T3 Code · Performance over time

Releases of the performance fork, oldest to newest. Each one is measured against the upstream code it is based on.

v0.1.9-perf.9: 0 regressed, 4 improved, 2 within noise (vs upstream base c2e54e3)

16 Nov 2026. Against its upstream base, built from source at c2e54e3. Linux x64, 50 agents. Tag on GitHub.

9 / 10
Agents
4improved beyond noise
0regressed beyond noise
2within noise
MetricUpstream base c2e54e3Fork perf.9worsebetterChangeVerdict
Event loop utilisation, mean (lower is better)0.960.17
−81.8%improvedimproved beyond noise
Reader lock wait p95 (lower is better)26.0 ms5.03 ms
−80.6%improvedimproved beyond noise
Main thread busy (lower is better)69.0%19.6%
−71.6%improvedimproved beyond noise
Turns completed in the window (higher is better)188250
+33.0%improvedimproved beyond noise
Peak RSS (lower is better)1,069 MiB1,101 MiB
+3.0%within noisewithin noise
Start to database ready, warm start (lower is better)1,535 ms1,536 ms
+0.1%within noisewithin noise
Dispatch p95 (lower is better)1,428 ms142 ms
−90.1%no verdictnoise band too wide for a verdict
Turn end to end p50 (lower is better)15,786 ms5,855 ms
−62.9%no verdictnoise band too wide for a verdict

Other metrics beyond noise

MetricUpstream base c2e54e3Fork perf.9worsebetterChangeVerdict
Process CPU per served agent (lower is better)0.10% per agent0.34% per agent
+234.0%regressedregressed beyond noise
Main thread voluntary context switches (lower is better)204/s536/s
+162.4%regressedregressed beyond noise
Main thread read syscalls (lower is better)844/s1,569/s
+86.0%regressedregressed beyond noise
Warning log lines (lower is better)232
regressedregressed beyond noise
Main thread minor faults (lower is better)29,590/s512/s
−98.3%improvedimproved beyond noise
MCP request authentication p95 (lower is better)39.8 ms1.32 ms
−96.7%improvedimproved beyond noise
Main thread system CPU (lower is better)41.9%1.51%
−96.4%improvedimproved beyond noise
Provider turn start p50 (lower is better)955 ms61.1 ms
−93.6%improvedimproved beyond noise
Child processes CPU (lower is better)98.1%8.44%
−91.4%improvedimproved beyond noise
Writer transaction, mean (lower is better)62.2 ms5.85 ms
−90.6%improvedimproved beyond noise
Reader statement p95 (lower is better)39.9 ms3.89 ms
−90.3%improvedimproved beyond noise
Writer transaction p95 (lower is better)171 ms16.9 ms
−90.1%improvedimproved beyond noise
Event loop delay max (server gauge) (lower is better)18 ms2 ms
−88.9%improvedimproved beyond noise
Process read syscalls (lower is better)44,165/s5,958/s
−86.5%improvedimproved beyond noise
Reader transaction, mean (lower is better)33.8 ms4.57 ms
−86.5%improvedimproved beyond noise
SQL statement time per second (lower is better)407 ms/s58.6 ms/s
−85.6%improvedimproved beyond noise
Reader transaction p95 (lower is better)76.7 ms12.4 ms
−83.9%improvedimproved beyond noise
Run update handling p95 (lower is better)160 ms26.2 ms
−83.6%improvedimproved beyond noise
Main thread involuntary context switches (lower is better)3.42/s0.56/s
−83.5%improvedimproved beyond noise
Reader lock wait, mean (lower is better)3.90 ms0.86 ms
−77.9%improvedimproved beyond noise
Thread shell read p95 (lower is better)67.5 ms17.1 ms
−74.7%improvedimproved beyond noise
Writer lock wait p95 (lower is better)20.5 ms5.53 ms
−73.0%improvedimproved beyond noise
SIGTERM to exit (lower is better)7.86 s2.27 s
−71.1%improvedimproved beyond noise
GC pause per second (lower is better)12.0 ms/s3.60 ms/s
−69.9%improvedimproved beyond noise
Process CPU, all threads (lower is better)104%32.6%
−68.7%improvedimproved beyond noise
Worker threads CPU (lower is better)28.1%9.56%
−66.0%improvedimproved beyond noise
Thread snapshot window p95 (lower is better)453 ms170 ms
−62.5%improvedimproved beyond noise
Event loop utilisation, max (lower is better)10.38
−62.3%improvedimproved beyond noise
Event loop delay max (preload histogram) (lower is better)67.6 ms33.8 ms
−50.1%improvedimproved beyond noise
Event loop delay p99 (preload histogram) (lower is better)22.4 ms11.5 ms
−48.5%improvedimproved beyond noise
Main thread run-queue wait (lower is better)0.08 ms/s0.04 ms/s
−42.3%improvedimproved beyond noise
Main thread user CPU (lower is better)27.0%18.1%
−33.1%improvedimproved beyond noise
Full snapshot GET of the largest thread (lower is better)20,740 ms15,547 ms
−25.0%improvedimproved beyond noise
Event loop delay p50 (preload histogram) (lower is better)13.2 ms10.1 ms
−23.8%improvedimproved beyond noise
Peak V8 heap used (lower is better)595 MiB456 MiB
−23.4%improvedimproved beyond noise

What moved.

Event loop utilisation, mean at 50 agents: 0.96 on the upstream base, 0.17 on perf.9, −81.8%, improved beyond noise. Reader lock wait p95 at 1 agent: 0.87 ms on the upstream base, 1.47 ms on perf.9, +69.0%, regressed beyond noise. Peak RSS at 1 agent: 538 MiB on the upstream base, 600 MiB on perf.9, +11.6%, regressed beyond noise.

Main thread minor faults at 50 agents: 29,590/s on the upstream base, 512/s on perf.9, −98.3%, improved beyond noise.

Table: Main thread minor faults, at each load level
AgentsUpstream base c2e54e3Fork perf.9
11,479 /s21.9 /s
5029,590 /s512 /s
10026,549 /s1,117 /s
15030,577 /s1,194 /s
20032,669 /s1,389 /s
25032,922 /s1,495 /s

Event loop utilisation, max at 50 agents: 1 on the upstream base, 0.38 on perf.9, −62.3%, improved beyond noise.

Table: Event loop utilisation, max, at each load level
AgentsUpstream base c2e54e3Fork perf.9
10.0640.035
5010.38
10010.74
15010.84
20010.90
25010.96

Event loop utilisation, mean at 50 agents: 0.96 on the upstream base, 0.17 on perf.9, −81.8%, improved beyond noise.

Table: Event loop utilisation, mean, at each load level
AgentsUpstream base c2e54e3Fork perf.9
10.0590.029
500.960.17
10010.36
15010.43
20010.53
25010.58

Main thread system CPU at 50 agents: 41.9% on the upstream base, 1.51% on perf.9, −96.4%, improved beyond noise.

Table: Main thread system CPU, at each load level
AgentsUpstream base c2e54e3Fork perf.9
11.74%0.32%
5041.9%1.51%
10049.6%2.55%
15047.5%3.06%
20046.6%3.55%
25046.9%4.20%

Main thread busy at 50 agents: 69.0% on the upstream base, 19.6% on perf.9, −71.6%, improved beyond noise.

Table: Main thread busy, at each load level
AgentsUpstream base c2e54e3Fork perf.9
14.98%3.16%
5069.0%19.6%
10069.2%32.7%
15072.1%42.0%
20074.1%51.0%
25074.4%62.3%

Process CPU per served agent: 0.10% per agent on the upstream base, 0.34% per agent on perf.9, +234.0%, regressed beyond noise.

Table: Process CPU per served agent, the last 30 records
DateTagMedianMin to maxValidvs up to 7 previousVerdict
18 Nov 2026v0.1.11-nightly.20261118.49030.10%0.10 to 0.101−13.4%within noise
17 Nov 2026v0.1.11-nightly.20261117.48510.11%0.11 to 0.111−26.9%improved beyond noise
16 Nov 2026v0.1.10-nightly.20261116.48000.10%0.10 to 0.101−42.6%improved beyond noise
15 Nov 2026v0.1.10-nightly.20261115.47490.12%0.12 to 0.121−44.4%improved beyond noise
14 Nov 2026v0.1.10-nightly.20261114.47090.095%0.095 to 0.0951−56.8%improved beyond noise
13 Nov 2026v0.1.10-nightly.20261113.46560.15%0.15 to 0.151−35.8%improved beyond noise
11 Nov 2026v0.1.10-nightly.20261111.45440.49%0.49 to 0.491+116.4%regressed beyond noise
10 Nov 2026v0.1.9-nightly.20261110.44910.99%0.99 to 0.991+338.5%regressed beyond noise
9 Nov 2026v0.1.9-nightly.20261109.44431.14%1.14 to 1.141+404.0%regressed beyond noise
8 Nov 2026v0.1.8-nightly.20261108.43800.56%0.56 to 0.561+149.1%regressed beyond noise
7 Nov 2026v0.1.8-nightly.20261107.43400.59%0.59 to 0.591+160.2%regressed beyond noise
6 Nov 2026v0.1.8-nightly.20261106.42881.04%1.04 to 1.041+361.1%regressed beyond noise
5 Nov 2026v0.1.8-nightly.20261105.42340.23%0.23 to 0.231+7.9%within noise, thin reference
4 Nov 2026v0.1.8-nightly.20261104.41920.21%0.21 to 0.211−5.3%within noise, thin reference
3 Nov 2026v0.1.7-nightly.20261103.41380.23%0.23 to 0.231no reference yet
2 Nov 2026v0.1.7-nightly.20261102.40931.14%1.14 to 1.141no reference yet
1 Nov 2026v0.1.7-nightly.20261101.40481.15%1.15 to 1.151no reference yet
31 Oct 2026v0.1.7-nightly.20261031.40041.16%1.16 to 1.161no reference yet
30 Oct 2026v0.1.6-nightly.20261030.39401.25%1.25 to 1.251no reference yet

Process CPU, all threads at 50 agents: 104% on the upstream base, 32.6% on perf.9, −68.7%, improved beyond noise.

Table: Process CPU, all threads, at each load level
AgentsUpstream base c2e54e3Fork perf.9
18.84%5.93%
50104%32.6%
10088.7%55.9%
15098.4%71.1%
200107%85.9%
250105%104%

Child processes CPU at 50 agents: 98.1% on the upstream base, 8.44% on perf.9, −91.4%, improved beyond noise.

Table: Child processes CPU, at each load level
AgentsUpstream base c2e54e3Fork perf.9
12.73%1.83%
5098.1%8.44%
100145%13.5%
150162%17.2%
200156%23.2%
250148%30.2%

From the fork's changelog

  • Queued-message delivery, completion wakes and restart recovery
  • Upstream T3 Code 0.1.9 at main of 2026-11-15, with upstream migrations 93 to 95; Node 26.13.1.
  • Migrations: upstream's migrations 93 to 95 apply in order, and the fork's migrations leave the message-ordinal index to upstream's migration 94.
  • Queued messages: a message queued while a run is busy steers into it after 60 seconds, whether a person, an automation or another thread queued it. Due messages merge into one steer, oldest first and within the provider's text limit, with slash commands sent on their own; a turn takes at most one automatic steer a minute. Send now is offered on a waiting run only when the server can end its wait.
  • Completion and pull request wakes: a thread's queued completion wakes, and its pull request watch wakes, merge into one wake of at most 25 results. A wake whose start failed, or that a restart cut, sends its results again with backoff, at most eight times; merged pull request news survives a restart.
  • Restart recovery holds only the messages a person queued (unattended work only when continuation is off) and wakes a delegated seat whose waiter the restart cut.
  • Run ends are handled per thread, in commit order and with bounded workers, so one busy thread no longer holds up the others; provider wakes dispatch in a pool of their own.
  • A full thread snapshot is capped at 32 MiB of JSON by default and refused when too long for one socket frame.
  • MCP tools in a wake turn wait for the turn's run (port of upstream #33330); a run that ends its wait early still captures its checkpoint.
  • Traces: a per-minute span aggregate is written beside the trace file.
  • Includes 0.1.9-perf.8's checkpoint cohort and descendant process fixes.
  • Bench parameters in tools/server-bench set to chosen defaults.

Causes are quoted from the changelog, never inferred from the charts.

Text written from the measured record and checked against it: 46 numbers, 8 charts, 0 findings. A post that fails the check ships with charts only.