feat: add shell skills, file-based instructions, and disable tracing

[John-Lin]

Summary

• Load agent instructions from instructions.md file instead of hardcoded DEFAULT_INSTRUCTIONS, with fail-fast behavior when file is missing
• Add ShellTool support with auto-discovered skills from skills/ directory, gated behind SHELL_SKILLS_ENABLED environment variable
• Disable OpenAI tracing via set_tracing_disabled(True)
• Reduce MAX_TURNS from 25 to 10 for consistency across all three bots (Telegram, Discord, Slack)

Changes

bot/agent.py

• Remove DEFAULT_INSTRUCTIONS constant
• Add _load_instructions() — reads instructions.md from working directory
• Add _parse_skill_description() — extracts description from SKILL.md YAML frontmatter
• Add _load_shell_skills() — auto-discovers skills under skills/ directory
• Add _shell_executor() — async shell command runner with timeout support
• Update OpenAIAgent.__init__ to accept tools parameter, make instructions required (no default)
• Update OpenAIAgent.from_dict to use file-based instructions and conditionally load shell skills

Tests (tests/test_agent.py)

• Add TestLoadShellSkills (9 tests) — env var gating, skill discovery, error handling
• Add TestParseSkillDescription (8 tests) — frontmatter parsing edge cases
• Add TestShellExecutor (9 tests) — command execution, timeouts, error handling
• Update TestInstructions for file-based loading
• Update TestHistoryTruncation to assert MAX_TURNS == 10
• Remove obsolete Azure/proxy client tests
• Total: 81 tests (up from 57)

Documentation

• Add instructions.md.example template
• Remove instructions field from servers_config.example.json
• Add .gitignore entry for instructions.md
• Update README: add Agent Instructions, Shell Skills sections; update Docker examples to mount instructions.md