PR 29971 real dev verification
Head: 4c466c96d23affc50dafa5f5962089715837173f
Baseline: github/main cefb544af3
Client: repository Playwright 1.60.0, /usr/bin/google-chrome, Linux, 1440x1000
Provider: Codex gpt-6.1-sol via local proxy, started from the web composer
Data: projects, thread, and turns created through the web app; no mocks/replay/fixtures
Only one dev server ran at a time. Main and PR used separate task-owned homes.

Main tracepoint counts, four sequential commands plus T3 list_thread_pull_requests:
{
  "resolveT3McpToolName": 8776,
  "workEntryDisplayIndicatesToolFailure": 1883,
  "workEntryIndicatesToolFailureFromOutput": 3007,
  "workEntryIndicatesToolFailure": 1124,
  "toolGroupAction": 1135
}

PR tracepoint counts, same prompt:
{
  "resolveT3McpToolName": 4260,
  "workEntryDisplayIndicatesToolFailure": 1893,
  "workEntryIndicatesToolFailureFromOutput": 1121,
  "toolGroupAction": 1160,
  "computeToolGroupAction": 463,
  "workEntryIndicatesToolFailure": 1127,
  "parseT3McpToolName": 0
}

Counts cover actual app/provider turns in a bounded thread, not a 2,000-item timing benchmark.
Debugger tracepoints only increment counters and never pause or alter function results.
Initial sampled profiler counts were discarded because optimized/inlined functions gave inconsistent totals.

Selected native provider completion events on PR:
{"type":"mcpToolCall","tool":"list_thread_pull_requests","status":"completed","error":null,"result":{"pullRequests":[],"chains":[]}}
{"type":"mcpToolCall","tool":"list_thread_pull_requests","status":"completed","error":null,"result":{"pullRequests":[],"chains":[]}}
{"type":"mcpToolCall","tool":"list_thread_pull_requests","status":"completed","error":null,"result":{"pullRequests":[],"chains":[]}}
{"type":"mcpToolCall","tool":"list_thread_pull_requests","status":"completed","error":null,"result":{"pullRequests":[],"chains":[]}}
{"type":"commandExecution","command":"/bin/bash -lc 'printf VERIFY_FINAL_ONE'","output":"VERIFY_FINAL_ONE","exitCode":0,"status":"completed"}
{"type":"commandExecution","command":"/bin/bash -lc 'printf VERIFY_FINAL_FAIL; exit 7'","output":"VERIFY_FINAL_FAIL","exitCode":7,"status":"failed"}
{"type":"commandExecution","command":"/bin/bash -lc 'printf VERIFY_FINAL_RECOVERED'","output":"VERIFY_FINAL_RECOVERED","exitCode":0,"status":"completed"}
{"type":"mcpToolCall","tool":"list_thread_pull_requests","status":"completed","error":null,"result":{"pullRequests":[],"chains":[]}}

Observed in client: live working row, completed groups, exit 7 marked red, subsequent successful output, T3 friendly label, fold/unfold, reload, interruption and later resumed turn.

Checks: 159 focused tests passed in four files on Blacksmith. check:changed against 61c3794c32b0eeebfac09add43e403618c65d2d7 passed format, lint, two knip passes, typecheck (34 packages), and 25 tests in three files.

Unverified: advertised 70% wall-clock improvement on 2,000 items; native desktop and iOS/Android; remote/relay/tunnel; non-Codex providers; 10,000-key eviction and >256-character ACP labels through a real ACP provider.