fix(web): unify activity logs and composer banners

[rynfar]

Closes DEF-9 from the upstream review ledger.

Pylon's composer grew a set of bespoke .chat-composer-* CSS rules that every
notice, badge and drawer positioned itself against by hand. Upstream replaced
them with ComposerSurface/ComposerBanner primitives and deleted 524 lines of
that CSS. This adopts the primitives and moves Pylon's fork-only surfaces onto
them rather than dropping them.

Adopted from upstream pingdotgg/t3code: #8734 (30175a8af), #8850
(9842518c9), #8855 (3f62e6fa6).

Why three commits and not four

The ledger recorded this as #8693. Upstream reverted #8693 twelve hours after
merging it (#8733), then re-landed a substantially different version as
#8734 — 979 insertions of difference across 18 files, including keeping the
timeline working row that #8693 had deleted. #8733 restores the pre-#8693
tree exactly, so the revert/re-land pair cancels and #8734 is a clean re-land
onto a tree shaped like Pylon's. Porting #8693 would have adopted a version
upstream itself abandoned.

What Pylon keeps

• ComposerTasksBadge takes upstream's banner shape but keeps waitingOn
  ("Needs your input" / "Waiting on agents" / "Waiting on external system"),
  per-turn dismiss with its turnId + step:waitingOn snapshot, delegated-agent
  counts, and Pylon's TaskProgressSegments/TaskStatusIndicator. Upstream's
  badge has none of these. Segments hide while the list is expanded, since the
  list already shows every step.
• The timeline working row and workingStepLabel.#8734 keeps the row;
  #8855 strips the step label so the timer shows only elapsed time. That one
  hunk is declined — the row still says what the agent is doing, not merely how
  long.
• BranchToolbar adopts ComposerSurface.ContextStrip while keeping the
  branch selector in the left run with Usage opposite, per the comment already
  there. Upstream's right-hand copy would have rendered it twice.
• ChatComposer keeps ThreadHandoffTab, the quick-question and
  session-resource dialogs, the attachment control, externalDrawerAttached and
  the collapsed-mobile stash toggle. The last two were dropped silently by the
  cherry-pick in non-conflicting hunks; typecheck caught them.
• session-logic keeps Pylon's activity skip list and adds
  turn.plan.updated.

What Pylon gives up

• The work-group overflow toggle.MAX_VISIBLE_WORK_LOG_ENTRIES and
  onlyToolEntries are gone in favour of upstream's simpler grouping, which
  excludes agent-spawn and error rows from grouping outright. That preserves the
  #7893 and #7999 findings by construction — a spawn row can no longer be
  inside a group, so it can never hide behind a "+N tool calls" toggle — and
  hasFailure keys off latestToolEntry, which is exactly the "#7893 retried
  successfully is not a failure" rule.
• The inline per-turn plan chip.deriveTurnPlans, TurnPlanTimelineRow and
  the turn-plan row kind are removed. Upstream folds plans into the composer
  tasks badge, which Pylon keeps and extends with waitingOn, so the
  transcript chip would show the same plan twice. Reversible if you disagree —
  it needs the row kind restored into upstream's rewritten grouping.

One behaviour note worth a look: a missing-response notice now renders as a
work-toggle rather than a plain work row. It stays outside the fold and a
lone non-tool entry surfaces its own label as the summary, so the message text is
still what the user reads.

Verified: vp run -F @t3tools/web typecheck (exit 0, no errors);
vp test run --dir apps/web apps/web/src/components/chat apps/web/src/session-logic.test.ts apps/web/src/components/ChatView.logic.test.ts apps/web/src/versionSkew.test.ts
(637 passed); vp fmt --check apps/web/src clean. Also checked that no JSX
still references a removed .chat-composer-* class.

Not exercised in a browser. This restructures the composer's banner, drawer
and shoulder-tab layout, which unit tests cannot judge — it wants a real-client
pass before merge.

Reviewed and integrated with Claude Opus 5 in Claude Code.

---

^{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	13.5 KiB	13.6 KiB	+38 B (+0.3%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	+11 B (+0.2%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.6 KiB	6.6 KiB	+27 B (+0.4%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	57.2 KiB	57.2 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	13.6 KiB	13.8 KiB	+198 B (+1.4%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	+1 B (+0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.6 KiB	6.8 KiB	+197 B (+2.9%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.1 KiB	59.5 KiB	+1.5 KiB (+2.5%)	66.4 KiB	✅
Claude	Live turn messages	10	11	+1 (+10.0%)	21	✅

Baseline: 35b8380 · PR result: 6279bf4 · 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: 109.5 KiB
• Claude decoded thread snapshot: 110.2 KiB

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

[rynfar]

Adversarial review found thirteen defects, and it was right to be blunt: several features this PR claimed to keep were wired to nothing. Typecheck stayed green throughout, because none of it is a type error. All fixed on the branch.

Claims in the original description that were false

• Per-turn tasks dismiss was dead. No onDismiss reached the badge, the banner row or the drawer, so dismissTasks, visibleTasksProgress, visibleTaskSteps and dismissedTasksSnapshot were all unused. The feature you chose to keep did nothing.
• Delegated-agent counts reached only the inline badge, not the tab, banner row or drawer.
• workingStepLabel was gone from the working row.#8855's removal applied through a non-conflicting hunk, so declining the conflict wasn't enough — the prop and memo dep were left behind while the row stopped rendering the step. Exactly the same silent-loss mode as externalDrawerAttached.

All three are now wired through every surface, and the task surfaces gate on the dismissal-aware values rather than the raw ones.

Rendered twice

• The collapsed-mobile prompt row.
• A second data-chat-composer-main-surface wrapper nested inside ComposerSurface.Main, repeating composerFrameClassName — in ultrathink mode that ran the infinite .ultrathink-frame::before rainbow animation on two nested elements. ComposerSurface.Main already carries the attribute, so the wrapper was pure duplication.

Other defects fixed

• The dock and inline stash badges were no longer mutually exclusive; both rendered when a banner was attached or the drawer was open.
• The inline stash badge wrapped itself in a full glass ComposerBanner.Root inside a compact toolbar. placement="inline" restored.
• The || 1.375 drawer-inset fallback was dropped. The custom property went with the deleted CSS, so parseFloat("") gives NaN — and a NaN never equals itself, so the settle loop would never converge.
• The composer's activity item was prepended, tying with background liveness at priority: "activity" and tucking its Stop button — the only stop affordance for a settled turn — behind the peek. Now appended; a test pins the ordering.
• Dismissing a failed update also wrote the permanent version-mismatch dismissal, burying the "update available" offer for that version pair forever.
• tasksAriaLabel hardcoded "Current task" for waiting, next and completed plans, and the per-step sr-only status moved after the step text.
• Groups of 2+ non-tool rows collapsed entirely to "Received N updates", hiding text that is the whole point of a runtime warning or missing-response notice. Any all-non-tool group now shows its newest label.
• ComposerBannerStack.test.tsx and ComposerStashBadge.test.tsx had been deleted by the same change that rewrote both. Replacements written.

Also removed dead imports and a dead waitingOwnerLabel left by the plan-chip removal — caught by vp lint, not typecheck.

Re-verified: @t3tools/web typecheck clean, vp lint clean on the touched files, vp fmt --check clean, 628 tests passing across apps/web/src/components/chat, session-logic and ChatView.logic.

The review's closing point stands and I'm acting on it: findings 1, 2, 4 and 5 are exactly what a single real-client pass catches. Doing that next before this merges.

[rynfar]

Real-client pass done. Ran the branch against an isolated dev environment (--home-dir on a temp dir seeded via VACUUM INTO from a read-only snapshot — the live ~/.pylon-code was never opened for writing), paired the browser, and asserted directly against the DOM.

The structural defects the review found are gone:

Check	Before	Now
[data-chat-composer-main-surface]	2 (nested)	1
[data-chat-composer-footer]	2	1
[data-chat-composer-actions="right"]	2	1
Orphaned .chat-composer-* classes in the DOM	—	0
Console errors / warnings / unhandled rejections	—	0

Also verified live:

• The composer renders and is interactive: clicking focuses the editor, typing lands text, and the Send button transitions to enabled. That exercises the primary-actions region where ComposerFooterPrimaryActions' props, canQueueFollowUp/onQueueFollowUp and supportedRuntimeModes were repaired — all five footer controls render ("Claude Fable 5", "High · 1M", "Full access").
• The empty ComposerBanner.Attachment collapses to 0×0 via empty:hidden, so no phantom gap above the composer when there is nothing to show.

Two honest gaps in this pass:

1. The preview tool's snapshot/screenshot kept failing, so there are no before/after images — the table above is DOM assertions instead. For these particular findings (duplicate nodes, orphaned classes) that is stronger evidence than a picture, but AGENTS.md asks for images on UI changes and I could not produce them.
2. The tasks badge, stash badge and drawer states were not exercised. They need a thread with an active plan or stashed prompts, and the seeded threads' projects do not resolve in an isolated environment. So the onDismiss wiring, delegated-agent counts and inline-vs-tab stash placement are verified by unit tests and code reading, not by clicking them. Worth a look on a live thread before merge.

[rynfar]

Closes DEF-9 from the upstream review ledger.

Pylon's composer grew a set of bespoke .chat-composer-* CSS rules that every
notice, badge and drawer positioned itself against by hand. Upstream replaced
them with ComposerSurface/ComposerBanner primitives and deleted 524 lines of
that CSS. This adopts the primitives and moves Pylon's fork-only surfaces onto
them rather than dropping them.

Adopted from upstream pingdotgg/t3code: #8734 (30175a8af), #8850
(9842518c9), #8855 (3f62e6fa6).

Why three commits and not four

The ledger recorded this as #8693. Upstream reverted #8693 twelve hours after
merging it (#8733), then re-landed a substantially different version as
#8734 — 979 insertions of difference across 18 files, including keeping the
timeline working row that #8693 had deleted. #8733 restores the pre-#8693
tree exactly, so the revert/re-land pair cancels and #8734 is a clean re-land
onto a tree shaped like Pylon's. Porting #8693 would have adopted a version
upstream itself abandoned.

What Pylon keeps

• ComposerTasksBadge takes upstream's banner shape but keeps waitingOn
  ("Needs your input" / "Waiting on agents" / "Waiting on external system"),
  per-turn dismiss with its turnId + step:waitingOn snapshot, delegated-agent
  counts, and Pylon's TaskProgressSegments/TaskStatusIndicator. Upstream's
  badge has none of these. Segments hide while the list is expanded, since the
  list already shows every step.
• The timeline working row and workingStepLabel.#8734 keeps the row;
  #8855 strips the step label so the timer shows only elapsed time. That one
  hunk is declined — the row still says what the agent is doing, not merely how
  long.
• BranchToolbar adopts ComposerSurface.ContextStrip while keeping the
  branch selector in the left run with Usage opposite, per the comment already
  there. Upstream's right-hand copy would have rendered it twice.
• ChatComposer keeps ThreadHandoffTab, the quick-question and
  session-resource dialogs, the attachment control, externalDrawerAttached and
  the collapsed-mobile stash toggle. The last two were dropped silently by the
  cherry-pick in non-conflicting hunks; typecheck caught them.
• session-logic keeps Pylon's activity skip list and adds
  turn.plan.updated.

What Pylon gives up

• The work-group overflow toggle.MAX_VISIBLE_WORK_LOG_ENTRIES and
  onlyToolEntries are gone in favour of upstream's simpler grouping, which
  excludes agent-spawn and error rows from grouping outright. That preserves the
  #7893 and #7999 findings by construction — a spawn row can no longer be
  inside a group, so it can never hide behind a "+N tool calls" toggle — and
  hasFailure keys off latestToolEntry, which is exactly the "#7893 retried
  successfully is not a failure" rule.
• The inline per-turn plan chip.deriveTurnPlans, TurnPlanTimelineRow and
  the turn-plan row kind are removed. Upstream folds plans into the composer
  tasks badge, which Pylon keeps and extends with waitingOn, so the
  transcript chip would show the same plan twice. Reversible if you disagree —
  it needs the row kind restored into upstream's rewritten grouping.

One behaviour note worth a look: a missing-response notice now renders as a
work-toggle rather than a plain work row. It stays outside the fold and a
lone non-tool entry surfaces its own label as the summary, so the message text is
still what the user reads.

Verified: vp run -F @t3tools/web typecheck (exit 0, no errors);
vp test run --dir apps/web apps/web/src/components/chat apps/web/src/session-logic.test.ts apps/web/src/components/ChatView.logic.test.ts apps/web/src/versionSkew.test.ts
(637 passed); vp fmt --check apps/web/src clean. Also checked that no JSX
still references a removed .chat-composer-* class.

Not exercised in a browser. This restructures the composer's banner, drawer
and shoulder-tab layout, which unit tests cannot judge — it wants a real-client
pass before merge.

Reviewed and integrated with Claude Opus 5 in Claude Code.

---

^{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	13.5 KiB	13.6 KiB	+38 B (+0.3%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	+11 B (+0.2%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.6 KiB	6.6 KiB	+27 B (+0.4%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	57.2 KiB	57.2 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	13.6 KiB	13.8 KiB	+198 B (+1.4%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	+1 B (+0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.6 KiB	6.8 KiB	+197 B (+2.9%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.1 KiB	59.5 KiB	+1.5 KiB (+2.5%)	66.4 KiB	✅
Claude	Live turn messages	10	11	+1 (+10.0%)	21	✅

Baseline: 35b8380 · PR result: 6279bf4 · 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: 109.5 KiB
• Claude decoded thread snapshot: 110.2 KiB

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

[rynfar]

Adversarial review found thirteen defects, and it was right to be blunt: several features this PR claimed to keep were wired to nothing. Typecheck stayed green throughout, because none of it is a type error. All fixed on the branch.

Claims in the original description that were false

• Per-turn tasks dismiss was dead. No onDismiss reached the badge, the banner row or the drawer, so dismissTasks, visibleTasksProgress, visibleTaskSteps and dismissedTasksSnapshot were all unused. The feature you chose to keep did nothing.
• Delegated-agent counts reached only the inline badge, not the tab, banner row or drawer.
• workingStepLabel was gone from the working row.#8855's removal applied through a non-conflicting hunk, so declining the conflict wasn't enough — the prop and memo dep were left behind while the row stopped rendering the step. Exactly the same silent-loss mode as externalDrawerAttached.

All three are now wired through every surface, and the task surfaces gate on the dismissal-aware values rather than the raw ones.

Rendered twice

• The collapsed-mobile prompt row.
• A second data-chat-composer-main-surface wrapper nested inside ComposerSurface.Main, repeating composerFrameClassName — in ultrathink mode that ran the infinite .ultrathink-frame::before rainbow animation on two nested elements. ComposerSurface.Main already carries the attribute, so the wrapper was pure duplication.

Other defects fixed

• The dock and inline stash badges were no longer mutually exclusive; both rendered when a banner was attached or the drawer was open.
• The inline stash badge wrapped itself in a full glass ComposerBanner.Root inside a compact toolbar. placement="inline" restored.
• The || 1.375 drawer-inset fallback was dropped. The custom property went with the deleted CSS, so parseFloat("") gives NaN — and a NaN never equals itself, so the settle loop would never converge.
• The composer's activity item was prepended, tying with background liveness at priority: "activity" and tucking its Stop button — the only stop affordance for a settled turn — behind the peek. Now appended; a test pins the ordering.
• Dismissing a failed update also wrote the permanent version-mismatch dismissal, burying the "update available" offer for that version pair forever.
• tasksAriaLabel hardcoded "Current task" for waiting, next and completed plans, and the per-step sr-only status moved after the step text.
• Groups of 2+ non-tool rows collapsed entirely to "Received N updates", hiding text that is the whole point of a runtime warning or missing-response notice. Any all-non-tool group now shows its newest label.
• ComposerBannerStack.test.tsx and ComposerStashBadge.test.tsx had been deleted by the same change that rewrote both. Replacements written.

Also removed dead imports and a dead waitingOwnerLabel left by the plan-chip removal — caught by vp lint, not typecheck.

Re-verified: @t3tools/web typecheck clean, vp lint clean on the touched files, vp fmt --check clean, 628 tests passing across apps/web/src/components/chat, session-logic and ChatView.logic.

The review's closing point stands and I'm acting on it: findings 1, 2, 4 and 5 are exactly what a single real-client pass catches. Doing that next before this merges.

[rynfar]

Real-client pass done. Ran the branch against an isolated dev environment (--home-dir on a temp dir seeded via VACUUM INTO from a read-only snapshot — the live ~/.pylon-code was never opened for writing), paired the browser, and asserted directly against the DOM.

The structural defects the review found are gone:

Check	Before	Now
[data-chat-composer-main-surface]	2 (nested)	1
[data-chat-composer-footer]	2	1
[data-chat-composer-actions="right"]	2	1
Orphaned .chat-composer-* classes in the DOM	—	0
Console errors / warnings / unhandled rejections	—	0

Also verified live:

• The composer renders and is interactive: clicking focuses the editor, typing lands text, and the Send button transitions to enabled. That exercises the primary-actions region where ComposerFooterPrimaryActions' props, canQueueFollowUp/onQueueFollowUp and supportedRuntimeModes were repaired — all five footer controls render ("Claude Fable 5", "High · 1M", "Full access").
• The empty ComposerBanner.Attachment collapses to 0×0 via empty:hidden, so no phantom gap above the composer when there is nothing to show.

Two honest gaps in this pass:

1. The preview tool's snapshot/screenshot kept failing, so there are no before/after images — the table above is DOM assertions instead. For these particular findings (duplicate nodes, orphaned classes) that is stronger evidence than a picture, but AGENTS.md asks for images on UI changes and I could not produce them.
2. The tasks badge, stash badge and drawer states were not exercised. They need a thread with an active plan or stashed prompts, and the seeded threads' projects do not resolve in an isolated environment. So the onDismiss wiring, delegated-agent counts and inline-vs-tab stash placement are verified by unit tests and code reading, not by clicking them. Worth a look on a live thread before merge.

[rynfar]

Closes DEF-9 from the upstream review ledger.

Pylon's composer grew a set of bespoke .chat-composer-* CSS rules that every
notice, badge and drawer positioned itself against by hand. Upstream replaced
them with ComposerSurface/ComposerBanner primitives and deleted 524 lines of
that CSS. This adopts the primitives and moves Pylon's fork-only surfaces onto
them rather than dropping them.

Adopted from upstream pingdotgg/t3code: #8734 (30175a8af), #8850
(9842518c9), #8855 (3f62e6fa6).

Why three commits and not four

The ledger recorded this as #8693. Upstream reverted #8693 twelve hours after
merging it (#8733), then re-landed a substantially different version as
#8734 — 979 insertions of difference across 18 files, including keeping the
timeline working row that #8693 had deleted. #8733 restores the pre-#8693
tree exactly, so the revert/re-land pair cancels and #8734 is a clean re-land
onto a tree shaped like Pylon's. Porting #8693 would have adopted a version
upstream itself abandoned.

What Pylon keeps

• ComposerTasksBadge takes upstream's banner shape but keeps waitingOn
  ("Needs your input" / "Waiting on agents" / "Waiting on external system"),
  per-turn dismiss with its turnId + step:waitingOn snapshot, delegated-agent
  counts, and Pylon's TaskProgressSegments/TaskStatusIndicator. Upstream's
  badge has none of these. Segments hide while the list is expanded, since the
  list already shows every step.
• The timeline working row and workingStepLabel.#8734 keeps the row;
  #8855 strips the step label so the timer shows only elapsed time. That one
  hunk is declined — the row still says what the agent is doing, not merely how
  long.
• BranchToolbar adopts ComposerSurface.ContextStrip while keeping the
  branch selector in the left run with Usage opposite, per the comment already
  there. Upstream's right-hand copy would have rendered it twice.
• ChatComposer keeps ThreadHandoffTab, the quick-question and
  session-resource dialogs, the attachment control, externalDrawerAttached and
  the collapsed-mobile stash toggle. The last two were dropped silently by the
  cherry-pick in non-conflicting hunks; typecheck caught them.
• session-logic keeps Pylon's activity skip list and adds
  turn.plan.updated.

What Pylon gives up

• The work-group overflow toggle.MAX_VISIBLE_WORK_LOG_ENTRIES and
  onlyToolEntries are gone in favour of upstream's simpler grouping, which
  excludes agent-spawn and error rows from grouping outright. That preserves the
  #7893 and #7999 findings by construction — a spawn row can no longer be
  inside a group, so it can never hide behind a "+N tool calls" toggle — and
  hasFailure keys off latestToolEntry, which is exactly the "#7893 retried
  successfully is not a failure" rule.
• The inline per-turn plan chip.deriveTurnPlans, TurnPlanTimelineRow and
  the turn-plan row kind are removed. Upstream folds plans into the composer
  tasks badge, which Pylon keeps and extends with waitingOn, so the
  transcript chip would show the same plan twice. Reversible if you disagree —
  it needs the row kind restored into upstream's rewritten grouping.

One behaviour note worth a look: a missing-response notice now renders as a
work-toggle rather than a plain work row. It stays outside the fold and a
lone non-tool entry surfaces its own label as the summary, so the message text is
still what the user reads.

Verified: vp run -F @t3tools/web typecheck (exit 0, no errors);
vp test run --dir apps/web apps/web/src/components/chat apps/web/src/session-logic.test.ts apps/web/src/components/ChatView.logic.test.ts apps/web/src/versionSkew.test.ts
(637 passed); vp fmt --check apps/web/src clean. Also checked that no JSX
still references a removed .chat-composer-* class.

Not exercised in a browser. This restructures the composer's banner, drawer
and shoulder-tab layout, which unit tests cannot judge — it wants a real-client
pass before merge.

Reviewed and integrated with Claude Opus 5 in Claude Code.

---

^{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	13.5 KiB	13.6 KiB	+38 B (+0.3%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	+11 B (+0.2%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.6 KiB	6.6 KiB	+27 B (+0.4%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	57.2 KiB	57.2 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	13.6 KiB	13.8 KiB	+198 B (+1.4%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	+1 B (+0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.6 KiB	6.8 KiB	+197 B (+2.9%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.1 KiB	59.5 KiB	+1.5 KiB (+2.5%)	66.4 KiB	✅
Claude	Live turn messages	10	11	+1 (+10.0%)	21	✅

Baseline: 35b8380 · PR result: 6279bf4 · 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: 109.5 KiB
• Claude decoded thread snapshot: 110.2 KiB

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

[rynfar]

Adversarial review found thirteen defects, and it was right to be blunt: several features this PR claimed to keep were wired to nothing. Typecheck stayed green throughout, because none of it is a type error. All fixed on the branch.

Claims in the original description that were false

• Per-turn tasks dismiss was dead. No onDismiss reached the badge, the banner row or the drawer, so dismissTasks, visibleTasksProgress, visibleTaskSteps and dismissedTasksSnapshot were all unused. The feature you chose to keep did nothing.
• Delegated-agent counts reached only the inline badge, not the tab, banner row or drawer.
• workingStepLabel was gone from the working row.#8855's removal applied through a non-conflicting hunk, so declining the conflict wasn't enough — the prop and memo dep were left behind while the row stopped rendering the step. Exactly the same silent-loss mode as externalDrawerAttached.

All three are now wired through every surface, and the task surfaces gate on the dismissal-aware values rather than the raw ones.

Rendered twice

• The collapsed-mobile prompt row.
• A second data-chat-composer-main-surface wrapper nested inside ComposerSurface.Main, repeating composerFrameClassName — in ultrathink mode that ran the infinite .ultrathink-frame::before rainbow animation on two nested elements. ComposerSurface.Main already carries the attribute, so the wrapper was pure duplication.

Other defects fixed

• The dock and inline stash badges were no longer mutually exclusive; both rendered when a banner was attached or the drawer was open.
• The inline stash badge wrapped itself in a full glass ComposerBanner.Root inside a compact toolbar. placement="inline" restored.
• The || 1.375 drawer-inset fallback was dropped. The custom property went with the deleted CSS, so parseFloat("") gives NaN — and a NaN never equals itself, so the settle loop would never converge.
• The composer's activity item was prepended, tying with background liveness at priority: "activity" and tucking its Stop button — the only stop affordance for a settled turn — behind the peek. Now appended; a test pins the ordering.
• Dismissing a failed update also wrote the permanent version-mismatch dismissal, burying the "update available" offer for that version pair forever.
• tasksAriaLabel hardcoded "Current task" for waiting, next and completed plans, and the per-step sr-only status moved after the step text.
• Groups of 2+ non-tool rows collapsed entirely to "Received N updates", hiding text that is the whole point of a runtime warning or missing-response notice. Any all-non-tool group now shows its newest label.
• ComposerBannerStack.test.tsx and ComposerStashBadge.test.tsx had been deleted by the same change that rewrote both. Replacements written.

Also removed dead imports and a dead waitingOwnerLabel left by the plan-chip removal — caught by vp lint, not typecheck.

Re-verified: @t3tools/web typecheck clean, vp lint clean on the touched files, vp fmt --check clean, 628 tests passing across apps/web/src/components/chat, session-logic and ChatView.logic.

The review's closing point stands and I'm acting on it: findings 1, 2, 4 and 5 are exactly what a single real-client pass catches. Doing that next before this merges.

[rynfar]

Real-client pass done. Ran the branch against an isolated dev environment (--home-dir on a temp dir seeded via VACUUM INTO from a read-only snapshot — the live ~/.pylon-code was never opened for writing), paired the browser, and asserted directly against the DOM.

The structural defects the review found are gone:

Check	Before	Now
[data-chat-composer-main-surface]	2 (nested)	1
[data-chat-composer-footer]	2	1
[data-chat-composer-actions="right"]	2	1
Orphaned .chat-composer-* classes in the DOM	—	0
Console errors / warnings / unhandled rejections	—	0

Also verified live:

• The composer renders and is interactive: clicking focuses the editor, typing lands text, and the Send button transitions to enabled. That exercises the primary-actions region where ComposerFooterPrimaryActions' props, canQueueFollowUp/onQueueFollowUp and supportedRuntimeModes were repaired — all five footer controls render ("Claude Fable 5", "High · 1M", "Full access").
• The empty ComposerBanner.Attachment collapses to 0×0 via empty:hidden, so no phantom gap above the composer when there is nothing to show.

Two honest gaps in this pass:

1. The preview tool's snapshot/screenshot kept failing, so there are no before/after images — the table above is DOM assertions instead. For these particular findings (duplicate nodes, orphaned classes) that is stronger evidence than a picture, but AGENTS.md asks for images on UI changes and I could not produce them.
2. The tasks badge, stash badge and drawer states were not exercised. They need a thread with an active plan or stashed prompts, and the seeded threads' projects do not resolve in an isolated environment. So the onDismiss wiring, delegated-agent counts and inline-vs-tab stash placement are verified by unit tests and code reading, not by clicking them. Worth a look on a live thread before merge.

[rynfar]

Closes DEF-9 from the upstream review ledger.

Pylon's composer grew a set of bespoke .chat-composer-* CSS rules that every
notice, badge and drawer positioned itself against by hand. Upstream replaced
them with ComposerSurface/ComposerBanner primitives and deleted 524 lines of
that CSS. This adopts the primitives and moves Pylon's fork-only surfaces onto
them rather than dropping them.

Adopted from upstream pingdotgg/t3code: #8734 (30175a8af), #8850
(9842518c9), #8855 (3f62e6fa6).

Why three commits and not four

The ledger recorded this as #8693. Upstream reverted #8693 twelve hours after
merging it (#8733), then re-landed a substantially different version as
#8734 — 979 insertions of difference across 18 files, including keeping the
timeline working row that #8693 had deleted. #8733 restores the pre-#8693
tree exactly, so the revert/re-land pair cancels and #8734 is a clean re-land
onto a tree shaped like Pylon's. Porting #8693 would have adopted a version
upstream itself abandoned.

What Pylon keeps

• ComposerTasksBadge takes upstream's banner shape but keeps waitingOn
  ("Needs your input" / "Waiting on agents" / "Waiting on external system"),
  per-turn dismiss with its turnId + step:waitingOn snapshot, delegated-agent
  counts, and Pylon's TaskProgressSegments/TaskStatusIndicator. Upstream's
  badge has none of these. Segments hide while the list is expanded, since the
  list already shows every step.
• The timeline working row and workingStepLabel.#8734 keeps the row;
  #8855 strips the step label so the timer shows only elapsed time. That one
  hunk is declined — the row still says what the agent is doing, not merely how
  long.
• BranchToolbar adopts ComposerSurface.ContextStrip while keeping the
  branch selector in the left run with Usage opposite, per the comment already
  there. Upstream's right-hand copy would have rendered it twice.
• ChatComposer keeps ThreadHandoffTab, the quick-question and
  session-resource dialogs, the attachment control, externalDrawerAttached and
  the collapsed-mobile stash toggle. The last two were dropped silently by the
  cherry-pick in non-conflicting hunks; typecheck caught them.
• session-logic keeps Pylon's activity skip list and adds
  turn.plan.updated.

What Pylon gives up

• The work-group overflow toggle.MAX_VISIBLE_WORK_LOG_ENTRIES and
  onlyToolEntries are gone in favour of upstream's simpler grouping, which
  excludes agent-spawn and error rows from grouping outright. That preserves the
  #7893 and #7999 findings by construction — a spawn row can no longer be
  inside a group, so it can never hide behind a "+N tool calls" toggle — and
  hasFailure keys off latestToolEntry, which is exactly the "#7893 retried
  successfully is not a failure" rule.
• The inline per-turn plan chip.deriveTurnPlans, TurnPlanTimelineRow and
  the turn-plan row kind are removed. Upstream folds plans into the composer
  tasks badge, which Pylon keeps and extends with waitingOn, so the
  transcript chip would show the same plan twice. Reversible if you disagree —
  it needs the row kind restored into upstream's rewritten grouping.

One behaviour note worth a look: a missing-response notice now renders as a
work-toggle rather than a plain work row. It stays outside the fold and a
lone non-tool entry surfaces its own label as the summary, so the message text is
still what the user reads.

Verified: vp run -F @t3tools/web typecheck (exit 0, no errors);
vp test run --dir apps/web apps/web/src/components/chat apps/web/src/session-logic.test.ts apps/web/src/components/ChatView.logic.test.ts apps/web/src/versionSkew.test.ts
(637 passed); vp fmt --check apps/web/src clean. Also checked that no JSX
still references a removed .chat-composer-* class.

Not exercised in a browser. This restructures the composer's banner, drawer
and shoulder-tab layout, which unit tests cannot judge — it wants a real-client
pass before merge.

Reviewed and integrated with Claude Opus 5 in Claude Code.

---

^{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	13.5 KiB	13.6 KiB	+38 B (+0.3%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	+11 B (+0.2%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.6 KiB	6.6 KiB	+27 B (+0.4%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	57.2 KiB	57.2 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	13.6 KiB	13.8 KiB	+198 B (+1.4%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	+1 B (+0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.6 KiB	6.8 KiB	+197 B (+2.9%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.1 KiB	59.5 KiB	+1.5 KiB (+2.5%)	66.4 KiB	✅
Claude	Live turn messages	10	11	+1 (+10.0%)	21	✅

Baseline: 35b8380 · PR result: 6279bf4 · 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: 109.5 KiB
• Claude decoded thread snapshot: 110.2 KiB

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

[rynfar]

Adversarial review found thirteen defects, and it was right to be blunt: several features this PR claimed to keep were wired to nothing. Typecheck stayed green throughout, because none of it is a type error. All fixed on the branch.

Claims in the original description that were false

• Per-turn tasks dismiss was dead. No onDismiss reached the badge, the banner row or the drawer, so dismissTasks, visibleTasksProgress, visibleTaskSteps and dismissedTasksSnapshot were all unused. The feature you chose to keep did nothing.
• Delegated-agent counts reached only the inline badge, not the tab, banner row or drawer.
• workingStepLabel was gone from the working row.#8855's removal applied through a non-conflicting hunk, so declining the conflict wasn't enough — the prop and memo dep were left behind while the row stopped rendering the step. Exactly the same silent-loss mode as externalDrawerAttached.

All three are now wired through every surface, and the task surfaces gate on the dismissal-aware values rather than the raw ones.

Rendered twice

• The collapsed-mobile prompt row.
• A second data-chat-composer-main-surface wrapper nested inside ComposerSurface.Main, repeating composerFrameClassName — in ultrathink mode that ran the infinite .ultrathink-frame::before rainbow animation on two nested elements. ComposerSurface.Main already carries the attribute, so the wrapper was pure duplication.

Other defects fixed

• The dock and inline stash badges were no longer mutually exclusive; both rendered when a banner was attached or the drawer was open.
• The inline stash badge wrapped itself in a full glass ComposerBanner.Root inside a compact toolbar. placement="inline" restored.
• The || 1.375 drawer-inset fallback was dropped. The custom property went with the deleted CSS, so parseFloat("") gives NaN — and a NaN never equals itself, so the settle loop would never converge.
• The composer's activity item was prepended, tying with background liveness at priority: "activity" and tucking its Stop button — the only stop affordance for a settled turn — behind the peek. Now appended; a test pins the ordering.
• Dismissing a failed update also wrote the permanent version-mismatch dismissal, burying the "update available" offer for that version pair forever.
• tasksAriaLabel hardcoded "Current task" for waiting, next and completed plans, and the per-step sr-only status moved after the step text.
• Groups of 2+ non-tool rows collapsed entirely to "Received N updates", hiding text that is the whole point of a runtime warning or missing-response notice. Any all-non-tool group now shows its newest label.
• ComposerBannerStack.test.tsx and ComposerStashBadge.test.tsx had been deleted by the same change that rewrote both. Replacements written.

Also removed dead imports and a dead waitingOwnerLabel left by the plan-chip removal — caught by vp lint, not typecheck.

Re-verified: @t3tools/web typecheck clean, vp lint clean on the touched files, vp fmt --check clean, 628 tests passing across apps/web/src/components/chat, session-logic and ChatView.logic.

The review's closing point stands and I'm acting on it: findings 1, 2, 4 and 5 are exactly what a single real-client pass catches. Doing that next before this merges.

[rynfar]

Real-client pass done. Ran the branch against an isolated dev environment (--home-dir on a temp dir seeded via VACUUM INTO from a read-only snapshot — the live ~/.pylon-code was never opened for writing), paired the browser, and asserted directly against the DOM.

The structural defects the review found are gone:

Check	Before	Now
[data-chat-composer-main-surface]	2 (nested)	1
[data-chat-composer-footer]	2	1
[data-chat-composer-actions="right"]	2	1
Orphaned .chat-composer-* classes in the DOM	—	0
Console errors / warnings / unhandled rejections	—	0

Also verified live:

• The composer renders and is interactive: clicking focuses the editor, typing lands text, and the Send button transitions to enabled. That exercises the primary-actions region where ComposerFooterPrimaryActions' props, canQueueFollowUp/onQueueFollowUp and supportedRuntimeModes were repaired — all five footer controls render ("Claude Fable 5", "High · 1M", "Full access").
• The empty ComposerBanner.Attachment collapses to 0×0 via empty:hidden, so no phantom gap above the composer when there is nothing to show.

Two honest gaps in this pass:

1. The preview tool's snapshot/screenshot kept failing, so there are no before/after images — the table above is DOM assertions instead. For these particular findings (duplicate nodes, orphaned classes) that is stronger evidence than a picture, but AGENTS.md asks for images on UI changes and I could not produce them.
2. The tasks badge, stash badge and drawer states were not exercised. They need a thread with an active plan or stashed prompts, and the seeded threads' projects do not resolve in an isolated environment. So the onDismiss wiring, delegated-agent counts and inline-vs-tab stash placement are verified by unit tests and code reading, not by clicking them. Worth a look on a live thread before merge.

[rynfar]

Closes DEF-9 from the upstream review ledger.

Pylon's composer grew a set of bespoke .chat-composer-* CSS rules that every
notice, badge and drawer positioned itself against by hand. Upstream replaced
them with ComposerSurface/ComposerBanner primitives and deleted 524 lines of
that CSS. This adopts the primitives and moves Pylon's fork-only surfaces onto
them rather than dropping them.

Adopted from upstream pingdotgg/t3code: #8734 (30175a8af), #8850
(9842518c9), #8855 (3f62e6fa6).

Why three commits and not four

The ledger recorded this as #8693. Upstream reverted #8693 twelve hours after
merging it (#8733), then re-landed a substantially different version as
#8734 — 979 insertions of difference across 18 files, including keeping the
timeline working row that #8693 had deleted. #8733 restores the pre-#8693
tree exactly, so the revert/re-land pair cancels and #8734 is a clean re-land
onto a tree shaped like Pylon's. Porting #8693 would have adopted a version
upstream itself abandoned.

What Pylon keeps

• ComposerTasksBadge takes upstream's banner shape but keeps waitingOn
  ("Needs your input" / "Waiting on agents" / "Waiting on external system"),
  per-turn dismiss with its turnId + step:waitingOn snapshot, delegated-agent
  counts, and Pylon's TaskProgressSegments/TaskStatusIndicator. Upstream's
  badge has none of these. Segments hide while the list is expanded, since the
  list already shows every step.
• The timeline working row and workingStepLabel.#8734 keeps the row;
  #8855 strips the step label so the timer shows only elapsed time. That one
  hunk is declined — the row still says what the agent is doing, not merely how
  long.
• BranchToolbar adopts ComposerSurface.ContextStrip while keeping the
  branch selector in the left run with Usage opposite, per the comment already
  there. Upstream's right-hand copy would have rendered it twice.
• ChatComposer keeps ThreadHandoffTab, the quick-question and
  session-resource dialogs, the attachment control, externalDrawerAttached and
  the collapsed-mobile stash toggle. The last two were dropped silently by the
  cherry-pick in non-conflicting hunks; typecheck caught them.
• session-logic keeps Pylon's activity skip list and adds
  turn.plan.updated.

What Pylon gives up

• The work-group overflow toggle.MAX_VISIBLE_WORK_LOG_ENTRIES and
  onlyToolEntries are gone in favour of upstream's simpler grouping, which
  excludes agent-spawn and error rows from grouping outright. That preserves the
  #7893 and #7999 findings by construction — a spawn row can no longer be
  inside a group, so it can never hide behind a "+N tool calls" toggle — and
  hasFailure keys off latestToolEntry, which is exactly the "#7893 retried
  successfully is not a failure" rule.
• The inline per-turn plan chip.deriveTurnPlans, TurnPlanTimelineRow and
  the turn-plan row kind are removed. Upstream folds plans into the composer
  tasks badge, which Pylon keeps and extends with waitingOn, so the
  transcript chip would show the same plan twice. Reversible if you disagree —
  it needs the row kind restored into upstream's rewritten grouping.

One behaviour note worth a look: a missing-response notice now renders as a
work-toggle rather than a plain work row. It stays outside the fold and a
lone non-tool entry surfaces its own label as the summary, so the message text is
still what the user reads.

Verified: vp run -F @t3tools/web typecheck (exit 0, no errors);
vp test run --dir apps/web apps/web/src/components/chat apps/web/src/session-logic.test.ts apps/web/src/components/ChatView.logic.test.ts apps/web/src/versionSkew.test.ts
(637 passed); vp fmt --check apps/web/src clean. Also checked that no JSX
still references a removed .chat-composer-* class.

Not exercised in a browser. This restructures the composer's banner, drawer
and shoulder-tab layout, which unit tests cannot judge — it wants a real-client
pass before merge.

Reviewed and integrated with Claude Opus 5 in Claude Code.

---

^{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	13.5 KiB	13.6 KiB	+38 B (+0.3%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	+11 B (+0.2%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.6 KiB	6.6 KiB	+27 B (+0.4%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	57.2 KiB	57.2 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	13.6 KiB	13.8 KiB	+198 B (+1.4%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	+1 B (+0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.6 KiB	6.8 KiB	+197 B (+2.9%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.1 KiB	59.5 KiB	+1.5 KiB (+2.5%)	66.4 KiB	✅
Claude	Live turn messages	10	11	+1 (+10.0%)	21	✅

Baseline: 35b8380 · PR result: 6279bf4 · 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: 109.5 KiB
• Claude decoded thread snapshot: 110.2 KiB

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

[rynfar]

Adversarial review found thirteen defects, and it was right to be blunt: several features this PR claimed to keep were wired to nothing. Typecheck stayed green throughout, because none of it is a type error. All fixed on the branch.

Claims in the original description that were false

• Per-turn tasks dismiss was dead. No onDismiss reached the badge, the banner row or the drawer, so dismissTasks, visibleTasksProgress, visibleTaskSteps and dismissedTasksSnapshot were all unused. The feature you chose to keep did nothing.
• Delegated-agent counts reached only the inline badge, not the tab, banner row or drawer.
• workingStepLabel was gone from the working row.#8855's removal applied through a non-conflicting hunk, so declining the conflict wasn't enough — the prop and memo dep were left behind while the row stopped rendering the step. Exactly the same silent-loss mode as externalDrawerAttached.

All three are now wired through every surface, and the task surfaces gate on the dismissal-aware values rather than the raw ones.

Rendered twice

• The collapsed-mobile prompt row.
• A second data-chat-composer-main-surface wrapper nested inside ComposerSurface.Main, repeating composerFrameClassName — in ultrathink mode that ran the infinite .ultrathink-frame::before rainbow animation on two nested elements. ComposerSurface.Main already carries the attribute, so the wrapper was pure duplication.

Other defects fixed

• The dock and inline stash badges were no longer mutually exclusive; both rendered when a banner was attached or the drawer was open.
• The inline stash badge wrapped itself in a full glass ComposerBanner.Root inside a compact toolbar. placement="inline" restored.
• The || 1.375 drawer-inset fallback was dropped. The custom property went with the deleted CSS, so parseFloat("") gives NaN — and a NaN never equals itself, so the settle loop would never converge.
• The composer's activity item was prepended, tying with background liveness at priority: "activity" and tucking its Stop button — the only stop affordance for a settled turn — behind the peek. Now appended; a test pins the ordering.
• Dismissing a failed update also wrote the permanent version-mismatch dismissal, burying the "update available" offer for that version pair forever.
• tasksAriaLabel hardcoded "Current task" for waiting, next and completed plans, and the per-step sr-only status moved after the step text.
• Groups of 2+ non-tool rows collapsed entirely to "Received N updates", hiding text that is the whole point of a runtime warning or missing-response notice. Any all-non-tool group now shows its newest label.
• ComposerBannerStack.test.tsx and ComposerStashBadge.test.tsx had been deleted by the same change that rewrote both. Replacements written.

Also removed dead imports and a dead waitingOwnerLabel left by the plan-chip removal — caught by vp lint, not typecheck.

Re-verified: @t3tools/web typecheck clean, vp lint clean on the touched files, vp fmt --check clean, 628 tests passing across apps/web/src/components/chat, session-logic and ChatView.logic.

The review's closing point stands and I'm acting on it: findings 1, 2, 4 and 5 are exactly what a single real-client pass catches. Doing that next before this merges.

[rynfar]

Real-client pass done. Ran the branch against an isolated dev environment (--home-dir on a temp dir seeded via VACUUM INTO from a read-only snapshot — the live ~/.pylon-code was never opened for writing), paired the browser, and asserted directly against the DOM.

The structural defects the review found are gone:

Check	Before	Now
[data-chat-composer-main-surface]	2 (nested)	1
[data-chat-composer-footer]	2	1
[data-chat-composer-actions="right"]	2	1
Orphaned .chat-composer-* classes in the DOM	—	0
Console errors / warnings / unhandled rejections	—	0

Also verified live:

• The composer renders and is interactive: clicking focuses the editor, typing lands text, and the Send button transitions to enabled. That exercises the primary-actions region where ComposerFooterPrimaryActions' props, canQueueFollowUp/onQueueFollowUp and supportedRuntimeModes were repaired — all five footer controls render ("Claude Fable 5", "High · 1M", "Full access").
• The empty ComposerBanner.Attachment collapses to 0×0 via empty:hidden, so no phantom gap above the composer when there is nothing to show.

Two honest gaps in this pass:

1. The preview tool's snapshot/screenshot kept failing, so there are no before/after images — the table above is DOM assertions instead. For these particular findings (duplicate nodes, orphaned classes) that is stronger evidence than a picture, but AGENTS.md asks for images on UI changes and I could not produce them.
2. The tasks badge, stash badge and drawer states were not exercised. They need a thread with an active plan or stashed prompts, and the seeded threads' projects do not resolve in an isolated environment. So the onDismiss wiring, delegated-agent counts and inline-vs-tab stash placement are verified by unit tests and code reading, not by clicking them. Worth a look on a live thread before merge.

[rynfar]

Closes DEF-9 from the upstream review ledger.

Pylon's composer grew a set of bespoke .chat-composer-* CSS rules that every
notice, badge and drawer positioned itself against by hand. Upstream replaced
them with ComposerSurface/ComposerBanner primitives and deleted 524 lines of
that CSS. This adopts the primitives and moves Pylon's fork-only surfaces onto
them rather than dropping them.

Adopted from upstream pingdotgg/t3code: #8734 (30175a8af), #8850
(9842518c9), #8855 (3f62e6fa6).

Why three commits and not four

The ledger recorded this as #8693. Upstream reverted #8693 twelve hours after
merging it (#8733), then re-landed a substantially different version as
#8734 — 979 insertions of difference across 18 files, including keeping the
timeline working row that #8693 had deleted. #8733 restores the pre-#8693
tree exactly, so the revert/re-land pair cancels and #8734 is a clean re-land
onto a tree shaped like Pylon's. Porting #8693 would have adopted a version
upstream itself abandoned.

What Pylon keeps

• ComposerTasksBadge takes upstream's banner shape but keeps waitingOn
  ("Needs your input" / "Waiting on agents" / "Waiting on external system"),
  per-turn dismiss with its turnId + step:waitingOn snapshot, delegated-agent
  counts, and Pylon's TaskProgressSegments/TaskStatusIndicator. Upstream's
  badge has none of these. Segments hide while the list is expanded, since the
  list already shows every step.
• The timeline working row and workingStepLabel.#8734 keeps the row;
  #8855 strips the step label so the timer shows only elapsed time. That one
  hunk is declined — the row still says what the agent is doing, not merely how
  long.
• BranchToolbar adopts ComposerSurface.ContextStrip while keeping the
  branch selector in the left run with Usage opposite, per the comment already
  there. Upstream's right-hand copy would have rendered it twice.
• ChatComposer keeps ThreadHandoffTab, the quick-question and
  session-resource dialogs, the attachment control, externalDrawerAttached and
  the collapsed-mobile stash toggle. The last two were dropped silently by the
  cherry-pick in non-conflicting hunks; typecheck caught them.
• session-logic keeps Pylon's activity skip list and adds
  turn.plan.updated.

What Pylon gives up

• The work-group overflow toggle.MAX_VISIBLE_WORK_LOG_ENTRIES and
  onlyToolEntries are gone in favour of upstream's simpler grouping, which
  excludes agent-spawn and error rows from grouping outright. That preserves the
  #7893 and #7999 findings by construction — a spawn row can no longer be
  inside a group, so it can never hide behind a "+N tool calls" toggle — and
  hasFailure keys off latestToolEntry, which is exactly the "#7893 retried
  successfully is not a failure" rule.
• The inline per-turn plan chip.deriveTurnPlans, TurnPlanTimelineRow and
  the turn-plan row kind are removed. Upstream folds plans into the composer
  tasks badge, which Pylon keeps and extends with waitingOn, so the
  transcript chip would show the same plan twice. Reversible if you disagree —
  it needs the row kind restored into upstream's rewritten grouping.

One behaviour note worth a look: a missing-response notice now renders as a
work-toggle rather than a plain work row. It stays outside the fold and a
lone non-tool entry surfaces its own label as the summary, so the message text is
still what the user reads.

Verified: vp run -F @t3tools/web typecheck (exit 0, no errors);
vp test run --dir apps/web apps/web/src/components/chat apps/web/src/session-logic.test.ts apps/web/src/components/ChatView.logic.test.ts apps/web/src/versionSkew.test.ts
(637 passed); vp fmt --check apps/web/src clean. Also checked that no JSX
still references a removed .chat-composer-* class.

Not exercised in a browser. This restructures the composer's banner, drawer
and shoulder-tab layout, which unit tests cannot judge — it wants a real-client
pass before merge.

Reviewed and integrated with Claude Opus 5 in Claude Code.

---

^{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	13.5 KiB	13.6 KiB	+38 B (+0.3%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	+11 B (+0.2%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.6 KiB	6.6 KiB	+27 B (+0.4%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	57.2 KiB	57.2 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	13.6 KiB	13.8 KiB	+198 B (+1.4%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	+1 B (+0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.6 KiB	6.8 KiB	+197 B (+2.9%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.1 KiB	59.5 KiB	+1.5 KiB (+2.5%)	66.4 KiB	✅
Claude	Live turn messages	10	11	+1 (+10.0%)	21	✅

Baseline: 35b8380 · PR result: 6279bf4 · 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: 109.5 KiB
• Claude decoded thread snapshot: 110.2 KiB

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

[rynfar]

Adversarial review found thirteen defects, and it was right to be blunt: several features this PR claimed to keep were wired to nothing. Typecheck stayed green throughout, because none of it is a type error. All fixed on the branch.

Claims in the original description that were false

• Per-turn tasks dismiss was dead. No onDismiss reached the badge, the banner row or the drawer, so dismissTasks, visibleTasksProgress, visibleTaskSteps and dismissedTasksSnapshot were all unused. The feature you chose to keep did nothing.
• Delegated-agent counts reached only the inline badge, not the tab, banner row or drawer.
• workingStepLabel was gone from the working row.#8855's removal applied through a non-conflicting hunk, so declining the conflict wasn't enough — the prop and memo dep were left behind while the row stopped rendering the step. Exactly the same silent-loss mode as externalDrawerAttached.

All three are now wired through every surface, and the task surfaces gate on the dismissal-aware values rather than the raw ones.

Rendered twice

• The collapsed-mobile prompt row.
• A second data-chat-composer-main-surface wrapper nested inside ComposerSurface.Main, repeating composerFrameClassName — in ultrathink mode that ran the infinite .ultrathink-frame::before rainbow animation on two nested elements. ComposerSurface.Main already carries the attribute, so the wrapper was pure duplication.

Other defects fixed

• The dock and inline stash badges were no longer mutually exclusive; both rendered when a banner was attached or the drawer was open.
• The inline stash badge wrapped itself in a full glass ComposerBanner.Root inside a compact toolbar. placement="inline" restored.
• The || 1.375 drawer-inset fallback was dropped. The custom property went with the deleted CSS, so parseFloat("") gives NaN — and a NaN never equals itself, so the settle loop would never converge.
• The composer's activity item was prepended, tying with background liveness at priority: "activity" and tucking its Stop button — the only stop affordance for a settled turn — behind the peek. Now appended; a test pins the ordering.
• Dismissing a failed update also wrote the permanent version-mismatch dismissal, burying the "update available" offer for that version pair forever.
• tasksAriaLabel hardcoded "Current task" for waiting, next and completed plans, and the per-step sr-only status moved after the step text.
• Groups of 2+ non-tool rows collapsed entirely to "Received N updates", hiding text that is the whole point of a runtime warning or missing-response notice. Any all-non-tool group now shows its newest label.
• ComposerBannerStack.test.tsx and ComposerStashBadge.test.tsx had been deleted by the same change that rewrote both. Replacements written.

Also removed dead imports and a dead waitingOwnerLabel left by the plan-chip removal — caught by vp lint, not typecheck.

Re-verified: @t3tools/web typecheck clean, vp lint clean on the touched files, vp fmt --check clean, 628 tests passing across apps/web/src/components/chat, session-logic and ChatView.logic.

The review's closing point stands and I'm acting on it: findings 1, 2, 4 and 5 are exactly what a single real-client pass catches. Doing that next before this merges.

[rynfar]

Real-client pass done. Ran the branch against an isolated dev environment (--home-dir on a temp dir seeded via VACUUM INTO from a read-only snapshot — the live ~/.pylon-code was never opened for writing), paired the browser, and asserted directly against the DOM.

The structural defects the review found are gone:

Check	Before	Now
[data-chat-composer-main-surface]	2 (nested)	1
[data-chat-composer-footer]	2	1
[data-chat-composer-actions="right"]	2	1
Orphaned .chat-composer-* classes in the DOM	—	0
Console errors / warnings / unhandled rejections	—	0

Also verified live:

• The composer renders and is interactive: clicking focuses the editor, typing lands text, and the Send button transitions to enabled. That exercises the primary-actions region where ComposerFooterPrimaryActions' props, canQueueFollowUp/onQueueFollowUp and supportedRuntimeModes were repaired — all five footer controls render ("Claude Fable 5", "High · 1M", "Full access").
• The empty ComposerBanner.Attachment collapses to 0×0 via empty:hidden, so no phantom gap above the composer when there is nothing to show.

Two honest gaps in this pass:

1. The preview tool's snapshot/screenshot kept failing, so there are no before/after images — the table above is DOM assertions instead. For these particular findings (duplicate nodes, orphaned classes) that is stronger evidence than a picture, but AGENTS.md asks for images on UI changes and I could not produce them.
2. The tasks badge, stash badge and drawer states were not exercised. They need a thread with an active plan or stashed prompts, and the seeded threads' projects do not resolve in an isolated environment. So the onDismiss wiring, delegated-agent counts and inline-vs-tab stash placement are verified by unit tests and code reading, not by clicking them. Worth a look on a live thread before merge.

[rynfar]

Closes DEF-9 from the upstream review ledger.

Pylon's composer grew a set of bespoke .chat-composer-* CSS rules that every
notice, badge and drawer positioned itself against by hand. Upstream replaced
them with ComposerSurface/ComposerBanner primitives and deleted 524 lines of
that CSS. This adopts the primitives and moves Pylon's fork-only surfaces onto
them rather than dropping them.

Adopted from upstream pingdotgg/t3code: #8734 (30175a8af), #8850
(9842518c9), #8855 (3f62e6fa6).

Why three commits and not four

The ledger recorded this as #8693. Upstream reverted #8693 twelve hours after
merging it (#8733), then re-landed a substantially different version as
#8734 — 979 insertions of difference across 18 files, including keeping the
timeline working row that #8693 had deleted. #8733 restores the pre-#8693
tree exactly, so the revert/re-land pair cancels and #8734 is a clean re-land
onto a tree shaped like Pylon's. Porting #8693 would have adopted a version
upstream itself abandoned.

What Pylon keeps

• ComposerTasksBadge takes upstream's banner shape but keeps waitingOn
  ("Needs your input" / "Waiting on agents" / "Waiting on external system"),
  per-turn dismiss with its turnId + step:waitingOn snapshot, delegated-agent
  counts, and Pylon's TaskProgressSegments/TaskStatusIndicator. Upstream's
  badge has none of these. Segments hide while the list is expanded, since the
  list already shows every step.
• The timeline working row and workingStepLabel.#8734 keeps the row;
  #8855 strips the step label so the timer shows only elapsed time. That one
  hunk is declined — the row still says what the agent is doing, not merely how
  long.
• BranchToolbar adopts ComposerSurface.ContextStrip while keeping the
  branch selector in the left run with Usage opposite, per the comment already
  there. Upstream's right-hand copy would have rendered it twice.
• ChatComposer keeps ThreadHandoffTab, the quick-question and
  session-resource dialogs, the attachment control, externalDrawerAttached and
  the collapsed-mobile stash toggle. The last two were dropped silently by the
  cherry-pick in non-conflicting hunks; typecheck caught them.
• session-logic keeps Pylon's activity skip list and adds
  turn.plan.updated.

What Pylon gives up

• The work-group overflow toggle.MAX_VISIBLE_WORK_LOG_ENTRIES and
  onlyToolEntries are gone in favour of upstream's simpler grouping, which
  excludes agent-spawn and error rows from grouping outright. That preserves the
  #7893 and #7999 findings by construction — a spawn row can no longer be
  inside a group, so it can never hide behind a "+N tool calls" toggle — and
  hasFailure keys off latestToolEntry, which is exactly the "#7893 retried
  successfully is not a failure" rule.
• The inline per-turn plan chip.deriveTurnPlans, TurnPlanTimelineRow and
  the turn-plan row kind are removed. Upstream folds plans into the composer
  tasks badge, which Pylon keeps and extends with waitingOn, so the
  transcript chip would show the same plan twice. Reversible if you disagree —
  it needs the row kind restored into upstream's rewritten grouping.

One behaviour note worth a look: a missing-response notice now renders as a
work-toggle rather than a plain work row. It stays outside the fold and a
lone non-tool entry surfaces its own label as the summary, so the message text is
still what the user reads.

Verified: vp run -F @t3tools/web typecheck (exit 0, no errors);
vp test run --dir apps/web apps/web/src/components/chat apps/web/src/session-logic.test.ts apps/web/src/components/ChatView.logic.test.ts apps/web/src/versionSkew.test.ts
(637 passed); vp fmt --check apps/web/src clean. Also checked that no JSX
still references a removed .chat-composer-* class.

Not exercised in a browser. This restructures the composer's banner, drawer
and shoulder-tab layout, which unit tests cannot judge — it wants a real-client
pass before merge.

Reviewed and integrated with Claude Opus 5 in Claude Code.

---

^{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	13.5 KiB	13.6 KiB	+38 B (+0.3%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	+11 B (+0.2%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.6 KiB	6.6 KiB	+27 B (+0.4%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	57.2 KiB	57.2 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	13.6 KiB	13.8 KiB	+198 B (+1.4%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	+1 B (+0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.6 KiB	6.8 KiB	+197 B (+2.9%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.1 KiB	59.5 KiB	+1.5 KiB (+2.5%)	66.4 KiB	✅
Claude	Live turn messages	10	11	+1 (+10.0%)	21	✅

Baseline: 35b8380 · PR result: 6279bf4 · 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: 109.5 KiB
• Claude decoded thread snapshot: 110.2 KiB

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

[rynfar]

Adversarial review found thirteen defects, and it was right to be blunt: several features this PR claimed to keep were wired to nothing. Typecheck stayed green throughout, because none of it is a type error. All fixed on the branch.

Claims in the original description that were false

• Per-turn tasks dismiss was dead. No onDismiss reached the badge, the banner row or the drawer, so dismissTasks, visibleTasksProgress, visibleTaskSteps and dismissedTasksSnapshot were all unused. The feature you chose to keep did nothing.
• Delegated-agent counts reached only the inline badge, not the tab, banner row or drawer.
• workingStepLabel was gone from the working row.#8855's removal applied through a non-conflicting hunk, so declining the conflict wasn't enough — the prop and memo dep were left behind while the row stopped rendering the step. Exactly the same silent-loss mode as externalDrawerAttached.

All three are now wired through every surface, and the task surfaces gate on the dismissal-aware values rather than the raw ones.

Rendered twice

• The collapsed-mobile prompt row.
• A second data-chat-composer-main-surface wrapper nested inside ComposerSurface.Main, repeating composerFrameClassName — in ultrathink mode that ran the infinite .ultrathink-frame::before rainbow animation on two nested elements. ComposerSurface.Main already carries the attribute, so the wrapper was pure duplication.

Other defects fixed

• The dock and inline stash badges were no longer mutually exclusive; both rendered when a banner was attached or the drawer was open.
• The inline stash badge wrapped itself in a full glass ComposerBanner.Root inside a compact toolbar. placement="inline" restored.
• The || 1.375 drawer-inset fallback was dropped. The custom property went with the deleted CSS, so parseFloat("") gives NaN — and a NaN never equals itself, so the settle loop would never converge.
• The composer's activity item was prepended, tying with background liveness at priority: "activity" and tucking its Stop button — the only stop affordance for a settled turn — behind the peek. Now appended; a test pins the ordering.
• Dismissing a failed update also wrote the permanent version-mismatch dismissal, burying the "update available" offer for that version pair forever.
• tasksAriaLabel hardcoded "Current task" for waiting, next and completed plans, and the per-step sr-only status moved after the step text.
• Groups of 2+ non-tool rows collapsed entirely to "Received N updates", hiding text that is the whole point of a runtime warning or missing-response notice. Any all-non-tool group now shows its newest label.
• ComposerBannerStack.test.tsx and ComposerStashBadge.test.tsx had been deleted by the same change that rewrote both. Replacements written.

Also removed dead imports and a dead waitingOwnerLabel left by the plan-chip removal — caught by vp lint, not typecheck.

Re-verified: @t3tools/web typecheck clean, vp lint clean on the touched files, vp fmt --check clean, 628 tests passing across apps/web/src/components/chat, session-logic and ChatView.logic.

The review's closing point stands and I'm acting on it: findings 1, 2, 4 and 5 are exactly what a single real-client pass catches. Doing that next before this merges.

[rynfar]

Real-client pass done. Ran the branch against an isolated dev environment (--home-dir on a temp dir seeded via VACUUM INTO from a read-only snapshot — the live ~/.pylon-code was never opened for writing), paired the browser, and asserted directly against the DOM.

The structural defects the review found are gone:

Check	Before	Now
[data-chat-composer-main-surface]	2 (nested)	1
[data-chat-composer-footer]	2	1
[data-chat-composer-actions="right"]	2	1
Orphaned .chat-composer-* classes in the DOM	—	0
Console errors / warnings / unhandled rejections	—	0

Also verified live:

• The composer renders and is interactive: clicking focuses the editor, typing lands text, and the Send button transitions to enabled. That exercises the primary-actions region where ComposerFooterPrimaryActions' props, canQueueFollowUp/onQueueFollowUp and supportedRuntimeModes were repaired — all five footer controls render ("Claude Fable 5", "High · 1M", "Full access").
• The empty ComposerBanner.Attachment collapses to 0×0 via empty:hidden, so no phantom gap above the composer when there is nothing to show.

Two honest gaps in this pass:

1. The preview tool's snapshot/screenshot kept failing, so there are no before/after images — the table above is DOM assertions instead. For these particular findings (duplicate nodes, orphaned classes) that is stronger evidence than a picture, but AGENTS.md asks for images on UI changes and I could not produce them.
2. The tasks badge, stash badge and drawer states were not exercised. They need a thread with an active plan or stashed prompts, and the seeded threads' projects do not resolve in an isolated environment. So the onDismiss wiring, delegated-agent counts and inline-vs-tab stash placement are verified by unit tests and code reading, not by clicking them. Worth a look on a live thread before merge.

[rynfar]

Closes DEF-9 from the upstream review ledger.

Pylon's composer grew a set of bespoke .chat-composer-* CSS rules that every
notice, badge and drawer positioned itself against by hand. Upstream replaced
them with ComposerSurface/ComposerBanner primitives and deleted 524 lines of
that CSS. This adopts the primitives and moves Pylon's fork-only surfaces onto
them rather than dropping them.

Adopted from upstream pingdotgg/t3code: #8734 (30175a8af), #8850
(9842518c9), #8855 (3f62e6fa6).

Why three commits and not four

The ledger recorded this as #8693. Upstream reverted #8693 twelve hours after
merging it (#8733), then re-landed a substantially different version as
#8734 — 979 insertions of difference across 18 files, including keeping the
timeline working row that #8693 had deleted. #8733 restores the pre-#8693
tree exactly, so the revert/re-land pair cancels and #8734 is a clean re-land
onto a tree shaped like Pylon's. Porting #8693 would have adopted a version
upstream itself abandoned.

What Pylon keeps

• ComposerTasksBadge takes upstream's banner shape but keeps waitingOn
  ("Needs your input" / "Waiting on agents" / "Waiting on external system"),
  per-turn dismiss with its turnId + step:waitingOn snapshot, delegated-agent
  counts, and Pylon's TaskProgressSegments/TaskStatusIndicator. Upstream's
  badge has none of these. Segments hide while the list is expanded, since the
  list already shows every step.
• The timeline working row and workingStepLabel.#8734 keeps the row;
  #8855 strips the step label so the timer shows only elapsed time. That one
  hunk is declined — the row still says what the agent is doing, not merely how
  long.
• BranchToolbar adopts ComposerSurface.ContextStrip while keeping the
  branch selector in the left run with Usage opposite, per the comment already
  there. Upstream's right-hand copy would have rendered it twice.
• ChatComposer keeps ThreadHandoffTab, the quick-question and
  session-resource dialogs, the attachment control, externalDrawerAttached and
  the collapsed-mobile stash toggle. The last two were dropped silently by the
  cherry-pick in non-conflicting hunks; typecheck caught them.
• session-logic keeps Pylon's activity skip list and adds
  turn.plan.updated.

What Pylon gives up

• The work-group overflow toggle.MAX_VISIBLE_WORK_LOG_ENTRIES and
  onlyToolEntries are gone in favour of upstream's simpler grouping, which
  excludes agent-spawn and error rows from grouping outright. That preserves the
  #7893 and #7999 findings by construction — a spawn row can no longer be
  inside a group, so it can never hide behind a "+N tool calls" toggle — and
  hasFailure keys off latestToolEntry, which is exactly the "#7893 retried
  successfully is not a failure" rule.
• The inline per-turn plan chip.deriveTurnPlans, TurnPlanTimelineRow and
  the turn-plan row kind are removed. Upstream folds plans into the composer
  tasks badge, which Pylon keeps and extends with waitingOn, so the
  transcript chip would show the same plan twice. Reversible if you disagree —
  it needs the row kind restored into upstream's rewritten grouping.

One behaviour note worth a look: a missing-response notice now renders as a
work-toggle rather than a plain work row. It stays outside the fold and a
lone non-tool entry surfaces its own label as the summary, so the message text is
still what the user reads.

Verified: vp run -F @t3tools/web typecheck (exit 0, no errors);
vp test run --dir apps/web apps/web/src/components/chat apps/web/src/session-logic.test.ts apps/web/src/components/ChatView.logic.test.ts apps/web/src/versionSkew.test.ts
(637 passed); vp fmt --check apps/web/src clean. Also checked that no JSX
still references a removed .chat-composer-* class.

Not exercised in a browser. This restructures the composer's banner, drawer
and shoulder-tab layout, which unit tests cannot judge — it wants a real-client
pass before merge.

Reviewed and integrated with Claude Opus 5 in Claude Code.

---

^{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	13.5 KiB	13.6 KiB	+38 B (+0.3%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	+11 B (+0.2%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.6 KiB	6.6 KiB	+27 B (+0.4%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	57.2 KiB	57.2 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	13.6 KiB	13.8 KiB	+198 B (+1.4%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	+1 B (+0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.6 KiB	6.8 KiB	+197 B (+2.9%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.1 KiB	59.5 KiB	+1.5 KiB (+2.5%)	66.4 KiB	✅
Claude	Live turn messages	10	11	+1 (+10.0%)	21	✅

Baseline: 35b8380 · PR result: 6279bf4 · 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: 109.5 KiB
• Claude decoded thread snapshot: 110.2 KiB

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

[rynfar]

Adversarial review found thirteen defects, and it was right to be blunt: several features this PR claimed to keep were wired to nothing. Typecheck stayed green throughout, because none of it is a type error. All fixed on the branch.

Claims in the original description that were false

• Per-turn tasks dismiss was dead. No onDismiss reached the badge, the banner row or the drawer, so dismissTasks, visibleTasksProgress, visibleTaskSteps and dismissedTasksSnapshot were all unused. The feature you chose to keep did nothing.
• Delegated-agent counts reached only the inline badge, not the tab, banner row or drawer.
• workingStepLabel was gone from the working row.#8855's removal applied through a non-conflicting hunk, so declining the conflict wasn't enough — the prop and memo dep were left behind while the row stopped rendering the step. Exactly the same silent-loss mode as externalDrawerAttached.

All three are now wired through every surface, and the task surfaces gate on the dismissal-aware values rather than the raw ones.

Rendered twice

• The collapsed-mobile prompt row.
• A second data-chat-composer-main-surface wrapper nested inside ComposerSurface.Main, repeating composerFrameClassName — in ultrathink mode that ran the infinite .ultrathink-frame::before rainbow animation on two nested elements. ComposerSurface.Main already carries the attribute, so the wrapper was pure duplication.

Other defects fixed

• The dock and inline stash badges were no longer mutually exclusive; both rendered when a banner was attached or the drawer was open.
• The inline stash badge wrapped itself in a full glass ComposerBanner.Root inside a compact toolbar. placement="inline" restored.
• The || 1.375 drawer-inset fallback was dropped. The custom property went with the deleted CSS, so parseFloat("") gives NaN — and a NaN never equals itself, so the settle loop would never converge.
• The composer's activity item was prepended, tying with background liveness at priority: "activity" and tucking its Stop button — the only stop affordance for a settled turn — behind the peek. Now appended; a test pins the ordering.
• Dismissing a failed update also wrote the permanent version-mismatch dismissal, burying the "update available" offer for that version pair forever.
• tasksAriaLabel hardcoded "Current task" for waiting, next and completed plans, and the per-step sr-only status moved after the step text.
• Groups of 2+ non-tool rows collapsed entirely to "Received N updates", hiding text that is the whole point of a runtime warning or missing-response notice. Any all-non-tool group now shows its newest label.
• ComposerBannerStack.test.tsx and ComposerStashBadge.test.tsx had been deleted by the same change that rewrote both. Replacements written.

Also removed dead imports and a dead waitingOwnerLabel left by the plan-chip removal — caught by vp lint, not typecheck.

Re-verified: @t3tools/web typecheck clean, vp lint clean on the touched files, vp fmt --check clean, 628 tests passing across apps/web/src/components/chat, session-logic and ChatView.logic.

The review's closing point stands and I'm acting on it: findings 1, 2, 4 and 5 are exactly what a single real-client pass catches. Doing that next before this merges.

[rynfar]

Real-client pass done. Ran the branch against an isolated dev environment (--home-dir on a temp dir seeded via VACUUM INTO from a read-only snapshot — the live ~/.pylon-code was never opened for writing), paired the browser, and asserted directly against the DOM.

The structural defects the review found are gone:

Check	Before	Now
[data-chat-composer-main-surface]	2 (nested)	1
[data-chat-composer-footer]	2	1
[data-chat-composer-actions="right"]	2	1
Orphaned .chat-composer-* classes in the DOM	—	0
Console errors / warnings / unhandled rejections	—	0

Also verified live:

• The composer renders and is interactive: clicking focuses the editor, typing lands text, and the Send button transitions to enabled. That exercises the primary-actions region where ComposerFooterPrimaryActions' props, canQueueFollowUp/onQueueFollowUp and supportedRuntimeModes were repaired — all five footer controls render ("Claude Fable 5", "High · 1M", "Full access").
• The empty ComposerBanner.Attachment collapses to 0×0 via empty:hidden, so no phantom gap above the composer when there is nothing to show.

Two honest gaps in this pass:

1. The preview tool's snapshot/screenshot kept failing, so there are no before/after images — the table above is DOM assertions instead. For these particular findings (duplicate nodes, orphaned classes) that is stronger evidence than a picture, but AGENTS.md asks for images on UI changes and I could not produce them.
2. The tasks badge, stash badge and drawer states were not exercised. They need a thread with an active plan or stashed prompts, and the seeded threads' projects do not resolve in an isolated environment. So the onDismiss wiring, delegated-agent counts and inline-vs-tab stash placement are verified by unit tests and code reading, not by clicking them. Worth a look on a live thread before merge.