feat(web): attach PDFs, ZIPs, and other files to a turn

[rynfar]

Problem

Pylon's server has accepted generic file uploads since #8235 landed as PR #140 — ChatFileAttachment in the contracts, fileAttachments.maxUploadBytes advertised per environment, adapter guards on every provider, and appendAttachmentPathLines in ProviderService handing each uploaded file's path to the agent. The web composer still only staged images, so the feature was half-shipped.

Fix

Adopts upstream #8236 (bcb855a63) and its typecheck fixup (d22709f75) from the f94a0d646..1f8ed54ad T3 review. The paperclip and paste accept any file up to the environment's advertised limit; uploads run while you type; files survive draft and stash moves; an interrupted upload rehydrates as an "Attach again" row rather than a stuck spinner; the timeline renders download rows, and unknown types render as an inert name with no link.

No server or contract change. packages/client-runtime gains only a ./state/attachments subpath export.

Conflict resolution

ChatComposer.tsx is ~4400 lines on Pylon with several fork-only controls, so git misaligned three regions that were re-inserted by hand: the draftTarget arguments to the upload queue, the file-attachment row, and the paperclip control. Upstream's fileCapabilityBlockReason / needs-reattach conditions were merged into Pylon's baseSendDisabledReason so both sets of conditions survive, with a file-capability block outranking Pylon's compaction message. ChatView.tsx had two adjacency conflicts beside the session-input-queue commands. composerDraftStore.ts auto-merged; COMPOSER_DRAFT_STORAGE_VERSION stays at 9 and Pylon's draft normalisation still runs. docs/user/composer.md gets the file-attachment paragraph in Pylon voice and states that mobile sends images only for now.

Two Pylon-specific notes: the paperclip is not shown in Pylon's collapsed-mobile composer row, consistent with that row having no image-attach affordance either (tapping it expands the full composer); and every provider, including Prime, receives generic files as prompt path lines — the adapter guards only prevent non-images being sent as base64 image blocks.

Verification

• vp test run across nine files (attachmentUploadQueue, attachmentUploadState, composerDraftStore, composerAttachmentFiles, promptStashStore, MessagesTimeline, ComposerStashMenu, ChatView.logic, client-runtime/state/attachments) — 267 passed.
• @t3tools/web and @t3tools/client-runtime typecheck clean; lint, format, and git diff --check clean on all 30 files.
• Independent review compared added/removed line multisets per file against upstream: all 30 match, ChatComposer.tsx has the same 42 hunks on both sides, and the hand-placed blocks sit at upstream's exact anchors (file row before ComposerPromptEditor; paperclip first in data-chat-composer-actions="right"). Reviewer also traced the reattach state machine (the upload store is not persisted, so a reload can never leave an entry stuck) and confirmed unknown file types carry no href.
• Not done: a real-client pass — attach a PDF, reload mid-upload and use "Attach again", stash and restore a file-bearing draft, and check footer crowding at narrow expanded widths now that the paperclip sits beside the inline badges.

---

Claude Fable 5 in Pylon.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.3 KiB	13.2 KiB	−111 B (−0.8%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	−2 B (−0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.3 KiB	−109 B (−1.7%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	55.5 KiB	54.7 KiB	−822 B (−1.4%)	66.4 KiB	✅
Codex	Live turn messages	9	9	0 (0.0%)	21	✅
Claude	Total thread wire	13.3 KiB	13.2 KiB	−150 B (−1.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	−14 B (−0.2%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.4 KiB	6.3 KiB	−136 B (−2.1%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	56.4 KiB	55.5 KiB	−884 B (−1.5%)	66.4 KiB	✅
Claude	Live turn messages	10	9	−1 (−10.0%)	21	✅

Baseline: abe70a7 · PR result: a7d79e8 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 109.4 KiB
• Claude decoded thread snapshot: 110.1 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[rynfar]

Problem

Pylon's server has accepted generic file uploads since #8235 landed as PR #140 — ChatFileAttachment in the contracts, fileAttachments.maxUploadBytes advertised per environment, adapter guards on every provider, and appendAttachmentPathLines in ProviderService handing each uploaded file's path to the agent. The web composer still only staged images, so the feature was half-shipped.

Fix

Adopts upstream #8236 (bcb855a63) and its typecheck fixup (d22709f75) from the f94a0d646..1f8ed54ad T3 review. The paperclip and paste accept any file up to the environment's advertised limit; uploads run while you type; files survive draft and stash moves; an interrupted upload rehydrates as an "Attach again" row rather than a stuck spinner; the timeline renders download rows, and unknown types render as an inert name with no link.

No server or contract change. packages/client-runtime gains only a ./state/attachments subpath export.

Conflict resolution

ChatComposer.tsx is ~4400 lines on Pylon with several fork-only controls, so git misaligned three regions that were re-inserted by hand: the draftTarget arguments to the upload queue, the file-attachment row, and the paperclip control. Upstream's fileCapabilityBlockReason / needs-reattach conditions were merged into Pylon's baseSendDisabledReason so both sets of conditions survive, with a file-capability block outranking Pylon's compaction message. ChatView.tsx had two adjacency conflicts beside the session-input-queue commands. composerDraftStore.ts auto-merged; COMPOSER_DRAFT_STORAGE_VERSION stays at 9 and Pylon's draft normalisation still runs. docs/user/composer.md gets the file-attachment paragraph in Pylon voice and states that mobile sends images only for now.

Two Pylon-specific notes: the paperclip is not shown in Pylon's collapsed-mobile composer row, consistent with that row having no image-attach affordance either (tapping it expands the full composer); and every provider, including Prime, receives generic files as prompt path lines — the adapter guards only prevent non-images being sent as base64 image blocks.

Verification

• vp test run across nine files (attachmentUploadQueue, attachmentUploadState, composerDraftStore, composerAttachmentFiles, promptStashStore, MessagesTimeline, ComposerStashMenu, ChatView.logic, client-runtime/state/attachments) — 267 passed.
• @t3tools/web and @t3tools/client-runtime typecheck clean; lint, format, and git diff --check clean on all 30 files.
• Independent review compared added/removed line multisets per file against upstream: all 30 match, ChatComposer.tsx has the same 42 hunks on both sides, and the hand-placed blocks sit at upstream's exact anchors (file row before ComposerPromptEditor; paperclip first in data-chat-composer-actions="right"). Reviewer also traced the reattach state machine (the upload store is not persisted, so a reload can never leave an entry stuck) and confirmed unknown file types carry no href.
• Not done: a real-client pass — attach a PDF, reload mid-upload and use "Attach again", stash and restore a file-bearing draft, and check footer crowding at narrow expanded widths now that the paperclip sits beside the inline badges.

---

Claude Fable 5 in Pylon.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.3 KiB	13.2 KiB	−111 B (−0.8%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	−2 B (−0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.3 KiB	−109 B (−1.7%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	55.5 KiB	54.7 KiB	−822 B (−1.4%)	66.4 KiB	✅
Codex	Live turn messages	9	9	0 (0.0%)	21	✅
Claude	Total thread wire	13.3 KiB	13.2 KiB	−150 B (−1.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	−14 B (−0.2%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.4 KiB	6.3 KiB	−136 B (−2.1%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	56.4 KiB	55.5 KiB	−884 B (−1.5%)	66.4 KiB	✅
Claude	Live turn messages	10	9	−1 (−10.0%)	21	✅

Baseline: abe70a7 · PR result: a7d79e8 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 109.4 KiB
• Claude decoded thread snapshot: 110.1 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[rynfar]

Problem

Pylon's server has accepted generic file uploads since #8235 landed as PR #140 — ChatFileAttachment in the contracts, fileAttachments.maxUploadBytes advertised per environment, adapter guards on every provider, and appendAttachmentPathLines in ProviderService handing each uploaded file's path to the agent. The web composer still only staged images, so the feature was half-shipped.

Fix

Adopts upstream #8236 (bcb855a63) and its typecheck fixup (d22709f75) from the f94a0d646..1f8ed54ad T3 review. The paperclip and paste accept any file up to the environment's advertised limit; uploads run while you type; files survive draft and stash moves; an interrupted upload rehydrates as an "Attach again" row rather than a stuck spinner; the timeline renders download rows, and unknown types render as an inert name with no link.

No server or contract change. packages/client-runtime gains only a ./state/attachments subpath export.

Conflict resolution

ChatComposer.tsx is ~4400 lines on Pylon with several fork-only controls, so git misaligned three regions that were re-inserted by hand: the draftTarget arguments to the upload queue, the file-attachment row, and the paperclip control. Upstream's fileCapabilityBlockReason / needs-reattach conditions were merged into Pylon's baseSendDisabledReason so both sets of conditions survive, with a file-capability block outranking Pylon's compaction message. ChatView.tsx had two adjacency conflicts beside the session-input-queue commands. composerDraftStore.ts auto-merged; COMPOSER_DRAFT_STORAGE_VERSION stays at 9 and Pylon's draft normalisation still runs. docs/user/composer.md gets the file-attachment paragraph in Pylon voice and states that mobile sends images only for now.

Two Pylon-specific notes: the paperclip is not shown in Pylon's collapsed-mobile composer row, consistent with that row having no image-attach affordance either (tapping it expands the full composer); and every provider, including Prime, receives generic files as prompt path lines — the adapter guards only prevent non-images being sent as base64 image blocks.

Verification

• vp test run across nine files (attachmentUploadQueue, attachmentUploadState, composerDraftStore, composerAttachmentFiles, promptStashStore, MessagesTimeline, ComposerStashMenu, ChatView.logic, client-runtime/state/attachments) — 267 passed.
• @t3tools/web and @t3tools/client-runtime typecheck clean; lint, format, and git diff --check clean on all 30 files.
• Independent review compared added/removed line multisets per file against upstream: all 30 match, ChatComposer.tsx has the same 42 hunks on both sides, and the hand-placed blocks sit at upstream's exact anchors (file row before ComposerPromptEditor; paperclip first in data-chat-composer-actions="right"). Reviewer also traced the reattach state machine (the upload store is not persisted, so a reload can never leave an entry stuck) and confirmed unknown file types carry no href.
• Not done: a real-client pass — attach a PDF, reload mid-upload and use "Attach again", stash and restore a file-bearing draft, and check footer crowding at narrow expanded widths now that the paperclip sits beside the inline badges.

---

Claude Fable 5 in Pylon.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.3 KiB	13.2 KiB	−111 B (−0.8%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	−2 B (−0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.3 KiB	−109 B (−1.7%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	55.5 KiB	54.7 KiB	−822 B (−1.4%)	66.4 KiB	✅
Codex	Live turn messages	9	9	0 (0.0%)	21	✅
Claude	Total thread wire	13.3 KiB	13.2 KiB	−150 B (−1.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	−14 B (−0.2%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.4 KiB	6.3 KiB	−136 B (−2.1%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	56.4 KiB	55.5 KiB	−884 B (−1.5%)	66.4 KiB	✅
Claude	Live turn messages	10	9	−1 (−10.0%)	21	✅

Baseline: abe70a7 · PR result: a7d79e8 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 109.4 KiB
• Claude decoded thread snapshot: 110.1 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[rynfar]

Problem

Pylon's server has accepted generic file uploads since #8235 landed as PR #140 — ChatFileAttachment in the contracts, fileAttachments.maxUploadBytes advertised per environment, adapter guards on every provider, and appendAttachmentPathLines in ProviderService handing each uploaded file's path to the agent. The web composer still only staged images, so the feature was half-shipped.

Fix

Adopts upstream #8236 (bcb855a63) and its typecheck fixup (d22709f75) from the f94a0d646..1f8ed54ad T3 review. The paperclip and paste accept any file up to the environment's advertised limit; uploads run while you type; files survive draft and stash moves; an interrupted upload rehydrates as an "Attach again" row rather than a stuck spinner; the timeline renders download rows, and unknown types render as an inert name with no link.

No server or contract change. packages/client-runtime gains only a ./state/attachments subpath export.

Conflict resolution

ChatComposer.tsx is ~4400 lines on Pylon with several fork-only controls, so git misaligned three regions that were re-inserted by hand: the draftTarget arguments to the upload queue, the file-attachment row, and the paperclip control. Upstream's fileCapabilityBlockReason / needs-reattach conditions were merged into Pylon's baseSendDisabledReason so both sets of conditions survive, with a file-capability block outranking Pylon's compaction message. ChatView.tsx had two adjacency conflicts beside the session-input-queue commands. composerDraftStore.ts auto-merged; COMPOSER_DRAFT_STORAGE_VERSION stays at 9 and Pylon's draft normalisation still runs. docs/user/composer.md gets the file-attachment paragraph in Pylon voice and states that mobile sends images only for now.

Two Pylon-specific notes: the paperclip is not shown in Pylon's collapsed-mobile composer row, consistent with that row having no image-attach affordance either (tapping it expands the full composer); and every provider, including Prime, receives generic files as prompt path lines — the adapter guards only prevent non-images being sent as base64 image blocks.

Verification

• vp test run across nine files (attachmentUploadQueue, attachmentUploadState, composerDraftStore, composerAttachmentFiles, promptStashStore, MessagesTimeline, ComposerStashMenu, ChatView.logic, client-runtime/state/attachments) — 267 passed.
• @t3tools/web and @t3tools/client-runtime typecheck clean; lint, format, and git diff --check clean on all 30 files.
• Independent review compared added/removed line multisets per file against upstream: all 30 match, ChatComposer.tsx has the same 42 hunks on both sides, and the hand-placed blocks sit at upstream's exact anchors (file row before ComposerPromptEditor; paperclip first in data-chat-composer-actions="right"). Reviewer also traced the reattach state machine (the upload store is not persisted, so a reload can never leave an entry stuck) and confirmed unknown file types carry no href.
• Not done: a real-client pass — attach a PDF, reload mid-upload and use "Attach again", stash and restore a file-bearing draft, and check footer crowding at narrow expanded widths now that the paperclip sits beside the inline badges.

---

Claude Fable 5 in Pylon.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.3 KiB	13.2 KiB	−111 B (−0.8%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	−2 B (−0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.3 KiB	−109 B (−1.7%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	55.5 KiB	54.7 KiB	−822 B (−1.4%)	66.4 KiB	✅
Codex	Live turn messages	9	9	0 (0.0%)	21	✅
Claude	Total thread wire	13.3 KiB	13.2 KiB	−150 B (−1.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	−14 B (−0.2%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.4 KiB	6.3 KiB	−136 B (−2.1%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	56.4 KiB	55.5 KiB	−884 B (−1.5%)	66.4 KiB	✅
Claude	Live turn messages	10	9	−1 (−10.0%)	21	✅

Baseline: abe70a7 · PR result: a7d79e8 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 109.4 KiB
• Claude decoded thread snapshot: 110.1 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[rynfar]

Problem

Pylon's server has accepted generic file uploads since #8235 landed as PR #140 — ChatFileAttachment in the contracts, fileAttachments.maxUploadBytes advertised per environment, adapter guards on every provider, and appendAttachmentPathLines in ProviderService handing each uploaded file's path to the agent. The web composer still only staged images, so the feature was half-shipped.

Fix

Adopts upstream #8236 (bcb855a63) and its typecheck fixup (d22709f75) from the f94a0d646..1f8ed54ad T3 review. The paperclip and paste accept any file up to the environment's advertised limit; uploads run while you type; files survive draft and stash moves; an interrupted upload rehydrates as an "Attach again" row rather than a stuck spinner; the timeline renders download rows, and unknown types render as an inert name with no link.

No server or contract change. packages/client-runtime gains only a ./state/attachments subpath export.

Conflict resolution

ChatComposer.tsx is ~4400 lines on Pylon with several fork-only controls, so git misaligned three regions that were re-inserted by hand: the draftTarget arguments to the upload queue, the file-attachment row, and the paperclip control. Upstream's fileCapabilityBlockReason / needs-reattach conditions were merged into Pylon's baseSendDisabledReason so both sets of conditions survive, with a file-capability block outranking Pylon's compaction message. ChatView.tsx had two adjacency conflicts beside the session-input-queue commands. composerDraftStore.ts auto-merged; COMPOSER_DRAFT_STORAGE_VERSION stays at 9 and Pylon's draft normalisation still runs. docs/user/composer.md gets the file-attachment paragraph in Pylon voice and states that mobile sends images only for now.

Two Pylon-specific notes: the paperclip is not shown in Pylon's collapsed-mobile composer row, consistent with that row having no image-attach affordance either (tapping it expands the full composer); and every provider, including Prime, receives generic files as prompt path lines — the adapter guards only prevent non-images being sent as base64 image blocks.

Verification

• vp test run across nine files (attachmentUploadQueue, attachmentUploadState, composerDraftStore, composerAttachmentFiles, promptStashStore, MessagesTimeline, ComposerStashMenu, ChatView.logic, client-runtime/state/attachments) — 267 passed.
• @t3tools/web and @t3tools/client-runtime typecheck clean; lint, format, and git diff --check clean on all 30 files.
• Independent review compared added/removed line multisets per file against upstream: all 30 match, ChatComposer.tsx has the same 42 hunks on both sides, and the hand-placed blocks sit at upstream's exact anchors (file row before ComposerPromptEditor; paperclip first in data-chat-composer-actions="right"). Reviewer also traced the reattach state machine (the upload store is not persisted, so a reload can never leave an entry stuck) and confirmed unknown file types carry no href.
• Not done: a real-client pass — attach a PDF, reload mid-upload and use "Attach again", stash and restore a file-bearing draft, and check footer crowding at narrow expanded widths now that the paperclip sits beside the inline badges.

---

Claude Fable 5 in Pylon.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.3 KiB	13.2 KiB	−111 B (−0.8%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	−2 B (−0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.3 KiB	−109 B (−1.7%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	55.5 KiB	54.7 KiB	−822 B (−1.4%)	66.4 KiB	✅
Codex	Live turn messages	9	9	0 (0.0%)	21	✅
Claude	Total thread wire	13.3 KiB	13.2 KiB	−150 B (−1.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	−14 B (−0.2%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.4 KiB	6.3 KiB	−136 B (−2.1%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	56.4 KiB	55.5 KiB	−884 B (−1.5%)	66.4 KiB	✅
Claude	Live turn messages	10	9	−1 (−10.0%)	21	✅

Baseline: abe70a7 · PR result: a7d79e8 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 109.4 KiB
• Claude decoded thread snapshot: 110.1 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[rynfar]

Problem

Pylon's server has accepted generic file uploads since #8235 landed as PR #140 — ChatFileAttachment in the contracts, fileAttachments.maxUploadBytes advertised per environment, adapter guards on every provider, and appendAttachmentPathLines in ProviderService handing each uploaded file's path to the agent. The web composer still only staged images, so the feature was half-shipped.

Fix

Adopts upstream #8236 (bcb855a63) and its typecheck fixup (d22709f75) from the f94a0d646..1f8ed54ad T3 review. The paperclip and paste accept any file up to the environment's advertised limit; uploads run while you type; files survive draft and stash moves; an interrupted upload rehydrates as an "Attach again" row rather than a stuck spinner; the timeline renders download rows, and unknown types render as an inert name with no link.

No server or contract change. packages/client-runtime gains only a ./state/attachments subpath export.

Conflict resolution

ChatComposer.tsx is ~4400 lines on Pylon with several fork-only controls, so git misaligned three regions that were re-inserted by hand: the draftTarget arguments to the upload queue, the file-attachment row, and the paperclip control. Upstream's fileCapabilityBlockReason / needs-reattach conditions were merged into Pylon's baseSendDisabledReason so both sets of conditions survive, with a file-capability block outranking Pylon's compaction message. ChatView.tsx had two adjacency conflicts beside the session-input-queue commands. composerDraftStore.ts auto-merged; COMPOSER_DRAFT_STORAGE_VERSION stays at 9 and Pylon's draft normalisation still runs. docs/user/composer.md gets the file-attachment paragraph in Pylon voice and states that mobile sends images only for now.

Two Pylon-specific notes: the paperclip is not shown in Pylon's collapsed-mobile composer row, consistent with that row having no image-attach affordance either (tapping it expands the full composer); and every provider, including Prime, receives generic files as prompt path lines — the adapter guards only prevent non-images being sent as base64 image blocks.

Verification

• vp test run across nine files (attachmentUploadQueue, attachmentUploadState, composerDraftStore, composerAttachmentFiles, promptStashStore, MessagesTimeline, ComposerStashMenu, ChatView.logic, client-runtime/state/attachments) — 267 passed.
• @t3tools/web and @t3tools/client-runtime typecheck clean; lint, format, and git diff --check clean on all 30 files.
• Independent review compared added/removed line multisets per file against upstream: all 30 match, ChatComposer.tsx has the same 42 hunks on both sides, and the hand-placed blocks sit at upstream's exact anchors (file row before ComposerPromptEditor; paperclip first in data-chat-composer-actions="right"). Reviewer also traced the reattach state machine (the upload store is not persisted, so a reload can never leave an entry stuck) and confirmed unknown file types carry no href.
• Not done: a real-client pass — attach a PDF, reload mid-upload and use "Attach again", stash and restore a file-bearing draft, and check footer crowding at narrow expanded widths now that the paperclip sits beside the inline badges.

---

Claude Fable 5 in Pylon.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.3 KiB	13.2 KiB	−111 B (−0.8%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	−2 B (−0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.3 KiB	−109 B (−1.7%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	55.5 KiB	54.7 KiB	−822 B (−1.4%)	66.4 KiB	✅
Codex	Live turn messages	9	9	0 (0.0%)	21	✅
Claude	Total thread wire	13.3 KiB	13.2 KiB	−150 B (−1.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	−14 B (−0.2%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.4 KiB	6.3 KiB	−136 B (−2.1%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	56.4 KiB	55.5 KiB	−884 B (−1.5%)	66.4 KiB	✅
Claude	Live turn messages	10	9	−1 (−10.0%)	21	✅

Baseline: abe70a7 · PR result: a7d79e8 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 109.4 KiB
• Claude decoded thread snapshot: 110.1 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[rynfar]

Problem

Pylon's server has accepted generic file uploads since #8235 landed as PR #140 — ChatFileAttachment in the contracts, fileAttachments.maxUploadBytes advertised per environment, adapter guards on every provider, and appendAttachmentPathLines in ProviderService handing each uploaded file's path to the agent. The web composer still only staged images, so the feature was half-shipped.

Fix

Adopts upstream #8236 (bcb855a63) and its typecheck fixup (d22709f75) from the f94a0d646..1f8ed54ad T3 review. The paperclip and paste accept any file up to the environment's advertised limit; uploads run while you type; files survive draft and stash moves; an interrupted upload rehydrates as an "Attach again" row rather than a stuck spinner; the timeline renders download rows, and unknown types render as an inert name with no link.

No server or contract change. packages/client-runtime gains only a ./state/attachments subpath export.

Conflict resolution

ChatComposer.tsx is ~4400 lines on Pylon with several fork-only controls, so git misaligned three regions that were re-inserted by hand: the draftTarget arguments to the upload queue, the file-attachment row, and the paperclip control. Upstream's fileCapabilityBlockReason / needs-reattach conditions were merged into Pylon's baseSendDisabledReason so both sets of conditions survive, with a file-capability block outranking Pylon's compaction message. ChatView.tsx had two adjacency conflicts beside the session-input-queue commands. composerDraftStore.ts auto-merged; COMPOSER_DRAFT_STORAGE_VERSION stays at 9 and Pylon's draft normalisation still runs. docs/user/composer.md gets the file-attachment paragraph in Pylon voice and states that mobile sends images only for now.

Two Pylon-specific notes: the paperclip is not shown in Pylon's collapsed-mobile composer row, consistent with that row having no image-attach affordance either (tapping it expands the full composer); and every provider, including Prime, receives generic files as prompt path lines — the adapter guards only prevent non-images being sent as base64 image blocks.

Verification

• vp test run across nine files (attachmentUploadQueue, attachmentUploadState, composerDraftStore, composerAttachmentFiles, promptStashStore, MessagesTimeline, ComposerStashMenu, ChatView.logic, client-runtime/state/attachments) — 267 passed.
• @t3tools/web and @t3tools/client-runtime typecheck clean; lint, format, and git diff --check clean on all 30 files.
• Independent review compared added/removed line multisets per file against upstream: all 30 match, ChatComposer.tsx has the same 42 hunks on both sides, and the hand-placed blocks sit at upstream's exact anchors (file row before ComposerPromptEditor; paperclip first in data-chat-composer-actions="right"). Reviewer also traced the reattach state machine (the upload store is not persisted, so a reload can never leave an entry stuck) and confirmed unknown file types carry no href.
• Not done: a real-client pass — attach a PDF, reload mid-upload and use "Attach again", stash and restore a file-bearing draft, and check footer crowding at narrow expanded widths now that the paperclip sits beside the inline badges.

---

Claude Fable 5 in Pylon.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.3 KiB	13.2 KiB	−111 B (−0.8%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	−2 B (−0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.3 KiB	−109 B (−1.7%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	55.5 KiB	54.7 KiB	−822 B (−1.4%)	66.4 KiB	✅
Codex	Live turn messages	9	9	0 (0.0%)	21	✅
Claude	Total thread wire	13.3 KiB	13.2 KiB	−150 B (−1.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	−14 B (−0.2%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.4 KiB	6.3 KiB	−136 B (−2.1%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	56.4 KiB	55.5 KiB	−884 B (−1.5%)	66.4 KiB	✅
Claude	Live turn messages	10	9	−1 (−10.0%)	21	✅

Baseline: abe70a7 · PR result: a7d79e8 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 109.4 KiB
• Claude decoded thread snapshot: 110.1 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[rynfar]

Problem

Pylon's server has accepted generic file uploads since #8235 landed as PR #140 — ChatFileAttachment in the contracts, fileAttachments.maxUploadBytes advertised per environment, adapter guards on every provider, and appendAttachmentPathLines in ProviderService handing each uploaded file's path to the agent. The web composer still only staged images, so the feature was half-shipped.

Fix

Adopts upstream #8236 (bcb855a63) and its typecheck fixup (d22709f75) from the f94a0d646..1f8ed54ad T3 review. The paperclip and paste accept any file up to the environment's advertised limit; uploads run while you type; files survive draft and stash moves; an interrupted upload rehydrates as an "Attach again" row rather than a stuck spinner; the timeline renders download rows, and unknown types render as an inert name with no link.

No server or contract change. packages/client-runtime gains only a ./state/attachments subpath export.

Conflict resolution

ChatComposer.tsx is ~4400 lines on Pylon with several fork-only controls, so git misaligned three regions that were re-inserted by hand: the draftTarget arguments to the upload queue, the file-attachment row, and the paperclip control. Upstream's fileCapabilityBlockReason / needs-reattach conditions were merged into Pylon's baseSendDisabledReason so both sets of conditions survive, with a file-capability block outranking Pylon's compaction message. ChatView.tsx had two adjacency conflicts beside the session-input-queue commands. composerDraftStore.ts auto-merged; COMPOSER_DRAFT_STORAGE_VERSION stays at 9 and Pylon's draft normalisation still runs. docs/user/composer.md gets the file-attachment paragraph in Pylon voice and states that mobile sends images only for now.

Two Pylon-specific notes: the paperclip is not shown in Pylon's collapsed-mobile composer row, consistent with that row having no image-attach affordance either (tapping it expands the full composer); and every provider, including Prime, receives generic files as prompt path lines — the adapter guards only prevent non-images being sent as base64 image blocks.

Verification

• vp test run across nine files (attachmentUploadQueue, attachmentUploadState, composerDraftStore, composerAttachmentFiles, promptStashStore, MessagesTimeline, ComposerStashMenu, ChatView.logic, client-runtime/state/attachments) — 267 passed.
• @t3tools/web and @t3tools/client-runtime typecheck clean; lint, format, and git diff --check clean on all 30 files.
• Independent review compared added/removed line multisets per file against upstream: all 30 match, ChatComposer.tsx has the same 42 hunks on both sides, and the hand-placed blocks sit at upstream's exact anchors (file row before ComposerPromptEditor; paperclip first in data-chat-composer-actions="right"). Reviewer also traced the reattach state machine (the upload store is not persisted, so a reload can never leave an entry stuck) and confirmed unknown file types carry no href.
• Not done: a real-client pass — attach a PDF, reload mid-upload and use "Attach again", stash and restore a file-bearing draft, and check footer crowding at narrow expanded widths now that the paperclip sits beside the inline badges.

---

Claude Fable 5 in Pylon.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.3 KiB	13.2 KiB	−111 B (−0.8%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	−2 B (−0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.3 KiB	−109 B (−1.7%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	55.5 KiB	54.7 KiB	−822 B (−1.4%)	66.4 KiB	✅
Codex	Live turn messages	9	9	0 (0.0%)	21	✅
Claude	Total thread wire	13.3 KiB	13.2 KiB	−150 B (−1.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	−14 B (−0.2%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.4 KiB	6.3 KiB	−136 B (−2.1%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	56.4 KiB	55.5 KiB	−884 B (−1.5%)	66.4 KiB	✅
Claude	Live turn messages	10	9	−1 (−10.0%)	21	✅

Baseline: abe70a7 · PR result: a7d79e8 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 109.4 KiB
• Claude decoded thread snapshot: 110.1 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.