feat: replace foundry-mcp/tavily with user-configurable skills system

[tylerburleigh]

Summary

• Remove hardcoded foundry-mcp and tavily-mcp from the sandbox Docker image and compose files, making sandboxes "empty" by default
• Add a user-configurable skills system (~/.sandboxes/skills.toml) that lets users define what tools, MCP servers, permissions, environment variables, and stubs to install into sandboxes
• Add CLI commands (cast skills list|show|init) for managing skill configurations

Details

Removed

• pip install foundry-mcp and npm install tavily-mcp from Dockerfile
• TAVILY_API_KEY, PERPLEXITY_API_KEY, SEMANTIC_SCHOLAR_API_KEY env vars from compose files
• prepopulate_foundry_global(), research provider config, and default hooks from foundry_plugin.py
• foundry_upgrade.py internals (stubbed to no-op)
• ensure_opencode_tavily_mcp() from tool_configs.py
• .foundry-mcp.toml and lib/python/ensure_opencode_tavily.py
• Hardcoded content from stubs/CLAUDE.md and stubs/AGENTS.md (now skill-controlled)

Added

• foundry_sandbox/skills.py — loads TOML config, resolves env vars, generates Docker mounts/env/permissions, installs MCP servers + stubs into containers
• foundry_sandbox/commands/skills_cmd.py — cast skills list|show|init
• --skill flag on cast new (replaces --pre-foundry), multi-select in wizard
• Per-skill permission merging in permissions.py
• HTML comment markers for idempotent stub installation
• Skills documentation in docs/configuration.md
• ~700 lines of new tests + all existing tests updated

Example skills.toml

[skills.my-research-tool]
mount = "~/GitHub/my-research-tool"
stub = "~/GitHub/my-research-tool/SANDBOX_STUB.md"

[skills.my-research-tool.mcp.my-server]
command = "python"
args = ["-m", "my_server"]

[skills.my-research-tool.permissions]
allow = ["Bash(python -m my_tool:*)"]

Test plan

• Run ./scripts/ci-local.sh to verify all unit tests pass
• Create a sandbox with cast new --skill <name> and verify skill installation
• Verify cast skills init creates a valid example config
• Verify old sandboxes (pre-skills) still start without errors

🤖 Generated with Claude Code