fix: filter LLM API keys from bash tool child env

[Darkroom4364]

• Adds sanitizedEnv() to strip sensitive keys before passing to execSync
• Keeps PATH, HOME, TARGET, and auth env vars

[Darkroom4364]

Good change, a few things:

Bug — missed call site (line ~1496)
The interactive shell session (create action) builds its env as { TARGET: …, ...this.buildAuthEnvVars() } — no sanitizedEnv(), so it inherits process.env via Node defaults and leaks every key this PR tries to strip.

Missing patterns

• MISTRAL_API, COHERE_API, HUGGINGFACE, DEEPSEEK_API, TOGETHER_API, GROQ_API — all common in the wild.
• _KEY / _SECRET suffix catch-all would be more future-proof than an allowlist of prefixes.

Nit
sanitizedEnv() rebuilds the filtered object on every call. Fine for now, but a module-level lazy cache would be free and cleaner.

Overall: scope is right (only tools.ts), diff is minimal. Fix the second call site and widen the patterns, then LGTM.