The plan has been written to .claude/pipeline-artifacts/plan.md. Here's a summary of the implementation plan:

Architecture: File-based signaling protocol. Stages write scaling request JSON files; the loop monitor checks them between iterations and spawns/dismisses agent tmux panes accordingly.

• scripts/sw-scaling.sh — Core scaling engine with 12 functions: trigger evaluation, spawn/dismiss, cooldown enforcement, budget checks, context transfer
• scripts/sw-scaling-test.sh — 18 test cases following the existing mock-based test harness pattern

• sw-loop.sh — Hot-add/remove agents between iterations; scaling monitor in multi-agent mode
• sw-pipeline.sh — Source scaling engine; evaluate triggers at stage boundaries; include scaling state in write_state()
• full.json / autonomous.json / standard.json — Add scaling config with trigger definitions
• sw-daemon.sh — Surface scale events in metrics/health checks
• .claude/CLAUDE.md / package.json — Documentation and test registration

1. Signal file (.claude/pipeline-artifacts/scaling-requests.json) — pipeline writes requests, loop reads them
2. 6 trigger types: build iteration threshold, coverage gap, security critical, idle agents, multi-module split, consecutive failures
3. Cooldown: 2-minute minimum between scale events
4. Budget reserve: Configurable % reserve (default 20%) — blocks spawns when budget is low
5. Context transfer: New agents get pipeline state, memory patterns, non-overlapping file assignments, error context, and git log