fix(ui): keep tool batch counts stable

[Astro-Han]

Summary

• Reconcile persisted tool stream evidence during multi-step turns instead of waiting for the whole turn to finish.
• Keep completed Bash output attached to its original tool row without counting it again in the next batch.
• Add UI projection and desktop handoff regression coverage.

Why

Completed tools with streamed output remained in the live projection until the turn became terminal. During a multi-step turn, those retained Bash rows were merged into later tool groups, so batch counts accumulated as 3, 7, 11, 14 even though the actual batches contained 3, 4, 4, 3 commands. The count snapped back only after persisted history replaced the stale live rows.

Scope

Changed:

• Live-turn reconciliation now removes evidence-only steps as soon as persisted history contains equivalent durable output.
• Desktop reconciliation runs while a later step is still active.
• Regression tests cover both the UI projection and desktop event-handler boundary.

Not included:

• Tool execution or persistence format changes.
• Tool summary copy, styling, or layout changes.
• Changes to incomplete or empty shell_run evidence retention.

Verification

• npm run -w @maka/ui test — 99 tests passed.
• npm run -w @maka/desktop test — 2,312 tests passed.
• Replayed the recorded four-batch session: counts remain 3, 4, 4, 3, with no residual live evidence.

User-facing impact

Tool counts remain stable before and after each call completes, while streamed output stays available on the original tool row. No docs, changelog, migrations, or breaking changes.

Reviewer notes

The review focus is the expansion from terminal-only reconciliation to active multi-step turns. durableStreamEvidence still prevents handoff when persisted terminal or shell_run output is empty, so live-only output is not discarded.

No new screenshot is included because this is a timing/state reconciliation fix with no visual styling change; the original recorded symptom is locked by deterministic projection and handoff tests.

Checklist

• Scope matches the PR title and excludes unrelated changes
• Verification lists commands/results
• User-facing impact and absence of docs/changelog/migration/breaking changes are noted
• Risk and review focus are called out
• UI screenshot/video applicability is explained

[Astro-Han]

Summary

• Reconcile persisted tool stream evidence during multi-step turns instead of waiting for the whole turn to finish.
• Keep completed Bash output attached to its original tool row without counting it again in the next batch.
• Add UI projection and desktop handoff regression coverage.

Why

Completed tools with streamed output remained in the live projection until the turn became terminal. During a multi-step turn, those retained Bash rows were merged into later tool groups, so batch counts accumulated as 3, 7, 11, 14 even though the actual batches contained 3, 4, 4, 3 commands. The count snapped back only after persisted history replaced the stale live rows.

Scope

Changed:

• Live-turn reconciliation now removes evidence-only steps as soon as persisted history contains equivalent durable output.
• Desktop reconciliation runs while a later step is still active.
• Regression tests cover both the UI projection and desktop event-handler boundary.

Not included:

• Tool execution or persistence format changes.
• Tool summary copy, styling, or layout changes.
• Changes to incomplete or empty shell_run evidence retention.

Verification

• npm run -w @maka/ui test — 99 tests passed.
• npm run -w @maka/desktop test — 2,312 tests passed.
• Replayed the recorded four-batch session: counts remain 3, 4, 4, 3, with no residual live evidence.

User-facing impact

Tool counts remain stable before and after each call completes, while streamed output stays available on the original tool row. No docs, changelog, migrations, or breaking changes.

Reviewer notes

The review focus is the expansion from terminal-only reconciliation to active multi-step turns. durableStreamEvidence still prevents handoff when persisted terminal or shell_run output is empty, so live-only output is not discarded.

No new screenshot is included because this is a timing/state reconciliation fix with no visual styling change; the original recorded symptom is locked by deterministic projection and handoff tests.

Checklist

• Scope matches the PR title and excludes unrelated changes
• Verification lists commands/results
• User-facing impact and absence of docs/changelog/migration/breaking changes are noted
• Risk and review focus are called out
• UI screenshot/video applicability is explained

[Astro-Han]

Summary

• Reconcile persisted tool stream evidence during multi-step turns instead of waiting for the whole turn to finish.
• Keep completed Bash output attached to its original tool row without counting it again in the next batch.
• Add UI projection and desktop handoff regression coverage.

Why

Completed tools with streamed output remained in the live projection until the turn became terminal. During a multi-step turn, those retained Bash rows were merged into later tool groups, so batch counts accumulated as 3, 7, 11, 14 even though the actual batches contained 3, 4, 4, 3 commands. The count snapped back only after persisted history replaced the stale live rows.

Scope

Changed:

• Live-turn reconciliation now removes evidence-only steps as soon as persisted history contains equivalent durable output.
• Desktop reconciliation runs while a later step is still active.
• Regression tests cover both the UI projection and desktop event-handler boundary.

Not included:

• Tool execution or persistence format changes.
• Tool summary copy, styling, or layout changes.
• Changes to incomplete or empty shell_run evidence retention.

Verification

• npm run -w @maka/ui test — 99 tests passed.
• npm run -w @maka/desktop test — 2,312 tests passed.
• Replayed the recorded four-batch session: counts remain 3, 4, 4, 3, with no residual live evidence.

User-facing impact

Tool counts remain stable before and after each call completes, while streamed output stays available on the original tool row. No docs, changelog, migrations, or breaking changes.

Reviewer notes

The review focus is the expansion from terminal-only reconciliation to active multi-step turns. durableStreamEvidence still prevents handoff when persisted terminal or shell_run output is empty, so live-only output is not discarded.

No new screenshot is included because this is a timing/state reconciliation fix with no visual styling change; the original recorded symptom is locked by deterministic projection and handoff tests.

Checklist

• Scope matches the PR title and excludes unrelated changes
• Verification lists commands/results
• User-facing impact and absence of docs/changelog/migration/breaking changes are noted
• Risk and review focus are called out
• UI screenshot/video applicability is explained

[Astro-Han]

Summary

• Reconcile persisted tool stream evidence during multi-step turns instead of waiting for the whole turn to finish.
• Keep completed Bash output attached to its original tool row without counting it again in the next batch.
• Add UI projection and desktop handoff regression coverage.

Why

Completed tools with streamed output remained in the live projection until the turn became terminal. During a multi-step turn, those retained Bash rows were merged into later tool groups, so batch counts accumulated as 3, 7, 11, 14 even though the actual batches contained 3, 4, 4, 3 commands. The count snapped back only after persisted history replaced the stale live rows.

Scope

Changed:

• Live-turn reconciliation now removes evidence-only steps as soon as persisted history contains equivalent durable output.
• Desktop reconciliation runs while a later step is still active.
• Regression tests cover both the UI projection and desktop event-handler boundary.

Not included:

• Tool execution or persistence format changes.
• Tool summary copy, styling, or layout changes.
• Changes to incomplete or empty shell_run evidence retention.

Verification

• npm run -w @maka/ui test — 99 tests passed.
• npm run -w @maka/desktop test — 2,312 tests passed.
• Replayed the recorded four-batch session: counts remain 3, 4, 4, 3, with no residual live evidence.

User-facing impact

Tool counts remain stable before and after each call completes, while streamed output stays available on the original tool row. No docs, changelog, migrations, or breaking changes.

Reviewer notes

The review focus is the expansion from terminal-only reconciliation to active multi-step turns. durableStreamEvidence still prevents handoff when persisted terminal or shell_run output is empty, so live-only output is not discarded.

No new screenshot is included because this is a timing/state reconciliation fix with no visual styling change; the original recorded symptom is locked by deterministic projection and handoff tests.

Checklist

• Scope matches the PR title and excludes unrelated changes
• Verification lists commands/results
• User-facing impact and absence of docs/changelog/migration/breaking changes are noted
• Risk and review focus are called out
• UI screenshot/video applicability is explained

[Astro-Han]

Summary

• Reconcile persisted tool stream evidence during multi-step turns instead of waiting for the whole turn to finish.
• Keep completed Bash output attached to its original tool row without counting it again in the next batch.
• Add UI projection and desktop handoff regression coverage.

Why

Completed tools with streamed output remained in the live projection until the turn became terminal. During a multi-step turn, those retained Bash rows were merged into later tool groups, so batch counts accumulated as 3, 7, 11, 14 even though the actual batches contained 3, 4, 4, 3 commands. The count snapped back only after persisted history replaced the stale live rows.

Scope

Changed:

• Live-turn reconciliation now removes evidence-only steps as soon as persisted history contains equivalent durable output.
• Desktop reconciliation runs while a later step is still active.
• Regression tests cover both the UI projection and desktop event-handler boundary.

Not included:

• Tool execution or persistence format changes.
• Tool summary copy, styling, or layout changes.
• Changes to incomplete or empty shell_run evidence retention.

Verification

• npm run -w @maka/ui test — 99 tests passed.
• npm run -w @maka/desktop test — 2,312 tests passed.
• Replayed the recorded four-batch session: counts remain 3, 4, 4, 3, with no residual live evidence.

User-facing impact

Tool counts remain stable before and after each call completes, while streamed output stays available on the original tool row. No docs, changelog, migrations, or breaking changes.

Reviewer notes

The review focus is the expansion from terminal-only reconciliation to active multi-step turns. durableStreamEvidence still prevents handoff when persisted terminal or shell_run output is empty, so live-only output is not discarded.

No new screenshot is included because this is a timing/state reconciliation fix with no visual styling change; the original recorded symptom is locked by deterministic projection and handoff tests.

Checklist

• Scope matches the PR title and excludes unrelated changes
• Verification lists commands/results
• User-facing impact and absence of docs/changelog/migration/breaking changes are noted
• Risk and review focus are called out
• UI screenshot/video applicability is explained

[Astro-Han]

Summary

• Reconcile persisted tool stream evidence during multi-step turns instead of waiting for the whole turn to finish.
• Keep completed Bash output attached to its original tool row without counting it again in the next batch.
• Add UI projection and desktop handoff regression coverage.

Why

Completed tools with streamed output remained in the live projection until the turn became terminal. During a multi-step turn, those retained Bash rows were merged into later tool groups, so batch counts accumulated as 3, 7, 11, 14 even though the actual batches contained 3, 4, 4, 3 commands. The count snapped back only after persisted history replaced the stale live rows.

Scope

Changed:

• Live-turn reconciliation now removes evidence-only steps as soon as persisted history contains equivalent durable output.
• Desktop reconciliation runs while a later step is still active.
• Regression tests cover both the UI projection and desktop event-handler boundary.

Not included:

• Tool execution or persistence format changes.
• Tool summary copy, styling, or layout changes.
• Changes to incomplete or empty shell_run evidence retention.

Verification

• npm run -w @maka/ui test — 99 tests passed.
• npm run -w @maka/desktop test — 2,312 tests passed.
• Replayed the recorded four-batch session: counts remain 3, 4, 4, 3, with no residual live evidence.

User-facing impact

Tool counts remain stable before and after each call completes, while streamed output stays available on the original tool row. No docs, changelog, migrations, or breaking changes.

Reviewer notes

The review focus is the expansion from terminal-only reconciliation to active multi-step turns. durableStreamEvidence still prevents handoff when persisted terminal or shell_run output is empty, so live-only output is not discarded.

No new screenshot is included because this is a timing/state reconciliation fix with no visual styling change; the original recorded symptom is locked by deterministic projection and handoff tests.

Checklist

• Scope matches the PR title and excludes unrelated changes
• Verification lists commands/results
• User-facing impact and absence of docs/changelog/migration/breaking changes are noted
• Risk and review focus are called out
• UI screenshot/video applicability is explained

[Astro-Han]

Summary

• Reconcile persisted tool stream evidence during multi-step turns instead of waiting for the whole turn to finish.
• Keep completed Bash output attached to its original tool row without counting it again in the next batch.
• Add UI projection and desktop handoff regression coverage.

Why

Completed tools with streamed output remained in the live projection until the turn became terminal. During a multi-step turn, those retained Bash rows were merged into later tool groups, so batch counts accumulated as 3, 7, 11, 14 even though the actual batches contained 3, 4, 4, 3 commands. The count snapped back only after persisted history replaced the stale live rows.

Scope

Changed:

• Live-turn reconciliation now removes evidence-only steps as soon as persisted history contains equivalent durable output.
• Desktop reconciliation runs while a later step is still active.
• Regression tests cover both the UI projection and desktop event-handler boundary.

Not included:

• Tool execution or persistence format changes.
• Tool summary copy, styling, or layout changes.
• Changes to incomplete or empty shell_run evidence retention.

Verification

• npm run -w @maka/ui test — 99 tests passed.
• npm run -w @maka/desktop test — 2,312 tests passed.
• Replayed the recorded four-batch session: counts remain 3, 4, 4, 3, with no residual live evidence.

User-facing impact

Tool counts remain stable before and after each call completes, while streamed output stays available on the original tool row. No docs, changelog, migrations, or breaking changes.

Reviewer notes

The review focus is the expansion from terminal-only reconciliation to active multi-step turns. durableStreamEvidence still prevents handoff when persisted terminal or shell_run output is empty, so live-only output is not discarded.

No new screenshot is included because this is a timing/state reconciliation fix with no visual styling change; the original recorded symptom is locked by deterministic projection and handoff tests.

Checklist

• Scope matches the PR title and excludes unrelated changes
• Verification lists commands/results
• User-facing impact and absence of docs/changelog/migration/breaking changes are noted
• Risk and review focus are called out
• UI screenshot/video applicability is explained

[Astro-Han]

Summary

• Reconcile persisted tool stream evidence during multi-step turns instead of waiting for the whole turn to finish.
• Keep completed Bash output attached to its original tool row without counting it again in the next batch.
• Add UI projection and desktop handoff regression coverage.

Why

Completed tools with streamed output remained in the live projection until the turn became terminal. During a multi-step turn, those retained Bash rows were merged into later tool groups, so batch counts accumulated as 3, 7, 11, 14 even though the actual batches contained 3, 4, 4, 3 commands. The count snapped back only after persisted history replaced the stale live rows.

Scope

Changed:

• Live-turn reconciliation now removes evidence-only steps as soon as persisted history contains equivalent durable output.
• Desktop reconciliation runs while a later step is still active.
• Regression tests cover both the UI projection and desktop event-handler boundary.

Not included:

• Tool execution or persistence format changes.
• Tool summary copy, styling, or layout changes.
• Changes to incomplete or empty shell_run evidence retention.

Verification

• npm run -w @maka/ui test — 99 tests passed.
• npm run -w @maka/desktop test — 2,312 tests passed.
• Replayed the recorded four-batch session: counts remain 3, 4, 4, 3, with no residual live evidence.

User-facing impact

Tool counts remain stable before and after each call completes, while streamed output stays available on the original tool row. No docs, changelog, migrations, or breaking changes.

Reviewer notes

The review focus is the expansion from terminal-only reconciliation to active multi-step turns. durableStreamEvidence still prevents handoff when persisted terminal or shell_run output is empty, so live-only output is not discarded.

No new screenshot is included because this is a timing/state reconciliation fix with no visual styling change; the original recorded symptom is locked by deterministic projection and handoff tests.

Checklist

• Scope matches the PR title and excludes unrelated changes
• Verification lists commands/results
• User-facing impact and absence of docs/changelog/migration/breaking changes are noted
• Risk and review focus are called out
• UI screenshot/video applicability is explained