Separate cognitive work from mechanical API execution to reduce
skill-related token consumption by ~76-83% during sprint workflows.
Changes:
- Add batch-execution.md skill with 4-phase protocol
- Promote mcp-tools-reference and batch-execution to frontmatter
for planner and orchestrator agents (auto-injected, zero re-read)
- Replace "Skills to Load" with phase-based "Skill Loading Protocol"
- Restructure planning-workflow.md Steps 8-10 for batch execution
- Update agent matrix in CLAUDE.md and docs/CONFIGURATION.md
- Add Phase-Based Skill Loading documentation section
- Clean up .gitignore (transient files, dev symlinks)
Token impact:
- 6-issue sprint planning: ~76% reduction
- 10-issue sprint planning: ~80% reduction
- 8-issue status updates: ~83% reduction
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Problem:
- Venvs in marketplace directory got deleted on every update
- Users had to manually run setup.sh and wait for full pip install
- This caused MCP servers to fail until manually fixed
Solution:
- Store venvs in external cache (~/.cache/claude-mcp-venvs/)
- Auto-repair symlinks via SessionStart hook (instant operation)
- Only run pip install on first use or when requirements change
Architecture:
Cache (runtime) → Marketplaces → External venv cache
The chain of symlinks ensures all three locations work:
1. ~/.claude/plugins/cache/.../mcp-servers/* (runtime)
2. ~/.claude/plugins/marketplaces/.../mcp-servers/* (install)
3. ~/.cache/claude-mcp-venvs/* (persistent venvs)
Performance:
- First install: ~2-3 min (unchanged)
- After marketplace update: 0.03 sec (was 2-3 min)
Files:
- scripts/venv-repair.sh: Fast symlink restoration for hooks
- scripts/setup-venvs.sh: Full setup with external cache
- plugins/projman/hooks/startup-check.sh: Auto-repair on session start
- .gitignore: Ignore .venv symlinks
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
