chore(coil): sync 182 upstream commits (2026-09-02)

[radroid]

Resolves#128. The daily sync had been failing on a pnpm-lock.yaml conflict, leaving the fork 182 upstream commits behind. This replays the fork's 175 patches onto 941acb4f9 and fixes the four fork behaviours upstream broke along the way.

The rebase

175 of 178 patches replayed. The three that went empty are the historical chore(deps): regenerate the lockfile commits — every intermediate lock conflict resolved to upstream's side, and the lock is regenerated once at the end instead, seeded from the fork's own pre-sync lock so the advisory sweep survives (verified: sharp 0.35.3, nanoid 3.3.17, tar 7.5.22, undici 6.28.0, form-data 4.0.6, hono 4.13.0, ip-address 10.4.0, path-to-regexp 6.3.0, shell-quote 1.10.0, fast-uri 3.1.5, builder-util-runtime 9.7.0).

What a clean rebase would have hidden

Four fork behaviours broke with, at most, one build error between them. Each is its own commit:

• The outbox drain lost its idle gate.fix(server): settle threads server-side pingdotgg/t3code#8600 moved thread settling server-side and deleted canSettle. The predicate is re-homed in the fork's own outbox module on top of hasQueuedTurnStart, which survived. The repair was the risk, not the break: canSnooze is the obvious substitute and deliberately permits a running session, so it would have dispatched the queued head on top of a live turn — the exact double-send the gate exists to stop.
• A server timer was destroying armed auto-resumes. The same commit added a one-minute sweep that dispatches thread.auto-settle through thread.settle's decider case, emitting an identical thread.settled with no provenance. threadIsGone() read that as "the user is done here", so a resume armed on Claude's seven-day window was cancelled on day three, logged as reason: "thread-gone" on a thread nobody settled. The existing test asserted the old behaviour and would have kept passing.
• The Queue button went dead on file attachments.feat(web): attach PDFs, ZIPs, and other files to a turn pingdotgg/t3code#8236 added a second attachment class. The fork's queue path counted only images, so a files-only composer rendered Queue enabled and then did nothing at all when pressed — and text-plus-a-PDF queued the text and destroyed the file with no warning.
• Return submitted a frozen composer.feat(mobile): add offline iPhone voice input pingdotgg/t3code#8614 added a read-only mode for settling dictation; the fork's Return-submits interception sat above upstream's guard and jumped it.

Also: the fork's auth mirror on /api/coil/auto-resume had drifted a now-optional argument, costing relay clients the DPoP failure reason every other endpoint reports.

Near misses worth recording

Three resolutions would have been silently wrong, and are called out in the ledger:

• Expo SDK 57 rewrites the exact overrides: lines the fork's advisory block sits between, and rewrites patchedDependencies wholesale. Keep-both would have re-pinned Expo 56 packages onto a 57 tree and named five patch files that no longer exist. Verified afterwards that the block is upstream's plus exactly the 13 fork keys, and that all 17 patch files are present.
• AppSymbol.tsx conflicted whole-file because fix(mobile): reduce dev-client reload and Metro startup cost pingdotgg/t3code#8694 de-barrelled the Tabler imports to stop Metro registering the entire icon set. The fork's two lines are re-applied as a deep import; the file has zero bare @tabler/icons-react-native imports.
• Electron — upstream moved to 43.4.1, two majors past the fork's 41.10.3 advisory pin, so the pin is dropped rather than defended. Taking --ours would have run Electron-43-targeted preview code on an Electron 41 runtime, and nothing would have caught it.

Verification

vp run typecheck and vp run lint are clean. The full suite is green except build-desktop-artifact.test.ts > skips the primary native probe for cross-architecture Windows payloads, which fails identically on pristine upstream files — it reads the real host architecture, so targetArch: "arm64" is not cross-arch on Apple Silicon. It passes on x64 CI.

apps/marketing had zero upstream churn, so the apps/coil-home parallel path has nothing to port. No new upstream workflow arrived active.

Seam ledger re-baselined: 53 files, +2609/-981 — same seam set, and the deletion count fell by 63, all of it accounted for.

Landing

This is the fork's patch series rebased onto new upstream, so main is not an ancestor of it and the merge button will not work. It lands by force-updating main to the reviewed tip, per docs/coil/sync-agent-runbook.md. Recovery tag: coil/pre-sync-20260902.

---

Claude Opus 5 (1M context) in Claude Code.

[radroid]

Resolves#128. The daily sync had been failing on a pnpm-lock.yaml conflict, leaving the fork 182 upstream commits behind. This replays the fork's 175 patches onto 941acb4f9 and fixes the four fork behaviours upstream broke along the way.

The rebase

175 of 178 patches replayed. The three that went empty are the historical chore(deps): regenerate the lockfile commits — every intermediate lock conflict resolved to upstream's side, and the lock is regenerated once at the end instead, seeded from the fork's own pre-sync lock so the advisory sweep survives (verified: sharp 0.35.3, nanoid 3.3.17, tar 7.5.22, undici 6.28.0, form-data 4.0.6, hono 4.13.0, ip-address 10.4.0, path-to-regexp 6.3.0, shell-quote 1.10.0, fast-uri 3.1.5, builder-util-runtime 9.7.0).

What a clean rebase would have hidden

Four fork behaviours broke with, at most, one build error between them. Each is its own commit:

• The outbox drain lost its idle gate.fix(server): settle threads server-side pingdotgg/t3code#8600 moved thread settling server-side and deleted canSettle. The predicate is re-homed in the fork's own outbox module on top of hasQueuedTurnStart, which survived. The repair was the risk, not the break: canSnooze is the obvious substitute and deliberately permits a running session, so it would have dispatched the queued head on top of a live turn — the exact double-send the gate exists to stop.
• A server timer was destroying armed auto-resumes. The same commit added a one-minute sweep that dispatches thread.auto-settle through thread.settle's decider case, emitting an identical thread.settled with no provenance. threadIsGone() read that as "the user is done here", so a resume armed on Claude's seven-day window was cancelled on day three, logged as reason: "thread-gone" on a thread nobody settled. The existing test asserted the old behaviour and would have kept passing.
• The Queue button went dead on file attachments.feat(web): attach PDFs, ZIPs, and other files to a turn pingdotgg/t3code#8236 added a second attachment class. The fork's queue path counted only images, so a files-only composer rendered Queue enabled and then did nothing at all when pressed — and text-plus-a-PDF queued the text and destroyed the file with no warning.
• Return submitted a frozen composer.feat(mobile): add offline iPhone voice input pingdotgg/t3code#8614 added a read-only mode for settling dictation; the fork's Return-submits interception sat above upstream's guard and jumped it.

Also: the fork's auth mirror on /api/coil/auto-resume had drifted a now-optional argument, costing relay clients the DPoP failure reason every other endpoint reports.

Near misses worth recording

Three resolutions would have been silently wrong, and are called out in the ledger:

• Expo SDK 57 rewrites the exact overrides: lines the fork's advisory block sits between, and rewrites patchedDependencies wholesale. Keep-both would have re-pinned Expo 56 packages onto a 57 tree and named five patch files that no longer exist. Verified afterwards that the block is upstream's plus exactly the 13 fork keys, and that all 17 patch files are present.
• AppSymbol.tsx conflicted whole-file because fix(mobile): reduce dev-client reload and Metro startup cost pingdotgg/t3code#8694 de-barrelled the Tabler imports to stop Metro registering the entire icon set. The fork's two lines are re-applied as a deep import; the file has zero bare @tabler/icons-react-native imports.
• Electron — upstream moved to 43.4.1, two majors past the fork's 41.10.3 advisory pin, so the pin is dropped rather than defended. Taking --ours would have run Electron-43-targeted preview code on an Electron 41 runtime, and nothing would have caught it.

Verification

vp run typecheck and vp run lint are clean. The full suite is green except build-desktop-artifact.test.ts > skips the primary native probe for cross-architecture Windows payloads, which fails identically on pristine upstream files — it reads the real host architecture, so targetArch: "arm64" is not cross-arch on Apple Silicon. It passes on x64 CI.

apps/marketing had zero upstream churn, so the apps/coil-home parallel path has nothing to port. No new upstream workflow arrived active.

Seam ledger re-baselined: 53 files, +2609/-981 — same seam set, and the deletion count fell by 63, all of it accounted for.

Landing

This is the fork's patch series rebased onto new upstream, so main is not an ancestor of it and the merge button will not work. It lands by force-updating main to the reviewed tip, per docs/coil/sync-agent-runbook.md. Recovery tag: coil/pre-sync-20260902.

---

Claude Opus 5 (1M context) in Claude Code.

[radroid]

Resolves#128. The daily sync had been failing on a pnpm-lock.yaml conflict, leaving the fork 182 upstream commits behind. This replays the fork's 175 patches onto 941acb4f9 and fixes the four fork behaviours upstream broke along the way.

The rebase

175 of 178 patches replayed. The three that went empty are the historical chore(deps): regenerate the lockfile commits — every intermediate lock conflict resolved to upstream's side, and the lock is regenerated once at the end instead, seeded from the fork's own pre-sync lock so the advisory sweep survives (verified: sharp 0.35.3, nanoid 3.3.17, tar 7.5.22, undici 6.28.0, form-data 4.0.6, hono 4.13.0, ip-address 10.4.0, path-to-regexp 6.3.0, shell-quote 1.10.0, fast-uri 3.1.5, builder-util-runtime 9.7.0).

What a clean rebase would have hidden

Four fork behaviours broke with, at most, one build error between them. Each is its own commit:

• The outbox drain lost its idle gate.fix(server): settle threads server-side pingdotgg/t3code#8600 moved thread settling server-side and deleted canSettle. The predicate is re-homed in the fork's own outbox module on top of hasQueuedTurnStart, which survived. The repair was the risk, not the break: canSnooze is the obvious substitute and deliberately permits a running session, so it would have dispatched the queued head on top of a live turn — the exact double-send the gate exists to stop.
• A server timer was destroying armed auto-resumes. The same commit added a one-minute sweep that dispatches thread.auto-settle through thread.settle's decider case, emitting an identical thread.settled with no provenance. threadIsGone() read that as "the user is done here", so a resume armed on Claude's seven-day window was cancelled on day three, logged as reason: "thread-gone" on a thread nobody settled. The existing test asserted the old behaviour and would have kept passing.
• The Queue button went dead on file attachments.feat(web): attach PDFs, ZIPs, and other files to a turn pingdotgg/t3code#8236 added a second attachment class. The fork's queue path counted only images, so a files-only composer rendered Queue enabled and then did nothing at all when pressed — and text-plus-a-PDF queued the text and destroyed the file with no warning.
• Return submitted a frozen composer.feat(mobile): add offline iPhone voice input pingdotgg/t3code#8614 added a read-only mode for settling dictation; the fork's Return-submits interception sat above upstream's guard and jumped it.

Also: the fork's auth mirror on /api/coil/auto-resume had drifted a now-optional argument, costing relay clients the DPoP failure reason every other endpoint reports.

Near misses worth recording

Three resolutions would have been silently wrong, and are called out in the ledger:

• Expo SDK 57 rewrites the exact overrides: lines the fork's advisory block sits between, and rewrites patchedDependencies wholesale. Keep-both would have re-pinned Expo 56 packages onto a 57 tree and named five patch files that no longer exist. Verified afterwards that the block is upstream's plus exactly the 13 fork keys, and that all 17 patch files are present.
• AppSymbol.tsx conflicted whole-file because fix(mobile): reduce dev-client reload and Metro startup cost pingdotgg/t3code#8694 de-barrelled the Tabler imports to stop Metro registering the entire icon set. The fork's two lines are re-applied as a deep import; the file has zero bare @tabler/icons-react-native imports.
• Electron — upstream moved to 43.4.1, two majors past the fork's 41.10.3 advisory pin, so the pin is dropped rather than defended. Taking --ours would have run Electron-43-targeted preview code on an Electron 41 runtime, and nothing would have caught it.

Verification

vp run typecheck and vp run lint are clean. The full suite is green except build-desktop-artifact.test.ts > skips the primary native probe for cross-architecture Windows payloads, which fails identically on pristine upstream files — it reads the real host architecture, so targetArch: "arm64" is not cross-arch on Apple Silicon. It passes on x64 CI.

apps/marketing had zero upstream churn, so the apps/coil-home parallel path has nothing to port. No new upstream workflow arrived active.

Seam ledger re-baselined: 53 files, +2609/-981 — same seam set, and the deletion count fell by 63, all of it accounted for.

Landing

This is the fork's patch series rebased onto new upstream, so main is not an ancestor of it and the merge button will not work. It lands by force-updating main to the reviewed tip, per docs/coil/sync-agent-runbook.md. Recovery tag: coil/pre-sync-20260902.

---

Claude Opus 5 (1M context) in Claude Code.

[radroid]

Resolves#128. The daily sync had been failing on a pnpm-lock.yaml conflict, leaving the fork 182 upstream commits behind. This replays the fork's 175 patches onto 941acb4f9 and fixes the four fork behaviours upstream broke along the way.

The rebase

175 of 178 patches replayed. The three that went empty are the historical chore(deps): regenerate the lockfile commits — every intermediate lock conflict resolved to upstream's side, and the lock is regenerated once at the end instead, seeded from the fork's own pre-sync lock so the advisory sweep survives (verified: sharp 0.35.3, nanoid 3.3.17, tar 7.5.22, undici 6.28.0, form-data 4.0.6, hono 4.13.0, ip-address 10.4.0, path-to-regexp 6.3.0, shell-quote 1.10.0, fast-uri 3.1.5, builder-util-runtime 9.7.0).

What a clean rebase would have hidden

Four fork behaviours broke with, at most, one build error between them. Each is its own commit:

• The outbox drain lost its idle gate.fix(server): settle threads server-side pingdotgg/t3code#8600 moved thread settling server-side and deleted canSettle. The predicate is re-homed in the fork's own outbox module on top of hasQueuedTurnStart, which survived. The repair was the risk, not the break: canSnooze is the obvious substitute and deliberately permits a running session, so it would have dispatched the queued head on top of a live turn — the exact double-send the gate exists to stop.
• A server timer was destroying armed auto-resumes. The same commit added a one-minute sweep that dispatches thread.auto-settle through thread.settle's decider case, emitting an identical thread.settled with no provenance. threadIsGone() read that as "the user is done here", so a resume armed on Claude's seven-day window was cancelled on day three, logged as reason: "thread-gone" on a thread nobody settled. The existing test asserted the old behaviour and would have kept passing.
• The Queue button went dead on file attachments.feat(web): attach PDFs, ZIPs, and other files to a turn pingdotgg/t3code#8236 added a second attachment class. The fork's queue path counted only images, so a files-only composer rendered Queue enabled and then did nothing at all when pressed — and text-plus-a-PDF queued the text and destroyed the file with no warning.
• Return submitted a frozen composer.feat(mobile): add offline iPhone voice input pingdotgg/t3code#8614 added a read-only mode for settling dictation; the fork's Return-submits interception sat above upstream's guard and jumped it.

Also: the fork's auth mirror on /api/coil/auto-resume had drifted a now-optional argument, costing relay clients the DPoP failure reason every other endpoint reports.

Near misses worth recording

Three resolutions would have been silently wrong, and are called out in the ledger:

• Expo SDK 57 rewrites the exact overrides: lines the fork's advisory block sits between, and rewrites patchedDependencies wholesale. Keep-both would have re-pinned Expo 56 packages onto a 57 tree and named five patch files that no longer exist. Verified afterwards that the block is upstream's plus exactly the 13 fork keys, and that all 17 patch files are present.
• AppSymbol.tsx conflicted whole-file because fix(mobile): reduce dev-client reload and Metro startup cost pingdotgg/t3code#8694 de-barrelled the Tabler imports to stop Metro registering the entire icon set. The fork's two lines are re-applied as a deep import; the file has zero bare @tabler/icons-react-native imports.
• Electron — upstream moved to 43.4.1, two majors past the fork's 41.10.3 advisory pin, so the pin is dropped rather than defended. Taking --ours would have run Electron-43-targeted preview code on an Electron 41 runtime, and nothing would have caught it.

Verification

vp run typecheck and vp run lint are clean. The full suite is green except build-desktop-artifact.test.ts > skips the primary native probe for cross-architecture Windows payloads, which fails identically on pristine upstream files — it reads the real host architecture, so targetArch: "arm64" is not cross-arch on Apple Silicon. It passes on x64 CI.

apps/marketing had zero upstream churn, so the apps/coil-home parallel path has nothing to port. No new upstream workflow arrived active.

Seam ledger re-baselined: 53 files, +2609/-981 — same seam set, and the deletion count fell by 63, all of it accounted for.

Landing

This is the fork's patch series rebased onto new upstream, so main is not an ancestor of it and the merge button will not work. It lands by force-updating main to the reviewed tip, per docs/coil/sync-agent-runbook.md. Recovery tag: coil/pre-sync-20260902.

---

Claude Opus 5 (1M context) in Claude Code.

[radroid]

Resolves#128. The daily sync had been failing on a pnpm-lock.yaml conflict, leaving the fork 182 upstream commits behind. This replays the fork's 175 patches onto 941acb4f9 and fixes the four fork behaviours upstream broke along the way.

The rebase

175 of 178 patches replayed. The three that went empty are the historical chore(deps): regenerate the lockfile commits — every intermediate lock conflict resolved to upstream's side, and the lock is regenerated once at the end instead, seeded from the fork's own pre-sync lock so the advisory sweep survives (verified: sharp 0.35.3, nanoid 3.3.17, tar 7.5.22, undici 6.28.0, form-data 4.0.6, hono 4.13.0, ip-address 10.4.0, path-to-regexp 6.3.0, shell-quote 1.10.0, fast-uri 3.1.5, builder-util-runtime 9.7.0).

What a clean rebase would have hidden

Four fork behaviours broke with, at most, one build error between them. Each is its own commit:

• The outbox drain lost its idle gate.fix(server): settle threads server-side pingdotgg/t3code#8600 moved thread settling server-side and deleted canSettle. The predicate is re-homed in the fork's own outbox module on top of hasQueuedTurnStart, which survived. The repair was the risk, not the break: canSnooze is the obvious substitute and deliberately permits a running session, so it would have dispatched the queued head on top of a live turn — the exact double-send the gate exists to stop.
• A server timer was destroying armed auto-resumes. The same commit added a one-minute sweep that dispatches thread.auto-settle through thread.settle's decider case, emitting an identical thread.settled with no provenance. threadIsGone() read that as "the user is done here", so a resume armed on Claude's seven-day window was cancelled on day three, logged as reason: "thread-gone" on a thread nobody settled. The existing test asserted the old behaviour and would have kept passing.
• The Queue button went dead on file attachments.feat(web): attach PDFs, ZIPs, and other files to a turn pingdotgg/t3code#8236 added a second attachment class. The fork's queue path counted only images, so a files-only composer rendered Queue enabled and then did nothing at all when pressed — and text-plus-a-PDF queued the text and destroyed the file with no warning.
• Return submitted a frozen composer.feat(mobile): add offline iPhone voice input pingdotgg/t3code#8614 added a read-only mode for settling dictation; the fork's Return-submits interception sat above upstream's guard and jumped it.

Also: the fork's auth mirror on /api/coil/auto-resume had drifted a now-optional argument, costing relay clients the DPoP failure reason every other endpoint reports.

Near misses worth recording

Three resolutions would have been silently wrong, and are called out in the ledger:

• Expo SDK 57 rewrites the exact overrides: lines the fork's advisory block sits between, and rewrites patchedDependencies wholesale. Keep-both would have re-pinned Expo 56 packages onto a 57 tree and named five patch files that no longer exist. Verified afterwards that the block is upstream's plus exactly the 13 fork keys, and that all 17 patch files are present.
• AppSymbol.tsx conflicted whole-file because fix(mobile): reduce dev-client reload and Metro startup cost pingdotgg/t3code#8694 de-barrelled the Tabler imports to stop Metro registering the entire icon set. The fork's two lines are re-applied as a deep import; the file has zero bare @tabler/icons-react-native imports.
• Electron — upstream moved to 43.4.1, two majors past the fork's 41.10.3 advisory pin, so the pin is dropped rather than defended. Taking --ours would have run Electron-43-targeted preview code on an Electron 41 runtime, and nothing would have caught it.

Verification

vp run typecheck and vp run lint are clean. The full suite is green except build-desktop-artifact.test.ts > skips the primary native probe for cross-architecture Windows payloads, which fails identically on pristine upstream files — it reads the real host architecture, so targetArch: "arm64" is not cross-arch on Apple Silicon. It passes on x64 CI.

apps/marketing had zero upstream churn, so the apps/coil-home parallel path has nothing to port. No new upstream workflow arrived active.

Seam ledger re-baselined: 53 files, +2609/-981 — same seam set, and the deletion count fell by 63, all of it accounted for.

Landing

This is the fork's patch series rebased onto new upstream, so main is not an ancestor of it and the merge button will not work. It lands by force-updating main to the reviewed tip, per docs/coil/sync-agent-runbook.md. Recovery tag: coil/pre-sync-20260902.

---

Claude Opus 5 (1M context) in Claude Code.

[radroid]

Resolves#128. The daily sync had been failing on a pnpm-lock.yaml conflict, leaving the fork 182 upstream commits behind. This replays the fork's 175 patches onto 941acb4f9 and fixes the four fork behaviours upstream broke along the way.

The rebase

175 of 178 patches replayed. The three that went empty are the historical chore(deps): regenerate the lockfile commits — every intermediate lock conflict resolved to upstream's side, and the lock is regenerated once at the end instead, seeded from the fork's own pre-sync lock so the advisory sweep survives (verified: sharp 0.35.3, nanoid 3.3.17, tar 7.5.22, undici 6.28.0, form-data 4.0.6, hono 4.13.0, ip-address 10.4.0, path-to-regexp 6.3.0, shell-quote 1.10.0, fast-uri 3.1.5, builder-util-runtime 9.7.0).

What a clean rebase would have hidden

Four fork behaviours broke with, at most, one build error between them. Each is its own commit:

• The outbox drain lost its idle gate.fix(server): settle threads server-side pingdotgg/t3code#8600 moved thread settling server-side and deleted canSettle. The predicate is re-homed in the fork's own outbox module on top of hasQueuedTurnStart, which survived. The repair was the risk, not the break: canSnooze is the obvious substitute and deliberately permits a running session, so it would have dispatched the queued head on top of a live turn — the exact double-send the gate exists to stop.
• A server timer was destroying armed auto-resumes. The same commit added a one-minute sweep that dispatches thread.auto-settle through thread.settle's decider case, emitting an identical thread.settled with no provenance. threadIsGone() read that as "the user is done here", so a resume armed on Claude's seven-day window was cancelled on day three, logged as reason: "thread-gone" on a thread nobody settled. The existing test asserted the old behaviour and would have kept passing.
• The Queue button went dead on file attachments.feat(web): attach PDFs, ZIPs, and other files to a turn pingdotgg/t3code#8236 added a second attachment class. The fork's queue path counted only images, so a files-only composer rendered Queue enabled and then did nothing at all when pressed — and text-plus-a-PDF queued the text and destroyed the file with no warning.
• Return submitted a frozen composer.feat(mobile): add offline iPhone voice input pingdotgg/t3code#8614 added a read-only mode for settling dictation; the fork's Return-submits interception sat above upstream's guard and jumped it.

Also: the fork's auth mirror on /api/coil/auto-resume had drifted a now-optional argument, costing relay clients the DPoP failure reason every other endpoint reports.

Near misses worth recording

Three resolutions would have been silently wrong, and are called out in the ledger:

• Expo SDK 57 rewrites the exact overrides: lines the fork's advisory block sits between, and rewrites patchedDependencies wholesale. Keep-both would have re-pinned Expo 56 packages onto a 57 tree and named five patch files that no longer exist. Verified afterwards that the block is upstream's plus exactly the 13 fork keys, and that all 17 patch files are present.
• AppSymbol.tsx conflicted whole-file because fix(mobile): reduce dev-client reload and Metro startup cost pingdotgg/t3code#8694 de-barrelled the Tabler imports to stop Metro registering the entire icon set. The fork's two lines are re-applied as a deep import; the file has zero bare @tabler/icons-react-native imports.
• Electron — upstream moved to 43.4.1, two majors past the fork's 41.10.3 advisory pin, so the pin is dropped rather than defended. Taking --ours would have run Electron-43-targeted preview code on an Electron 41 runtime, and nothing would have caught it.

Verification

vp run typecheck and vp run lint are clean. The full suite is green except build-desktop-artifact.test.ts > skips the primary native probe for cross-architecture Windows payloads, which fails identically on pristine upstream files — it reads the real host architecture, so targetArch: "arm64" is not cross-arch on Apple Silicon. It passes on x64 CI.

apps/marketing had zero upstream churn, so the apps/coil-home parallel path has nothing to port. No new upstream workflow arrived active.

Seam ledger re-baselined: 53 files, +2609/-981 — same seam set, and the deletion count fell by 63, all of it accounted for.

Landing

This is the fork's patch series rebased onto new upstream, so main is not an ancestor of it and the merge button will not work. It lands by force-updating main to the reviewed tip, per docs/coil/sync-agent-runbook.md. Recovery tag: coil/pre-sync-20260902.

---

Claude Opus 5 (1M context) in Claude Code.

[radroid]

Resolves#128. The daily sync had been failing on a pnpm-lock.yaml conflict, leaving the fork 182 upstream commits behind. This replays the fork's 175 patches onto 941acb4f9 and fixes the four fork behaviours upstream broke along the way.

The rebase

175 of 178 patches replayed. The three that went empty are the historical chore(deps): regenerate the lockfile commits — every intermediate lock conflict resolved to upstream's side, and the lock is regenerated once at the end instead, seeded from the fork's own pre-sync lock so the advisory sweep survives (verified: sharp 0.35.3, nanoid 3.3.17, tar 7.5.22, undici 6.28.0, form-data 4.0.6, hono 4.13.0, ip-address 10.4.0, path-to-regexp 6.3.0, shell-quote 1.10.0, fast-uri 3.1.5, builder-util-runtime 9.7.0).

What a clean rebase would have hidden

Four fork behaviours broke with, at most, one build error between them. Each is its own commit:

• The outbox drain lost its idle gate.fix(server): settle threads server-side pingdotgg/t3code#8600 moved thread settling server-side and deleted canSettle. The predicate is re-homed in the fork's own outbox module on top of hasQueuedTurnStart, which survived. The repair was the risk, not the break: canSnooze is the obvious substitute and deliberately permits a running session, so it would have dispatched the queued head on top of a live turn — the exact double-send the gate exists to stop.
• A server timer was destroying armed auto-resumes. The same commit added a one-minute sweep that dispatches thread.auto-settle through thread.settle's decider case, emitting an identical thread.settled with no provenance. threadIsGone() read that as "the user is done here", so a resume armed on Claude's seven-day window was cancelled on day three, logged as reason: "thread-gone" on a thread nobody settled. The existing test asserted the old behaviour and would have kept passing.
• The Queue button went dead on file attachments.feat(web): attach PDFs, ZIPs, and other files to a turn pingdotgg/t3code#8236 added a second attachment class. The fork's queue path counted only images, so a files-only composer rendered Queue enabled and then did nothing at all when pressed — and text-plus-a-PDF queued the text and destroyed the file with no warning.
• Return submitted a frozen composer.feat(mobile): add offline iPhone voice input pingdotgg/t3code#8614 added a read-only mode for settling dictation; the fork's Return-submits interception sat above upstream's guard and jumped it.

Also: the fork's auth mirror on /api/coil/auto-resume had drifted a now-optional argument, costing relay clients the DPoP failure reason every other endpoint reports.

Near misses worth recording

Three resolutions would have been silently wrong, and are called out in the ledger:

• Expo SDK 57 rewrites the exact overrides: lines the fork's advisory block sits between, and rewrites patchedDependencies wholesale. Keep-both would have re-pinned Expo 56 packages onto a 57 tree and named five patch files that no longer exist. Verified afterwards that the block is upstream's plus exactly the 13 fork keys, and that all 17 patch files are present.
• AppSymbol.tsx conflicted whole-file because fix(mobile): reduce dev-client reload and Metro startup cost pingdotgg/t3code#8694 de-barrelled the Tabler imports to stop Metro registering the entire icon set. The fork's two lines are re-applied as a deep import; the file has zero bare @tabler/icons-react-native imports.
• Electron — upstream moved to 43.4.1, two majors past the fork's 41.10.3 advisory pin, so the pin is dropped rather than defended. Taking --ours would have run Electron-43-targeted preview code on an Electron 41 runtime, and nothing would have caught it.

Verification

vp run typecheck and vp run lint are clean. The full suite is green except build-desktop-artifact.test.ts > skips the primary native probe for cross-architecture Windows payloads, which fails identically on pristine upstream files — it reads the real host architecture, so targetArch: "arm64" is not cross-arch on Apple Silicon. It passes on x64 CI.

apps/marketing had zero upstream churn, so the apps/coil-home parallel path has nothing to port. No new upstream workflow arrived active.

Seam ledger re-baselined: 53 files, +2609/-981 — same seam set, and the deletion count fell by 63, all of it accounted for.

Landing

This is the fork's patch series rebased onto new upstream, so main is not an ancestor of it and the merge button will not work. It lands by force-updating main to the reviewed tip, per docs/coil/sync-agent-runbook.md. Recovery tag: coil/pre-sync-20260902.

---

Claude Opus 5 (1M context) in Claude Code.

[radroid]

Resolves#128. The daily sync had been failing on a pnpm-lock.yaml conflict, leaving the fork 182 upstream commits behind. This replays the fork's 175 patches onto 941acb4f9 and fixes the four fork behaviours upstream broke along the way.

The rebase

175 of 178 patches replayed. The three that went empty are the historical chore(deps): regenerate the lockfile commits — every intermediate lock conflict resolved to upstream's side, and the lock is regenerated once at the end instead, seeded from the fork's own pre-sync lock so the advisory sweep survives (verified: sharp 0.35.3, nanoid 3.3.17, tar 7.5.22, undici 6.28.0, form-data 4.0.6, hono 4.13.0, ip-address 10.4.0, path-to-regexp 6.3.0, shell-quote 1.10.0, fast-uri 3.1.5, builder-util-runtime 9.7.0).

What a clean rebase would have hidden

Four fork behaviours broke with, at most, one build error between them. Each is its own commit:

• The outbox drain lost its idle gate.fix(server): settle threads server-side pingdotgg/t3code#8600 moved thread settling server-side and deleted canSettle. The predicate is re-homed in the fork's own outbox module on top of hasQueuedTurnStart, which survived. The repair was the risk, not the break: canSnooze is the obvious substitute and deliberately permits a running session, so it would have dispatched the queued head on top of a live turn — the exact double-send the gate exists to stop.
• A server timer was destroying armed auto-resumes. The same commit added a one-minute sweep that dispatches thread.auto-settle through thread.settle's decider case, emitting an identical thread.settled with no provenance. threadIsGone() read that as "the user is done here", so a resume armed on Claude's seven-day window was cancelled on day three, logged as reason: "thread-gone" on a thread nobody settled. The existing test asserted the old behaviour and would have kept passing.
• The Queue button went dead on file attachments.feat(web): attach PDFs, ZIPs, and other files to a turn pingdotgg/t3code#8236 added a second attachment class. The fork's queue path counted only images, so a files-only composer rendered Queue enabled and then did nothing at all when pressed — and text-plus-a-PDF queued the text and destroyed the file with no warning.
• Return submitted a frozen composer.feat(mobile): add offline iPhone voice input pingdotgg/t3code#8614 added a read-only mode for settling dictation; the fork's Return-submits interception sat above upstream's guard and jumped it.

Also: the fork's auth mirror on /api/coil/auto-resume had drifted a now-optional argument, costing relay clients the DPoP failure reason every other endpoint reports.

Near misses worth recording

Three resolutions would have been silently wrong, and are called out in the ledger:

• Expo SDK 57 rewrites the exact overrides: lines the fork's advisory block sits between, and rewrites patchedDependencies wholesale. Keep-both would have re-pinned Expo 56 packages onto a 57 tree and named five patch files that no longer exist. Verified afterwards that the block is upstream's plus exactly the 13 fork keys, and that all 17 patch files are present.
• AppSymbol.tsx conflicted whole-file because fix(mobile): reduce dev-client reload and Metro startup cost pingdotgg/t3code#8694 de-barrelled the Tabler imports to stop Metro registering the entire icon set. The fork's two lines are re-applied as a deep import; the file has zero bare @tabler/icons-react-native imports.
• Electron — upstream moved to 43.4.1, two majors past the fork's 41.10.3 advisory pin, so the pin is dropped rather than defended. Taking --ours would have run Electron-43-targeted preview code on an Electron 41 runtime, and nothing would have caught it.

Verification

vp run typecheck and vp run lint are clean. The full suite is green except build-desktop-artifact.test.ts > skips the primary native probe for cross-architecture Windows payloads, which fails identically on pristine upstream files — it reads the real host architecture, so targetArch: "arm64" is not cross-arch on Apple Silicon. It passes on x64 CI.

apps/marketing had zero upstream churn, so the apps/coil-home parallel path has nothing to port. No new upstream workflow arrived active.

Seam ledger re-baselined: 53 files, +2609/-981 — same seam set, and the deletion count fell by 63, all of it accounted for.

Landing

This is the fork's patch series rebased onto new upstream, so main is not an ancestor of it and the merge button will not work. It lands by force-updating main to the reviewed tip, per docs/coil/sync-agent-runbook.md. Recovery tag: coil/pre-sync-20260902.

---

Claude Opus 5 (1M context) in Claude Code.