Before: main 2ed92618e6
Environment: fs-16 Linux x64; real T3 dev server, web client, Claude Code 2.1.303, claude-opus-5-5
Launch arguments: --settings '{"env":{"ANTHROPIC_VERTEX_PROJECT_ID":"issue-25017-marker"}}'
Configured the instance in Settings > Providers, selected it as text generation model, sent "Reply with OK only. Do not use tools or change any files."
Observed thread title: Quick OK response check
Recorded actual Claude argv:
[
  "-p",
  "--output-format",
  "json",
  "--json-schema",
  "{\"type\":\"object\",\"properties\":{\"title\":{\"type\":\"string\"},\"needsRefinement\":{\"type\":\"boolean\"}},\"required\":[\"title\",\"needsRefinement\"],\"additionalProperties\":false}",
  "--model",
  "claude-opus-5-5",
  "--effort",
  "medium",
  "--settings",
  "{\"disableAllHooks\":true}",
  "--tools",
  "",
  "--disable-slash-commands",
  "--strict-mcp-config",
  "--permission-mode",
  "dontAsk"
]
Actual settings: {"disableAllHooks": true}
Project marker absent: True
The probe records argv and subprocess cwd, then execs the real Claude CLI. No provider response was stubbed.

After: bd4851187a
Inline JSON: clicked thread actions > Regenerate title. Result: Quick OK response check.
Actual argv:
[
  "-p",
  "--output-format",
  "json",
  "--json-schema",
  "{\"type\":\"object\",\"properties\":{\"title\":{\"type\":\"string\"},\"needsRefinement\":{\"type\":\"boolean\"}},\"required\":[\"title\",\"needsRefinement\"],\"additionalProperties\":false}",
  "--model",
  "claude-opus-5-5",
  "--effort",
  "medium",
  "--settings",
  "{\"env\":{\"ANTHROPIC_VERTEX_PROJECT_ID\":\"issue-25017-marker\"},\"disableAllHooks\":true}",
  "--tools",
  "",
  "--disable-slash-commands",
  "--strict-mcp-config",
  "--permission-mode",
  "dontAsk"
]
Relative file: request cwd /home/ubuntu/wt/issue-25017/apps/server
Launch args: --settings ../../../../../../tmp/issue-25017/launch-settings.json
File contents: {"env": {"ANTHROPIC_VERTEX_PROJECT_ID": "issue-25017-file-marker"}, "disableAllHooks": false}
Sent same message in a new thread. Result title: Agent OK response check; agent replied OK.
Actual argv:
[
  "-p",
  "--output-format",
  "json",
  "--json-schema",
  "{\"type\":\"object\",\"properties\":{\"title\":{\"type\":\"string\"},\"needsRefinement\":{\"type\":\"boolean\"}},\"required\":[\"title\",\"needsRefinement\"],\"additionalProperties\":false}",
  "--model",
  "claude-opus-5-5",
  "--effort",
  "medium",
  "--settings",
  "{\"env\":{\"ANTHROPIC_VERTEX_PROJECT_ID\":\"issue-25017-file-marker\"},\"disableAllHooks\":true}",
  "--tools",
  "",
  "--disable-slash-commands",
  "--strict-mcp-config",
  "--permission-mode",
  "dontAsk"
]
Actual title subprocess working directories (inline regeneration then relative file):
{"cwd": "/tmp/t3code-claude-title-locAJC"}
{"cwd": "/tmp/t3code-claude-title-NM3G3N"}

Test proof:
At f77c5a127b, vp test run src/textGeneration/ClaudeTextGeneration.test.ts: 2 failed, 15 passed. Both cases fail with: TextGenerationError: Claude CLI command failed: --settings was {"disableAllHooks":true,"fastMode":true}
At bd4851187a, same focused suite: 19 passed (15 existing, 2 regression, 2 invalid settings cases).
check:changed: format/lint/graphql/typecheck/tests pass; knip fails on base-only unused agent-device in scripts/package.json:15:6.
Server-only change: capture matrix []. No rendering changes.