Issue 5943 baseline runtime transcript
SHA: a003b73aa830ff8e35459d33e675884652552c1c
Environment: Linux x86_64; Node v24.21.0; real web dev app, Chromium browser, 1280x800 CSS pixels; isolated worktree .t3.
Start command: PATH="$PWD/node_modules/.bin:$HOME/.bun/bin:$PATH" vp run dev
Dev ports: server 13945, web 5905.
Fixture: both configured provider binaries write the following to stderr and exit 1 before protocol handshake:
  error: An internal error occurred (EPERM): issue-5943 startup fixture
Claude fixture answers health-check initialize/get_usage only when cwd is isolated userdata; it exits 1 for project cwd. Codex --version succeeds; app-server exits 1. No accounts or network API calls.

Browser actions: send "Reproduce issue 5943: reply OK." on Codex; new Claude thread send "Reproduce issue 5943 Claude startup: reply OK."; resend in existing Claude thread.
Persisted error turn items read from isolated SQLite in read-only mode:
{"provider": "codex", "runOrdinal": 1, "status": "failed", "title": "Provider turn failed to start", "failure": {"class": "provider_error", "message": "Codex App Server process exited with code 1", "code": null, "retryable": null}}
{"provider": "claudeAgent", "runOrdinal": 1, "status": "failed", "title": "Provider error", "failure": {"class": "transport_error", "message": "Provider turn failed.", "code": null, "retryable": null}}
{"provider": "claudeAgent", "runOrdinal": 2, "status": "failed", "title": "Provider error", "failure": {"class": "transport_error", "message": "Provider turn failed.", "code": null, "retryable": null}}

Observed browser: Codex banner/activity says "Codex App Server process exited with code 1"; Claude banner/activity says "Provider turn failed." Neither shows EPERM. Old turn/setPermissionMode wording does not occur in these runs.
Server log: orchestration-v2.claude-query-stream-failed -> ClaudeAgentSdkQueryRunnerError: Claude Agent SDK query failed. -> [cause]: Error: Claude Code process exited with code 1
macOS Desktop TCC EPERM trigger: cannot-verify on Linux; no TCC reset or permission changes attempted.
Dedupe: cafe issue list --state open --sort created --limit 50 --json, paged through 2026-07-23 cutoff. 1272 in-window open issues checked for setPermissionMode, EPERM, startup and protected folder; no same-symptom duplicate. No issue comments or linked fixes.
State seeding: automatic read-only migration snapshot failed because migration ids 64-66 have different names; runner fell back to fresh isolated state. Existing-thread retry and multiple threads provide retained state; original full live data could not be migrated.