PR 30331 runtime verification
main: 7c22e24e3f (github/main fetched this run)
PR: da38aaebf74359d0291ee53bff40b5bbd35060cc

Environment: Linux Boat VM, /usr/bin/google-chrome driven by the repository's Playwright 1.60.0.
One development server at a time. Main ports 8693/16733; PR ports 5733/13773.
Final PR startup explicitly used --home-dir /home/user/t3code/.t3.
The first root-checkout startup selected the shared home; it was stopped and corrected.
No application source edits, mocks, fixtures, or fabricated provider output.

Created /home/user/pr30331-scenario as a real git repository, added it through the app,
and used Codex gpt-6.1-sol turns to edit README.md.
Selected the Claude source-control writer through Settings > Source control > Advanced.
Used custom model claude-haiku-4-5-20251001 through the local authenticated proxy.
A transparent CLI launcher executed the installed Claude Code 2.1.317 and added
--max-budget-usd 0.0000001 to print-mode requests, then removed that limit for recovery.
The launcher did not synthesize stdout or stderr.

Real CLI result included:
  is_error: true
  subtype: error_max_budget_usd
  errors: ["Reached maximum budget ($1e-7)"]
  no result field

Main, app Commit action:
  Text generation failed in generateCommitMessage: Claude CLI command failed:
  {"duration_api_ms":1747,"stop_reason":null,...raw CLI JSON...}
  The toast truncated the JSON. README.md remained modified and uncommitted.

PR, same project, thread, diff, model, CLI and budget:
  Text generation failed in generateCommitMessage: Claude CLI command failed:
  Reached maximum budget ($1e-7)
  README.md remained modified and uncommitted. No raw JSON appeared in the toast.

PR recovery, budget limit removed, app Commit action:
  Committed d6169b6
  docs: note budget error verification

$ git -C /home/user/pr30331-scenario log -2 --oneline
d6169b6 docs: note budget error verification
bc0a4d6 docs: note source control generation verification
$ git -C /home/user/pr30331-scenario status --short
(empty)

PR Codex gpt-6.1-sol follow-up turn read README.md and replied:
  verified source control generation. budget error verification.

Blacksmith remote check:
  vp i && cd apps/server && vp test run src/textGeneration/ClaudeTextGeneration.test.ts
  1 test file passed; 21 tests passed.
Existing PR CI: 30 successful checks, 0 failing, 0 pending.

Not exercised live: error_max_structured_output_retries itself, telemetry filtering,
verbose-output arrays, PR-content generation/PR publication, native desktop/mobile,
Windows/macOS, relay/tunnel connections. The focused tests cover errors-only results,
verbose output, telemetry filtering, and existing error fallbacks.