v0.2.6

What's New

Self-Improvement System

• Reflector agent — post-engagement reflection, analyzes outcomes, generates learned skills
• Knowledge store — cross-engagement memory at ~/.pentestcode/knowledge/global.json (tool effectiveness, FP patterns)
• /reflect slash command — spawns reflector on current or named engagement
• Auto-generated skills saved to ~/.pentestcode/skills/user/learned/

CyberStrike Gap Closure

• inject_probe tool — structured HTTP injection probing (XSS, SSTI, CMDi, SQLi, NoSQLi, LFI, SSRF, LDAP, XPath), up to 120 payloads/call, WAF detection, boolean differential analysis
• ensure_tools tool — check/install 21 common pentest tools with platform-aware installers
• os_hook tool — structured post-exploitation command generator (Windows 14 + Linux 12 submodules)
• ReAct discipline + evidence quality gates + confidence tagging in all agent prompts

Jinja Crash Fix (thanks @adi-IL — #12)

• Consolidate leading system messages for OpenAI-compatible endpoints (Ollama, vLLM, Qwen, llama.cpp)
• Lower non-leading system messages to <system-update> blocks after compaction
• Preserves split system blocks for Anthropic prompt caching
• Fixes permanent session crash on local models after auto-compaction (#11)

Other

• Double Ctrl+C required to exit TUI (prevents accidental exits)
• 503 resilience, context carry, severity inflation fix
• Skills trimmed for token efficiency (−1189 lines across 15 skill files)

Install

curl -fsSL https://raw.githubusercontent.com/s0ld13rr/pentestcode/main/install.sh | bash