verification: github #24024 / https://git.cafe/pingdotgg/t3code/pulls/24024
main: 8031991591e22d76830e4acf6358c4fd78ffcf60
pr: a3c90a1cb109f3adaeb8c1f99d3f61ea699c1e59
linux boat vm; real dev server, playwright 1.60.0, google chrome 151.
project and provider configured through the web UI; real git workspace /home/user/terminal-verification.
provider: @agentclientprotocol/codex-acp 2.3.0 -> real codex 0.164.0 -> local proxy -> gpt-6.1-sol.
no mock provider, replay, fixtures, or seeded data used for runtime verification.

user command (same on main and pr):
node -e 'let i=0;let t=setInterval(()=>{for(let j=0;j<64;j++){i++;process.stdout.write(String(i).padStart(8,"0")+" "+"x".repeat(54)+"\n")}if(i===16384){clearInterval(t);console.log("TERMINAL-END-16384")}},40)'

read-only sqlite query on each dev state:
select count(*) events, max(length(json_extract(payload_json,'$.output'))) maxOutputChars,
       sum(length(payload_json)) payloadChars
from orchestration_events
where event_type='turn-item.updated'
  and json_extract(payload_json,'$.type')='command_execution'
  and payload_json like '%acpRegistry_codexv2%';
{
  "main": {
    "events": 1101,
    "maxOutputChars": 1036307,
    "payloadChars": 364438885
  },
  "pr": {
    "events": 1717,
    "maxOutputChars": 327680,
    "payloadChars": 488741967
  }
}

main reproduced an unbounded projected row: 1,036,307 characters, above 327,680.
pr: every persisted command output <=327,680 characters.
main/pr display equality: true; characters=65491; sha256=5139bba00e792cd8b962fe413d64dc41c7122aad9e1966d1d8f6140eaa46ba75.
both showed Earlier output not shown and TERMINAL-END-16384; real provider exited 0.
pr reload/reopen: same 65,491-character displayed output, byte-identical to main.
pr short command: node -e 'console.log("SHORT-BEGIN");console.log("SHORT-END");process.exit(7)'
real provider exited 7; stored command status failed; output exactly SHORT-BEGIN\nSHORT-END\n.

important limits:
main had 1,096 output chunks; pr had 1,712. actual received framing differed despite identical shell commands.
aggregate stored payload was HIGHER on pr in this run (488,741,967 vs 364,438,885 chars).
this verifies the per-row cap and unchanged display, not an aggregate-storage or CPU/throughput improvement.
main browser reads stalled and the UI reported Worked for 18m 27s; the 10-minute stall threshold was exceeded.
recording overhead and varying framing make the observed UI durations unsuitable as a controlled benchmark.

checks reproduced this turn through blacksmith:
focused long-output regression: 1 passed, 66 filtered out.
runtimeModel + AcpRegistryAdapterV2 + CommandOutputHub: 69 passed.
current pr CI: 28 successful, 0 failing, 0 pending on a3c90a1cb1.
no code edits, commits, or pushes made.

not verified: native desktop/mobile, macOS/Windows, remote relay/tunnel, other vendor agents above the cap,
production builds, CPU or memory improvements, outputs above 16 MiB.
claude ACP 0.98.0 short/nonzero command worked on main; its SDK clipped long output at 128,000 chars,
so that vendor could not exercise this cap. native codex is a separate adapter and was checked only as a neighbor on main.