Research-Stack/.opencode/skills/break-glass/SKILL.md
allaun ba226a28d9 feat(agent): Gemma4-12B as primary model, free and local
Gemma4-12B is now the primary model for all questions:
- Endpoint: http://127.0.0.1:8081/v1
- Model ID: gemma4-12b
- Speed: ~40 tokens/sec
- Cost: free (local)

Escalation: Gemma4-12B → DeepSeek V4 Pro → Claude Opus → GPT-5.5 → all 8
2026-06-22 22:45:46 -05:00

2.4 KiB

SKILL: break-glass

THIS IS FOR UNSOLVABLE PROBLEMS ONLY

DO NOT USE THIS SKILL UNLESS ALL OF THE FOLLOWING ARE TRUE:

  1. The user has no solution — they do not know the answer
  2. All free models have been tried and failed (including local Gemma4-12B)
  3. All cheap models (< $0.50/call) have been tried and failed
  4. The problem is genuinely novel (not a rehash of known issues)
  5. The problem is blocking progress (not a nice-to-have)

This skill costs real money. Each invocation = $0.10-$3.00.

If the problem is solvable by switching models, just switch models. If the problem is solvable by reading docs, just read docs. If the problem is solvable by asking the user, just ask the user. If a free model can solve it, use the free model. If the local Gemma4-12B can solve it, use Gemma4-12B.

This skill is the last resort before giving up.


Primary Model

Gemma4-12B (local, free, fast) — use this first for all questions.

  • Endpoint: http://127.0.0.1:8081/v1
  • Model ID: gemma4-12b
  • API key: none (any dummy string)
  • Speed: ~40 tokens/sec generation
  • Strengths: mathematical reasoning, code, Lean 4 proofs

Trigger

break glass: <problem statement>

What It Does

Selects the optimal model panel using dual quaternion chiral ratio. Each model is a dual quaternion (real = compressive, dual = anti-compressive). χ = |real|² / (|real|² + |dual|²) determines selection.

Model Selection (dual quaternion χ at threshold 0.99)

Problem Models Cost
lean_proof DeepSeek V4 Pro $0.10-0.50
code_bug Kimi K2.6 $0.10-0.30
architecture Claude Opus $0.50-2.00
math_novel DeepSeek V4 Pro $0.10-0.50
infra_debug GPT 5.5 + DeepSeek Flash $0.30-1.00
hard_wall Claude Opus + DeepSeek Flash $0.50-3.00

Escalation

  1. Gemma4-12B (local, free) — try first
  2. DeepSeek V4 Pro — if Gemma stuck ($0.27/$1.10)
  3. Claude Opus — if DeepSeek stuck ($15/$75)
  4. GPT-5.5 — if Opus stuck ($10/$30)
  5. All 8 models — hard wall ($5-10 total)

Files

File Purpose
4-Infrastructure/shim/dual_quat_model_selector.py Dual quaternion model selector
4-Infrastructure/shim/model_theorem_matrix.py Degenerate sector analysis
6-Documentation/docs/specs/break_glass_problem_statement.md Problem statement
6-Documentation/docs/specs/break_glass_test_list.md Test list