PR 28643 real development verification, 2026-11-01, Linux Boat VM

Baseline: github/main at 2b1021ce41
PR head: 3a6b3e8d5cc953fea3c148b286fab6f4ecd9bc47
No source modifications or pushes.

Startup commands:
git worktree add ../main github/main
cd ../main && vp i
setsid nohup vp run dev > /tmp/main-dev.log 2>&1 &
main: serverPort=16733 webPort=8693, isolated /home/user/main/.t3
Stopped main using captured launcher PID before starting PR.
git switch --detach 3a6b3e8d5cc953fea3c148b286fab6f4ecd9bc47
vp i
setsid nohup vp run dev > /tmp/dev.log 2>&1 &
PR: serverPort=13773 webPort=5733, isolated /home/user/t3code/.t3
Commands were bounded with timeout. Claude used ANTHROPIC_BASE_URL pointing
to the VM's real local provider proxy and its existing credential.

Browser: repository Playwright dependency, /usr/bin/google-chrome.
Created local decoder-verification git projects through Add project > New project.
Selected models and submitted all prompts through the web composer.
No replay provider, mocked SDK, seeded database, or fabricated provider frames.

Affected path: Claude Haiku 5.5 > foreground Bash > real task_started
(task_type=local_bash, is_backgrounded=false) > actual .output file >
tailClaudeTaskOutput > running server > live web command output.

Command submitted to Claude, foreground Bash, timeout=60000:
node -e 'const fs=require("fs"),p=require("path");const prefix=Buffer.concat([Buffer.from("pr28643:caf"),Buffer.from([195])]);process.stdout.write(prefix);function walk(d){try{return fs.readdirSync(d,{withFileTypes:true}).flatMap(e=>e.isDirectory()?walk(p.join(d,e.name)):e.name.endsWith(".output")?[p.join(d,e.name)]:[])}catch{return []}}setTimeout(()=>{const files=walk("/tmp/claude-"+process.getuid());const target=files.find(f=>{try{return fs.readFileSync(f).equals(prefix)}catch{return false}});if(!target){console.error("NO_MATCHING_TASK_OUTPUT");process.exitCode=2;return}fs.writeFileSync(target,"ok\n");setTimeout(()=>process.stdout.write("append café 😀\n"),15000);setTimeout(()=>process.stdout.write("finish-ok\n"),25000)},6000)'

This intentionally makes a real provider task rewrite its own output file.
It does not establish how frequently Claude spontaneously truncates output.

Baseline observed live web output:
pr28643:caf
then, following the rewrite:
pr28643:caf�ok
then:
pr28643:caf�ok
append café 😀

PR observed live web output with the same command:
pr28643:caf
then, following the rewrite:
pr28643:cafok
then:
pr28643:cafok
append café 😀

Final tool output on the PR, also displayed in the conversation:
ok
append café 😀
finish-ok

Neighbor 1, same Claude thread, foreground Bash, timeout=30000:
node -e 'process.stdout.write(Buffer.concat([Buffer.from("split-neighbor:caf"),Buffer.from([195])]));setTimeout(()=>{process.stdout.write(Buffer.from([169]));process.stdout.write(Buffer.from([240,159]));},6000);setTimeout(()=>{process.stdout.write(Buffer.from([152,128]));process.stdout.write("\nfinish-ok\n");},12000)'
Observed live and final output, with no replacement characters:
split-neighbor:café😀
finish-ok

Neighbor 2, real Codex gpt-6.1-sol through the web composer on both revisions:
timeout 30 node -e 'process.stdout.write("codex-neighbor: café 😀\n");setTimeout(()=>process.stdout.write("append-ok\n"),1000)'
Observed output:
codex-neighbor: café 😀
append-ok
exit code: 0

Blacksmith, exact PR head, after remote vp i:
cd apps/server && vp test run src/orchestration-v2/Adapters/ClaudeTaskOutputTail.test.ts
Test Files 1 passed (1)
Tests 5 passed (5)

Not exercised: Electron shell; native iOS/Android; macOS/Windows;
relay/tunnel connections; spontaneous provider-managed truncation.
The nine-second MP4 is a 1 fps capture of real UI states, not a timing benchmark.