Promote expbkmain: bkt3 performance work (PRs #214–#239)

[tusharbhardwaj-bk]

Do not merge without warning the team. A bkmain deploy restarts t3-bkmain.service and ends every live session on bkt3. The human merges this.

What this promotes

36 commits from PRs #214–#223 and #225–#234, all merged to expbkmain and verified on expbkt3. Nothing else: git log --no-merges origin/bkmain..origin/expbkmain shows only these, and bkmain has no commits that expbkmain lacks. The 9 commits by other authors are the upstream cherry-picks in #222.

Server: fewer statements, less event-loop blocking

PR	Change	Measured
#214	Drop duplicate provider status events	10 identical pings → 1 event; real turn on expbkt3: 0 duplicates, against 58% before (77% in the prod audit)
#215	Startup reconcile reads only waiting threads; per-phase startup timing logs	28.7 s cold / 0.9 s warm → 0.4 ms on a prod copy; expbkt3 boots to Listening in 3–4 s
#227	Execution overlay batched	588 → 3 statements per shell read; 61 → 22 ms on a prod copy
#228	Shell resume filters visibility once	~1,576 → ~11 statements per team-mode reconnect (with #227)
#229	Reaper decodes only live bindings	1,355 → 3 rows; 37–67 ms → 1–4 ms (prod p90 was 1.86 s); expbkt3 bindingsMs: 1
#219	Thread cost reads only its own transcripts; 23 MB cache write debounced	per-thread file listing; cache rewrite at most every 5 min
#218	PRAGMA optimize every 6 h (never at startup: the first run is 2.8 s)	plans unchanged for the 7 hot queries
#234	Summary reactors skip activity payloads	58–206 → 33–40 ms per load (feature is off in prod)
#222	9 upstream cherry-picks	no provider refresh per connect, git process cap, bounded provider logs, cheaper checkpoints, etc.

New fork APIs for the Linear bridge (switched off until the gated steps below)

• feat(server): event feed for followers #216 GET /api/orchestration/events: an event feed that replaces the bridge's 2 s shell polling.
• feat(server): accept pull-request state from an external syncer #217 POST /api/orchestration/pull-request-state, the external-sync:write scope with --with-scope, and the T3_EXTERNAL_PR_SYNC switch (not set).
• fix(server): a retried pull-request-state delivery never fails with 500 #225 A retried PR-state delivery never returns 500. A stress run found 225 errors in 1,841 writes; after the fix there are 0 in 1,912.

Client (reaches desktop users only with the next desktop build)

• perf(web): window focus probes instead of resyncing; desktop wake reconnects at once #220 A quick alt-tab only probes the connection; desktop resume/unlock reconnects at once.
• fix(client-runtime): a busy thread stream no longer runs out of retries #230 A busy thread stream no longer goes dormant until reconnect: the retry budget resets after 10 s of health.
• perf(client-runtime): context-window updates stop re-sorting a thread's history #231 Context-window updates stop re-sorting the whole history: replay of 1,000 events on a 5.5k-activity thread goes from ~355 ms to ~50 ms.
• perf(web): history deepening waits for a quiet, visible thread #232 History deepening pauses while a thread runs, while the window is hidden, and for 10 s after a thread goes live.

CI and deploy after the repository move to iamtushar324/bkt3code

• fix(deploy): auto-deploy follows repository redirects #223 auto-deploy.sh (expbkt3, bkt3 and t3.dev) follows the repository-move redirect and waits on odd responses instead of crashing.
• ci: run every workflow on GitHub-hosted runners #226 All workflows run on GitHub-hosted runners (Blacksmith isn't installed on the new account).
• ci: fork-marker check ignores lines identical to upstream #221 The fork-marker check accepts lines identical to upstream (needed for cherry-pick -x).
• test(mobile): de-flake the cold Shiki highlighting test #233 De-flakes a timing-sensitive mobile Shiki test.

Verified on expbkt3 (ed3c051fd, deployed 2026-09-27 07:01 UTC)

• Deploy timer: installs new builds automatically again (it deferred correctly while validation ran).
• Startup: 3 s to Listening, with every phase timed in the journal.
• Event feed: paging, limit, wait, 410, 400 and 401 all correct.
• PR-state: no match, bad body, and a replay ignored as unchanged all correct.
• 3-minute stress run:
  • load: 40 long-polling followers, 15 commands/s, and 10 concurrent PR-state writers;
  • results: 99,031 feed pages with 0 gaps, 0 duplicates, 0 ordering errors and 0 failed requests; 1,912 of 1,912 PR-state writes returned 200; 511 of 511 commands succeeded;
  • memory returns to baseline after the load (412 MB after ~2 min idle; not a leak).
• Real Claude turn: completed in 15 s, with 0 duplicate session events.

After merging (human steps)

1. bkt3 won't auto-deploy this merge. bkmain's current auto-deploy.sh still has the redirect bug that fix(deploy): auto-deploy follows repository redirects #223 fixes. Once the deploy-bkt3.yml push run for the merge commit is green, run one manual deploy from a human shell or the expbkt3 instance, never from a bkt3 session:

   T3_DEPLOY_GH_TOKEN_FILE=/home/ubuntu/.config/t3-deploy/gh-token \
     /home/ubuntu/.local/bin/t3-deploy-with-github-token \
     /home/ubuntu/repos/t3code-bkmain/deploy/bkt3/deploy.sh <merge-sha> <run-id>

   The same deploy fixes the t3.dev timer, whose script runs from this checkout.

2. Biggest remaining prod win, gated:

   • give the bridge a token with external-sync:write (docs/operations/external-pr-sync.md);
   • deploy the bridge's event-feed follower and PR-state sender (session B);
   • then set T3_EXTERNAL_PR_SYNC=1 in deploy/bkt3/start.sh.

   That removes the per-minute PR sweep (92% of git spawns on bkt3) and the bridge's shell polling.

3. Cut a desktop build, so desktop users get perf(web): window focus probes instead of resyncing; desktop wake reconnects at once #220, perf: cherry-pick upstream connection, git and checkpoint fixes #222 (keep-alive), fix(client-runtime): a busy thread stream no longer runs out of retries #230, perf(client-runtime): context-window updates stop re-sorting a thread's history #231 and perf(web): history deepening waits for a quiet, visible thread #232.

Model/harness: Claude Opus 5.5 via Claude Code in T3 Code.

Added 2026-09-27: reconnect and catch-up lag (#236–#239)

Tushar reported that returning to the desktop app and syncing new messages still wasn't instant. Measured on expbkt3 and fixed:

PR	What was slow	Before	After
#236	Catch-up replay went one event per WebSocket frame, each waiting for the client's ack	500 missed events: 17.9 s at a 35 ms round trip (493 frames)	0.3 s (6 frames)
#237	Desktop window open waited on the bundled local backend before connecting to the primary	~10.5 s gap on 2 of 3 opens in the traces	at most 1.5 s
#238	One return to the window resynced twice (focus + visibility), and a return after 60 s away only reached the first subscriber	shell and 4 threads resubscribed twice, 2 s apart	once, for every subscriber
#239	An execution frame batched with events was dropped, which could leave a chat stuck on "Running"	dropped	applied

This promotion also carries #235 (worktree setup progress row). The desktop fixes (#237, #238, #239) reach users through a desktop build; staging build v0.0.41-staging-nightly.20260927.8 has them.

🤖 Generated with Claude Code

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

ℹ️ No successful main baseline artifact is available yet. This run establishes the initial measurement.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	—	13.8 KiB	—	15.1 KiB	✅
Codex	Thread snapshot wire	—	7.3 KiB	—	7.8 KiB	✅
Codex	Live turn WebSocket wire	—	6.5 KiB	—	7.8 KiB	✅
Codex	Live turn WebSocket decoded	—	57.1 KiB	—	66.4 KiB	✅
Codex	Live turn messages	—	9	—	21	✅
Claude	Total thread wire	—	13.8 KiB	—	15.1 KiB	✅
Claude	Thread snapshot wire	—	7.3 KiB	—	7.8 KiB	✅
Claude	Live turn WebSocket wire	—	6.5 KiB	—	7.8 KiB	✅
Claude	Live turn WebSocket decoded	—	57.9 KiB	—	66.4 KiB	✅
Claude	Live turn messages	—	9	—	21	✅

Baseline: unavailable · PR result: e1d692c · 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: 114.9 KiB
• Claude decoded thread snapshot: 115.6 KiB

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