fix(orchestrator): Keep Home, steers, and errors honest when work ends

[mwolson]

What Changed

You can steer a turn while the agent is still working. The run could settle anyway, and the later answer vanished. That later text now stays on the thread.

A provider that never started could leave the thread looking alive forever. Exhausted retries now fail the run.

You settle a thread, and Home still treats it like live work. Missing snapshots keep retrying, archived threads drift back, and Waiting keeps refreshing stale rows. Settled now stays settled. A gone thread stops retrying.

You dismiss a thread error, and a later one never appears, or the old one never leaves. Dismissal lasts until a later error. Plan mode off still left the Build/Plan toggle in the composer after the v2 chat rewrite. The toggle is hidden in that case.

Why

Settle meant the useful work was over in the UI while the agent, the outbox, or a later error was still in play. Home, Waiting, and the composer should follow that same rule.

Julius's bounded thread history on current CTM stays as shipped.

UI Changes

Steered output stays in the thread. Failed starts no longer hang as if they were still starting. Home and Waiting settled state is quieter. Thread error banners dismiss until a later error. Build/Plan is hidden when plan mode is disabled.

No before/after screenshots.

Checklist

• This PR is small and focused
• I explained what changed and why
• I included before/after screenshots for any UI changes
• I included a video for animation/interaction changes

---

Note

High Risk
Changes core orchestration paths—settlement merging, command commit guards, Claude turn lifecycle, and process-loss reconciliation—with race-sensitive behavior that can leave runs stuck or drop recovery if callers mishandle committed: false.

Overview
Threads gain settledOverrideAt so settle/keep-active pins have a stable establishment time that survives renames and is exposed on shells and mobile projections.

Settlement is tightened end-to-end: settle is blocked on queued runs, pending runtime requests, and provider background work; archived threads cannot settle/unsettle; pin/unsettle paths maintain the new timestamp; message dispatch clears overrides via a shared helper. Settled/unsettled events are merged through reduceThreadSettlementEvent in memory and in SQL so stale full-thread payloads cannot overwrite title or other metadata. Provider ingest tests document when pending thread-scoped requests should emit thread.unsettled without waking keep-active or non-activity updates.

Claude steering tracks per-turn steer counts and frame counts, detects steering handoff results before terminalizing, marks in-flight tools interrupted, and bounds silent handoffs with a timeout that fails the turn if no new frames arrive. Turn finalization is claimed up front to avoid duplicate terminals when the handoff fiber races the frame handler; new turns wait while a turn is finalizing.

Recovery adds optional updatedBefore filters on outbox cancel/reconcile and expectedThreadUpdatedAt on commitCommand so stale recovery writes no-op when the thread has already advanced.

^{Reviewed by Cursor Bugbot for commit 57fca87. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add Orchestration V2 system with provider adapters, persistence, and V2 UI migration

• Introduces the complete Orchestration V2 architecture: contracts, event store, projection store, command receipts, effect outbox/worker, runtime recovery, and database migrations 044–052
• Adds V2 provider adapters for Claude, Codex, Cursor, Grok, OpenCode, and ACP Registry through the ProviderAdapterV2 service contract and driver registry, with replay testkits for deterministic testing
• Implements thread lifecycle services for launch, fork, settlement, checkpoint capture/rollback, context handoff, continuation, run finalization, and provider-session management with idle timeout and shutdown
• Migrates client-runtime, web, and mobile state from legacy orchestration types to V2 thread projections, bounded shell snapshots, runtime summaries, and progressive history paging
• Adds MCP orchestrator and worktree toolkits, scheduled-task service with interval and fixed-time scheduling, T3 orchestration instructions, and a legacy V1 thread importer with idempotent shell and transcript hydration
• Risk: migrations 044–052 create new schema tables and copy existing V2/project data into a shared application event log; EnvironmentThreadState in threadState.ts replaces legacy thread/page fields with OrchestrationV2ThreadProjection and ThreadHistoryMeta, removing threadHasOlderTurns; RPC authorization scopes in RpcAuthorization.ts are remapped to V2 method constants

^{Macroscope summarized 674b0e4.}

[coderabbitai]

Important

Review skipped

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

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 7726a859-eb6b-4efc-9e83-8c6c368e6e50

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

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

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

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[macroscopeapp]

One convention issue found in the new provider-ingest settlement code: a statically known tagged failure is recovered with Effect.catchTag instead of Effect.catchTags.

^{Posted via Macroscope — Effect Service Conventions}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This is a 964-file Orchestration V2 architecture and persistence migration that adds provider integrations, new workflows, protocol/schema changes, and product-default changes across server, web, and mobile. Unresolved projection and provider-lifecycle race findings add material consistency risk beyond the scope of an auto-approvable fix.

Not approved because:

• 3 blocking correctness issues found at or above your repo's Minimum Blocking Severity

^{Adjust the Minimum Blocking Severity for this repo — including turning it Off — in Settings. You can add or adjust custom eligibility rules. Learn more.}

[macroscopeapp]

Effect service conventions review of the settle-lifecycle changes. The previously flagged Effect.catchTag usage in ProviderEventIngestor.ts is now Effect.catchTags({ ... }), and the new/changed service code (ProjectionStore, EventSink, EffectOutbox, ProviderRuntimeRecoveryService, ProviderEventIngestor, shared settlement helpers) follows the import, layer, and dependency-acquisition rules. Two small change-discipline findings in packages/client-runtime/src/state/threadSettled.ts.

^{Posted via Macroscope — Effect Service Conventions}

[macroscopeapp]

🟡 Mediumorchestration-v2/ProjectionStore.ts:1261

thread.unsettled events lose event.payload.unsettledAt in the durable projection, so shell snapshots report unsettledAt as the previous value (often null) instead of when the thread was explicitly kept active. Include unsettledAt in the settlement merge.

 settlement: {
+ unsettledAt: event.payload.unsettledAt ?? null,

🤖 Copy this AI Prompt to have your agent fix this:

In file @apps/server/src/orchestration-v2/ProjectionStore.ts around line 1261:
`thread.unsettled` events lose `event.payload.unsettledAt` in the durable projection, so shell snapshots report `unsettledAt` as the previous value (often `null`) instead of when the thread was explicitly kept active. Include `unsettledAt` in the settlement merge.

[macroscopeapp]

🟡 Mediumorchestration-v2/ProviderEventIngestor.ts:356

Archived threads receive a thread.unsettled event for pending non-auth runtime-request.updated activity, clearing their settledOverride instead of leaving archived settlement state unchanged. getThreadProjection returns the still-present archived projection, and this path does not reject it before appending the synthetic event; restrict candidate generation to active threads.

🤖 Copy this AI Prompt to have your agent fix this:

In file @apps/server/src/orchestration-v2/ProviderEventIngestor.ts around line 356:
Archived threads receive a `thread.unsettled` event for pending non-auth `runtime-request.updated` activity, clearing their `settledOverride` instead of leaving archived settlement state unchanged. `getThreadProjection` returns the still-present archived projection, and this path does not reject it before appending the synthetic event; restrict candidate generation to active threads.

[macroscopeapp]

🟡 Mediumstate/orchestrationV2Projection.ts:207

A delayed thread.unsettled rewinds projection.updatedAt to event.occurredAt, even when the projection already has a newer timestamp from a concurrent metadata update. Although reduceThreadSettlementEvent preserves the newer thread.updatedAt, spreading base here still makes client ordering/activity views treat the thread as stale; preserve the later of the existing projection timestamp and event.occurredAt, as the server reducer does.

🤖 Copy this AI Prompt to have your agent fix this:

In file @packages/client-runtime/src/state/orchestrationV2Projection.ts around line 207:
A delayed `thread.unsettled` rewinds `projection.updatedAt` to `event.occurredAt`, even when the projection already has a newer timestamp from a concurrent metadata update. Although `reduceThreadSettlementEvent` preserves the newer `thread.updatedAt`, spreading `base` here still makes client ordering/activity views treat the thread as stale; preserve the later of the existing projection timestamp and `event.occurredAt`, as the server reducer does.

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 57fca87. Configure here.}

[cursor]

Finalize claim races new turns

Medium Severity

finalizeActiveTurn nulls activeTurn and only then adds the id to finalizingTurnIds. startTurn checks those refs separately, so a concurrent start (the handoff timeout is forked) can observe both empty and open a second provider turn while the first is still emitting terminals.

Additional Locations (1)

• apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts#L5197-L5211

^{Reviewed by Cursor Bugbot for commit 57fca87. Configure here.}

[macroscopeapp]

Scope note: only apps/web/src/** TS/TSX/CSS changes were reviewed. Findings are the newly added global CSS in index.css (dead/duplicated rules) and the composer context strip's motion contract, which changed lines break while the code that depends on it is retained.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

The strip's compact/expand FLIP animation still keys on [data-composer-context-control] (BranchToolbar.tsx:232, consumed by useLabelsOverflow), but this PR removes that attribute from the environment selector, the env-mode selector (locked and select paths) and the branch selector. The only remaining carrier is the vertical Separator at BranchToolbar.tsx:545, so collapsing now animates a 1px separator while the actual controls jump to their new positions. Consider keeping the attribute on the toolbar-mode triggers (it is inert in panel mode), or removing the FLIP pass as well if the motion is intentionally gone.

Suggested change

	aria-label="Run on"
	aria-label="Run on"
	data-composer-context-control

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

This whole .glass-opacity-slider block (through the forced-colors rules) is unreferenced — the glass-opacity input in SettingsPanels.tsx:1111 uses settings-slider with --settings-slider-progress, and no literal or composed glass-opacity-slider class exists anywhere in the repo. It is also a verbatim copy of .settings-slider below with only the custom-property names changed, so it adds a second owner for the same slider treatment. Suggest dropping it (or, if a distinct slider is intended, wiring a call site and deriving it from .settings-slider rather than duplicating the vendor pseudo-element rules).

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

alert-glass is already defined as @utility alert-glass (line 284) and is what ProviderStatusBanner consumes. This second definition in @layer components duplicates it and omits the utility's @supports not ((-webkit-backdrop-filter: blur(1px)) or (backdrop-filter: blur(1px))) opaque fallback; because !important declarations resolve in reverse layer order, this components-layer background: ... !important outranks the utility's important fallback, so browsers without backdrop-filter get a translucent, unblurred banner instead of the solid background. Suggest deleting this block and keeping the @utility as the single owner.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

.surface-subheader has no consumer: all four subheaders (DiffPanelShell, FileBrowserPanel, FilePreviewPanel, PreviewChromeRow) still carry the same geometry inline and only set data-surface-subheader. The inline-mode rule below is also redundant — every carrier already has in-data-[preview-panel-mode=inline]:h-7/min-h-7/mb-3/border-b-transparent, and those utilities sort after @layer components, so they win wherever this rule matches. Either apply surface-subheader at those call sites and drop the duplicated utility strings, or drop both rules so the shared geometry keeps one owner.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

The label previously animated transform/opacity on an inner data-composer-label-motion span with motion-reduce:transform-none motion-reduce:transition-opacity; this replaces it with a max-width transition and no reduced-motion guard, so reduced-motion users now get an animated layout change on every collapse. Same class appears at line 79 and in BranchToolbarEnvModeSelector/BranchToolbarBranchSelector.

Suggested change

	className="min-w-0 max-w-[240px] truncate transition-[max-width,opacity] duration-300 ease-out group-data-[compact]/composer-context:max-w-0 group-data-[compact]/composer-context:opacity-0"
	className="min-w-0 max-w-[240px] truncate transition-[max-width,opacity] duration-300 ease-out group-data-[compact]/composer-context:max-w-0 group-data-[compact]/composer-context:opacity-0 motion-reduce:transition-none"

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

These row classes override Button's core contract at the call site — height (h-9 sm:h-9), radius, padding, and base/hover/pressed background, with ! needed to beat the ghost variant's own hover colors and raw black/white alpha instead of semantic tokens. Since this treatment is now shared by every panel row (select rows, link rows, split halves, icon actions), a named Button size/variant (e.g. size="panel-row" plus a muted-ghost variant) would express it in the primitive and drop the !important fights and the sm: re-statements; contextual width/color could stay here.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

Macroscope skipped reviewing this pull request. Per-review cost limit exceeded (workspace setting).

This review would cost an estimated $250.50, which exceeds your per-review limit of $10.00.

The top 3 files driving up this estimate:

File	Diff Size	Estimate
apps/server/src/orchestration-v2/Orchestrator.ts	280.33KB	$14.02
apps/server/src/orchestration-v2/Adapters/AcpAdapterV2.ts	272.17KB	$13.61
apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts	238.01KB	$11.90

Tip

To get this pull request reviewed, you can:

1. Comment @macroscope-app on this PR to request a manual review (monthly spend limits still apply).
2. Exclude the file(s) above from review by adding a pattern to your .macroscope/ignore.md — note that creating this file replaces Macroscope's built-in default ignores rather than extending them.
3. Raise your cost limit in your workspace billing settings.

^{Turn off this reminder going forward}

[mwolson]

What Changed

You can steer a turn while the agent is still working. The run could settle anyway, and the later answer vanished. That later text now stays on the thread.

A provider that never started could leave the thread looking alive forever. Exhausted retries now fail the run.

You settle a thread, and Home still treats it like live work. Missing snapshots keep retrying, archived threads drift back, and Waiting keeps refreshing stale rows. Settled now stays settled. A gone thread stops retrying.

You dismiss a thread error, and a later one never appears, or the old one never leaves. Dismissal lasts until a later error. Plan mode off still left the Build/Plan toggle in the composer after the v2 chat rewrite. The toggle is hidden in that case.

Why

Settle meant the useful work was over in the UI while the agent, the outbox, or a later error was still in play. Home, Waiting, and the composer should follow that same rule.

Julius's bounded thread history on current CTM stays as shipped.

UI Changes

Steered output stays in the thread. Failed starts no longer hang as if they were still starting. Home and Waiting settled state is quieter. Thread error banners dismiss until a later error. Build/Plan is hidden when plan mode is disabled.

No before/after screenshots.

Checklist

• This PR is small and focused
• I explained what changed and why
• I included before/after screenshots for any UI changes
• I included a video for animation/interaction changes

---

Note

High Risk
Changes core orchestration paths—settlement merging, command commit guards, Claude turn lifecycle, and process-loss reconciliation—with race-sensitive behavior that can leave runs stuck or drop recovery if callers mishandle committed: false.

Overview
Threads gain settledOverrideAt so settle/keep-active pins have a stable establishment time that survives renames and is exposed on shells and mobile projections.

Settlement is tightened end-to-end: settle is blocked on queued runs, pending runtime requests, and provider background work; archived threads cannot settle/unsettle; pin/unsettle paths maintain the new timestamp; message dispatch clears overrides via a shared helper. Settled/unsettled events are merged through reduceThreadSettlementEvent in memory and in SQL so stale full-thread payloads cannot overwrite title or other metadata. Provider ingest tests document when pending thread-scoped requests should emit thread.unsettled without waking keep-active or non-activity updates.

Claude steering tracks per-turn steer counts and frame counts, detects steering handoff results before terminalizing, marks in-flight tools interrupted, and bounds silent handoffs with a timeout that fails the turn if no new frames arrive. Turn finalization is claimed up front to avoid duplicate terminals when the handoff fiber races the frame handler; new turns wait while a turn is finalizing.

Recovery adds optional updatedBefore filters on outbox cancel/reconcile and expectedThreadUpdatedAt on commitCommand so stale recovery writes no-op when the thread has already advanced.

^{Reviewed by Cursor Bugbot for commit 57fca87. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add Orchestration V2 system with provider adapters, persistence, and V2 UI migration

• Introduces the complete Orchestration V2 architecture: contracts, event store, projection store, command receipts, effect outbox/worker, runtime recovery, and database migrations 044–052
• Adds V2 provider adapters for Claude, Codex, Cursor, Grok, OpenCode, and ACP Registry through the ProviderAdapterV2 service contract and driver registry, with replay testkits for deterministic testing
• Implements thread lifecycle services for launch, fork, settlement, checkpoint capture/rollback, context handoff, continuation, run finalization, and provider-session management with idle timeout and shutdown
• Migrates client-runtime, web, and mobile state from legacy orchestration types to V2 thread projections, bounded shell snapshots, runtime summaries, and progressive history paging
• Adds MCP orchestrator and worktree toolkits, scheduled-task service with interval and fixed-time scheduling, T3 orchestration instructions, and a legacy V1 thread importer with idempotent shell and transcript hydration
• Risk: migrations 044–052 create new schema tables and copy existing V2/project data into a shared application event log; EnvironmentThreadState in threadState.ts replaces legacy thread/page fields with OrchestrationV2ThreadProjection and ThreadHistoryMeta, removing threadHasOlderTurns; RPC authorization scopes in RpcAuthorization.ts are remapped to V2 method constants

^{Macroscope summarized 674b0e4.}

[coderabbitai]

Important

Review skipped

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

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 7726a859-eb6b-4efc-9e83-8c6c368e6e50

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

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

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

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[macroscopeapp]

One convention issue found in the new provider-ingest settlement code: a statically known tagged failure is recovered with Effect.catchTag instead of Effect.catchTags.

^{Posted via Macroscope — Effect Service Conventions}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This is a 964-file Orchestration V2 architecture and persistence migration that adds provider integrations, new workflows, protocol/schema changes, and product-default changes across server, web, and mobile. Unresolved projection and provider-lifecycle race findings add material consistency risk beyond the scope of an auto-approvable fix.

Not approved because:

• 3 blocking correctness issues found at or above your repo's Minimum Blocking Severity

^{Adjust the Minimum Blocking Severity for this repo — including turning it Off — in Settings. You can add or adjust custom eligibility rules. Learn more.}

[macroscopeapp]

Effect service conventions review of the settle-lifecycle changes. The previously flagged Effect.catchTag usage in ProviderEventIngestor.ts is now Effect.catchTags({ ... }), and the new/changed service code (ProjectionStore, EventSink, EffectOutbox, ProviderRuntimeRecoveryService, ProviderEventIngestor, shared settlement helpers) follows the import, layer, and dependency-acquisition rules. Two small change-discipline findings in packages/client-runtime/src/state/threadSettled.ts.

^{Posted via Macroscope — Effect Service Conventions}

[macroscopeapp]

🟡 Mediumorchestration-v2/ProjectionStore.ts:1261

thread.unsettled events lose event.payload.unsettledAt in the durable projection, so shell snapshots report unsettledAt as the previous value (often null) instead of when the thread was explicitly kept active. Include unsettledAt in the settlement merge.

 settlement: {
+ unsettledAt: event.payload.unsettledAt ?? null,

🤖 Copy this AI Prompt to have your agent fix this:

In file @apps/server/src/orchestration-v2/ProjectionStore.ts around line 1261:
`thread.unsettled` events lose `event.payload.unsettledAt` in the durable projection, so shell snapshots report `unsettledAt` as the previous value (often `null`) instead of when the thread was explicitly kept active. Include `unsettledAt` in the settlement merge.

[macroscopeapp]

🟡 Mediumorchestration-v2/ProviderEventIngestor.ts:356

Archived threads receive a thread.unsettled event for pending non-auth runtime-request.updated activity, clearing their settledOverride instead of leaving archived settlement state unchanged. getThreadProjection returns the still-present archived projection, and this path does not reject it before appending the synthetic event; restrict candidate generation to active threads.

🤖 Copy this AI Prompt to have your agent fix this:

In file @apps/server/src/orchestration-v2/ProviderEventIngestor.ts around line 356:
Archived threads receive a `thread.unsettled` event for pending non-auth `runtime-request.updated` activity, clearing their `settledOverride` instead of leaving archived settlement state unchanged. `getThreadProjection` returns the still-present archived projection, and this path does not reject it before appending the synthetic event; restrict candidate generation to active threads.

[macroscopeapp]

🟡 Mediumstate/orchestrationV2Projection.ts:207

A delayed thread.unsettled rewinds projection.updatedAt to event.occurredAt, even when the projection already has a newer timestamp from a concurrent metadata update. Although reduceThreadSettlementEvent preserves the newer thread.updatedAt, spreading base here still makes client ordering/activity views treat the thread as stale; preserve the later of the existing projection timestamp and event.occurredAt, as the server reducer does.

🤖 Copy this AI Prompt to have your agent fix this:

In file @packages/client-runtime/src/state/orchestrationV2Projection.ts around line 207:
A delayed `thread.unsettled` rewinds `projection.updatedAt` to `event.occurredAt`, even when the projection already has a newer timestamp from a concurrent metadata update. Although `reduceThreadSettlementEvent` preserves the newer `thread.updatedAt`, spreading `base` here still makes client ordering/activity views treat the thread as stale; preserve the later of the existing projection timestamp and `event.occurredAt`, as the server reducer does.

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 57fca87. Configure here.}

[cursor]

Finalize claim races new turns

Medium Severity

finalizeActiveTurn nulls activeTurn and only then adds the id to finalizingTurnIds. startTurn checks those refs separately, so a concurrent start (the handoff timeout is forked) can observe both empty and open a second provider turn while the first is still emitting terminals.

Additional Locations (1)

• apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts#L5197-L5211

^{Reviewed by Cursor Bugbot for commit 57fca87. Configure here.}

[macroscopeapp]

Scope note: only apps/web/src/** TS/TSX/CSS changes were reviewed. Findings are the newly added global CSS in index.css (dead/duplicated rules) and the composer context strip's motion contract, which changed lines break while the code that depends on it is retained.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

The strip's compact/expand FLIP animation still keys on [data-composer-context-control] (BranchToolbar.tsx:232, consumed by useLabelsOverflow), but this PR removes that attribute from the environment selector, the env-mode selector (locked and select paths) and the branch selector. The only remaining carrier is the vertical Separator at BranchToolbar.tsx:545, so collapsing now animates a 1px separator while the actual controls jump to their new positions. Consider keeping the attribute on the toolbar-mode triggers (it is inert in panel mode), or removing the FLIP pass as well if the motion is intentionally gone.

Suggested change

	aria-label="Run on"
	aria-label="Run on"
	data-composer-context-control

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

This whole .glass-opacity-slider block (through the forced-colors rules) is unreferenced — the glass-opacity input in SettingsPanels.tsx:1111 uses settings-slider with --settings-slider-progress, and no literal or composed glass-opacity-slider class exists anywhere in the repo. It is also a verbatim copy of .settings-slider below with only the custom-property names changed, so it adds a second owner for the same slider treatment. Suggest dropping it (or, if a distinct slider is intended, wiring a call site and deriving it from .settings-slider rather than duplicating the vendor pseudo-element rules).

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

alert-glass is already defined as @utility alert-glass (line 284) and is what ProviderStatusBanner consumes. This second definition in @layer components duplicates it and omits the utility's @supports not ((-webkit-backdrop-filter: blur(1px)) or (backdrop-filter: blur(1px))) opaque fallback; because !important declarations resolve in reverse layer order, this components-layer background: ... !important outranks the utility's important fallback, so browsers without backdrop-filter get a translucent, unblurred banner instead of the solid background. Suggest deleting this block and keeping the @utility as the single owner.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

.surface-subheader has no consumer: all four subheaders (DiffPanelShell, FileBrowserPanel, FilePreviewPanel, PreviewChromeRow) still carry the same geometry inline and only set data-surface-subheader. The inline-mode rule below is also redundant — every carrier already has in-data-[preview-panel-mode=inline]:h-7/min-h-7/mb-3/border-b-transparent, and those utilities sort after @layer components, so they win wherever this rule matches. Either apply surface-subheader at those call sites and drop the duplicated utility strings, or drop both rules so the shared geometry keeps one owner.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

The label previously animated transform/opacity on an inner data-composer-label-motion span with motion-reduce:transform-none motion-reduce:transition-opacity; this replaces it with a max-width transition and no reduced-motion guard, so reduced-motion users now get an animated layout change on every collapse. Same class appears at line 79 and in BranchToolbarEnvModeSelector/BranchToolbarBranchSelector.

Suggested change

	className="min-w-0 max-w-[240px] truncate transition-[max-width,opacity] duration-300 ease-out group-data-[compact]/composer-context:max-w-0 group-data-[compact]/composer-context:opacity-0"
	className="min-w-0 max-w-[240px] truncate transition-[max-width,opacity] duration-300 ease-out group-data-[compact]/composer-context:max-w-0 group-data-[compact]/composer-context:opacity-0 motion-reduce:transition-none"

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

These row classes override Button's core contract at the call site — height (h-9 sm:h-9), radius, padding, and base/hover/pressed background, with ! needed to beat the ghost variant's own hover colors and raw black/white alpha instead of semantic tokens. Since this treatment is now shared by every panel row (select rows, link rows, split halves, icon actions), a named Button size/variant (e.g. size="panel-row" plus a muted-ghost variant) would express it in the primitive and drop the !important fights and the sm: re-statements; contextual width/color could stay here.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

Macroscope skipped reviewing this pull request. Per-review cost limit exceeded (workspace setting).

This review would cost an estimated $250.50, which exceeds your per-review limit of $10.00.

The top 3 files driving up this estimate:

File	Diff Size	Estimate
apps/server/src/orchestration-v2/Orchestrator.ts	280.33KB	$14.02
apps/server/src/orchestration-v2/Adapters/AcpAdapterV2.ts	272.17KB	$13.61
apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts	238.01KB	$11.90

Tip

To get this pull request reviewed, you can:

1. Comment @macroscope-app on this PR to request a manual review (monthly spend limits still apply).
2. Exclude the file(s) above from review by adding a pattern to your .macroscope/ignore.md — note that creating this file replaces Macroscope's built-in default ignores rather than extending them.
3. Raise your cost limit in your workspace billing settings.

^{Turn off this reminder going forward}

[mwolson]

What Changed

You can steer a turn while the agent is still working. The run could settle anyway, and the later answer vanished. That later text now stays on the thread.

A provider that never started could leave the thread looking alive forever. Exhausted retries now fail the run.

You settle a thread, and Home still treats it like live work. Missing snapshots keep retrying, archived threads drift back, and Waiting keeps refreshing stale rows. Settled now stays settled. A gone thread stops retrying.

You dismiss a thread error, and a later one never appears, or the old one never leaves. Dismissal lasts until a later error. Plan mode off still left the Build/Plan toggle in the composer after the v2 chat rewrite. The toggle is hidden in that case.

Why

Settle meant the useful work was over in the UI while the agent, the outbox, or a later error was still in play. Home, Waiting, and the composer should follow that same rule.

Julius's bounded thread history on current CTM stays as shipped.

UI Changes

Steered output stays in the thread. Failed starts no longer hang as if they were still starting. Home and Waiting settled state is quieter. Thread error banners dismiss until a later error. Build/Plan is hidden when plan mode is disabled.

No before/after screenshots.

Checklist

• This PR is small and focused
• I explained what changed and why
• I included before/after screenshots for any UI changes
• I included a video for animation/interaction changes

---

Note

High Risk
Changes core orchestration paths—settlement merging, command commit guards, Claude turn lifecycle, and process-loss reconciliation—with race-sensitive behavior that can leave runs stuck or drop recovery if callers mishandle committed: false.

Overview
Threads gain settledOverrideAt so settle/keep-active pins have a stable establishment time that survives renames and is exposed on shells and mobile projections.

Settlement is tightened end-to-end: settle is blocked on queued runs, pending runtime requests, and provider background work; archived threads cannot settle/unsettle; pin/unsettle paths maintain the new timestamp; message dispatch clears overrides via a shared helper. Settled/unsettled events are merged through reduceThreadSettlementEvent in memory and in SQL so stale full-thread payloads cannot overwrite title or other metadata. Provider ingest tests document when pending thread-scoped requests should emit thread.unsettled without waking keep-active or non-activity updates.

Claude steering tracks per-turn steer counts and frame counts, detects steering handoff results before terminalizing, marks in-flight tools interrupted, and bounds silent handoffs with a timeout that fails the turn if no new frames arrive. Turn finalization is claimed up front to avoid duplicate terminals when the handoff fiber races the frame handler; new turns wait while a turn is finalizing.

Recovery adds optional updatedBefore filters on outbox cancel/reconcile and expectedThreadUpdatedAt on commitCommand so stale recovery writes no-op when the thread has already advanced.

^{Reviewed by Cursor Bugbot for commit 57fca87. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add Orchestration V2 system with provider adapters, persistence, and V2 UI migration

• Introduces the complete Orchestration V2 architecture: contracts, event store, projection store, command receipts, effect outbox/worker, runtime recovery, and database migrations 044–052
• Adds V2 provider adapters for Claude, Codex, Cursor, Grok, OpenCode, and ACP Registry through the ProviderAdapterV2 service contract and driver registry, with replay testkits for deterministic testing
• Implements thread lifecycle services for launch, fork, settlement, checkpoint capture/rollback, context handoff, continuation, run finalization, and provider-session management with idle timeout and shutdown
• Migrates client-runtime, web, and mobile state from legacy orchestration types to V2 thread projections, bounded shell snapshots, runtime summaries, and progressive history paging
• Adds MCP orchestrator and worktree toolkits, scheduled-task service with interval and fixed-time scheduling, T3 orchestration instructions, and a legacy V1 thread importer with idempotent shell and transcript hydration
• Risk: migrations 044–052 create new schema tables and copy existing V2/project data into a shared application event log; EnvironmentThreadState in threadState.ts replaces legacy thread/page fields with OrchestrationV2ThreadProjection and ThreadHistoryMeta, removing threadHasOlderTurns; RPC authorization scopes in RpcAuthorization.ts are remapped to V2 method constants

^{Macroscope summarized 674b0e4.}

[coderabbitai]

Important

Review skipped

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

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 7726a859-eb6b-4efc-9e83-8c6c368e6e50

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

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

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

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[macroscopeapp]

One convention issue found in the new provider-ingest settlement code: a statically known tagged failure is recovered with Effect.catchTag instead of Effect.catchTags.

^{Posted via Macroscope — Effect Service Conventions}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This is a 964-file Orchestration V2 architecture and persistence migration that adds provider integrations, new workflows, protocol/schema changes, and product-default changes across server, web, and mobile. Unresolved projection and provider-lifecycle race findings add material consistency risk beyond the scope of an auto-approvable fix.

Not approved because:

• 3 blocking correctness issues found at or above your repo's Minimum Blocking Severity

^{Adjust the Minimum Blocking Severity for this repo — including turning it Off — in Settings. You can add or adjust custom eligibility rules. Learn more.}

[macroscopeapp]

Effect service conventions review of the settle-lifecycle changes. The previously flagged Effect.catchTag usage in ProviderEventIngestor.ts is now Effect.catchTags({ ... }), and the new/changed service code (ProjectionStore, EventSink, EffectOutbox, ProviderRuntimeRecoveryService, ProviderEventIngestor, shared settlement helpers) follows the import, layer, and dependency-acquisition rules. Two small change-discipline findings in packages/client-runtime/src/state/threadSettled.ts.

^{Posted via Macroscope — Effect Service Conventions}

[macroscopeapp]

🟡 Mediumorchestration-v2/ProjectionStore.ts:1261

thread.unsettled events lose event.payload.unsettledAt in the durable projection, so shell snapshots report unsettledAt as the previous value (often null) instead of when the thread was explicitly kept active. Include unsettledAt in the settlement merge.

 settlement: {
+ unsettledAt: event.payload.unsettledAt ?? null,

🤖 Copy this AI Prompt to have your agent fix this:

In file @apps/server/src/orchestration-v2/ProjectionStore.ts around line 1261:
`thread.unsettled` events lose `event.payload.unsettledAt` in the durable projection, so shell snapshots report `unsettledAt` as the previous value (often `null`) instead of when the thread was explicitly kept active. Include `unsettledAt` in the settlement merge.

[macroscopeapp]

🟡 Mediumorchestration-v2/ProviderEventIngestor.ts:356

Archived threads receive a thread.unsettled event for pending non-auth runtime-request.updated activity, clearing their settledOverride instead of leaving archived settlement state unchanged. getThreadProjection returns the still-present archived projection, and this path does not reject it before appending the synthetic event; restrict candidate generation to active threads.

🤖 Copy this AI Prompt to have your agent fix this:

In file @apps/server/src/orchestration-v2/ProviderEventIngestor.ts around line 356:
Archived threads receive a `thread.unsettled` event for pending non-auth `runtime-request.updated` activity, clearing their `settledOverride` instead of leaving archived settlement state unchanged. `getThreadProjection` returns the still-present archived projection, and this path does not reject it before appending the synthetic event; restrict candidate generation to active threads.

[macroscopeapp]

🟡 Mediumstate/orchestrationV2Projection.ts:207

A delayed thread.unsettled rewinds projection.updatedAt to event.occurredAt, even when the projection already has a newer timestamp from a concurrent metadata update. Although reduceThreadSettlementEvent preserves the newer thread.updatedAt, spreading base here still makes client ordering/activity views treat the thread as stale; preserve the later of the existing projection timestamp and event.occurredAt, as the server reducer does.

🤖 Copy this AI Prompt to have your agent fix this:

In file @packages/client-runtime/src/state/orchestrationV2Projection.ts around line 207:
A delayed `thread.unsettled` rewinds `projection.updatedAt` to `event.occurredAt`, even when the projection already has a newer timestamp from a concurrent metadata update. Although `reduceThreadSettlementEvent` preserves the newer `thread.updatedAt`, spreading `base` here still makes client ordering/activity views treat the thread as stale; preserve the later of the existing projection timestamp and `event.occurredAt`, as the server reducer does.

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 57fca87. Configure here.}

[cursor]

Finalize claim races new turns

Medium Severity

finalizeActiveTurn nulls activeTurn and only then adds the id to finalizingTurnIds. startTurn checks those refs separately, so a concurrent start (the handoff timeout is forked) can observe both empty and open a second provider turn while the first is still emitting terminals.

Additional Locations (1)

• apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts#L5197-L5211

^{Reviewed by Cursor Bugbot for commit 57fca87. Configure here.}

[macroscopeapp]

Scope note: only apps/web/src/** TS/TSX/CSS changes were reviewed. Findings are the newly added global CSS in index.css (dead/duplicated rules) and the composer context strip's motion contract, which changed lines break while the code that depends on it is retained.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

The strip's compact/expand FLIP animation still keys on [data-composer-context-control] (BranchToolbar.tsx:232, consumed by useLabelsOverflow), but this PR removes that attribute from the environment selector, the env-mode selector (locked and select paths) and the branch selector. The only remaining carrier is the vertical Separator at BranchToolbar.tsx:545, so collapsing now animates a 1px separator while the actual controls jump to their new positions. Consider keeping the attribute on the toolbar-mode triggers (it is inert in panel mode), or removing the FLIP pass as well if the motion is intentionally gone.

Suggested change

	aria-label="Run on"
	aria-label="Run on"
	data-composer-context-control

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

This whole .glass-opacity-slider block (through the forced-colors rules) is unreferenced — the glass-opacity input in SettingsPanels.tsx:1111 uses settings-slider with --settings-slider-progress, and no literal or composed glass-opacity-slider class exists anywhere in the repo. It is also a verbatim copy of .settings-slider below with only the custom-property names changed, so it adds a second owner for the same slider treatment. Suggest dropping it (or, if a distinct slider is intended, wiring a call site and deriving it from .settings-slider rather than duplicating the vendor pseudo-element rules).

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

alert-glass is already defined as @utility alert-glass (line 284) and is what ProviderStatusBanner consumes. This second definition in @layer components duplicates it and omits the utility's @supports not ((-webkit-backdrop-filter: blur(1px)) or (backdrop-filter: blur(1px))) opaque fallback; because !important declarations resolve in reverse layer order, this components-layer background: ... !important outranks the utility's important fallback, so browsers without backdrop-filter get a translucent, unblurred banner instead of the solid background. Suggest deleting this block and keeping the @utility as the single owner.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

.surface-subheader has no consumer: all four subheaders (DiffPanelShell, FileBrowserPanel, FilePreviewPanel, PreviewChromeRow) still carry the same geometry inline and only set data-surface-subheader. The inline-mode rule below is also redundant — every carrier already has in-data-[preview-panel-mode=inline]:h-7/min-h-7/mb-3/border-b-transparent, and those utilities sort after @layer components, so they win wherever this rule matches. Either apply surface-subheader at those call sites and drop the duplicated utility strings, or drop both rules so the shared geometry keeps one owner.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

The label previously animated transform/opacity on an inner data-composer-label-motion span with motion-reduce:transform-none motion-reduce:transition-opacity; this replaces it with a max-width transition and no reduced-motion guard, so reduced-motion users now get an animated layout change on every collapse. Same class appears at line 79 and in BranchToolbarEnvModeSelector/BranchToolbarBranchSelector.

Suggested change

	className="min-w-0 max-w-[240px] truncate transition-[max-width,opacity] duration-300 ease-out group-data-[compact]/composer-context:max-w-0 group-data-[compact]/composer-context:opacity-0"
	className="min-w-0 max-w-[240px] truncate transition-[max-width,opacity] duration-300 ease-out group-data-[compact]/composer-context:max-w-0 group-data-[compact]/composer-context:opacity-0 motion-reduce:transition-none"

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

These row classes override Button's core contract at the call site — height (h-9 sm:h-9), radius, padding, and base/hover/pressed background, with ! needed to beat the ghost variant's own hover colors and raw black/white alpha instead of semantic tokens. Since this treatment is now shared by every panel row (select rows, link rows, split halves, icon actions), a named Button size/variant (e.g. size="panel-row" plus a muted-ghost variant) would express it in the primitive and drop the !important fights and the sm: re-statements; contextual width/color could stay here.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

Macroscope skipped reviewing this pull request. Per-review cost limit exceeded (workspace setting).

This review would cost an estimated $250.50, which exceeds your per-review limit of $10.00.

The top 3 files driving up this estimate:

File	Diff Size	Estimate
apps/server/src/orchestration-v2/Orchestrator.ts	280.33KB	$14.02
apps/server/src/orchestration-v2/Adapters/AcpAdapterV2.ts	272.17KB	$13.61
apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts	238.01KB	$11.90

Tip

To get this pull request reviewed, you can:

1. Comment @macroscope-app on this PR to request a manual review (monthly spend limits still apply).
2. Exclude the file(s) above from review by adding a pattern to your .macroscope/ignore.md — note that creating this file replaces Macroscope's built-in default ignores rather than extending them.
3. Raise your cost limit in your workspace billing settings.

^{Turn off this reminder going forward}

[mwolson]

What Changed

You can steer a turn while the agent is still working. The run could settle anyway, and the later answer vanished. That later text now stays on the thread.

A provider that never started could leave the thread looking alive forever. Exhausted retries now fail the run.

You settle a thread, and Home still treats it like live work. Missing snapshots keep retrying, archived threads drift back, and Waiting keeps refreshing stale rows. Settled now stays settled. A gone thread stops retrying.

You dismiss a thread error, and a later one never appears, or the old one never leaves. Dismissal lasts until a later error. Plan mode off still left the Build/Plan toggle in the composer after the v2 chat rewrite. The toggle is hidden in that case.

Why

Settle meant the useful work was over in the UI while the agent, the outbox, or a later error was still in play. Home, Waiting, and the composer should follow that same rule.

Julius's bounded thread history on current CTM stays as shipped.

UI Changes

Steered output stays in the thread. Failed starts no longer hang as if they were still starting. Home and Waiting settled state is quieter. Thread error banners dismiss until a later error. Build/Plan is hidden when plan mode is disabled.

No before/after screenshots.

Checklist

• This PR is small and focused
• I explained what changed and why
• I included before/after screenshots for any UI changes
• I included a video for animation/interaction changes

---

Note

High Risk
Changes core orchestration paths—settlement merging, command commit guards, Claude turn lifecycle, and process-loss reconciliation—with race-sensitive behavior that can leave runs stuck or drop recovery if callers mishandle committed: false.

Overview
Threads gain settledOverrideAt so settle/keep-active pins have a stable establishment time that survives renames and is exposed on shells and mobile projections.

Settlement is tightened end-to-end: settle is blocked on queued runs, pending runtime requests, and provider background work; archived threads cannot settle/unsettle; pin/unsettle paths maintain the new timestamp; message dispatch clears overrides via a shared helper. Settled/unsettled events are merged through reduceThreadSettlementEvent in memory and in SQL so stale full-thread payloads cannot overwrite title or other metadata. Provider ingest tests document when pending thread-scoped requests should emit thread.unsettled without waking keep-active or non-activity updates.

Claude steering tracks per-turn steer counts and frame counts, detects steering handoff results before terminalizing, marks in-flight tools interrupted, and bounds silent handoffs with a timeout that fails the turn if no new frames arrive. Turn finalization is claimed up front to avoid duplicate terminals when the handoff fiber races the frame handler; new turns wait while a turn is finalizing.

Recovery adds optional updatedBefore filters on outbox cancel/reconcile and expectedThreadUpdatedAt on commitCommand so stale recovery writes no-op when the thread has already advanced.

^{Reviewed by Cursor Bugbot for commit 57fca87. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add Orchestration V2 system with provider adapters, persistence, and V2 UI migration

• Introduces the complete Orchestration V2 architecture: contracts, event store, projection store, command receipts, effect outbox/worker, runtime recovery, and database migrations 044–052
• Adds V2 provider adapters for Claude, Codex, Cursor, Grok, OpenCode, and ACP Registry through the ProviderAdapterV2 service contract and driver registry, with replay testkits for deterministic testing
• Implements thread lifecycle services for launch, fork, settlement, checkpoint capture/rollback, context handoff, continuation, run finalization, and provider-session management with idle timeout and shutdown
• Migrates client-runtime, web, and mobile state from legacy orchestration types to V2 thread projections, bounded shell snapshots, runtime summaries, and progressive history paging
• Adds MCP orchestrator and worktree toolkits, scheduled-task service with interval and fixed-time scheduling, T3 orchestration instructions, and a legacy V1 thread importer with idempotent shell and transcript hydration
• Risk: migrations 044–052 create new schema tables and copy existing V2/project data into a shared application event log; EnvironmentThreadState in threadState.ts replaces legacy thread/page fields with OrchestrationV2ThreadProjection and ThreadHistoryMeta, removing threadHasOlderTurns; RPC authorization scopes in RpcAuthorization.ts are remapped to V2 method constants

^{Macroscope summarized 674b0e4.}

[coderabbitai]

Important

Review skipped

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

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 7726a859-eb6b-4efc-9e83-8c6c368e6e50

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

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

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

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[macroscopeapp]

One convention issue found in the new provider-ingest settlement code: a statically known tagged failure is recovered with Effect.catchTag instead of Effect.catchTags.

^{Posted via Macroscope — Effect Service Conventions}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This is a 964-file Orchestration V2 architecture and persistence migration that adds provider integrations, new workflows, protocol/schema changes, and product-default changes across server, web, and mobile. Unresolved projection and provider-lifecycle race findings add material consistency risk beyond the scope of an auto-approvable fix.

Not approved because:

• 3 blocking correctness issues found at or above your repo's Minimum Blocking Severity

^{Adjust the Minimum Blocking Severity for this repo — including turning it Off — in Settings. You can add or adjust custom eligibility rules. Learn more.}

[macroscopeapp]

Effect service conventions review of the settle-lifecycle changes. The previously flagged Effect.catchTag usage in ProviderEventIngestor.ts is now Effect.catchTags({ ... }), and the new/changed service code (ProjectionStore, EventSink, EffectOutbox, ProviderRuntimeRecoveryService, ProviderEventIngestor, shared settlement helpers) follows the import, layer, and dependency-acquisition rules. Two small change-discipline findings in packages/client-runtime/src/state/threadSettled.ts.

^{Posted via Macroscope — Effect Service Conventions}

[macroscopeapp]

🟡 Mediumorchestration-v2/ProjectionStore.ts:1261

thread.unsettled events lose event.payload.unsettledAt in the durable projection, so shell snapshots report unsettledAt as the previous value (often null) instead of when the thread was explicitly kept active. Include unsettledAt in the settlement merge.

 settlement: {
+ unsettledAt: event.payload.unsettledAt ?? null,

🤖 Copy this AI Prompt to have your agent fix this:

In file @apps/server/src/orchestration-v2/ProjectionStore.ts around line 1261:
`thread.unsettled` events lose `event.payload.unsettledAt` in the durable projection, so shell snapshots report `unsettledAt` as the previous value (often `null`) instead of when the thread was explicitly kept active. Include `unsettledAt` in the settlement merge.

[macroscopeapp]

🟡 Mediumorchestration-v2/ProviderEventIngestor.ts:356

Archived threads receive a thread.unsettled event for pending non-auth runtime-request.updated activity, clearing their settledOverride instead of leaving archived settlement state unchanged. getThreadProjection returns the still-present archived projection, and this path does not reject it before appending the synthetic event; restrict candidate generation to active threads.

🤖 Copy this AI Prompt to have your agent fix this:

In file @apps/server/src/orchestration-v2/ProviderEventIngestor.ts around line 356:
Archived threads receive a `thread.unsettled` event for pending non-auth `runtime-request.updated` activity, clearing their `settledOverride` instead of leaving archived settlement state unchanged. `getThreadProjection` returns the still-present archived projection, and this path does not reject it before appending the synthetic event; restrict candidate generation to active threads.

[macroscopeapp]

🟡 Mediumstate/orchestrationV2Projection.ts:207

A delayed thread.unsettled rewinds projection.updatedAt to event.occurredAt, even when the projection already has a newer timestamp from a concurrent metadata update. Although reduceThreadSettlementEvent preserves the newer thread.updatedAt, spreading base here still makes client ordering/activity views treat the thread as stale; preserve the later of the existing projection timestamp and event.occurredAt, as the server reducer does.

🤖 Copy this AI Prompt to have your agent fix this:

In file @packages/client-runtime/src/state/orchestrationV2Projection.ts around line 207:
A delayed `thread.unsettled` rewinds `projection.updatedAt` to `event.occurredAt`, even when the projection already has a newer timestamp from a concurrent metadata update. Although `reduceThreadSettlementEvent` preserves the newer `thread.updatedAt`, spreading `base` here still makes client ordering/activity views treat the thread as stale; preserve the later of the existing projection timestamp and `event.occurredAt`, as the server reducer does.

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 57fca87. Configure here.}

[cursor]

Finalize claim races new turns

Medium Severity

finalizeActiveTurn nulls activeTurn and only then adds the id to finalizingTurnIds. startTurn checks those refs separately, so a concurrent start (the handoff timeout is forked) can observe both empty and open a second provider turn while the first is still emitting terminals.

Additional Locations (1)

• apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts#L5197-L5211

^{Reviewed by Cursor Bugbot for commit 57fca87. Configure here.}

[macroscopeapp]

Scope note: only apps/web/src/** TS/TSX/CSS changes were reviewed. Findings are the newly added global CSS in index.css (dead/duplicated rules) and the composer context strip's motion contract, which changed lines break while the code that depends on it is retained.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

The strip's compact/expand FLIP animation still keys on [data-composer-context-control] (BranchToolbar.tsx:232, consumed by useLabelsOverflow), but this PR removes that attribute from the environment selector, the env-mode selector (locked and select paths) and the branch selector. The only remaining carrier is the vertical Separator at BranchToolbar.tsx:545, so collapsing now animates a 1px separator while the actual controls jump to their new positions. Consider keeping the attribute on the toolbar-mode triggers (it is inert in panel mode), or removing the FLIP pass as well if the motion is intentionally gone.

Suggested change

	aria-label="Run on"
	aria-label="Run on"
	data-composer-context-control

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

This whole .glass-opacity-slider block (through the forced-colors rules) is unreferenced — the glass-opacity input in SettingsPanels.tsx:1111 uses settings-slider with --settings-slider-progress, and no literal or composed glass-opacity-slider class exists anywhere in the repo. It is also a verbatim copy of .settings-slider below with only the custom-property names changed, so it adds a second owner for the same slider treatment. Suggest dropping it (or, if a distinct slider is intended, wiring a call site and deriving it from .settings-slider rather than duplicating the vendor pseudo-element rules).

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

alert-glass is already defined as @utility alert-glass (line 284) and is what ProviderStatusBanner consumes. This second definition in @layer components duplicates it and omits the utility's @supports not ((-webkit-backdrop-filter: blur(1px)) or (backdrop-filter: blur(1px))) opaque fallback; because !important declarations resolve in reverse layer order, this components-layer background: ... !important outranks the utility's important fallback, so browsers without backdrop-filter get a translucent, unblurred banner instead of the solid background. Suggest deleting this block and keeping the @utility as the single owner.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

.surface-subheader has no consumer: all four subheaders (DiffPanelShell, FileBrowserPanel, FilePreviewPanel, PreviewChromeRow) still carry the same geometry inline and only set data-surface-subheader. The inline-mode rule below is also redundant — every carrier already has in-data-[preview-panel-mode=inline]:h-7/min-h-7/mb-3/border-b-transparent, and those utilities sort after @layer components, so they win wherever this rule matches. Either apply surface-subheader at those call sites and drop the duplicated utility strings, or drop both rules so the shared geometry keeps one owner.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

The label previously animated transform/opacity on an inner data-composer-label-motion span with motion-reduce:transform-none motion-reduce:transition-opacity; this replaces it with a max-width transition and no reduced-motion guard, so reduced-motion users now get an animated layout change on every collapse. Same class appears at line 79 and in BranchToolbarEnvModeSelector/BranchToolbarBranchSelector.

Suggested change

	className="min-w-0 max-w-[240px] truncate transition-[max-width,opacity] duration-300 ease-out group-data-[compact]/composer-context:max-w-0 group-data-[compact]/composer-context:opacity-0"
	className="min-w-0 max-w-[240px] truncate transition-[max-width,opacity] duration-300 ease-out group-data-[compact]/composer-context:max-w-0 group-data-[compact]/composer-context:opacity-0 motion-reduce:transition-none"

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

These row classes override Button's core contract at the call site — height (h-9 sm:h-9), radius, padding, and base/hover/pressed background, with ! needed to beat the ghost variant's own hover colors and raw black/white alpha instead of semantic tokens. Since this treatment is now shared by every panel row (select rows, link rows, split halves, icon actions), a named Button size/variant (e.g. size="panel-row" plus a muted-ghost variant) would express it in the primitive and drop the !important fights and the sm: re-statements; contextual width/color could stay here.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

Macroscope skipped reviewing this pull request. Per-review cost limit exceeded (workspace setting).

This review would cost an estimated $250.50, which exceeds your per-review limit of $10.00.

The top 3 files driving up this estimate:

File	Diff Size	Estimate
apps/server/src/orchestration-v2/Orchestrator.ts	280.33KB	$14.02
apps/server/src/orchestration-v2/Adapters/AcpAdapterV2.ts	272.17KB	$13.61
apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts	238.01KB	$11.90

Tip

To get this pull request reviewed, you can:

1. Comment @macroscope-app on this PR to request a manual review (monthly spend limits still apply).
2. Exclude the file(s) above from review by adding a pattern to your .macroscope/ignore.md — note that creating this file replaces Macroscope's built-in default ignores rather than extending them.
3. Raise your cost limit in your workspace billing settings.

^{Turn off this reminder going forward}

[mwolson]

What Changed

You can steer a turn while the agent is still working. The run could settle anyway, and the later answer vanished. That later text now stays on the thread.

A provider that never started could leave the thread looking alive forever. Exhausted retries now fail the run.

You settle a thread, and Home still treats it like live work. Missing snapshots keep retrying, archived threads drift back, and Waiting keeps refreshing stale rows. Settled now stays settled. A gone thread stops retrying.

You dismiss a thread error, and a later one never appears, or the old one never leaves. Dismissal lasts until a later error. Plan mode off still left the Build/Plan toggle in the composer after the v2 chat rewrite. The toggle is hidden in that case.

Why

Settle meant the useful work was over in the UI while the agent, the outbox, or a later error was still in play. Home, Waiting, and the composer should follow that same rule.

Julius's bounded thread history on current CTM stays as shipped.

UI Changes

Steered output stays in the thread. Failed starts no longer hang as if they were still starting. Home and Waiting settled state is quieter. Thread error banners dismiss until a later error. Build/Plan is hidden when plan mode is disabled.

No before/after screenshots.

Checklist

• This PR is small and focused
• I explained what changed and why
• I included before/after screenshots for any UI changes
• I included a video for animation/interaction changes

---

Note

High Risk
Changes core orchestration paths—settlement merging, command commit guards, Claude turn lifecycle, and process-loss reconciliation—with race-sensitive behavior that can leave runs stuck or drop recovery if callers mishandle committed: false.

Overview
Threads gain settledOverrideAt so settle/keep-active pins have a stable establishment time that survives renames and is exposed on shells and mobile projections.

Settlement is tightened end-to-end: settle is blocked on queued runs, pending runtime requests, and provider background work; archived threads cannot settle/unsettle; pin/unsettle paths maintain the new timestamp; message dispatch clears overrides via a shared helper. Settled/unsettled events are merged through reduceThreadSettlementEvent in memory and in SQL so stale full-thread payloads cannot overwrite title or other metadata. Provider ingest tests document when pending thread-scoped requests should emit thread.unsettled without waking keep-active or non-activity updates.

Claude steering tracks per-turn steer counts and frame counts, detects steering handoff results before terminalizing, marks in-flight tools interrupted, and bounds silent handoffs with a timeout that fails the turn if no new frames arrive. Turn finalization is claimed up front to avoid duplicate terminals when the handoff fiber races the frame handler; new turns wait while a turn is finalizing.

Recovery adds optional updatedBefore filters on outbox cancel/reconcile and expectedThreadUpdatedAt on commitCommand so stale recovery writes no-op when the thread has already advanced.

^{Reviewed by Cursor Bugbot for commit 57fca87. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add Orchestration V2 system with provider adapters, persistence, and V2 UI migration

• Introduces the complete Orchestration V2 architecture: contracts, event store, projection store, command receipts, effect outbox/worker, runtime recovery, and database migrations 044–052
• Adds V2 provider adapters for Claude, Codex, Cursor, Grok, OpenCode, and ACP Registry through the ProviderAdapterV2 service contract and driver registry, with replay testkits for deterministic testing
• Implements thread lifecycle services for launch, fork, settlement, checkpoint capture/rollback, context handoff, continuation, run finalization, and provider-session management with idle timeout and shutdown
• Migrates client-runtime, web, and mobile state from legacy orchestration types to V2 thread projections, bounded shell snapshots, runtime summaries, and progressive history paging
• Adds MCP orchestrator and worktree toolkits, scheduled-task service with interval and fixed-time scheduling, T3 orchestration instructions, and a legacy V1 thread importer with idempotent shell and transcript hydration
• Risk: migrations 044–052 create new schema tables and copy existing V2/project data into a shared application event log; EnvironmentThreadState in threadState.ts replaces legacy thread/page fields with OrchestrationV2ThreadProjection and ThreadHistoryMeta, removing threadHasOlderTurns; RPC authorization scopes in RpcAuthorization.ts are remapped to V2 method constants

^{Macroscope summarized 674b0e4.}

[coderabbitai]

Important

Review skipped

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

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 7726a859-eb6b-4efc-9e83-8c6c368e6e50

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

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

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

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[macroscopeapp]

One convention issue found in the new provider-ingest settlement code: a statically known tagged failure is recovered with Effect.catchTag instead of Effect.catchTags.

^{Posted via Macroscope — Effect Service Conventions}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This is a 964-file Orchestration V2 architecture and persistence migration that adds provider integrations, new workflows, protocol/schema changes, and product-default changes across server, web, and mobile. Unresolved projection and provider-lifecycle race findings add material consistency risk beyond the scope of an auto-approvable fix.

Not approved because:

• 3 blocking correctness issues found at or above your repo's Minimum Blocking Severity

^{Adjust the Minimum Blocking Severity for this repo — including turning it Off — in Settings. You can add or adjust custom eligibility rules. Learn more.}

[macroscopeapp]

Effect service conventions review of the settle-lifecycle changes. The previously flagged Effect.catchTag usage in ProviderEventIngestor.ts is now Effect.catchTags({ ... }), and the new/changed service code (ProjectionStore, EventSink, EffectOutbox, ProviderRuntimeRecoveryService, ProviderEventIngestor, shared settlement helpers) follows the import, layer, and dependency-acquisition rules. Two small change-discipline findings in packages/client-runtime/src/state/threadSettled.ts.

^{Posted via Macroscope — Effect Service Conventions}

[macroscopeapp]

🟡 Mediumorchestration-v2/ProjectionStore.ts:1261

thread.unsettled events lose event.payload.unsettledAt in the durable projection, so shell snapshots report unsettledAt as the previous value (often null) instead of when the thread was explicitly kept active. Include unsettledAt in the settlement merge.

 settlement: {
+ unsettledAt: event.payload.unsettledAt ?? null,

🤖 Copy this AI Prompt to have your agent fix this:

In file @apps/server/src/orchestration-v2/ProjectionStore.ts around line 1261:
`thread.unsettled` events lose `event.payload.unsettledAt` in the durable projection, so shell snapshots report `unsettledAt` as the previous value (often `null`) instead of when the thread was explicitly kept active. Include `unsettledAt` in the settlement merge.

[macroscopeapp]

🟡 Mediumorchestration-v2/ProviderEventIngestor.ts:356

Archived threads receive a thread.unsettled event for pending non-auth runtime-request.updated activity, clearing their settledOverride instead of leaving archived settlement state unchanged. getThreadProjection returns the still-present archived projection, and this path does not reject it before appending the synthetic event; restrict candidate generation to active threads.

🤖 Copy this AI Prompt to have your agent fix this:

In file @apps/server/src/orchestration-v2/ProviderEventIngestor.ts around line 356:
Archived threads receive a `thread.unsettled` event for pending non-auth `runtime-request.updated` activity, clearing their `settledOverride` instead of leaving archived settlement state unchanged. `getThreadProjection` returns the still-present archived projection, and this path does not reject it before appending the synthetic event; restrict candidate generation to active threads.

[macroscopeapp]

🟡 Mediumstate/orchestrationV2Projection.ts:207

A delayed thread.unsettled rewinds projection.updatedAt to event.occurredAt, even when the projection already has a newer timestamp from a concurrent metadata update. Although reduceThreadSettlementEvent preserves the newer thread.updatedAt, spreading base here still makes client ordering/activity views treat the thread as stale; preserve the later of the existing projection timestamp and event.occurredAt, as the server reducer does.

🤖 Copy this AI Prompt to have your agent fix this:

In file @packages/client-runtime/src/state/orchestrationV2Projection.ts around line 207:
A delayed `thread.unsettled` rewinds `projection.updatedAt` to `event.occurredAt`, even when the projection already has a newer timestamp from a concurrent metadata update. Although `reduceThreadSettlementEvent` preserves the newer `thread.updatedAt`, spreading `base` here still makes client ordering/activity views treat the thread as stale; preserve the later of the existing projection timestamp and `event.occurredAt`, as the server reducer does.

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 57fca87. Configure here.}

[cursor]

Finalize claim races new turns

Medium Severity

finalizeActiveTurn nulls activeTurn and only then adds the id to finalizingTurnIds. startTurn checks those refs separately, so a concurrent start (the handoff timeout is forked) can observe both empty and open a second provider turn while the first is still emitting terminals.

Additional Locations (1)

• apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts#L5197-L5211

^{Reviewed by Cursor Bugbot for commit 57fca87. Configure here.}

[macroscopeapp]

Scope note: only apps/web/src/** TS/TSX/CSS changes were reviewed. Findings are the newly added global CSS in index.css (dead/duplicated rules) and the composer context strip's motion contract, which changed lines break while the code that depends on it is retained.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

The strip's compact/expand FLIP animation still keys on [data-composer-context-control] (BranchToolbar.tsx:232, consumed by useLabelsOverflow), but this PR removes that attribute from the environment selector, the env-mode selector (locked and select paths) and the branch selector. The only remaining carrier is the vertical Separator at BranchToolbar.tsx:545, so collapsing now animates a 1px separator while the actual controls jump to their new positions. Consider keeping the attribute on the toolbar-mode triggers (it is inert in panel mode), or removing the FLIP pass as well if the motion is intentionally gone.

Suggested change

	aria-label="Run on"
	aria-label="Run on"
	data-composer-context-control

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

This whole .glass-opacity-slider block (through the forced-colors rules) is unreferenced — the glass-opacity input in SettingsPanels.tsx:1111 uses settings-slider with --settings-slider-progress, and no literal or composed glass-opacity-slider class exists anywhere in the repo. It is also a verbatim copy of .settings-slider below with only the custom-property names changed, so it adds a second owner for the same slider treatment. Suggest dropping it (or, if a distinct slider is intended, wiring a call site and deriving it from .settings-slider rather than duplicating the vendor pseudo-element rules).

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

alert-glass is already defined as @utility alert-glass (line 284) and is what ProviderStatusBanner consumes. This second definition in @layer components duplicates it and omits the utility's @supports not ((-webkit-backdrop-filter: blur(1px)) or (backdrop-filter: blur(1px))) opaque fallback; because !important declarations resolve in reverse layer order, this components-layer background: ... !important outranks the utility's important fallback, so browsers without backdrop-filter get a translucent, unblurred banner instead of the solid background. Suggest deleting this block and keeping the @utility as the single owner.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

.surface-subheader has no consumer: all four subheaders (DiffPanelShell, FileBrowserPanel, FilePreviewPanel, PreviewChromeRow) still carry the same geometry inline and only set data-surface-subheader. The inline-mode rule below is also redundant — every carrier already has in-data-[preview-panel-mode=inline]:h-7/min-h-7/mb-3/border-b-transparent, and those utilities sort after @layer components, so they win wherever this rule matches. Either apply surface-subheader at those call sites and drop the duplicated utility strings, or drop both rules so the shared geometry keeps one owner.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

The label previously animated transform/opacity on an inner data-composer-label-motion span with motion-reduce:transform-none motion-reduce:transition-opacity; this replaces it with a max-width transition and no reduced-motion guard, so reduced-motion users now get an animated layout change on every collapse. Same class appears at line 79 and in BranchToolbarEnvModeSelector/BranchToolbarBranchSelector.

Suggested change

	className="min-w-0 max-w-[240px] truncate transition-[max-width,opacity] duration-300 ease-out group-data-[compact]/composer-context:max-w-0 group-data-[compact]/composer-context:opacity-0"
	className="min-w-0 max-w-[240px] truncate transition-[max-width,opacity] duration-300 ease-out group-data-[compact]/composer-context:max-w-0 group-data-[compact]/composer-context:opacity-0 motion-reduce:transition-none"

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

These row classes override Button's core contract at the call site — height (h-9 sm:h-9), radius, padding, and base/hover/pressed background, with ! needed to beat the ghost variant's own hover colors and raw black/white alpha instead of semantic tokens. Since this treatment is now shared by every panel row (select rows, link rows, split halves, icon actions), a named Button size/variant (e.g. size="panel-row" plus a muted-ghost variant) would express it in the primitive and drop the !important fights and the sm: re-statements; contextual width/color could stay here.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

Macroscope skipped reviewing this pull request. Per-review cost limit exceeded (workspace setting).

This review would cost an estimated $250.50, which exceeds your per-review limit of $10.00.

The top 3 files driving up this estimate:

File	Diff Size	Estimate
apps/server/src/orchestration-v2/Orchestrator.ts	280.33KB	$14.02
apps/server/src/orchestration-v2/Adapters/AcpAdapterV2.ts	272.17KB	$13.61
apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts	238.01KB	$11.90

Tip

To get this pull request reviewed, you can:

1. Comment @macroscope-app on this PR to request a manual review (monthly spend limits still apply).
2. Exclude the file(s) above from review by adding a pattern to your .macroscope/ignore.md — note that creating this file replaces Macroscope's built-in default ignores rather than extending them.
3. Raise your cost limit in your workspace billing settings.

^{Turn off this reminder going forward}

[mwolson]

What Changed

You can steer a turn while the agent is still working. The run could settle anyway, and the later answer vanished. That later text now stays on the thread.

A provider that never started could leave the thread looking alive forever. Exhausted retries now fail the run.

You settle a thread, and Home still treats it like live work. Missing snapshots keep retrying, archived threads drift back, and Waiting keeps refreshing stale rows. Settled now stays settled. A gone thread stops retrying.

You dismiss a thread error, and a later one never appears, or the old one never leaves. Dismissal lasts until a later error. Plan mode off still left the Build/Plan toggle in the composer after the v2 chat rewrite. The toggle is hidden in that case.

Why

Settle meant the useful work was over in the UI while the agent, the outbox, or a later error was still in play. Home, Waiting, and the composer should follow that same rule.

Julius's bounded thread history on current CTM stays as shipped.

UI Changes

Steered output stays in the thread. Failed starts no longer hang as if they were still starting. Home and Waiting settled state is quieter. Thread error banners dismiss until a later error. Build/Plan is hidden when plan mode is disabled.

No before/after screenshots.

Checklist

• This PR is small and focused
• I explained what changed and why
• I included before/after screenshots for any UI changes
• I included a video for animation/interaction changes

---

Note

High Risk
Changes core orchestration paths—settlement merging, command commit guards, Claude turn lifecycle, and process-loss reconciliation—with race-sensitive behavior that can leave runs stuck or drop recovery if callers mishandle committed: false.

Overview
Threads gain settledOverrideAt so settle/keep-active pins have a stable establishment time that survives renames and is exposed on shells and mobile projections.

Settlement is tightened end-to-end: settle is blocked on queued runs, pending runtime requests, and provider background work; archived threads cannot settle/unsettle; pin/unsettle paths maintain the new timestamp; message dispatch clears overrides via a shared helper. Settled/unsettled events are merged through reduceThreadSettlementEvent in memory and in SQL so stale full-thread payloads cannot overwrite title or other metadata. Provider ingest tests document when pending thread-scoped requests should emit thread.unsettled without waking keep-active or non-activity updates.

Claude steering tracks per-turn steer counts and frame counts, detects steering handoff results before terminalizing, marks in-flight tools interrupted, and bounds silent handoffs with a timeout that fails the turn if no new frames arrive. Turn finalization is claimed up front to avoid duplicate terminals when the handoff fiber races the frame handler; new turns wait while a turn is finalizing.

Recovery adds optional updatedBefore filters on outbox cancel/reconcile and expectedThreadUpdatedAt on commitCommand so stale recovery writes no-op when the thread has already advanced.

^{Reviewed by Cursor Bugbot for commit 57fca87. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add Orchestration V2 system with provider adapters, persistence, and V2 UI migration

• Introduces the complete Orchestration V2 architecture: contracts, event store, projection store, command receipts, effect outbox/worker, runtime recovery, and database migrations 044–052
• Adds V2 provider adapters for Claude, Codex, Cursor, Grok, OpenCode, and ACP Registry through the ProviderAdapterV2 service contract and driver registry, with replay testkits for deterministic testing
• Implements thread lifecycle services for launch, fork, settlement, checkpoint capture/rollback, context handoff, continuation, run finalization, and provider-session management with idle timeout and shutdown
• Migrates client-runtime, web, and mobile state from legacy orchestration types to V2 thread projections, bounded shell snapshots, runtime summaries, and progressive history paging
• Adds MCP orchestrator and worktree toolkits, scheduled-task service with interval and fixed-time scheduling, T3 orchestration instructions, and a legacy V1 thread importer with idempotent shell and transcript hydration
• Risk: migrations 044–052 create new schema tables and copy existing V2/project data into a shared application event log; EnvironmentThreadState in threadState.ts replaces legacy thread/page fields with OrchestrationV2ThreadProjection and ThreadHistoryMeta, removing threadHasOlderTurns; RPC authorization scopes in RpcAuthorization.ts are remapped to V2 method constants

^{Macroscope summarized 674b0e4.}

[coderabbitai]

Important

Review skipped

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

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 7726a859-eb6b-4efc-9e83-8c6c368e6e50

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

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

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

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[macroscopeapp]

One convention issue found in the new provider-ingest settlement code: a statically known tagged failure is recovered with Effect.catchTag instead of Effect.catchTags.

^{Posted via Macroscope — Effect Service Conventions}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This is a 964-file Orchestration V2 architecture and persistence migration that adds provider integrations, new workflows, protocol/schema changes, and product-default changes across server, web, and mobile. Unresolved projection and provider-lifecycle race findings add material consistency risk beyond the scope of an auto-approvable fix.

Not approved because:

• 3 blocking correctness issues found at or above your repo's Minimum Blocking Severity

^{Adjust the Minimum Blocking Severity for this repo — including turning it Off — in Settings. You can add or adjust custom eligibility rules. Learn more.}

[macroscopeapp]

Effect service conventions review of the settle-lifecycle changes. The previously flagged Effect.catchTag usage in ProviderEventIngestor.ts is now Effect.catchTags({ ... }), and the new/changed service code (ProjectionStore, EventSink, EffectOutbox, ProviderRuntimeRecoveryService, ProviderEventIngestor, shared settlement helpers) follows the import, layer, and dependency-acquisition rules. Two small change-discipline findings in packages/client-runtime/src/state/threadSettled.ts.

^{Posted via Macroscope — Effect Service Conventions}

[macroscopeapp]

🟡 Mediumorchestration-v2/ProjectionStore.ts:1261

thread.unsettled events lose event.payload.unsettledAt in the durable projection, so shell snapshots report unsettledAt as the previous value (often null) instead of when the thread was explicitly kept active. Include unsettledAt in the settlement merge.

 settlement: {
+ unsettledAt: event.payload.unsettledAt ?? null,

🤖 Copy this AI Prompt to have your agent fix this:

In file @apps/server/src/orchestration-v2/ProjectionStore.ts around line 1261:
`thread.unsettled` events lose `event.payload.unsettledAt` in the durable projection, so shell snapshots report `unsettledAt` as the previous value (often `null`) instead of when the thread was explicitly kept active. Include `unsettledAt` in the settlement merge.

[macroscopeapp]

🟡 Mediumorchestration-v2/ProviderEventIngestor.ts:356

Archived threads receive a thread.unsettled event for pending non-auth runtime-request.updated activity, clearing their settledOverride instead of leaving archived settlement state unchanged. getThreadProjection returns the still-present archived projection, and this path does not reject it before appending the synthetic event; restrict candidate generation to active threads.

🤖 Copy this AI Prompt to have your agent fix this:

In file @apps/server/src/orchestration-v2/ProviderEventIngestor.ts around line 356:
Archived threads receive a `thread.unsettled` event for pending non-auth `runtime-request.updated` activity, clearing their `settledOverride` instead of leaving archived settlement state unchanged. `getThreadProjection` returns the still-present archived projection, and this path does not reject it before appending the synthetic event; restrict candidate generation to active threads.

[macroscopeapp]

🟡 Mediumstate/orchestrationV2Projection.ts:207

A delayed thread.unsettled rewinds projection.updatedAt to event.occurredAt, even when the projection already has a newer timestamp from a concurrent metadata update. Although reduceThreadSettlementEvent preserves the newer thread.updatedAt, spreading base here still makes client ordering/activity views treat the thread as stale; preserve the later of the existing projection timestamp and event.occurredAt, as the server reducer does.

🤖 Copy this AI Prompt to have your agent fix this:

In file @packages/client-runtime/src/state/orchestrationV2Projection.ts around line 207:
A delayed `thread.unsettled` rewinds `projection.updatedAt` to `event.occurredAt`, even when the projection already has a newer timestamp from a concurrent metadata update. Although `reduceThreadSettlementEvent` preserves the newer `thread.updatedAt`, spreading `base` here still makes client ordering/activity views treat the thread as stale; preserve the later of the existing projection timestamp and `event.occurredAt`, as the server reducer does.

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 57fca87. Configure here.}

[cursor]

Finalize claim races new turns

Medium Severity

finalizeActiveTurn nulls activeTurn and only then adds the id to finalizingTurnIds. startTurn checks those refs separately, so a concurrent start (the handoff timeout is forked) can observe both empty and open a second provider turn while the first is still emitting terminals.

Additional Locations (1)

• apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts#L5197-L5211

^{Reviewed by Cursor Bugbot for commit 57fca87. Configure here.}

[macroscopeapp]

Scope note: only apps/web/src/** TS/TSX/CSS changes were reviewed. Findings are the newly added global CSS in index.css (dead/duplicated rules) and the composer context strip's motion contract, which changed lines break while the code that depends on it is retained.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

The strip's compact/expand FLIP animation still keys on [data-composer-context-control] (BranchToolbar.tsx:232, consumed by useLabelsOverflow), but this PR removes that attribute from the environment selector, the env-mode selector (locked and select paths) and the branch selector. The only remaining carrier is the vertical Separator at BranchToolbar.tsx:545, so collapsing now animates a 1px separator while the actual controls jump to their new positions. Consider keeping the attribute on the toolbar-mode triggers (it is inert in panel mode), or removing the FLIP pass as well if the motion is intentionally gone.

Suggested change

	aria-label="Run on"
	aria-label="Run on"
	data-composer-context-control

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

This whole .glass-opacity-slider block (through the forced-colors rules) is unreferenced — the glass-opacity input in SettingsPanels.tsx:1111 uses settings-slider with --settings-slider-progress, and no literal or composed glass-opacity-slider class exists anywhere in the repo. It is also a verbatim copy of .settings-slider below with only the custom-property names changed, so it adds a second owner for the same slider treatment. Suggest dropping it (or, if a distinct slider is intended, wiring a call site and deriving it from .settings-slider rather than duplicating the vendor pseudo-element rules).

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

alert-glass is already defined as @utility alert-glass (line 284) and is what ProviderStatusBanner consumes. This second definition in @layer components duplicates it and omits the utility's @supports not ((-webkit-backdrop-filter: blur(1px)) or (backdrop-filter: blur(1px))) opaque fallback; because !important declarations resolve in reverse layer order, this components-layer background: ... !important outranks the utility's important fallback, so browsers without backdrop-filter get a translucent, unblurred banner instead of the solid background. Suggest deleting this block and keeping the @utility as the single owner.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

.surface-subheader has no consumer: all four subheaders (DiffPanelShell, FileBrowserPanel, FilePreviewPanel, PreviewChromeRow) still carry the same geometry inline and only set data-surface-subheader. The inline-mode rule below is also redundant — every carrier already has in-data-[preview-panel-mode=inline]:h-7/min-h-7/mb-3/border-b-transparent, and those utilities sort after @layer components, so they win wherever this rule matches. Either apply surface-subheader at those call sites and drop the duplicated utility strings, or drop both rules so the shared geometry keeps one owner.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

The label previously animated transform/opacity on an inner data-composer-label-motion span with motion-reduce:transform-none motion-reduce:transition-opacity; this replaces it with a max-width transition and no reduced-motion guard, so reduced-motion users now get an animated layout change on every collapse. Same class appears at line 79 and in BranchToolbarEnvModeSelector/BranchToolbarBranchSelector.

Suggested change

	className="min-w-0 max-w-[240px] truncate transition-[max-width,opacity] duration-300 ease-out group-data-[compact]/composer-context:max-w-0 group-data-[compact]/composer-context:opacity-0"
	className="min-w-0 max-w-[240px] truncate transition-[max-width,opacity] duration-300 ease-out group-data-[compact]/composer-context:max-w-0 group-data-[compact]/composer-context:opacity-0 motion-reduce:transition-none"

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

These row classes override Button's core contract at the call site — height (h-9 sm:h-9), radius, padding, and base/hover/pressed background, with ! needed to beat the ghost variant's own hover colors and raw black/white alpha instead of semantic tokens. Since this treatment is now shared by every panel row (select rows, link rows, split halves, icon actions), a named Button size/variant (e.g. size="panel-row" plus a muted-ghost variant) would express it in the primitive and drop the !important fights and the sm: re-statements; contextual width/color could stay here.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

Macroscope skipped reviewing this pull request. Per-review cost limit exceeded (workspace setting).

This review would cost an estimated $250.50, which exceeds your per-review limit of $10.00.

The top 3 files driving up this estimate:

File	Diff Size	Estimate
apps/server/src/orchestration-v2/Orchestrator.ts	280.33KB	$14.02
apps/server/src/orchestration-v2/Adapters/AcpAdapterV2.ts	272.17KB	$13.61
apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts	238.01KB	$11.90

Tip

To get this pull request reviewed, you can:

1. Comment @macroscope-app on this PR to request a manual review (monthly spend limits still apply).
2. Exclude the file(s) above from review by adding a pattern to your .macroscope/ignore.md — note that creating this file replaces Macroscope's built-in default ignores rather than extending them.
3. Raise your cost limit in your workspace billing settings.

^{Turn off this reminder going forward}

[mwolson]

What Changed

You can steer a turn while the agent is still working. The run could settle anyway, and the later answer vanished. That later text now stays on the thread.

A provider that never started could leave the thread looking alive forever. Exhausted retries now fail the run.

You settle a thread, and Home still treats it like live work. Missing snapshots keep retrying, archived threads drift back, and Waiting keeps refreshing stale rows. Settled now stays settled. A gone thread stops retrying.

You dismiss a thread error, and a later one never appears, or the old one never leaves. Dismissal lasts until a later error. Plan mode off still left the Build/Plan toggle in the composer after the v2 chat rewrite. The toggle is hidden in that case.

Why

Settle meant the useful work was over in the UI while the agent, the outbox, or a later error was still in play. Home, Waiting, and the composer should follow that same rule.

Julius's bounded thread history on current CTM stays as shipped.

UI Changes

Steered output stays in the thread. Failed starts no longer hang as if they were still starting. Home and Waiting settled state is quieter. Thread error banners dismiss until a later error. Build/Plan is hidden when plan mode is disabled.

No before/after screenshots.

Checklist

• This PR is small and focused
• I explained what changed and why
• I included before/after screenshots for any UI changes
• I included a video for animation/interaction changes

---

Note

High Risk
Changes core orchestration paths—settlement merging, command commit guards, Claude turn lifecycle, and process-loss reconciliation—with race-sensitive behavior that can leave runs stuck or drop recovery if callers mishandle committed: false.

Overview
Threads gain settledOverrideAt so settle/keep-active pins have a stable establishment time that survives renames and is exposed on shells and mobile projections.

Settlement is tightened end-to-end: settle is blocked on queued runs, pending runtime requests, and provider background work; archived threads cannot settle/unsettle; pin/unsettle paths maintain the new timestamp; message dispatch clears overrides via a shared helper. Settled/unsettled events are merged through reduceThreadSettlementEvent in memory and in SQL so stale full-thread payloads cannot overwrite title or other metadata. Provider ingest tests document when pending thread-scoped requests should emit thread.unsettled without waking keep-active or non-activity updates.

Claude steering tracks per-turn steer counts and frame counts, detects steering handoff results before terminalizing, marks in-flight tools interrupted, and bounds silent handoffs with a timeout that fails the turn if no new frames arrive. Turn finalization is claimed up front to avoid duplicate terminals when the handoff fiber races the frame handler; new turns wait while a turn is finalizing.

Recovery adds optional updatedBefore filters on outbox cancel/reconcile and expectedThreadUpdatedAt on commitCommand so stale recovery writes no-op when the thread has already advanced.

^{Reviewed by Cursor Bugbot for commit 57fca87. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add Orchestration V2 system with provider adapters, persistence, and V2 UI migration

• Introduces the complete Orchestration V2 architecture: contracts, event store, projection store, command receipts, effect outbox/worker, runtime recovery, and database migrations 044–052
• Adds V2 provider adapters for Claude, Codex, Cursor, Grok, OpenCode, and ACP Registry through the ProviderAdapterV2 service contract and driver registry, with replay testkits for deterministic testing
• Implements thread lifecycle services for launch, fork, settlement, checkpoint capture/rollback, context handoff, continuation, run finalization, and provider-session management with idle timeout and shutdown
• Migrates client-runtime, web, and mobile state from legacy orchestration types to V2 thread projections, bounded shell snapshots, runtime summaries, and progressive history paging
• Adds MCP orchestrator and worktree toolkits, scheduled-task service with interval and fixed-time scheduling, T3 orchestration instructions, and a legacy V1 thread importer with idempotent shell and transcript hydration
• Risk: migrations 044–052 create new schema tables and copy existing V2/project data into a shared application event log; EnvironmentThreadState in threadState.ts replaces legacy thread/page fields with OrchestrationV2ThreadProjection and ThreadHistoryMeta, removing threadHasOlderTurns; RPC authorization scopes in RpcAuthorization.ts are remapped to V2 method constants

^{Macroscope summarized 674b0e4.}

[coderabbitai]

Important

Review skipped

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

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 7726a859-eb6b-4efc-9e83-8c6c368e6e50

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

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

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

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[macroscopeapp]

One convention issue found in the new provider-ingest settlement code: a statically known tagged failure is recovered with Effect.catchTag instead of Effect.catchTags.

^{Posted via Macroscope — Effect Service Conventions}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This is a 964-file Orchestration V2 architecture and persistence migration that adds provider integrations, new workflows, protocol/schema changes, and product-default changes across server, web, and mobile. Unresolved projection and provider-lifecycle race findings add material consistency risk beyond the scope of an auto-approvable fix.

Not approved because:

• 3 blocking correctness issues found at or above your repo's Minimum Blocking Severity

^{Adjust the Minimum Blocking Severity for this repo — including turning it Off — in Settings. You can add or adjust custom eligibility rules. Learn more.}

[macroscopeapp]

Effect service conventions review of the settle-lifecycle changes. The previously flagged Effect.catchTag usage in ProviderEventIngestor.ts is now Effect.catchTags({ ... }), and the new/changed service code (ProjectionStore, EventSink, EffectOutbox, ProviderRuntimeRecoveryService, ProviderEventIngestor, shared settlement helpers) follows the import, layer, and dependency-acquisition rules. Two small change-discipline findings in packages/client-runtime/src/state/threadSettled.ts.

^{Posted via Macroscope — Effect Service Conventions}

[macroscopeapp]

🟡 Mediumorchestration-v2/ProjectionStore.ts:1261

thread.unsettled events lose event.payload.unsettledAt in the durable projection, so shell snapshots report unsettledAt as the previous value (often null) instead of when the thread was explicitly kept active. Include unsettledAt in the settlement merge.

 settlement: {
+ unsettledAt: event.payload.unsettledAt ?? null,

🤖 Copy this AI Prompt to have your agent fix this:

In file @apps/server/src/orchestration-v2/ProjectionStore.ts around line 1261:
`thread.unsettled` events lose `event.payload.unsettledAt` in the durable projection, so shell snapshots report `unsettledAt` as the previous value (often `null`) instead of when the thread was explicitly kept active. Include `unsettledAt` in the settlement merge.

[macroscopeapp]

🟡 Mediumorchestration-v2/ProviderEventIngestor.ts:356

Archived threads receive a thread.unsettled event for pending non-auth runtime-request.updated activity, clearing their settledOverride instead of leaving archived settlement state unchanged. getThreadProjection returns the still-present archived projection, and this path does not reject it before appending the synthetic event; restrict candidate generation to active threads.

🤖 Copy this AI Prompt to have your agent fix this:

In file @apps/server/src/orchestration-v2/ProviderEventIngestor.ts around line 356:
Archived threads receive a `thread.unsettled` event for pending non-auth `runtime-request.updated` activity, clearing their `settledOverride` instead of leaving archived settlement state unchanged. `getThreadProjection` returns the still-present archived projection, and this path does not reject it before appending the synthetic event; restrict candidate generation to active threads.

[macroscopeapp]

🟡 Mediumstate/orchestrationV2Projection.ts:207

A delayed thread.unsettled rewinds projection.updatedAt to event.occurredAt, even when the projection already has a newer timestamp from a concurrent metadata update. Although reduceThreadSettlementEvent preserves the newer thread.updatedAt, spreading base here still makes client ordering/activity views treat the thread as stale; preserve the later of the existing projection timestamp and event.occurredAt, as the server reducer does.

🤖 Copy this AI Prompt to have your agent fix this:

In file @packages/client-runtime/src/state/orchestrationV2Projection.ts around line 207:
A delayed `thread.unsettled` rewinds `projection.updatedAt` to `event.occurredAt`, even when the projection already has a newer timestamp from a concurrent metadata update. Although `reduceThreadSettlementEvent` preserves the newer `thread.updatedAt`, spreading `base` here still makes client ordering/activity views treat the thread as stale; preserve the later of the existing projection timestamp and `event.occurredAt`, as the server reducer does.

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 57fca87. Configure here.}

[cursor]

Finalize claim races new turns

Medium Severity

finalizeActiveTurn nulls activeTurn and only then adds the id to finalizingTurnIds. startTurn checks those refs separately, so a concurrent start (the handoff timeout is forked) can observe both empty and open a second provider turn while the first is still emitting terminals.

Additional Locations (1)

• apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts#L5197-L5211

^{Reviewed by Cursor Bugbot for commit 57fca87. Configure here.}

[macroscopeapp]

Scope note: only apps/web/src/** TS/TSX/CSS changes were reviewed. Findings are the newly added global CSS in index.css (dead/duplicated rules) and the composer context strip's motion contract, which changed lines break while the code that depends on it is retained.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

The strip's compact/expand FLIP animation still keys on [data-composer-context-control] (BranchToolbar.tsx:232, consumed by useLabelsOverflow), but this PR removes that attribute from the environment selector, the env-mode selector (locked and select paths) and the branch selector. The only remaining carrier is the vertical Separator at BranchToolbar.tsx:545, so collapsing now animates a 1px separator while the actual controls jump to their new positions. Consider keeping the attribute on the toolbar-mode triggers (it is inert in panel mode), or removing the FLIP pass as well if the motion is intentionally gone.

Suggested change

	aria-label="Run on"
	aria-label="Run on"
	data-composer-context-control

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

This whole .glass-opacity-slider block (through the forced-colors rules) is unreferenced — the glass-opacity input in SettingsPanels.tsx:1111 uses settings-slider with --settings-slider-progress, and no literal or composed glass-opacity-slider class exists anywhere in the repo. It is also a verbatim copy of .settings-slider below with only the custom-property names changed, so it adds a second owner for the same slider treatment. Suggest dropping it (or, if a distinct slider is intended, wiring a call site and deriving it from .settings-slider rather than duplicating the vendor pseudo-element rules).

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

alert-glass is already defined as @utility alert-glass (line 284) and is what ProviderStatusBanner consumes. This second definition in @layer components duplicates it and omits the utility's @supports not ((-webkit-backdrop-filter: blur(1px)) or (backdrop-filter: blur(1px))) opaque fallback; because !important declarations resolve in reverse layer order, this components-layer background: ... !important outranks the utility's important fallback, so browsers without backdrop-filter get a translucent, unblurred banner instead of the solid background. Suggest deleting this block and keeping the @utility as the single owner.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

.surface-subheader has no consumer: all four subheaders (DiffPanelShell, FileBrowserPanel, FilePreviewPanel, PreviewChromeRow) still carry the same geometry inline and only set data-surface-subheader. The inline-mode rule below is also redundant — every carrier already has in-data-[preview-panel-mode=inline]:h-7/min-h-7/mb-3/border-b-transparent, and those utilities sort after @layer components, so they win wherever this rule matches. Either apply surface-subheader at those call sites and drop the duplicated utility strings, or drop both rules so the shared geometry keeps one owner.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

The label previously animated transform/opacity on an inner data-composer-label-motion span with motion-reduce:transform-none motion-reduce:transition-opacity; this replaces it with a max-width transition and no reduced-motion guard, so reduced-motion users now get an animated layout change on every collapse. Same class appears at line 79 and in BranchToolbarEnvModeSelector/BranchToolbarBranchSelector.

Suggested change

	className="min-w-0 max-w-[240px] truncate transition-[max-width,opacity] duration-300 ease-out group-data-[compact]/composer-context:max-w-0 group-data-[compact]/composer-context:opacity-0"
	className="min-w-0 max-w-[240px] truncate transition-[max-width,opacity] duration-300 ease-out group-data-[compact]/composer-context:max-w-0 group-data-[compact]/composer-context:opacity-0 motion-reduce:transition-none"

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

These row classes override Button's core contract at the call site — height (h-9 sm:h-9), radius, padding, and base/hover/pressed background, with ! needed to beat the ghost variant's own hover colors and raw black/white alpha instead of semantic tokens. Since this treatment is now shared by every panel row (select rows, link rows, split halves, icon actions), a named Button size/variant (e.g. size="panel-row" plus a muted-ghost variant) would express it in the primitive and drop the !important fights and the sm: re-statements; contextual width/color could stay here.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

Macroscope skipped reviewing this pull request. Per-review cost limit exceeded (workspace setting).

This review would cost an estimated $250.50, which exceeds your per-review limit of $10.00.

The top 3 files driving up this estimate:

File	Diff Size	Estimate
apps/server/src/orchestration-v2/Orchestrator.ts	280.33KB	$14.02
apps/server/src/orchestration-v2/Adapters/AcpAdapterV2.ts	272.17KB	$13.61
apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts	238.01KB	$11.90

Tip

To get this pull request reviewed, you can:

1. Comment @macroscope-app on this PR to request a manual review (monthly spend limits still apply).
2. Exclude the file(s) above from review by adding a pattern to your .macroscope/ignore.md — note that creating this file replaces Macroscope's built-in default ignores rather than extending them.
3. Raise your cost limit in your workspace billing settings.

^{Turn off this reminder going forward}

[mwolson]

What Changed

You can steer a turn while the agent is still working. The run could settle anyway, and the later answer vanished. That later text now stays on the thread.

A provider that never started could leave the thread looking alive forever. Exhausted retries now fail the run.

You settle a thread, and Home still treats it like live work. Missing snapshots keep retrying, archived threads drift back, and Waiting keeps refreshing stale rows. Settled now stays settled. A gone thread stops retrying.

You dismiss a thread error, and a later one never appears, or the old one never leaves. Dismissal lasts until a later error. Plan mode off still left the Build/Plan toggle in the composer after the v2 chat rewrite. The toggle is hidden in that case.

Why

Settle meant the useful work was over in the UI while the agent, the outbox, or a later error was still in play. Home, Waiting, and the composer should follow that same rule.

Julius's bounded thread history on current CTM stays as shipped.

UI Changes

Steered output stays in the thread. Failed starts no longer hang as if they were still starting. Home and Waiting settled state is quieter. Thread error banners dismiss until a later error. Build/Plan is hidden when plan mode is disabled.

No before/after screenshots.

Checklist

• This PR is small and focused
• I explained what changed and why
• I included before/after screenshots for any UI changes
• I included a video for animation/interaction changes

---

Note

High Risk
Changes core orchestration paths—settlement merging, command commit guards, Claude turn lifecycle, and process-loss reconciliation—with race-sensitive behavior that can leave runs stuck or drop recovery if callers mishandle committed: false.

Overview
Threads gain settledOverrideAt so settle/keep-active pins have a stable establishment time that survives renames and is exposed on shells and mobile projections.

Settlement is tightened end-to-end: settle is blocked on queued runs, pending runtime requests, and provider background work; archived threads cannot settle/unsettle; pin/unsettle paths maintain the new timestamp; message dispatch clears overrides via a shared helper. Settled/unsettled events are merged through reduceThreadSettlementEvent in memory and in SQL so stale full-thread payloads cannot overwrite title or other metadata. Provider ingest tests document when pending thread-scoped requests should emit thread.unsettled without waking keep-active or non-activity updates.

Claude steering tracks per-turn steer counts and frame counts, detects steering handoff results before terminalizing, marks in-flight tools interrupted, and bounds silent handoffs with a timeout that fails the turn if no new frames arrive. Turn finalization is claimed up front to avoid duplicate terminals when the handoff fiber races the frame handler; new turns wait while a turn is finalizing.

Recovery adds optional updatedBefore filters on outbox cancel/reconcile and expectedThreadUpdatedAt on commitCommand so stale recovery writes no-op when the thread has already advanced.

^{Reviewed by Cursor Bugbot for commit 57fca87. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add Orchestration V2 system with provider adapters, persistence, and V2 UI migration

• Introduces the complete Orchestration V2 architecture: contracts, event store, projection store, command receipts, effect outbox/worker, runtime recovery, and database migrations 044–052
• Adds V2 provider adapters for Claude, Codex, Cursor, Grok, OpenCode, and ACP Registry through the ProviderAdapterV2 service contract and driver registry, with replay testkits for deterministic testing
• Implements thread lifecycle services for launch, fork, settlement, checkpoint capture/rollback, context handoff, continuation, run finalization, and provider-session management with idle timeout and shutdown
• Migrates client-runtime, web, and mobile state from legacy orchestration types to V2 thread projections, bounded shell snapshots, runtime summaries, and progressive history paging
• Adds MCP orchestrator and worktree toolkits, scheduled-task service with interval and fixed-time scheduling, T3 orchestration instructions, and a legacy V1 thread importer with idempotent shell and transcript hydration
• Risk: migrations 044–052 create new schema tables and copy existing V2/project data into a shared application event log; EnvironmentThreadState in threadState.ts replaces legacy thread/page fields with OrchestrationV2ThreadProjection and ThreadHistoryMeta, removing threadHasOlderTurns; RPC authorization scopes in RpcAuthorization.ts are remapped to V2 method constants

^{Macroscope summarized 674b0e4.}

[coderabbitai]

Important

Review skipped

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

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 7726a859-eb6b-4efc-9e83-8c6c368e6e50

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

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

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

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[macroscopeapp]

One convention issue found in the new provider-ingest settlement code: a statically known tagged failure is recovered with Effect.catchTag instead of Effect.catchTags.

^{Posted via Macroscope — Effect Service Conventions}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This is a 964-file Orchestration V2 architecture and persistence migration that adds provider integrations, new workflows, protocol/schema changes, and product-default changes across server, web, and mobile. Unresolved projection and provider-lifecycle race findings add material consistency risk beyond the scope of an auto-approvable fix.

Not approved because:

• 3 blocking correctness issues found at or above your repo's Minimum Blocking Severity

^{Adjust the Minimum Blocking Severity for this repo — including turning it Off — in Settings. You can add or adjust custom eligibility rules. Learn more.}

[macroscopeapp]

Effect service conventions review of the settle-lifecycle changes. The previously flagged Effect.catchTag usage in ProviderEventIngestor.ts is now Effect.catchTags({ ... }), and the new/changed service code (ProjectionStore, EventSink, EffectOutbox, ProviderRuntimeRecoveryService, ProviderEventIngestor, shared settlement helpers) follows the import, layer, and dependency-acquisition rules. Two small change-discipline findings in packages/client-runtime/src/state/threadSettled.ts.

^{Posted via Macroscope — Effect Service Conventions}

[macroscopeapp]

🟡 Mediumorchestration-v2/ProjectionStore.ts:1261

thread.unsettled events lose event.payload.unsettledAt in the durable projection, so shell snapshots report unsettledAt as the previous value (often null) instead of when the thread was explicitly kept active. Include unsettledAt in the settlement merge.

 settlement: {
+ unsettledAt: event.payload.unsettledAt ?? null,

🤖 Copy this AI Prompt to have your agent fix this:

In file @apps/server/src/orchestration-v2/ProjectionStore.ts around line 1261:
`thread.unsettled` events lose `event.payload.unsettledAt` in the durable projection, so shell snapshots report `unsettledAt` as the previous value (often `null`) instead of when the thread was explicitly kept active. Include `unsettledAt` in the settlement merge.

[macroscopeapp]

🟡 Mediumorchestration-v2/ProviderEventIngestor.ts:356

Archived threads receive a thread.unsettled event for pending non-auth runtime-request.updated activity, clearing their settledOverride instead of leaving archived settlement state unchanged. getThreadProjection returns the still-present archived projection, and this path does not reject it before appending the synthetic event; restrict candidate generation to active threads.

🤖 Copy this AI Prompt to have your agent fix this:

In file @apps/server/src/orchestration-v2/ProviderEventIngestor.ts around line 356:
Archived threads receive a `thread.unsettled` event for pending non-auth `runtime-request.updated` activity, clearing their `settledOverride` instead of leaving archived settlement state unchanged. `getThreadProjection` returns the still-present archived projection, and this path does not reject it before appending the synthetic event; restrict candidate generation to active threads.

[macroscopeapp]

🟡 Mediumstate/orchestrationV2Projection.ts:207

A delayed thread.unsettled rewinds projection.updatedAt to event.occurredAt, even when the projection already has a newer timestamp from a concurrent metadata update. Although reduceThreadSettlementEvent preserves the newer thread.updatedAt, spreading base here still makes client ordering/activity views treat the thread as stale; preserve the later of the existing projection timestamp and event.occurredAt, as the server reducer does.

🤖 Copy this AI Prompt to have your agent fix this:

In file @packages/client-runtime/src/state/orchestrationV2Projection.ts around line 207:
A delayed `thread.unsettled` rewinds `projection.updatedAt` to `event.occurredAt`, even when the projection already has a newer timestamp from a concurrent metadata update. Although `reduceThreadSettlementEvent` preserves the newer `thread.updatedAt`, spreading `base` here still makes client ordering/activity views treat the thread as stale; preserve the later of the existing projection timestamp and `event.occurredAt`, as the server reducer does.

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 57fca87. Configure here.}

[cursor]

Finalize claim races new turns

Medium Severity

finalizeActiveTurn nulls activeTurn and only then adds the id to finalizingTurnIds. startTurn checks those refs separately, so a concurrent start (the handoff timeout is forked) can observe both empty and open a second provider turn while the first is still emitting terminals.

Additional Locations (1)

• apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts#L5197-L5211

^{Reviewed by Cursor Bugbot for commit 57fca87. Configure here.}

[macroscopeapp]

Scope note: only apps/web/src/** TS/TSX/CSS changes were reviewed. Findings are the newly added global CSS in index.css (dead/duplicated rules) and the composer context strip's motion contract, which changed lines break while the code that depends on it is retained.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

The strip's compact/expand FLIP animation still keys on [data-composer-context-control] (BranchToolbar.tsx:232, consumed by useLabelsOverflow), but this PR removes that attribute from the environment selector, the env-mode selector (locked and select paths) and the branch selector. The only remaining carrier is the vertical Separator at BranchToolbar.tsx:545, so collapsing now animates a 1px separator while the actual controls jump to their new positions. Consider keeping the attribute on the toolbar-mode triggers (it is inert in panel mode), or removing the FLIP pass as well if the motion is intentionally gone.

Suggested change

	aria-label="Run on"
	aria-label="Run on"
	data-composer-context-control

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

This whole .glass-opacity-slider block (through the forced-colors rules) is unreferenced — the glass-opacity input in SettingsPanels.tsx:1111 uses settings-slider with --settings-slider-progress, and no literal or composed glass-opacity-slider class exists anywhere in the repo. It is also a verbatim copy of .settings-slider below with only the custom-property names changed, so it adds a second owner for the same slider treatment. Suggest dropping it (or, if a distinct slider is intended, wiring a call site and deriving it from .settings-slider rather than duplicating the vendor pseudo-element rules).

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

alert-glass is already defined as @utility alert-glass (line 284) and is what ProviderStatusBanner consumes. This second definition in @layer components duplicates it and omits the utility's @supports not ((-webkit-backdrop-filter: blur(1px)) or (backdrop-filter: blur(1px))) opaque fallback; because !important declarations resolve in reverse layer order, this components-layer background: ... !important outranks the utility's important fallback, so browsers without backdrop-filter get a translucent, unblurred banner instead of the solid background. Suggest deleting this block and keeping the @utility as the single owner.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

.surface-subheader has no consumer: all four subheaders (DiffPanelShell, FileBrowserPanel, FilePreviewPanel, PreviewChromeRow) still carry the same geometry inline and only set data-surface-subheader. The inline-mode rule below is also redundant — every carrier already has in-data-[preview-panel-mode=inline]:h-7/min-h-7/mb-3/border-b-transparent, and those utilities sort after @layer components, so they win wherever this rule matches. Either apply surface-subheader at those call sites and drop the duplicated utility strings, or drop both rules so the shared geometry keeps one owner.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

The label previously animated transform/opacity on an inner data-composer-label-motion span with motion-reduce:transform-none motion-reduce:transition-opacity; this replaces it with a max-width transition and no reduced-motion guard, so reduced-motion users now get an animated layout change on every collapse. Same class appears at line 79 and in BranchToolbarEnvModeSelector/BranchToolbarBranchSelector.

Suggested change

	className="min-w-0 max-w-[240px] truncate transition-[max-width,opacity] duration-300 ease-out group-data-[compact]/composer-context:max-w-0 group-data-[compact]/composer-context:opacity-0"
	className="min-w-0 max-w-[240px] truncate transition-[max-width,opacity] duration-300 ease-out group-data-[compact]/composer-context:max-w-0 group-data-[compact]/composer-context:opacity-0 motion-reduce:transition-none"

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

These row classes override Button's core contract at the call site — height (h-9 sm:h-9), radius, padding, and base/hover/pressed background, with ! needed to beat the ghost variant's own hover colors and raw black/white alpha instead of semantic tokens. Since this treatment is now shared by every panel row (select rows, link rows, split halves, icon actions), a named Button size/variant (e.g. size="panel-row" plus a muted-ghost variant) would express it in the primitive and drop the !important fights and the sm: re-statements; contextual width/color could stay here.

^{Posted via Macroscope — UI Consistency}

[macroscopeapp]

Macroscope skipped reviewing this pull request. Per-review cost limit exceeded (workspace setting).

This review would cost an estimated $250.50, which exceeds your per-review limit of $10.00.

The top 3 files driving up this estimate:

File	Diff Size	Estimate
apps/server/src/orchestration-v2/Orchestrator.ts	280.33KB	$14.02
apps/server/src/orchestration-v2/Adapters/AcpAdapterV2.ts	272.17KB	$13.61
apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts	238.01KB	$11.90

Tip

To get this pull request reviewed, you can:

1. Comment @macroscope-app on this PR to request a manual review (monthly spend limits still apply).
2. Exclude the file(s) above from review by adding a pattern to your .macroscope/ignore.md — note that creating this file replaces Macroscope's built-in default ignores rather than extending them.
3. Raise your cost limit in your workspace billing settings.

^{Turn off this reminder going forward}