send-label mirror also still matches — upstream's only mobile composer commit this range was theming.
Parallel path reviewed, nothing ported.apps/coil-home/ is a copy of apps/marketing/, so it drifts without ever conflicting. Both upstream marketing commits were checked and deliberately skipped: the Mac-chip detection (fix(marketing): detect Mac chip on homepage download button pingdotgg/t3code#4197) has nothing to choose between, since the release matrix builds only darwin-arm64 and win32-x64; and the Grok-mark mobile fix (fix(marketing): keep Grok mark clear of mobile hero copy pingdotgg/t3code#4542) repositions absolutely-positioned marks the fork already replaced with an in-flow row, so porting it would regress the fork's own fix.
Lockfile untouched. Upstream changed no dependency in this range — its only manifest edit adds two exports entries to packages/shared/package.json. pnpm-lock.yaml is byte-identical to pre-sync main, so the dependency sweep is intact and no regeneration was needed. (This is the step that destroyed the sweep on 08-08.)

Verification

vp run typecheck — exit 0, 17/17 packages.
vp run lint — exit 0, no error-level findings.
vp run test --testTimeout=120000 --hookTimeout=120000 — one fully green run (2708 passed | 7 skipped, exit 0).
Patch integrity: 166/166 fork patches replayed, subject-for-subject identical to pre-sync main. Nothing dropped, nothing absorbed.

Two caveats, both honest:

Later full-suite runs each hit one intermittently failing test — composerInlineTokens (a < 1000ms wall-clock assertion) and nativeReviewDiffHighlighter. Both are upstream-owned, untouched by the fork, and pass 3/3 in isolation; they fail only under full-suite parallel load on this machine. CI is the arbiter.
build-desktop-artifact's cross-arch Windows probe test fails locally for an environment reason, not a code one: the probe does {...process.env}, and this machine's shell has ELECTRON_RUN_AS_NODE=1 because the agent runs inside Electron-as-Node. Verified by control run — pristine upstream code fails it identically here — and it passes with the variable unset. It will pass in CI.

Not included

#118 (auto-resume never armed) is open and unrelated to the sync; it is untouched here.

radroid · 2026-08-17T18:12:26Z

CI is green — run 32052894658: Typecheck ✅ Lint ✅ Test ✅.

That settles both caveats in the description. Neither of the intermittently-failing tests (composerInlineTokens, nativeReviewDiffHighlighter) failed on CI, and neither did the build-desktop-artifact cross-arch probe test — confirming all three were local environment artifacts (parallel-load contention, and this machine's ambient ELECTRON_RUN_AS_NODE=1) rather than anything this branch changed.

coderabbitai · 2026-08-17T20:02:01Z

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: 1a471a64-016a-4afb-a024-f7b5c6541362

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

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

radroid · 2026-08-17T18:01:51Z

Closes#117. Absorbs the 113-commit upstream range that three consecutive daily syncs (08-15, 08-16, 08-17) could not, and replays all 166 fork patches on top of cd096b9ad.

Important

Do not use the GitHub merge button. This is the fork's patch series rebased onto new upstream, so main is not an ancestor of it and GitHub reports it CONFLICTING. That is expected and is not a defect in this branch. Land it by force-updating main:

git fetch origin
OLD_MAIN=$(git rev-parse origin/main)# save: needed to rebase in-flight branches
git push --force-with-lease origin origin/coil/sync-20260817:main

Rollback point is coil/last-good-20260817-083055 (verified == the pre-sync origin/main). If you'd rather have a genuinely button-mergeable PR, say so and I'll re-shape this as a merge commit instead — the tree would be identical.

Why the automation kept failing

Each daily run stops at the first conflict and escalates. Two different ones were in the way, so fixing either alone would not have unblocked it:

Day	Conflicted file	Upstream culprit
08-15	`apps/desktop/src/ipc/DesktopIpcHandlers.ts`	`57a299a7` open remote environments over SSH (pingdotgg#6572)
08-16, 08-17	`apps/server/src/sourceControl/SourceControlProviderDiscovery.ts`	`d715c2e56` raise the discovery probe budget (pingdotgg#6223)

Six conflicts surfaced in total across the replay. The three interesting ones:

1. Upstream independently fixed half of fork issue #4 — and did it better.d715c2e56 added a per-spec probeTimeoutMs with az at 20s; the fork carried a global 15s CLI_PROBE_TIMEOUT_MS. The fork's constant is deleted and the timeout ceded entirely to upstream. What upstream still does not do is the behavioural half — classifying only VcsProcessSpawnError as "missing", so a slow-but-present CLI stays available and the auth probe still runs — so that is all the fork keeps. Seam: +20/-8 → +12/-5.

2. The stale-hoist hazard on ComposerPrimaryActions.tsx is retired. Upstream 7afa184a9 (pingdotgg#4781) hoisted its inline send button to a const — the exact hoist the fork had been carrying since #35, and the one the ledger has flagged since 08-08 as able to silently revert an upstream restyle, because a copy cannot conflict. The two copies turned out byte-identical apart from one aria-label branch, so the fork's copy is deleted and upstream's kept. The fork's queue/steer branch was then folded into upstream's new bottom dispatch rather than left as the early return it used to be — as an early return it would have pre-empted upstream's whole mobile path and silently reverted pingdotgg#4781. Seam: +128/-55 → +74/-5; the -55 → -5 is the real signal, the fork now extends upstream's dispatch instead of displacing it.

3. manifest.webmanifest became an add/add seam. Upstream now ships its own. Upstream's has no name/short_name/description and no maskable icon — an installable PWA needs all four, which is what Web Push (#23) rides on — so the fork's fields are unioned onto upstream's, dropping the fork's duplicate apple-touch-icon.

Adapting to upstream, beyond the conflicts

Four commits sit on top of the replayed series:

no-native-title-tooltip — upstream added this oxlint rule in fee10def1 (feat(lint): ban native title tooltips and migrate to styled Tooltip pingdotgg/t3code#7209) and migrated its own call sites. The fork's UpdateToast.tsx was not in that sweep, so a new upstream rule turned fork lint red on a file upstream has never seen. Migrated to the same Tooltip/TooltipTrigger/TooltipPopup idiom.
DMG title assertion — ad117235b (feat(desktop): add signal macOS DMG installer background pingdotgg/t3code#6201) added a deepStrictEqual over the whole mac.dmg config, including a title built from resolveDesktopProductName, which Rename the fork's identity from t3x to coil (everything except the landing site) #71 renamed. Same one-word substitution the fork already makes ~300 lines above in that file.
Running-send aria-label — fix(web): keep send reachable while a turn is running on mobile pingdotgg/t3code#4781's new test expects aria-label="Send message"; the fork lengthens exactly that label to "Send message to the running turn" (Thread outbox: UI polish, queue images, send-and-interrupt — and decide whether the queue should survive #35), because on a steer-capable driver the submit folds into the work in progress and nothing else says so. The label stays the fork's; upstream's actual subject (a submit button beside Stop, at the larger size) is still pinned by the untouched assertions around it.
Seam ledger re-baseline — both halves against cd096b9ad, including the churn/risk columns, which are measured over the 60 days before the merge-base and so go stale every time it moves.

The checks CI cannot do

Logic mirrors re-checked. The steer allowlist has no capability flag behind it — if upstream changed an adapter to stop steering, nothing fails, and the fork would keep sending mid-turn into a provider that drops it. All four allowlisted drivers still steer (three still covered by upstream's own "steers a running turn..." tests, Grok still reuses ctx.activeTurnId), and grep -rn 'turn/steer' apps/server is still empty, so the deliberate Codex exclusion holds. The mobile send-label mirror also still matches — upstream's only mobile composer commit this range was theming.
Parallel path reviewed, nothing ported.apps/coil-home/ is a copy of apps/marketing/, so it drifts without ever conflicting. Both upstream marketing commits were checked and deliberately skipped: the Mac-chip detection (fix(marketing): detect Mac chip on homepage download button pingdotgg/t3code#4197) has nothing to choose between, since the release matrix builds only darwin-arm64 and win32-x64; and the Grok-mark mobile fix (fix(marketing): keep Grok mark clear of mobile hero copy pingdotgg/t3code#4542) repositions absolutely-positioned marks the fork already replaced with an in-flow row, so porting it would regress the fork's own fix.
Lockfile untouched. Upstream changed no dependency in this range — its only manifest edit adds two exports entries to packages/shared/package.json. pnpm-lock.yaml is byte-identical to pre-sync main, so the dependency sweep is intact and no regeneration was needed. (This is the step that destroyed the sweep on 08-08.)

Verification

vp run typecheck — exit 0, 17/17 packages.
vp run lint — exit 0, no error-level findings.
vp run test --testTimeout=120000 --hookTimeout=120000 — one fully green run (2708 passed | 7 skipped, exit 0).
Patch integrity: 166/166 fork patches replayed, subject-for-subject identical to pre-sync main. Nothing dropped, nothing absorbed.

Two caveats, both honest:

Later full-suite runs each hit one intermittently failing test — composerInlineTokens (a < 1000ms wall-clock assertion) and nativeReviewDiffHighlighter. Both are upstream-owned, untouched by the fork, and pass 3/3 in isolation; they fail only under full-suite parallel load on this machine. CI is the arbiter.
build-desktop-artifact's cross-arch Windows probe test fails locally for an environment reason, not a code one: the probe does {...process.env}, and this machine's shell has ELECTRON_RUN_AS_NODE=1 because the agent runs inside Electron-as-Node. Verified by control run — pristine upstream code fails it identically here — and it passes with the variable unset. It will pass in CI.

Not included

#118 (auto-resume never armed) is open and unrelated to the sync; it is untouched here.

radroid · 2026-08-17T18:12:26Z

CI is green — run 32052894658: Typecheck ✅ Lint ✅ Test ✅.

That settles both caveats in the description. Neither of the intermittently-failing tests (composerInlineTokens, nativeReviewDiffHighlighter) failed on CI, and neither did the build-desktop-artifact cross-arch probe test — confirming all three were local environment artifacts (parallel-load contention, and this machine's ambient ELECTRON_RUN_AS_NODE=1) rather than anything this branch changed.

coderabbitai · 2026-08-17T20:02:01Z

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: 1a471a64-016a-4afb-a024-f7b5c6541362

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

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

radroid · 2026-08-17T18:01:51Z

Closes#117. Absorbs the 113-commit upstream range that three consecutive daily syncs (08-15, 08-16, 08-17) could not, and replays all 166 fork patches on top of cd096b9ad.

Important

Do not use the GitHub merge button. This is the fork's patch series rebased onto new upstream, so main is not an ancestor of it and GitHub reports it CONFLICTING. That is expected and is not a defect in this branch. Land it by force-updating main:

git fetch origin
OLD_MAIN=$(git rev-parse origin/main)# save: needed to rebase in-flight branches
git push --force-with-lease origin origin/coil/sync-20260817:main

Rollback point is coil/last-good-20260817-083055 (verified == the pre-sync origin/main). If you'd rather have a genuinely button-mergeable PR, say so and I'll re-shape this as a merge commit instead — the tree would be identical.

Why the automation kept failing

Each daily run stops at the first conflict and escalates. Two different ones were in the way, so fixing either alone would not have unblocked it:

Day	Conflicted file	Upstream culprit
08-15	`apps/desktop/src/ipc/DesktopIpcHandlers.ts`	`57a299a7` open remote environments over SSH (pingdotgg#6572)
08-16, 08-17	`apps/server/src/sourceControl/SourceControlProviderDiscovery.ts`	`d715c2e56` raise the discovery probe budget (pingdotgg#6223)

Six conflicts surfaced in total across the replay. The three interesting ones:

1. Upstream independently fixed half of fork issue #4 — and did it better.d715c2e56 added a per-spec probeTimeoutMs with az at 20s; the fork carried a global 15s CLI_PROBE_TIMEOUT_MS. The fork's constant is deleted and the timeout ceded entirely to upstream. What upstream still does not do is the behavioural half — classifying only VcsProcessSpawnError as "missing", so a slow-but-present CLI stays available and the auth probe still runs — so that is all the fork keeps. Seam: +20/-8 → +12/-5.

2. The stale-hoist hazard on ComposerPrimaryActions.tsx is retired. Upstream 7afa184a9 (pingdotgg#4781) hoisted its inline send button to a const — the exact hoist the fork had been carrying since #35, and the one the ledger has flagged since 08-08 as able to silently revert an upstream restyle, because a copy cannot conflict. The two copies turned out byte-identical apart from one aria-label branch, so the fork's copy is deleted and upstream's kept. The fork's queue/steer branch was then folded into upstream's new bottom dispatch rather than left as the early return it used to be — as an early return it would have pre-empted upstream's whole mobile path and silently reverted pingdotgg#4781. Seam: +128/-55 → +74/-5; the -55 → -5 is the real signal, the fork now extends upstream's dispatch instead of displacing it.

3. manifest.webmanifest became an add/add seam. Upstream now ships its own. Upstream's has no name/short_name/description and no maskable icon — an installable PWA needs all four, which is what Web Push (#23) rides on — so the fork's fields are unioned onto upstream's, dropping the fork's duplicate apple-touch-icon.

Adapting to upstream, beyond the conflicts

Four commits sit on top of the replayed series:

no-native-title-tooltip — upstream added this oxlint rule in fee10def1 (feat(lint): ban native title tooltips and migrate to styled Tooltip pingdotgg/t3code#7209) and migrated its own call sites. The fork's UpdateToast.tsx was not in that sweep, so a new upstream rule turned fork lint red on a file upstream has never seen. Migrated to the same Tooltip/TooltipTrigger/TooltipPopup idiom.
DMG title assertion — ad117235b (feat(desktop): add signal macOS DMG installer background pingdotgg/t3code#6201) added a deepStrictEqual over the whole mac.dmg config, including a title built from resolveDesktopProductName, which Rename the fork's identity from t3x to coil (everything except the landing site) #71 renamed. Same one-word substitution the fork already makes ~300 lines above in that file.
Running-send aria-label — fix(web): keep send reachable while a turn is running on mobile pingdotgg/t3code#4781's new test expects aria-label="Send message"; the fork lengthens exactly that label to "Send message to the running turn" (Thread outbox: UI polish, queue images, send-and-interrupt — and decide whether the queue should survive #35), because on a steer-capable driver the submit folds into the work in progress and nothing else says so. The label stays the fork's; upstream's actual subject (a submit button beside Stop, at the larger size) is still pinned by the untouched assertions around it.
Seam ledger re-baseline — both halves against cd096b9ad, including the churn/risk columns, which are measured over the 60 days before the merge-base and so go stale every time it moves.

The checks CI cannot do

Logic mirrors re-checked. The steer allowlist has no capability flag behind it — if upstream changed an adapter to stop steering, nothing fails, and the fork would keep sending mid-turn into a provider that drops it. All four allowlisted drivers still steer (three still covered by upstream's own "steers a running turn..." tests, Grok still reuses ctx.activeTurnId), and grep -rn 'turn/steer' apps/server is still empty, so the deliberate Codex exclusion holds. The mobile send-label mirror also still matches — upstream's only mobile composer commit this range was theming.
Parallel path reviewed, nothing ported.apps/coil-home/ is a copy of apps/marketing/, so it drifts without ever conflicting. Both upstream marketing commits were checked and deliberately skipped: the Mac-chip detection (fix(marketing): detect Mac chip on homepage download button pingdotgg/t3code#4197) has nothing to choose between, since the release matrix builds only darwin-arm64 and win32-x64; and the Grok-mark mobile fix (fix(marketing): keep Grok mark clear of mobile hero copy pingdotgg/t3code#4542) repositions absolutely-positioned marks the fork already replaced with an in-flow row, so porting it would regress the fork's own fix.
Lockfile untouched. Upstream changed no dependency in this range — its only manifest edit adds two exports entries to packages/shared/package.json. pnpm-lock.yaml is byte-identical to pre-sync main, so the dependency sweep is intact and no regeneration was needed. (This is the step that destroyed the sweep on 08-08.)

Verification

vp run typecheck — exit 0, 17/17 packages.
vp run lint — exit 0, no error-level findings.
vp run test --testTimeout=120000 --hookTimeout=120000 — one fully green run (2708 passed | 7 skipped, exit 0).
Patch integrity: 166/166 fork patches replayed, subject-for-subject identical to pre-sync main. Nothing dropped, nothing absorbed.

Two caveats, both honest:

Later full-suite runs each hit one intermittently failing test — composerInlineTokens (a < 1000ms wall-clock assertion) and nativeReviewDiffHighlighter. Both are upstream-owned, untouched by the fork, and pass 3/3 in isolation; they fail only under full-suite parallel load on this machine. CI is the arbiter.
build-desktop-artifact's cross-arch Windows probe test fails locally for an environment reason, not a code one: the probe does {...process.env}, and this machine's shell has ELECTRON_RUN_AS_NODE=1 because the agent runs inside Electron-as-Node. Verified by control run — pristine upstream code fails it identically here — and it passes with the variable unset. It will pass in CI.

Not included

#118 (auto-resume never armed) is open and unrelated to the sync; it is untouched here.

radroid · 2026-08-17T18:12:26Z

CI is green — run 32052894658: Typecheck ✅ Lint ✅ Test ✅.

That settles both caveats in the description. Neither of the intermittently-failing tests (composerInlineTokens, nativeReviewDiffHighlighter) failed on CI, and neither did the build-desktop-artifact cross-arch probe test — confirming all three were local environment artifacts (parallel-load contention, and this machine's ambient ELECTRON_RUN_AS_NODE=1) rather than anything this branch changed.

coderabbitai · 2026-08-17T20:02:01Z

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: 1a471a64-016a-4afb-a024-f7b5c6541362

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

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

radroid mentioned this pull request Aug 17, 2026
[coil-sync] daily rebase needs attention (conflict) #117
Closed

radroid force-pushed the coil/sync-20260817 branch from 6aafe6f to f6355f0Compare August 17, 2026 20:01

radroid merged commit f6355f0 into mainAug 17, 2026
2 checks passed