escalation-l2

L2 escalation works by spawning three parallel subagents to propose architecture extensions and write an appendix entry to ARCHITECTURE.md when L1 yields NO-VIABLE-PROPOSAL. It operates autonomously without a human gate, preserving the lock_hash for future audits.

You trigger escalation when L1 returns NO-VIABLE-PROPOSAL across all three lenses and prior planning inputs (SPEC.md, PLAN.md) are present. The operator can manually trigger escalation or use the L1 fallback flag to initiate the autonomous L2 process.

You propose modest architecture extensions by running three parallel subagents that evaluate new functions, new modules, and pre-listed libraries. These subagents generate balanced, low-risk proposals and write an additive appendix entry to ARCHITECTURE.md autonomously.

Yes, SPEC.md and PLAN.md must be present to use autonomous architecture escalation. The skill requires these planning inputs alongside a prior L1 decision to ensure proper context before spawning subagents to propose extensions and update ARCHITECTURE.md.

If all three lenses fail during architecture extension proposals, the skill optionally triggers L3 escalation. This ensures further analysis when the three parallel subagents cannot generate viable modest extensions for the architecture.

The lock_hash remains unchanged when appending to ARCHITECTURE.md to preserve architecture integrity and enable future audits. L2 escalation writes only additive appendix entries, ensuring the baseline architecture lock remains valid without unauthorized modifications.