test(server): measure shell, second client, and reconnect transfer

[t3dotgg]

The transfer budget measured one thread subscription during one turn. Real clients also hold the sidebar (shell) subscription, users often have a desktop and a phone connected at once, and clients reconnect. Nobody had measured those, and the larger perf proposals (sharing the sidebar projector across clients, sharing git services across connections, incremental history) are gated on knowing whether they matter.

This PR measures them in the same test and prints them next to the existing capped rows. The single-thread measurement and its caps are unchanged.

What it adds:

• A shell-only socket and a second client socket holding both the thread and shell subscriptions during the measured turn. The report shows per-socket wire bytes, decoded bytes, message counts, and total server egress.
• A reconnect after the turn. The second client disconnects and resubscribes each subscription from the cursor it held before the turn. The report shows catch-up bytes and whether the server chose replay or a snapshot reset.
• A SQL statement count from sql.execute spans. A counting tracer is installed on the harness runtime and the HTTP/WS app so reactor writes and subscription reads land in one total. No new instrumentation in src/.
• openMeasuredWsClient opens each RPC socket on its own child scope so clients disconnect independently.

The new rows have no caps. The shell stream coalesces on a 50 ms window, so the second client's message count moved between 9 and 10 across runs and its wire bytes between 6.7 and 7.1 KiB. SQL statement counts moved between 1096 and 1167 for the same reason. The capped thread rows stayed within a few bytes.

Measured (local, one run):

Provider	Phase	Metric	Observed	Budget	Result
codex	thread snapshot	HTTP wire	6.9 KiB (7,049 B)	7.3 KiB (7,500 B)	PASS
codex	measured turn	WebSocket wire	6.3 KiB (6,401 B)	7.8 KiB (8,000 B)	PASS
codex	measured turn	WebSocket decoded	54.7 KiB (55,974 B)	66.4 KiB (68,000 B)	PASS
codex	measured turn	WebSocket messages	8	21	PASS
codex	shell snapshot	HTTP wire	829 B	none	INFO
codex	measured turn, shell	WebSocket wire	628 B	none	INFO
codex	measured turn, shell	WebSocket decoded	3.7 KiB (3,793 B)	none	INFO
codex	measured turn, shell	WebSocket messages	3	none	INFO
codex	measured turn, second client	WebSocket wire	7.0 KiB (7,144 B)	none	INFO
codex	measured turn, second client	WebSocket decoded	58.3 KiB (59,723 B)	none	INFO
codex	measured turn, second client	WebSocket messages	10	none	INFO
codex	measured turn	server egress wire	13.8 KiB (14,173 B)	none	INFO
codex	measured turn	SQL statements	1096	none	INFO
codex	reconnect, thread (replay)	WebSocket wire	6.2 KiB (6,325 B)	none	INFO
codex	reconnect, thread (replay)	WebSocket decoded	54.4 KiB (55,707 B)	none	INFO
codex	reconnect, thread (replay)	WebSocket messages	2	none	INFO
codex	reconnect, shell (replay)	WebSocket wire	386 B	none	INFO
codex	reconnect, shell (replay)	WebSocket decoded	1.3 KiB (1,355 B)	none	INFO
codex	reconnect, shell (replay)	WebSocket messages	2	none	INFO
codex	reconnect	SQL statements	10	none	INFO
claudeAgent	thread snapshot	HTTP wire	6.9 KiB (7,059 B)	7.3 KiB (7,500 B)	PASS
claudeAgent	measured turn	WebSocket wire	6.3 KiB (6,409 B)	7.8 KiB (8,000 B)	PASS
claudeAgent	measured turn	WebSocket decoded	55.5 KiB (56,806 B)	66.4 KiB (68,000 B)	PASS
claudeAgent	measured turn	WebSocket messages	8	21	PASS
claudeAgent	shell snapshot	HTTP wire	829 B	none	INFO
claudeAgent	measured turn, shell	WebSocket wire	626 B	none	INFO
claudeAgent	measured turn, shell	WebSocket decoded	3.8 KiB (3,877 B)	none	INFO
claudeAgent	measured turn, shell	WebSocket messages	3	none	INFO
claudeAgent	measured turn, second client	WebSocket wire	7.0 KiB (7,161 B)	none	INFO
claudeAgent	measured turn, second client	WebSocket decoded	59.2 KiB (60,602 B)	none	INFO
claudeAgent	measured turn, second client	WebSocket messages	10	none	INFO
claudeAgent	measured turn	server egress wire	13.9 KiB (14,196 B)	none	INFO
claudeAgent	measured turn	SQL statements	1096	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket wire	6.2 KiB (6,342 B)	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket decoded	55.2 KiB (56,539 B)	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket messages	2	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket wire	379 B	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket decoded	1.4 KiB (1,383 B)	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket messages	2	none	INFO
claudeAgent	reconnect	SQL statements	10	none	INFO

Reading the numbers:

• The sidebar costs about 630 B wire and 3 messages per turn, about 10% of the thread stream. A second device costs one more thread stream plus one more shell stream, about 7 KiB. Egress scales linearly with connected clients and there is no hidden multiplier.
• Reconnecting from the pre-turn cursor replays about the same bytes as live delivery (6.3 KiB thread, 380 B shell) in 2 messages each and 10 SQL statements total. Replay was chosen every time at this gap size. Snapshot fallback did not trigger.
• This fixture mocks git, so it says nothing about sharing git services across connections.

Change authored by Claude Fable 5.1 running in Claude Code.

---

Note

Low Risk
Changes are confined to integration tests and reporting helpers; production server code is not modified.

Overview
Extends the thread transfer budget integration test so CI reports more than a single thread WebSocket during one turn. Existing capped thread snapshot and measured-turn rows are unchanged.

The scenario now opens three measured WebSocket clients (thread-only, shell-only, and a second device with both subscriptions), records shell HTTP snapshot size, and sums server egress across the three sockets. After the turn, the second client closes and resubscribes from its pre-turn cursors; the report records catch-up wire/decoded/message totals and whether catch-up used replay or snapshot.

Test harness plumbing adds openMeasuredWsClient (per-client forked scope + close), WebSocket recorder improvements (awaitOpen, wire totals after close), shared SQL statement counting via a counting Tracer on the orchestration harness and HTTP/WS app (sql.execute spans), and helpers moved into TransferBudgetScenario for thread/shell subscribe and sync detection. New report fields are INFO-only (no new budgets).

^{Reviewed by Cursor Bugbot for commit 7aff2b1. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Expand transfer-budget test to measure shell, second client, and reconnect transfer

• Reworks the live transfer-budget integration test in server.test.ts to open three concurrent measured clients (thread-only, shell-only, second-client), capture per-socket WebSocket transfer and server-wide sql.execute deltas, then close and reconnect the second client to measure replay-or-snapshot catch-up
• Adds openMeasuredWsClient, a factory that forks a per-client child scope, builds a counting WebSocket protocol, waits for open (10s timeout defect), and returns an explicit close effect
• Updates makeWebSocketTransferRecorder to retain the underlying transport at open time so wire-byte totals remain readable after socket closure
• Adds makeSqlStatementCounter, a tracer that increments a counter for spans named exactly sql.execute while returning native spans for all operations
• Extends formatTransferBudgetReport with informational (uncapped) rows for shell snapshot, shell-only and second-client WebSocket diagnostics, aggregate server egress, SQL counts, and reconnect catch-up modes alongside existing capped thread rows
• Adds subscription helpers subscribeThreadItems, subscribeShellItems, and awaitSubscriptionSynchronized in TransferBudgetScenario.integration.ts that run scoped subscriptions and classify initial catch-up as replay or snapshot
• Risk: makeWebSocketTransferRecorder open wait fails with a timeout defect after 10 seconds; awaitSubscriptionSynchronized similarly defects after 10 seconds if the synchronization marker never arrives

^{Macroscope summarized 7aff2b1.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.3 KiB	13.3 KiB	+15 B (+0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	+6 B (+0.1%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.4 KiB	+9 B (+0.1%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	55.6 KiB	55.6 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	13.3 KiB	13.3 KiB	+13 B (+0.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	−6 B (−0.1%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.4 KiB	6.4 KiB	+19 B (+0.3%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	56.4 KiB	56.4 KiB	0 B (0.0%)	66.4 KiB	✅
Claude	Live turn messages	10	10	0 (0.0%)	21	✅

Baseline: 80c708a · PR result: 7aff2b1 · Source CI: success

Scenario and decoded snapshot size

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

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

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

[macroscopeapp]

Approvability

Verdict: Approved at 7aff2b1

Macroscope's review found this PR approvable — This PR expands a server integration test to measure shell traffic, multiple clients, reconnects, and SQL activity. All changes are confined to test and reporting infrastructure, with no shipped production behavior or product-default changes.

Notes:

• No code objects were reviewed. Approvability was decided on eligibility alone.

^{You can add or adjust custom eligibility rules. Learn more.}

[t3dotgg]

The transfer budget measured one thread subscription during one turn. Real clients also hold the sidebar (shell) subscription, users often have a desktop and a phone connected at once, and clients reconnect. Nobody had measured those, and the larger perf proposals (sharing the sidebar projector across clients, sharing git services across connections, incremental history) are gated on knowing whether they matter.

This PR measures them in the same test and prints them next to the existing capped rows. The single-thread measurement and its caps are unchanged.

What it adds:

• A shell-only socket and a second client socket holding both the thread and shell subscriptions during the measured turn. The report shows per-socket wire bytes, decoded bytes, message counts, and total server egress.
• A reconnect after the turn. The second client disconnects and resubscribes each subscription from the cursor it held before the turn. The report shows catch-up bytes and whether the server chose replay or a snapshot reset.
• A SQL statement count from sql.execute spans. A counting tracer is installed on the harness runtime and the HTTP/WS app so reactor writes and subscription reads land in one total. No new instrumentation in src/.
• openMeasuredWsClient opens each RPC socket on its own child scope so clients disconnect independently.

The new rows have no caps. The shell stream coalesces on a 50 ms window, so the second client's message count moved between 9 and 10 across runs and its wire bytes between 6.7 and 7.1 KiB. SQL statement counts moved between 1096 and 1167 for the same reason. The capped thread rows stayed within a few bytes.

Measured (local, one run):

Provider	Phase	Metric	Observed	Budget	Result
codex	thread snapshot	HTTP wire	6.9 KiB (7,049 B)	7.3 KiB (7,500 B)	PASS
codex	measured turn	WebSocket wire	6.3 KiB (6,401 B)	7.8 KiB (8,000 B)	PASS
codex	measured turn	WebSocket decoded	54.7 KiB (55,974 B)	66.4 KiB (68,000 B)	PASS
codex	measured turn	WebSocket messages	8	21	PASS
codex	shell snapshot	HTTP wire	829 B	none	INFO
codex	measured turn, shell	WebSocket wire	628 B	none	INFO
codex	measured turn, shell	WebSocket decoded	3.7 KiB (3,793 B)	none	INFO
codex	measured turn, shell	WebSocket messages	3	none	INFO
codex	measured turn, second client	WebSocket wire	7.0 KiB (7,144 B)	none	INFO
codex	measured turn, second client	WebSocket decoded	58.3 KiB (59,723 B)	none	INFO
codex	measured turn, second client	WebSocket messages	10	none	INFO
codex	measured turn	server egress wire	13.8 KiB (14,173 B)	none	INFO
codex	measured turn	SQL statements	1096	none	INFO
codex	reconnect, thread (replay)	WebSocket wire	6.2 KiB (6,325 B)	none	INFO
codex	reconnect, thread (replay)	WebSocket decoded	54.4 KiB (55,707 B)	none	INFO
codex	reconnect, thread (replay)	WebSocket messages	2	none	INFO
codex	reconnect, shell (replay)	WebSocket wire	386 B	none	INFO
codex	reconnect, shell (replay)	WebSocket decoded	1.3 KiB (1,355 B)	none	INFO
codex	reconnect, shell (replay)	WebSocket messages	2	none	INFO
codex	reconnect	SQL statements	10	none	INFO
claudeAgent	thread snapshot	HTTP wire	6.9 KiB (7,059 B)	7.3 KiB (7,500 B)	PASS
claudeAgent	measured turn	WebSocket wire	6.3 KiB (6,409 B)	7.8 KiB (8,000 B)	PASS
claudeAgent	measured turn	WebSocket decoded	55.5 KiB (56,806 B)	66.4 KiB (68,000 B)	PASS
claudeAgent	measured turn	WebSocket messages	8	21	PASS
claudeAgent	shell snapshot	HTTP wire	829 B	none	INFO
claudeAgent	measured turn, shell	WebSocket wire	626 B	none	INFO
claudeAgent	measured turn, shell	WebSocket decoded	3.8 KiB (3,877 B)	none	INFO
claudeAgent	measured turn, shell	WebSocket messages	3	none	INFO
claudeAgent	measured turn, second client	WebSocket wire	7.0 KiB (7,161 B)	none	INFO
claudeAgent	measured turn, second client	WebSocket decoded	59.2 KiB (60,602 B)	none	INFO
claudeAgent	measured turn, second client	WebSocket messages	10	none	INFO
claudeAgent	measured turn	server egress wire	13.9 KiB (14,196 B)	none	INFO
claudeAgent	measured turn	SQL statements	1096	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket wire	6.2 KiB (6,342 B)	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket decoded	55.2 KiB (56,539 B)	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket messages	2	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket wire	379 B	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket decoded	1.4 KiB (1,383 B)	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket messages	2	none	INFO
claudeAgent	reconnect	SQL statements	10	none	INFO

Reading the numbers:

• The sidebar costs about 630 B wire and 3 messages per turn, about 10% of the thread stream. A second device costs one more thread stream plus one more shell stream, about 7 KiB. Egress scales linearly with connected clients and there is no hidden multiplier.
• Reconnecting from the pre-turn cursor replays about the same bytes as live delivery (6.3 KiB thread, 380 B shell) in 2 messages each and 10 SQL statements total. Replay was chosen every time at this gap size. Snapshot fallback did not trigger.
• This fixture mocks git, so it says nothing about sharing git services across connections.

Change authored by Claude Fable 5.1 running in Claude Code.

---

Note

Low Risk
Changes are confined to integration tests and reporting helpers; production server code is not modified.

Overview
Extends the thread transfer budget integration test so CI reports more than a single thread WebSocket during one turn. Existing capped thread snapshot and measured-turn rows are unchanged.

The scenario now opens three measured WebSocket clients (thread-only, shell-only, and a second device with both subscriptions), records shell HTTP snapshot size, and sums server egress across the three sockets. After the turn, the second client closes and resubscribes from its pre-turn cursors; the report records catch-up wire/decoded/message totals and whether catch-up used replay or snapshot.

Test harness plumbing adds openMeasuredWsClient (per-client forked scope + close), WebSocket recorder improvements (awaitOpen, wire totals after close), shared SQL statement counting via a counting Tracer on the orchestration harness and HTTP/WS app (sql.execute spans), and helpers moved into TransferBudgetScenario for thread/shell subscribe and sync detection. New report fields are INFO-only (no new budgets).

^{Reviewed by Cursor Bugbot for commit 7aff2b1. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Expand transfer-budget test to measure shell, second client, and reconnect transfer

• Reworks the live transfer-budget integration test in server.test.ts to open three concurrent measured clients (thread-only, shell-only, second-client), capture per-socket WebSocket transfer and server-wide sql.execute deltas, then close and reconnect the second client to measure replay-or-snapshot catch-up
• Adds openMeasuredWsClient, a factory that forks a per-client child scope, builds a counting WebSocket protocol, waits for open (10s timeout defect), and returns an explicit close effect
• Updates makeWebSocketTransferRecorder to retain the underlying transport at open time so wire-byte totals remain readable after socket closure
• Adds makeSqlStatementCounter, a tracer that increments a counter for spans named exactly sql.execute while returning native spans for all operations
• Extends formatTransferBudgetReport with informational (uncapped) rows for shell snapshot, shell-only and second-client WebSocket diagnostics, aggregate server egress, SQL counts, and reconnect catch-up modes alongside existing capped thread rows
• Adds subscription helpers subscribeThreadItems, subscribeShellItems, and awaitSubscriptionSynchronized in TransferBudgetScenario.integration.ts that run scoped subscriptions and classify initial catch-up as replay or snapshot
• Risk: makeWebSocketTransferRecorder open wait fails with a timeout defect after 10 seconds; awaitSubscriptionSynchronized similarly defects after 10 seconds if the synchronization marker never arrives

^{Macroscope summarized 7aff2b1.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.3 KiB	13.3 KiB	+15 B (+0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	+6 B (+0.1%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.4 KiB	+9 B (+0.1%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	55.6 KiB	55.6 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	13.3 KiB	13.3 KiB	+13 B (+0.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	−6 B (−0.1%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.4 KiB	6.4 KiB	+19 B (+0.3%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	56.4 KiB	56.4 KiB	0 B (0.0%)	66.4 KiB	✅
Claude	Live turn messages	10	10	0 (0.0%)	21	✅

Baseline: 80c708a · PR result: 7aff2b1 · Source CI: success

Scenario and decoded snapshot size

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

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

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

[macroscopeapp]

Approvability

Verdict: Approved at 7aff2b1

Macroscope's review found this PR approvable — This PR expands a server integration test to measure shell traffic, multiple clients, reconnects, and SQL activity. All changes are confined to test and reporting infrastructure, with no shipped production behavior or product-default changes.

Notes:

• No code objects were reviewed. Approvability was decided on eligibility alone.

^{You can add or adjust custom eligibility rules. Learn more.}

[t3dotgg]

The transfer budget measured one thread subscription during one turn. Real clients also hold the sidebar (shell) subscription, users often have a desktop and a phone connected at once, and clients reconnect. Nobody had measured those, and the larger perf proposals (sharing the sidebar projector across clients, sharing git services across connections, incremental history) are gated on knowing whether they matter.

This PR measures them in the same test and prints them next to the existing capped rows. The single-thread measurement and its caps are unchanged.

What it adds:

• A shell-only socket and a second client socket holding both the thread and shell subscriptions during the measured turn. The report shows per-socket wire bytes, decoded bytes, message counts, and total server egress.
• A reconnect after the turn. The second client disconnects and resubscribes each subscription from the cursor it held before the turn. The report shows catch-up bytes and whether the server chose replay or a snapshot reset.
• A SQL statement count from sql.execute spans. A counting tracer is installed on the harness runtime and the HTTP/WS app so reactor writes and subscription reads land in one total. No new instrumentation in src/.
• openMeasuredWsClient opens each RPC socket on its own child scope so clients disconnect independently.

The new rows have no caps. The shell stream coalesces on a 50 ms window, so the second client's message count moved between 9 and 10 across runs and its wire bytes between 6.7 and 7.1 KiB. SQL statement counts moved between 1096 and 1167 for the same reason. The capped thread rows stayed within a few bytes.

Measured (local, one run):

Provider	Phase	Metric	Observed	Budget	Result
codex	thread snapshot	HTTP wire	6.9 KiB (7,049 B)	7.3 KiB (7,500 B)	PASS
codex	measured turn	WebSocket wire	6.3 KiB (6,401 B)	7.8 KiB (8,000 B)	PASS
codex	measured turn	WebSocket decoded	54.7 KiB (55,974 B)	66.4 KiB (68,000 B)	PASS
codex	measured turn	WebSocket messages	8	21	PASS
codex	shell snapshot	HTTP wire	829 B	none	INFO
codex	measured turn, shell	WebSocket wire	628 B	none	INFO
codex	measured turn, shell	WebSocket decoded	3.7 KiB (3,793 B)	none	INFO
codex	measured turn, shell	WebSocket messages	3	none	INFO
codex	measured turn, second client	WebSocket wire	7.0 KiB (7,144 B)	none	INFO
codex	measured turn, second client	WebSocket decoded	58.3 KiB (59,723 B)	none	INFO
codex	measured turn, second client	WebSocket messages	10	none	INFO
codex	measured turn	server egress wire	13.8 KiB (14,173 B)	none	INFO
codex	measured turn	SQL statements	1096	none	INFO
codex	reconnect, thread (replay)	WebSocket wire	6.2 KiB (6,325 B)	none	INFO
codex	reconnect, thread (replay)	WebSocket decoded	54.4 KiB (55,707 B)	none	INFO
codex	reconnect, thread (replay)	WebSocket messages	2	none	INFO
codex	reconnect, shell (replay)	WebSocket wire	386 B	none	INFO
codex	reconnect, shell (replay)	WebSocket decoded	1.3 KiB (1,355 B)	none	INFO
codex	reconnect, shell (replay)	WebSocket messages	2	none	INFO
codex	reconnect	SQL statements	10	none	INFO
claudeAgent	thread snapshot	HTTP wire	6.9 KiB (7,059 B)	7.3 KiB (7,500 B)	PASS
claudeAgent	measured turn	WebSocket wire	6.3 KiB (6,409 B)	7.8 KiB (8,000 B)	PASS
claudeAgent	measured turn	WebSocket decoded	55.5 KiB (56,806 B)	66.4 KiB (68,000 B)	PASS
claudeAgent	measured turn	WebSocket messages	8	21	PASS
claudeAgent	shell snapshot	HTTP wire	829 B	none	INFO
claudeAgent	measured turn, shell	WebSocket wire	626 B	none	INFO
claudeAgent	measured turn, shell	WebSocket decoded	3.8 KiB (3,877 B)	none	INFO
claudeAgent	measured turn, shell	WebSocket messages	3	none	INFO
claudeAgent	measured turn, second client	WebSocket wire	7.0 KiB (7,161 B)	none	INFO
claudeAgent	measured turn, second client	WebSocket decoded	59.2 KiB (60,602 B)	none	INFO
claudeAgent	measured turn, second client	WebSocket messages	10	none	INFO
claudeAgent	measured turn	server egress wire	13.9 KiB (14,196 B)	none	INFO
claudeAgent	measured turn	SQL statements	1096	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket wire	6.2 KiB (6,342 B)	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket decoded	55.2 KiB (56,539 B)	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket messages	2	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket wire	379 B	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket decoded	1.4 KiB (1,383 B)	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket messages	2	none	INFO
claudeAgent	reconnect	SQL statements	10	none	INFO

Reading the numbers:

• The sidebar costs about 630 B wire and 3 messages per turn, about 10% of the thread stream. A second device costs one more thread stream plus one more shell stream, about 7 KiB. Egress scales linearly with connected clients and there is no hidden multiplier.
• Reconnecting from the pre-turn cursor replays about the same bytes as live delivery (6.3 KiB thread, 380 B shell) in 2 messages each and 10 SQL statements total. Replay was chosen every time at this gap size. Snapshot fallback did not trigger.
• This fixture mocks git, so it says nothing about sharing git services across connections.

Change authored by Claude Fable 5.1 running in Claude Code.

---

Note

Low Risk
Changes are confined to integration tests and reporting helpers; production server code is not modified.

Overview
Extends the thread transfer budget integration test so CI reports more than a single thread WebSocket during one turn. Existing capped thread snapshot and measured-turn rows are unchanged.

The scenario now opens three measured WebSocket clients (thread-only, shell-only, and a second device with both subscriptions), records shell HTTP snapshot size, and sums server egress across the three sockets. After the turn, the second client closes and resubscribes from its pre-turn cursors; the report records catch-up wire/decoded/message totals and whether catch-up used replay or snapshot.

Test harness plumbing adds openMeasuredWsClient (per-client forked scope + close), WebSocket recorder improvements (awaitOpen, wire totals after close), shared SQL statement counting via a counting Tracer on the orchestration harness and HTTP/WS app (sql.execute spans), and helpers moved into TransferBudgetScenario for thread/shell subscribe and sync detection. New report fields are INFO-only (no new budgets).

^{Reviewed by Cursor Bugbot for commit 7aff2b1. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Expand transfer-budget test to measure shell, second client, and reconnect transfer

• Reworks the live transfer-budget integration test in server.test.ts to open three concurrent measured clients (thread-only, shell-only, second-client), capture per-socket WebSocket transfer and server-wide sql.execute deltas, then close and reconnect the second client to measure replay-or-snapshot catch-up
• Adds openMeasuredWsClient, a factory that forks a per-client child scope, builds a counting WebSocket protocol, waits for open (10s timeout defect), and returns an explicit close effect
• Updates makeWebSocketTransferRecorder to retain the underlying transport at open time so wire-byte totals remain readable after socket closure
• Adds makeSqlStatementCounter, a tracer that increments a counter for spans named exactly sql.execute while returning native spans for all operations
• Extends formatTransferBudgetReport with informational (uncapped) rows for shell snapshot, shell-only and second-client WebSocket diagnostics, aggregate server egress, SQL counts, and reconnect catch-up modes alongside existing capped thread rows
• Adds subscription helpers subscribeThreadItems, subscribeShellItems, and awaitSubscriptionSynchronized in TransferBudgetScenario.integration.ts that run scoped subscriptions and classify initial catch-up as replay or snapshot
• Risk: makeWebSocketTransferRecorder open wait fails with a timeout defect after 10 seconds; awaitSubscriptionSynchronized similarly defects after 10 seconds if the synchronization marker never arrives

^{Macroscope summarized 7aff2b1.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.3 KiB	13.3 KiB	+15 B (+0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	+6 B (+0.1%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.4 KiB	+9 B (+0.1%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	55.6 KiB	55.6 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	13.3 KiB	13.3 KiB	+13 B (+0.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	−6 B (−0.1%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.4 KiB	6.4 KiB	+19 B (+0.3%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	56.4 KiB	56.4 KiB	0 B (0.0%)	66.4 KiB	✅
Claude	Live turn messages	10	10	0 (0.0%)	21	✅

Baseline: 80c708a · PR result: 7aff2b1 · Source CI: success

Scenario and decoded snapshot size

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

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

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

[macroscopeapp]

Approvability

Verdict: Approved at 7aff2b1

Macroscope's review found this PR approvable — This PR expands a server integration test to measure shell traffic, multiple clients, reconnects, and SQL activity. All changes are confined to test and reporting infrastructure, with no shipped production behavior or product-default changes.

Notes:

• No code objects were reviewed. Approvability was decided on eligibility alone.

^{You can add or adjust custom eligibility rules. Learn more.}

[t3dotgg]

The transfer budget measured one thread subscription during one turn. Real clients also hold the sidebar (shell) subscription, users often have a desktop and a phone connected at once, and clients reconnect. Nobody had measured those, and the larger perf proposals (sharing the sidebar projector across clients, sharing git services across connections, incremental history) are gated on knowing whether they matter.

This PR measures them in the same test and prints them next to the existing capped rows. The single-thread measurement and its caps are unchanged.

What it adds:

• A shell-only socket and a second client socket holding both the thread and shell subscriptions during the measured turn. The report shows per-socket wire bytes, decoded bytes, message counts, and total server egress.
• A reconnect after the turn. The second client disconnects and resubscribes each subscription from the cursor it held before the turn. The report shows catch-up bytes and whether the server chose replay or a snapshot reset.
• A SQL statement count from sql.execute spans. A counting tracer is installed on the harness runtime and the HTTP/WS app so reactor writes and subscription reads land in one total. No new instrumentation in src/.
• openMeasuredWsClient opens each RPC socket on its own child scope so clients disconnect independently.

The new rows have no caps. The shell stream coalesces on a 50 ms window, so the second client's message count moved between 9 and 10 across runs and its wire bytes between 6.7 and 7.1 KiB. SQL statement counts moved between 1096 and 1167 for the same reason. The capped thread rows stayed within a few bytes.

Measured (local, one run):

Provider	Phase	Metric	Observed	Budget	Result
codex	thread snapshot	HTTP wire	6.9 KiB (7,049 B)	7.3 KiB (7,500 B)	PASS
codex	measured turn	WebSocket wire	6.3 KiB (6,401 B)	7.8 KiB (8,000 B)	PASS
codex	measured turn	WebSocket decoded	54.7 KiB (55,974 B)	66.4 KiB (68,000 B)	PASS
codex	measured turn	WebSocket messages	8	21	PASS
codex	shell snapshot	HTTP wire	829 B	none	INFO
codex	measured turn, shell	WebSocket wire	628 B	none	INFO
codex	measured turn, shell	WebSocket decoded	3.7 KiB (3,793 B)	none	INFO
codex	measured turn, shell	WebSocket messages	3	none	INFO
codex	measured turn, second client	WebSocket wire	7.0 KiB (7,144 B)	none	INFO
codex	measured turn, second client	WebSocket decoded	58.3 KiB (59,723 B)	none	INFO
codex	measured turn, second client	WebSocket messages	10	none	INFO
codex	measured turn	server egress wire	13.8 KiB (14,173 B)	none	INFO
codex	measured turn	SQL statements	1096	none	INFO
codex	reconnect, thread (replay)	WebSocket wire	6.2 KiB (6,325 B)	none	INFO
codex	reconnect, thread (replay)	WebSocket decoded	54.4 KiB (55,707 B)	none	INFO
codex	reconnect, thread (replay)	WebSocket messages	2	none	INFO
codex	reconnect, shell (replay)	WebSocket wire	386 B	none	INFO
codex	reconnect, shell (replay)	WebSocket decoded	1.3 KiB (1,355 B)	none	INFO
codex	reconnect, shell (replay)	WebSocket messages	2	none	INFO
codex	reconnect	SQL statements	10	none	INFO
claudeAgent	thread snapshot	HTTP wire	6.9 KiB (7,059 B)	7.3 KiB (7,500 B)	PASS
claudeAgent	measured turn	WebSocket wire	6.3 KiB (6,409 B)	7.8 KiB (8,000 B)	PASS
claudeAgent	measured turn	WebSocket decoded	55.5 KiB (56,806 B)	66.4 KiB (68,000 B)	PASS
claudeAgent	measured turn	WebSocket messages	8	21	PASS
claudeAgent	shell snapshot	HTTP wire	829 B	none	INFO
claudeAgent	measured turn, shell	WebSocket wire	626 B	none	INFO
claudeAgent	measured turn, shell	WebSocket decoded	3.8 KiB (3,877 B)	none	INFO
claudeAgent	measured turn, shell	WebSocket messages	3	none	INFO
claudeAgent	measured turn, second client	WebSocket wire	7.0 KiB (7,161 B)	none	INFO
claudeAgent	measured turn, second client	WebSocket decoded	59.2 KiB (60,602 B)	none	INFO
claudeAgent	measured turn, second client	WebSocket messages	10	none	INFO
claudeAgent	measured turn	server egress wire	13.9 KiB (14,196 B)	none	INFO
claudeAgent	measured turn	SQL statements	1096	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket wire	6.2 KiB (6,342 B)	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket decoded	55.2 KiB (56,539 B)	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket messages	2	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket wire	379 B	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket decoded	1.4 KiB (1,383 B)	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket messages	2	none	INFO
claudeAgent	reconnect	SQL statements	10	none	INFO

Reading the numbers:

• The sidebar costs about 630 B wire and 3 messages per turn, about 10% of the thread stream. A second device costs one more thread stream plus one more shell stream, about 7 KiB. Egress scales linearly with connected clients and there is no hidden multiplier.
• Reconnecting from the pre-turn cursor replays about the same bytes as live delivery (6.3 KiB thread, 380 B shell) in 2 messages each and 10 SQL statements total. Replay was chosen every time at this gap size. Snapshot fallback did not trigger.
• This fixture mocks git, so it says nothing about sharing git services across connections.

Change authored by Claude Fable 5.1 running in Claude Code.

---

Note

Low Risk
Changes are confined to integration tests and reporting helpers; production server code is not modified.

Overview
Extends the thread transfer budget integration test so CI reports more than a single thread WebSocket during one turn. Existing capped thread snapshot and measured-turn rows are unchanged.

The scenario now opens three measured WebSocket clients (thread-only, shell-only, and a second device with both subscriptions), records shell HTTP snapshot size, and sums server egress across the three sockets. After the turn, the second client closes and resubscribes from its pre-turn cursors; the report records catch-up wire/decoded/message totals and whether catch-up used replay or snapshot.

Test harness plumbing adds openMeasuredWsClient (per-client forked scope + close), WebSocket recorder improvements (awaitOpen, wire totals after close), shared SQL statement counting via a counting Tracer on the orchestration harness and HTTP/WS app (sql.execute spans), and helpers moved into TransferBudgetScenario for thread/shell subscribe and sync detection. New report fields are INFO-only (no new budgets).

^{Reviewed by Cursor Bugbot for commit 7aff2b1. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Expand transfer-budget test to measure shell, second client, and reconnect transfer

• Reworks the live transfer-budget integration test in server.test.ts to open three concurrent measured clients (thread-only, shell-only, second-client), capture per-socket WebSocket transfer and server-wide sql.execute deltas, then close and reconnect the second client to measure replay-or-snapshot catch-up
• Adds openMeasuredWsClient, a factory that forks a per-client child scope, builds a counting WebSocket protocol, waits for open (10s timeout defect), and returns an explicit close effect
• Updates makeWebSocketTransferRecorder to retain the underlying transport at open time so wire-byte totals remain readable after socket closure
• Adds makeSqlStatementCounter, a tracer that increments a counter for spans named exactly sql.execute while returning native spans for all operations
• Extends formatTransferBudgetReport with informational (uncapped) rows for shell snapshot, shell-only and second-client WebSocket diagnostics, aggregate server egress, SQL counts, and reconnect catch-up modes alongside existing capped thread rows
• Adds subscription helpers subscribeThreadItems, subscribeShellItems, and awaitSubscriptionSynchronized in TransferBudgetScenario.integration.ts that run scoped subscriptions and classify initial catch-up as replay or snapshot
• Risk: makeWebSocketTransferRecorder open wait fails with a timeout defect after 10 seconds; awaitSubscriptionSynchronized similarly defects after 10 seconds if the synchronization marker never arrives

^{Macroscope summarized 7aff2b1.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.3 KiB	13.3 KiB	+15 B (+0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	+6 B (+0.1%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.4 KiB	+9 B (+0.1%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	55.6 KiB	55.6 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	13.3 KiB	13.3 KiB	+13 B (+0.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	−6 B (−0.1%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.4 KiB	6.4 KiB	+19 B (+0.3%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	56.4 KiB	56.4 KiB	0 B (0.0%)	66.4 KiB	✅
Claude	Live turn messages	10	10	0 (0.0%)	21	✅

Baseline: 80c708a · PR result: 7aff2b1 · Source CI: success

Scenario and decoded snapshot size

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

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

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

[macroscopeapp]

Approvability

Verdict: Approved at 7aff2b1

Macroscope's review found this PR approvable — This PR expands a server integration test to measure shell traffic, multiple clients, reconnects, and SQL activity. All changes are confined to test and reporting infrastructure, with no shipped production behavior or product-default changes.

Notes:

• No code objects were reviewed. Approvability was decided on eligibility alone.

^{You can add or adjust custom eligibility rules. Learn more.}

[t3dotgg]

The transfer budget measured one thread subscription during one turn. Real clients also hold the sidebar (shell) subscription, users often have a desktop and a phone connected at once, and clients reconnect. Nobody had measured those, and the larger perf proposals (sharing the sidebar projector across clients, sharing git services across connections, incremental history) are gated on knowing whether they matter.

This PR measures them in the same test and prints them next to the existing capped rows. The single-thread measurement and its caps are unchanged.

What it adds:

• A shell-only socket and a second client socket holding both the thread and shell subscriptions during the measured turn. The report shows per-socket wire bytes, decoded bytes, message counts, and total server egress.
• A reconnect after the turn. The second client disconnects and resubscribes each subscription from the cursor it held before the turn. The report shows catch-up bytes and whether the server chose replay or a snapshot reset.
• A SQL statement count from sql.execute spans. A counting tracer is installed on the harness runtime and the HTTP/WS app so reactor writes and subscription reads land in one total. No new instrumentation in src/.
• openMeasuredWsClient opens each RPC socket on its own child scope so clients disconnect independently.

The new rows have no caps. The shell stream coalesces on a 50 ms window, so the second client's message count moved between 9 and 10 across runs and its wire bytes between 6.7 and 7.1 KiB. SQL statement counts moved between 1096 and 1167 for the same reason. The capped thread rows stayed within a few bytes.

Measured (local, one run):

Provider	Phase	Metric	Observed	Budget	Result
codex	thread snapshot	HTTP wire	6.9 KiB (7,049 B)	7.3 KiB (7,500 B)	PASS
codex	measured turn	WebSocket wire	6.3 KiB (6,401 B)	7.8 KiB (8,000 B)	PASS
codex	measured turn	WebSocket decoded	54.7 KiB (55,974 B)	66.4 KiB (68,000 B)	PASS
codex	measured turn	WebSocket messages	8	21	PASS
codex	shell snapshot	HTTP wire	829 B	none	INFO
codex	measured turn, shell	WebSocket wire	628 B	none	INFO
codex	measured turn, shell	WebSocket decoded	3.7 KiB (3,793 B)	none	INFO
codex	measured turn, shell	WebSocket messages	3	none	INFO
codex	measured turn, second client	WebSocket wire	7.0 KiB (7,144 B)	none	INFO
codex	measured turn, second client	WebSocket decoded	58.3 KiB (59,723 B)	none	INFO
codex	measured turn, second client	WebSocket messages	10	none	INFO
codex	measured turn	server egress wire	13.8 KiB (14,173 B)	none	INFO
codex	measured turn	SQL statements	1096	none	INFO
codex	reconnect, thread (replay)	WebSocket wire	6.2 KiB (6,325 B)	none	INFO
codex	reconnect, thread (replay)	WebSocket decoded	54.4 KiB (55,707 B)	none	INFO
codex	reconnect, thread (replay)	WebSocket messages	2	none	INFO
codex	reconnect, shell (replay)	WebSocket wire	386 B	none	INFO
codex	reconnect, shell (replay)	WebSocket decoded	1.3 KiB (1,355 B)	none	INFO
codex	reconnect, shell (replay)	WebSocket messages	2	none	INFO
codex	reconnect	SQL statements	10	none	INFO
claudeAgent	thread snapshot	HTTP wire	6.9 KiB (7,059 B)	7.3 KiB (7,500 B)	PASS
claudeAgent	measured turn	WebSocket wire	6.3 KiB (6,409 B)	7.8 KiB (8,000 B)	PASS
claudeAgent	measured turn	WebSocket decoded	55.5 KiB (56,806 B)	66.4 KiB (68,000 B)	PASS
claudeAgent	measured turn	WebSocket messages	8	21	PASS
claudeAgent	shell snapshot	HTTP wire	829 B	none	INFO
claudeAgent	measured turn, shell	WebSocket wire	626 B	none	INFO
claudeAgent	measured turn, shell	WebSocket decoded	3.8 KiB (3,877 B)	none	INFO
claudeAgent	measured turn, shell	WebSocket messages	3	none	INFO
claudeAgent	measured turn, second client	WebSocket wire	7.0 KiB (7,161 B)	none	INFO
claudeAgent	measured turn, second client	WebSocket decoded	59.2 KiB (60,602 B)	none	INFO
claudeAgent	measured turn, second client	WebSocket messages	10	none	INFO
claudeAgent	measured turn	server egress wire	13.9 KiB (14,196 B)	none	INFO
claudeAgent	measured turn	SQL statements	1096	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket wire	6.2 KiB (6,342 B)	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket decoded	55.2 KiB (56,539 B)	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket messages	2	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket wire	379 B	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket decoded	1.4 KiB (1,383 B)	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket messages	2	none	INFO
claudeAgent	reconnect	SQL statements	10	none	INFO

Reading the numbers:

• The sidebar costs about 630 B wire and 3 messages per turn, about 10% of the thread stream. A second device costs one more thread stream plus one more shell stream, about 7 KiB. Egress scales linearly with connected clients and there is no hidden multiplier.
• Reconnecting from the pre-turn cursor replays about the same bytes as live delivery (6.3 KiB thread, 380 B shell) in 2 messages each and 10 SQL statements total. Replay was chosen every time at this gap size. Snapshot fallback did not trigger.
• This fixture mocks git, so it says nothing about sharing git services across connections.

Change authored by Claude Fable 5.1 running in Claude Code.

---

Note

Low Risk
Changes are confined to integration tests and reporting helpers; production server code is not modified.

Overview
Extends the thread transfer budget integration test so CI reports more than a single thread WebSocket during one turn. Existing capped thread snapshot and measured-turn rows are unchanged.

The scenario now opens three measured WebSocket clients (thread-only, shell-only, and a second device with both subscriptions), records shell HTTP snapshot size, and sums server egress across the three sockets. After the turn, the second client closes and resubscribes from its pre-turn cursors; the report records catch-up wire/decoded/message totals and whether catch-up used replay or snapshot.

Test harness plumbing adds openMeasuredWsClient (per-client forked scope + close), WebSocket recorder improvements (awaitOpen, wire totals after close), shared SQL statement counting via a counting Tracer on the orchestration harness and HTTP/WS app (sql.execute spans), and helpers moved into TransferBudgetScenario for thread/shell subscribe and sync detection. New report fields are INFO-only (no new budgets).

^{Reviewed by Cursor Bugbot for commit 7aff2b1. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Expand transfer-budget test to measure shell, second client, and reconnect transfer

• Reworks the live transfer-budget integration test in server.test.ts to open three concurrent measured clients (thread-only, shell-only, second-client), capture per-socket WebSocket transfer and server-wide sql.execute deltas, then close and reconnect the second client to measure replay-or-snapshot catch-up
• Adds openMeasuredWsClient, a factory that forks a per-client child scope, builds a counting WebSocket protocol, waits for open (10s timeout defect), and returns an explicit close effect
• Updates makeWebSocketTransferRecorder to retain the underlying transport at open time so wire-byte totals remain readable after socket closure
• Adds makeSqlStatementCounter, a tracer that increments a counter for spans named exactly sql.execute while returning native spans for all operations
• Extends formatTransferBudgetReport with informational (uncapped) rows for shell snapshot, shell-only and second-client WebSocket diagnostics, aggregate server egress, SQL counts, and reconnect catch-up modes alongside existing capped thread rows
• Adds subscription helpers subscribeThreadItems, subscribeShellItems, and awaitSubscriptionSynchronized in TransferBudgetScenario.integration.ts that run scoped subscriptions and classify initial catch-up as replay or snapshot
• Risk: makeWebSocketTransferRecorder open wait fails with a timeout defect after 10 seconds; awaitSubscriptionSynchronized similarly defects after 10 seconds if the synchronization marker never arrives

^{Macroscope summarized 7aff2b1.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.3 KiB	13.3 KiB	+15 B (+0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	+6 B (+0.1%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.4 KiB	+9 B (+0.1%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	55.6 KiB	55.6 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	13.3 KiB	13.3 KiB	+13 B (+0.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	−6 B (−0.1%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.4 KiB	6.4 KiB	+19 B (+0.3%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	56.4 KiB	56.4 KiB	0 B (0.0%)	66.4 KiB	✅
Claude	Live turn messages	10	10	0 (0.0%)	21	✅

Baseline: 80c708a · PR result: 7aff2b1 · Source CI: success

Scenario and decoded snapshot size

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

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

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

[macroscopeapp]

Approvability

Verdict: Approved at 7aff2b1

Macroscope's review found this PR approvable — This PR expands a server integration test to measure shell traffic, multiple clients, reconnects, and SQL activity. All changes are confined to test and reporting infrastructure, with no shipped production behavior or product-default changes.

Notes:

• No code objects were reviewed. Approvability was decided on eligibility alone.

^{You can add or adjust custom eligibility rules. Learn more.}

[t3dotgg]

The transfer budget measured one thread subscription during one turn. Real clients also hold the sidebar (shell) subscription, users often have a desktop and a phone connected at once, and clients reconnect. Nobody had measured those, and the larger perf proposals (sharing the sidebar projector across clients, sharing git services across connections, incremental history) are gated on knowing whether they matter.

This PR measures them in the same test and prints them next to the existing capped rows. The single-thread measurement and its caps are unchanged.

What it adds:

• A shell-only socket and a second client socket holding both the thread and shell subscriptions during the measured turn. The report shows per-socket wire bytes, decoded bytes, message counts, and total server egress.
• A reconnect after the turn. The second client disconnects and resubscribes each subscription from the cursor it held before the turn. The report shows catch-up bytes and whether the server chose replay or a snapshot reset.
• A SQL statement count from sql.execute spans. A counting tracer is installed on the harness runtime and the HTTP/WS app so reactor writes and subscription reads land in one total. No new instrumentation in src/.
• openMeasuredWsClient opens each RPC socket on its own child scope so clients disconnect independently.

The new rows have no caps. The shell stream coalesces on a 50 ms window, so the second client's message count moved between 9 and 10 across runs and its wire bytes between 6.7 and 7.1 KiB. SQL statement counts moved between 1096 and 1167 for the same reason. The capped thread rows stayed within a few bytes.

Measured (local, one run):

Provider	Phase	Metric	Observed	Budget	Result
codex	thread snapshot	HTTP wire	6.9 KiB (7,049 B)	7.3 KiB (7,500 B)	PASS
codex	measured turn	WebSocket wire	6.3 KiB (6,401 B)	7.8 KiB (8,000 B)	PASS
codex	measured turn	WebSocket decoded	54.7 KiB (55,974 B)	66.4 KiB (68,000 B)	PASS
codex	measured turn	WebSocket messages	8	21	PASS
codex	shell snapshot	HTTP wire	829 B	none	INFO
codex	measured turn, shell	WebSocket wire	628 B	none	INFO
codex	measured turn, shell	WebSocket decoded	3.7 KiB (3,793 B)	none	INFO
codex	measured turn, shell	WebSocket messages	3	none	INFO
codex	measured turn, second client	WebSocket wire	7.0 KiB (7,144 B)	none	INFO
codex	measured turn, second client	WebSocket decoded	58.3 KiB (59,723 B)	none	INFO
codex	measured turn, second client	WebSocket messages	10	none	INFO
codex	measured turn	server egress wire	13.8 KiB (14,173 B)	none	INFO
codex	measured turn	SQL statements	1096	none	INFO
codex	reconnect, thread (replay)	WebSocket wire	6.2 KiB (6,325 B)	none	INFO
codex	reconnect, thread (replay)	WebSocket decoded	54.4 KiB (55,707 B)	none	INFO
codex	reconnect, thread (replay)	WebSocket messages	2	none	INFO
codex	reconnect, shell (replay)	WebSocket wire	386 B	none	INFO
codex	reconnect, shell (replay)	WebSocket decoded	1.3 KiB (1,355 B)	none	INFO
codex	reconnect, shell (replay)	WebSocket messages	2	none	INFO
codex	reconnect	SQL statements	10	none	INFO
claudeAgent	thread snapshot	HTTP wire	6.9 KiB (7,059 B)	7.3 KiB (7,500 B)	PASS
claudeAgent	measured turn	WebSocket wire	6.3 KiB (6,409 B)	7.8 KiB (8,000 B)	PASS
claudeAgent	measured turn	WebSocket decoded	55.5 KiB (56,806 B)	66.4 KiB (68,000 B)	PASS
claudeAgent	measured turn	WebSocket messages	8	21	PASS
claudeAgent	shell snapshot	HTTP wire	829 B	none	INFO
claudeAgent	measured turn, shell	WebSocket wire	626 B	none	INFO
claudeAgent	measured turn, shell	WebSocket decoded	3.8 KiB (3,877 B)	none	INFO
claudeAgent	measured turn, shell	WebSocket messages	3	none	INFO
claudeAgent	measured turn, second client	WebSocket wire	7.0 KiB (7,161 B)	none	INFO
claudeAgent	measured turn, second client	WebSocket decoded	59.2 KiB (60,602 B)	none	INFO
claudeAgent	measured turn, second client	WebSocket messages	10	none	INFO
claudeAgent	measured turn	server egress wire	13.9 KiB (14,196 B)	none	INFO
claudeAgent	measured turn	SQL statements	1096	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket wire	6.2 KiB (6,342 B)	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket decoded	55.2 KiB (56,539 B)	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket messages	2	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket wire	379 B	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket decoded	1.4 KiB (1,383 B)	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket messages	2	none	INFO
claudeAgent	reconnect	SQL statements	10	none	INFO

Reading the numbers:

• The sidebar costs about 630 B wire and 3 messages per turn, about 10% of the thread stream. A second device costs one more thread stream plus one more shell stream, about 7 KiB. Egress scales linearly with connected clients and there is no hidden multiplier.
• Reconnecting from the pre-turn cursor replays about the same bytes as live delivery (6.3 KiB thread, 380 B shell) in 2 messages each and 10 SQL statements total. Replay was chosen every time at this gap size. Snapshot fallback did not trigger.
• This fixture mocks git, so it says nothing about sharing git services across connections.

Change authored by Claude Fable 5.1 running in Claude Code.

---

Note

Low Risk
Changes are confined to integration tests and reporting helpers; production server code is not modified.

Overview
Extends the thread transfer budget integration test so CI reports more than a single thread WebSocket during one turn. Existing capped thread snapshot and measured-turn rows are unchanged.

The scenario now opens three measured WebSocket clients (thread-only, shell-only, and a second device with both subscriptions), records shell HTTP snapshot size, and sums server egress across the three sockets. After the turn, the second client closes and resubscribes from its pre-turn cursors; the report records catch-up wire/decoded/message totals and whether catch-up used replay or snapshot.

Test harness plumbing adds openMeasuredWsClient (per-client forked scope + close), WebSocket recorder improvements (awaitOpen, wire totals after close), shared SQL statement counting via a counting Tracer on the orchestration harness and HTTP/WS app (sql.execute spans), and helpers moved into TransferBudgetScenario for thread/shell subscribe and sync detection. New report fields are INFO-only (no new budgets).

^{Reviewed by Cursor Bugbot for commit 7aff2b1. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Expand transfer-budget test to measure shell, second client, and reconnect transfer

• Reworks the live transfer-budget integration test in server.test.ts to open three concurrent measured clients (thread-only, shell-only, second-client), capture per-socket WebSocket transfer and server-wide sql.execute deltas, then close and reconnect the second client to measure replay-or-snapshot catch-up
• Adds openMeasuredWsClient, a factory that forks a per-client child scope, builds a counting WebSocket protocol, waits for open (10s timeout defect), and returns an explicit close effect
• Updates makeWebSocketTransferRecorder to retain the underlying transport at open time so wire-byte totals remain readable after socket closure
• Adds makeSqlStatementCounter, a tracer that increments a counter for spans named exactly sql.execute while returning native spans for all operations
• Extends formatTransferBudgetReport with informational (uncapped) rows for shell snapshot, shell-only and second-client WebSocket diagnostics, aggregate server egress, SQL counts, and reconnect catch-up modes alongside existing capped thread rows
• Adds subscription helpers subscribeThreadItems, subscribeShellItems, and awaitSubscriptionSynchronized in TransferBudgetScenario.integration.ts that run scoped subscriptions and classify initial catch-up as replay or snapshot
• Risk: makeWebSocketTransferRecorder open wait fails with a timeout defect after 10 seconds; awaitSubscriptionSynchronized similarly defects after 10 seconds if the synchronization marker never arrives

^{Macroscope summarized 7aff2b1.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.3 KiB	13.3 KiB	+15 B (+0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	+6 B (+0.1%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.4 KiB	+9 B (+0.1%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	55.6 KiB	55.6 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	13.3 KiB	13.3 KiB	+13 B (+0.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	−6 B (−0.1%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.4 KiB	6.4 KiB	+19 B (+0.3%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	56.4 KiB	56.4 KiB	0 B (0.0%)	66.4 KiB	✅
Claude	Live turn messages	10	10	0 (0.0%)	21	✅

Baseline: 80c708a · PR result: 7aff2b1 · Source CI: success

Scenario and decoded snapshot size

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

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

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

[macroscopeapp]

Approvability

Verdict: Approved at 7aff2b1

Macroscope's review found this PR approvable — This PR expands a server integration test to measure shell traffic, multiple clients, reconnects, and SQL activity. All changes are confined to test and reporting infrastructure, with no shipped production behavior or product-default changes.

Notes:

• No code objects were reviewed. Approvability was decided on eligibility alone.

^{You can add or adjust custom eligibility rules. Learn more.}

[t3dotgg]

The transfer budget measured one thread subscription during one turn. Real clients also hold the sidebar (shell) subscription, users often have a desktop and a phone connected at once, and clients reconnect. Nobody had measured those, and the larger perf proposals (sharing the sidebar projector across clients, sharing git services across connections, incremental history) are gated on knowing whether they matter.

This PR measures them in the same test and prints them next to the existing capped rows. The single-thread measurement and its caps are unchanged.

What it adds:

• A shell-only socket and a second client socket holding both the thread and shell subscriptions during the measured turn. The report shows per-socket wire bytes, decoded bytes, message counts, and total server egress.
• A reconnect after the turn. The second client disconnects and resubscribes each subscription from the cursor it held before the turn. The report shows catch-up bytes and whether the server chose replay or a snapshot reset.
• A SQL statement count from sql.execute spans. A counting tracer is installed on the harness runtime and the HTTP/WS app so reactor writes and subscription reads land in one total. No new instrumentation in src/.
• openMeasuredWsClient opens each RPC socket on its own child scope so clients disconnect independently.

The new rows have no caps. The shell stream coalesces on a 50 ms window, so the second client's message count moved between 9 and 10 across runs and its wire bytes between 6.7 and 7.1 KiB. SQL statement counts moved between 1096 and 1167 for the same reason. The capped thread rows stayed within a few bytes.

Measured (local, one run):

Provider	Phase	Metric	Observed	Budget	Result
codex	thread snapshot	HTTP wire	6.9 KiB (7,049 B)	7.3 KiB (7,500 B)	PASS
codex	measured turn	WebSocket wire	6.3 KiB (6,401 B)	7.8 KiB (8,000 B)	PASS
codex	measured turn	WebSocket decoded	54.7 KiB (55,974 B)	66.4 KiB (68,000 B)	PASS
codex	measured turn	WebSocket messages	8	21	PASS
codex	shell snapshot	HTTP wire	829 B	none	INFO
codex	measured turn, shell	WebSocket wire	628 B	none	INFO
codex	measured turn, shell	WebSocket decoded	3.7 KiB (3,793 B)	none	INFO
codex	measured turn, shell	WebSocket messages	3	none	INFO
codex	measured turn, second client	WebSocket wire	7.0 KiB (7,144 B)	none	INFO
codex	measured turn, second client	WebSocket decoded	58.3 KiB (59,723 B)	none	INFO
codex	measured turn, second client	WebSocket messages	10	none	INFO
codex	measured turn	server egress wire	13.8 KiB (14,173 B)	none	INFO
codex	measured turn	SQL statements	1096	none	INFO
codex	reconnect, thread (replay)	WebSocket wire	6.2 KiB (6,325 B)	none	INFO
codex	reconnect, thread (replay)	WebSocket decoded	54.4 KiB (55,707 B)	none	INFO
codex	reconnect, thread (replay)	WebSocket messages	2	none	INFO
codex	reconnect, shell (replay)	WebSocket wire	386 B	none	INFO
codex	reconnect, shell (replay)	WebSocket decoded	1.3 KiB (1,355 B)	none	INFO
codex	reconnect, shell (replay)	WebSocket messages	2	none	INFO
codex	reconnect	SQL statements	10	none	INFO
claudeAgent	thread snapshot	HTTP wire	6.9 KiB (7,059 B)	7.3 KiB (7,500 B)	PASS
claudeAgent	measured turn	WebSocket wire	6.3 KiB (6,409 B)	7.8 KiB (8,000 B)	PASS
claudeAgent	measured turn	WebSocket decoded	55.5 KiB (56,806 B)	66.4 KiB (68,000 B)	PASS
claudeAgent	measured turn	WebSocket messages	8	21	PASS
claudeAgent	shell snapshot	HTTP wire	829 B	none	INFO
claudeAgent	measured turn, shell	WebSocket wire	626 B	none	INFO
claudeAgent	measured turn, shell	WebSocket decoded	3.8 KiB (3,877 B)	none	INFO
claudeAgent	measured turn, shell	WebSocket messages	3	none	INFO
claudeAgent	measured turn, second client	WebSocket wire	7.0 KiB (7,161 B)	none	INFO
claudeAgent	measured turn, second client	WebSocket decoded	59.2 KiB (60,602 B)	none	INFO
claudeAgent	measured turn, second client	WebSocket messages	10	none	INFO
claudeAgent	measured turn	server egress wire	13.9 KiB (14,196 B)	none	INFO
claudeAgent	measured turn	SQL statements	1096	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket wire	6.2 KiB (6,342 B)	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket decoded	55.2 KiB (56,539 B)	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket messages	2	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket wire	379 B	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket decoded	1.4 KiB (1,383 B)	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket messages	2	none	INFO
claudeAgent	reconnect	SQL statements	10	none	INFO

Reading the numbers:

• The sidebar costs about 630 B wire and 3 messages per turn, about 10% of the thread stream. A second device costs one more thread stream plus one more shell stream, about 7 KiB. Egress scales linearly with connected clients and there is no hidden multiplier.
• Reconnecting from the pre-turn cursor replays about the same bytes as live delivery (6.3 KiB thread, 380 B shell) in 2 messages each and 10 SQL statements total. Replay was chosen every time at this gap size. Snapshot fallback did not trigger.
• This fixture mocks git, so it says nothing about sharing git services across connections.

Change authored by Claude Fable 5.1 running in Claude Code.

---

Note

Low Risk
Changes are confined to integration tests and reporting helpers; production server code is not modified.

Overview
Extends the thread transfer budget integration test so CI reports more than a single thread WebSocket during one turn. Existing capped thread snapshot and measured-turn rows are unchanged.

The scenario now opens three measured WebSocket clients (thread-only, shell-only, and a second device with both subscriptions), records shell HTTP snapshot size, and sums server egress across the three sockets. After the turn, the second client closes and resubscribes from its pre-turn cursors; the report records catch-up wire/decoded/message totals and whether catch-up used replay or snapshot.

Test harness plumbing adds openMeasuredWsClient (per-client forked scope + close), WebSocket recorder improvements (awaitOpen, wire totals after close), shared SQL statement counting via a counting Tracer on the orchestration harness and HTTP/WS app (sql.execute spans), and helpers moved into TransferBudgetScenario for thread/shell subscribe and sync detection. New report fields are INFO-only (no new budgets).

^{Reviewed by Cursor Bugbot for commit 7aff2b1. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Expand transfer-budget test to measure shell, second client, and reconnect transfer

• Reworks the live transfer-budget integration test in server.test.ts to open three concurrent measured clients (thread-only, shell-only, second-client), capture per-socket WebSocket transfer and server-wide sql.execute deltas, then close and reconnect the second client to measure replay-or-snapshot catch-up
• Adds openMeasuredWsClient, a factory that forks a per-client child scope, builds a counting WebSocket protocol, waits for open (10s timeout defect), and returns an explicit close effect
• Updates makeWebSocketTransferRecorder to retain the underlying transport at open time so wire-byte totals remain readable after socket closure
• Adds makeSqlStatementCounter, a tracer that increments a counter for spans named exactly sql.execute while returning native spans for all operations
• Extends formatTransferBudgetReport with informational (uncapped) rows for shell snapshot, shell-only and second-client WebSocket diagnostics, aggregate server egress, SQL counts, and reconnect catch-up modes alongside existing capped thread rows
• Adds subscription helpers subscribeThreadItems, subscribeShellItems, and awaitSubscriptionSynchronized in TransferBudgetScenario.integration.ts that run scoped subscriptions and classify initial catch-up as replay or snapshot
• Risk: makeWebSocketTransferRecorder open wait fails with a timeout defect after 10 seconds; awaitSubscriptionSynchronized similarly defects after 10 seconds if the synchronization marker never arrives

^{Macroscope summarized 7aff2b1.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.3 KiB	13.3 KiB	+15 B (+0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	+6 B (+0.1%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.4 KiB	+9 B (+0.1%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	55.6 KiB	55.6 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	13.3 KiB	13.3 KiB	+13 B (+0.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	−6 B (−0.1%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.4 KiB	6.4 KiB	+19 B (+0.3%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	56.4 KiB	56.4 KiB	0 B (0.0%)	66.4 KiB	✅
Claude	Live turn messages	10	10	0 (0.0%)	21	✅

Baseline: 80c708a · PR result: 7aff2b1 · Source CI: success

Scenario and decoded snapshot size

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

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

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

[macroscopeapp]

Approvability

Verdict: Approved at 7aff2b1

Macroscope's review found this PR approvable — This PR expands a server integration test to measure shell traffic, multiple clients, reconnects, and SQL activity. All changes are confined to test and reporting infrastructure, with no shipped production behavior or product-default changes.

Notes:

• No code objects were reviewed. Approvability was decided on eligibility alone.

^{You can add or adjust custom eligibility rules. Learn more.}

[t3dotgg]

The transfer budget measured one thread subscription during one turn. Real clients also hold the sidebar (shell) subscription, users often have a desktop and a phone connected at once, and clients reconnect. Nobody had measured those, and the larger perf proposals (sharing the sidebar projector across clients, sharing git services across connections, incremental history) are gated on knowing whether they matter.

This PR measures them in the same test and prints them next to the existing capped rows. The single-thread measurement and its caps are unchanged.

What it adds:

• A shell-only socket and a second client socket holding both the thread and shell subscriptions during the measured turn. The report shows per-socket wire bytes, decoded bytes, message counts, and total server egress.
• A reconnect after the turn. The second client disconnects and resubscribes each subscription from the cursor it held before the turn. The report shows catch-up bytes and whether the server chose replay or a snapshot reset.
• A SQL statement count from sql.execute spans. A counting tracer is installed on the harness runtime and the HTTP/WS app so reactor writes and subscription reads land in one total. No new instrumentation in src/.
• openMeasuredWsClient opens each RPC socket on its own child scope so clients disconnect independently.

The new rows have no caps. The shell stream coalesces on a 50 ms window, so the second client's message count moved between 9 and 10 across runs and its wire bytes between 6.7 and 7.1 KiB. SQL statement counts moved between 1096 and 1167 for the same reason. The capped thread rows stayed within a few bytes.

Measured (local, one run):

Provider	Phase	Metric	Observed	Budget	Result
codex	thread snapshot	HTTP wire	6.9 KiB (7,049 B)	7.3 KiB (7,500 B)	PASS
codex	measured turn	WebSocket wire	6.3 KiB (6,401 B)	7.8 KiB (8,000 B)	PASS
codex	measured turn	WebSocket decoded	54.7 KiB (55,974 B)	66.4 KiB (68,000 B)	PASS
codex	measured turn	WebSocket messages	8	21	PASS
codex	shell snapshot	HTTP wire	829 B	none	INFO
codex	measured turn, shell	WebSocket wire	628 B	none	INFO
codex	measured turn, shell	WebSocket decoded	3.7 KiB (3,793 B)	none	INFO
codex	measured turn, shell	WebSocket messages	3	none	INFO
codex	measured turn, second client	WebSocket wire	7.0 KiB (7,144 B)	none	INFO
codex	measured turn, second client	WebSocket decoded	58.3 KiB (59,723 B)	none	INFO
codex	measured turn, second client	WebSocket messages	10	none	INFO
codex	measured turn	server egress wire	13.8 KiB (14,173 B)	none	INFO
codex	measured turn	SQL statements	1096	none	INFO
codex	reconnect, thread (replay)	WebSocket wire	6.2 KiB (6,325 B)	none	INFO
codex	reconnect, thread (replay)	WebSocket decoded	54.4 KiB (55,707 B)	none	INFO
codex	reconnect, thread (replay)	WebSocket messages	2	none	INFO
codex	reconnect, shell (replay)	WebSocket wire	386 B	none	INFO
codex	reconnect, shell (replay)	WebSocket decoded	1.3 KiB (1,355 B)	none	INFO
codex	reconnect, shell (replay)	WebSocket messages	2	none	INFO
codex	reconnect	SQL statements	10	none	INFO
claudeAgent	thread snapshot	HTTP wire	6.9 KiB (7,059 B)	7.3 KiB (7,500 B)	PASS
claudeAgent	measured turn	WebSocket wire	6.3 KiB (6,409 B)	7.8 KiB (8,000 B)	PASS
claudeAgent	measured turn	WebSocket decoded	55.5 KiB (56,806 B)	66.4 KiB (68,000 B)	PASS
claudeAgent	measured turn	WebSocket messages	8	21	PASS
claudeAgent	shell snapshot	HTTP wire	829 B	none	INFO
claudeAgent	measured turn, shell	WebSocket wire	626 B	none	INFO
claudeAgent	measured turn, shell	WebSocket decoded	3.8 KiB (3,877 B)	none	INFO
claudeAgent	measured turn, shell	WebSocket messages	3	none	INFO
claudeAgent	measured turn, second client	WebSocket wire	7.0 KiB (7,161 B)	none	INFO
claudeAgent	measured turn, second client	WebSocket decoded	59.2 KiB (60,602 B)	none	INFO
claudeAgent	measured turn, second client	WebSocket messages	10	none	INFO
claudeAgent	measured turn	server egress wire	13.9 KiB (14,196 B)	none	INFO
claudeAgent	measured turn	SQL statements	1096	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket wire	6.2 KiB (6,342 B)	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket decoded	55.2 KiB (56,539 B)	none	INFO
claudeAgent	reconnect, thread (replay)	WebSocket messages	2	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket wire	379 B	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket decoded	1.4 KiB (1,383 B)	none	INFO
claudeAgent	reconnect, shell (replay)	WebSocket messages	2	none	INFO
claudeAgent	reconnect	SQL statements	10	none	INFO

Reading the numbers:

• The sidebar costs about 630 B wire and 3 messages per turn, about 10% of the thread stream. A second device costs one more thread stream plus one more shell stream, about 7 KiB. Egress scales linearly with connected clients and there is no hidden multiplier.
• Reconnecting from the pre-turn cursor replays about the same bytes as live delivery (6.3 KiB thread, 380 B shell) in 2 messages each and 10 SQL statements total. Replay was chosen every time at this gap size. Snapshot fallback did not trigger.
• This fixture mocks git, so it says nothing about sharing git services across connections.

Change authored by Claude Fable 5.1 running in Claude Code.

---

Note

Low Risk
Changes are confined to integration tests and reporting helpers; production server code is not modified.

Overview
Extends the thread transfer budget integration test so CI reports more than a single thread WebSocket during one turn. Existing capped thread snapshot and measured-turn rows are unchanged.

The scenario now opens three measured WebSocket clients (thread-only, shell-only, and a second device with both subscriptions), records shell HTTP snapshot size, and sums server egress across the three sockets. After the turn, the second client closes and resubscribes from its pre-turn cursors; the report records catch-up wire/decoded/message totals and whether catch-up used replay or snapshot.

Test harness plumbing adds openMeasuredWsClient (per-client forked scope + close), WebSocket recorder improvements (awaitOpen, wire totals after close), shared SQL statement counting via a counting Tracer on the orchestration harness and HTTP/WS app (sql.execute spans), and helpers moved into TransferBudgetScenario for thread/shell subscribe and sync detection. New report fields are INFO-only (no new budgets).

^{Reviewed by Cursor Bugbot for commit 7aff2b1. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Expand transfer-budget test to measure shell, second client, and reconnect transfer

• Reworks the live transfer-budget integration test in server.test.ts to open three concurrent measured clients (thread-only, shell-only, second-client), capture per-socket WebSocket transfer and server-wide sql.execute deltas, then close and reconnect the second client to measure replay-or-snapshot catch-up
• Adds openMeasuredWsClient, a factory that forks a per-client child scope, builds a counting WebSocket protocol, waits for open (10s timeout defect), and returns an explicit close effect
• Updates makeWebSocketTransferRecorder to retain the underlying transport at open time so wire-byte totals remain readable after socket closure
• Adds makeSqlStatementCounter, a tracer that increments a counter for spans named exactly sql.execute while returning native spans for all operations
• Extends formatTransferBudgetReport with informational (uncapped) rows for shell snapshot, shell-only and second-client WebSocket diagnostics, aggregate server egress, SQL counts, and reconnect catch-up modes alongside existing capped thread rows
• Adds subscription helpers subscribeThreadItems, subscribeShellItems, and awaitSubscriptionSynchronized in TransferBudgetScenario.integration.ts that run scoped subscriptions and classify initial catch-up as replay or snapshot
• Risk: makeWebSocketTransferRecorder open wait fails with a timeout defect after 10 seconds; awaitSubscriptionSynchronized similarly defects after 10 seconds if the synchronization marker never arrives

^{Macroscope summarized 7aff2b1.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.3 KiB	13.3 KiB	+15 B (+0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	+6 B (+0.1%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.4 KiB	+9 B (+0.1%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	55.6 KiB	55.6 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	13.3 KiB	13.3 KiB	+13 B (+0.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	−6 B (−0.1%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.4 KiB	6.4 KiB	+19 B (+0.3%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	56.4 KiB	56.4 KiB	0 B (0.0%)	66.4 KiB	✅
Claude	Live turn messages	10	10	0 (0.0%)	21	✅

Baseline: 80c708a · PR result: 7aff2b1 · Source CI: success

Scenario and decoded snapshot size

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

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

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

[macroscopeapp]

Approvability

Verdict: Approved at 7aff2b1

Macroscope's review found this PR approvable — This PR expands a server integration test to measure shell traffic, multiple clients, reconnects, and SQL activity. All changes are confined to test and reporting infrastructure, with no shipped production behavior or product-default changes.

Notes:

• No code objects were reviewed. Approvability was decided on eligibility alone.

^{You can add or adjust custom eligibility rules. Learn more.}