PR #19958 real-environment verification, 2026-10-28 UTC
GitCafe: https://git.cafe/pingdotgg/t3code/pulls/19958
Main: 48a292364a (github/main at start)
PR: c563f06ec11d06519d1c4d62f8d242f9ba7f9642

Setup:
  git worktree add ../main github/main
  cd ../main && vp i
  background: timeout 6600 vp run dev --no-seed
  observed main ports: server=16733 web=8693
  stopped main process group 15292 before starting PR
  git switch maria/takeover-pr-19958-verify (exact PR head)
  vp i
  background: timeout 6000 vp run dev --home-dir /home/user/t3code/.t3 --no-seed
  observed PR ports: server=13773 web=5733

Browser:
  repo Playwright, executablePath=/usr/bin/google-chrome, 1440x1000
  local folder project added through UI: /home/user/usage-verification-repo
  real git repo, no mocks/demo data/fixture writes
  Codex gpt-6.1-sol, low effort, full access, local proxy
  identical first prompt: remember the codeword lantern-19958. reply with exactly:
  lantern-19958 acknowledged. do not run tools or change files.
  main and PR both displayed: lantern-19958 acknowledged.

Read-only SQL:
  select ordinal,status,json_extract(payload_json,'$.turnTokenUsage')
  from orchestration_v2_projection_provider_turns order by started_at;

main:
1|completed|{"usageScope":"main_agent","cachedInputTokens":11264,"cacheCreationTokens":0,"reasoningTokens":0,"hasSubagents":false,"usageStatus":"complete","inputTokens":16238,"outputTokens":11}

PR:
1|completed|{"usageScope":"main_agent","cachedInputTokens":11264,"cacheCreationTokens":0,"reasoningTokens":0,"hasSubagents":false,"usageStatus":"complete","inputTokens":16148,"outputTokens":11}
2|completed|{"usageScope":"main_agent","cachedInputTokens":16000,"cacheCreationTokens":0,"reasoningTokens":93,"hasSubagents":false,"usageStatus":"complete","inputTokens":16188,"outputTokens":119}
3|interrupted|{"usageScope":"main_agent","hasSubagents":false,"usageStatus":"unavailable"}
4|completed|{"usageScope":"main_agent","cachedInputTokens":16128,"cacheCreationTokens":0,"reasoningTokens":0,"hasSubagents":false,"usageStatus":"complete","inputTokens":17654,"outputTokens":9}

Neighboring behavior:
  turn 2 completed before Stop could be clicked; provider declined 10000 lines.
  turn 3: clicked actual Stop generation during compiler prompt;
  UI displayed Run interrupted / Stopped by user; SQL interrupted/unavailable.
  stopped PR process group 21126 and restarted against same isolated state.
  turn 4: asked remembered codeword after full server restart;
  UI displayed lantern-19958; SQL completed/complete.
  native Codex thread id identical before/after restart:
  01a1824d-ef19-7b40-8c13-170994af8763
  stopped turn remained interrupted/unavailable after restart and follow-up.

Focused checks, reproduced through blacksmith-check:
  vp i
  vp test run apps/server/src/orchestration-v2/Adapters/AcpAdapterV2.test.ts packages/provider-acp packages/provider-acp-registry
  13 files, 218 tests passed across adapter, provider-acp and registry.
  Explicit registry-only selection also passed 55 tests in 5 files.

GitHub CI:
  7 check contexts successful (one skipped); CI workflow run 40376121475
  conclusion action_required; no claim of full CI success.

Limitations:
  No glm or glm-acp-agent executable/configured account on VM.
  PR captures usage only for registry glm-acp-agent. Native Codex uses a
  different adapter. Codex evidence does not reproduce the original ACP bug
  on main or prove the GLM fix, per-turn semantics, clamps or Stop races.
  No live undeclared ACP provider, native desktop/mobile, macOS or Windows run.
  One 18s CPU overload warning occurred during cold dev startup; it recovered.
  No production source changes, commits, pushes, labels or merges.