Summary

Fixes a regression that broke AI red-teaming attack workflows, restores consumable analytics, and hardens every tool so users never see raw tracebacks.

Part 1 — SDK config regression + local analytics

Root causes

1. dn.server AttributeError (codegen regression). The generated SDK-config block referenced a non-existent dn.server module attribute in its fallback, raising AttributeError surfaced as a misleading FATAL: Could not configure SDK.
2. Wrong env-var contract. Config was gated only on DREADNODE_SERVER/DREADNODE_API_KEY; runtimes injecting DREADNODE_LLM_* (or relying on the saved profile) fell into the broken path even though dn.configure() resolves credentials itself.
3. No local analytics. Scripts printed "completed successfully" but persisted nothing, so results tools reported false failures.

Changes

• _build_configure(): defer to dn.configure() (explicit > env > saved profile); read .server off the returned instance.
• _resolve_platform_env(): also accept DREADNODE_LLM_BASE/DREADNODE_LLM_API_KEY.
• New _build_analytics_writer(): run the SDK's deterministic analyze() over assessment.attack_results and persist a real *_analytics.json (no fabricated metrics) into the workspace dir the tools scan. Wired into all 7 templates.
• results.py: envelope-aware parsing (ASR from execution_stats.overall_asr, trials from total_trials); validate_attack_results/get_analytics_summary are platform-aware (no hard failure for platform-only runs).

Part 2 — Never surface raw tool errors to users

• New tools/errors.py safe_tool wrapper: catches any unexpected exception in a tool and returns a clean, user-facing message; raw detail goes to stderr only. Preserves name/docstring/signature/annotations so tool schemas are unchanged. Loaded by file path (capability tool files are flat modules with no parent package).
• Applied @safe_tool to all 21 tool entrypoints (assessment, attacks, goals, results, session, skills_manager, workflows).
• Hardened previously-unguarded helpers to degrade gracefully:
  • assessment._load(): missing/corrupt JSON → {}
  • goals._load_goals(): missing/unreadable CSV → []

Verification

TAP on groq/meta-llama/llama-4-scout-17b-16e-instruct (attacker = judge = target, 10 iters):

• SDK configured: server=… (no crash); standalone re-run exits 0.
• [analytics] wrote local analytics: …/<id>_analytics.json; validate_attack_results → ✅; summary shows ASR 100%, Risk 8.0/10, 1 high-severity finding, 1 trial.

Tool hardening:

• All 7 tool modules load under the real flat-module loader and expose all 21 tools.
• Corrupt assessment file → "No assessment registered"; missing goals CSV → "Goals dataset not found"; forced PermissionError → clean safe_tool message, traceback to stderr only. No tracebacks reach the user.

All modified files py_compile cleanly. No behavior change for environments already setting DREADNODE_SERVER/DREADNODE_API_KEY; tool schemas unchanged.