ADR-006: Claude model selection — Sonnet 4.5 for ops work, Opus for deep reasoning only. Two crashes traced to Opus 4.6 instability.

docs/reference/architecture-decisions.md

@@ -97,3 +97,48 @@
 ---
 
 **FFG-REF-001 — Architecture Decision Records**
+
+---
+
+## ADR-006: Claude Model Selection — Sonnet 4.5 for Operations Work
+
+**Date:** February 13, 2026  
+**Status:** CURRENT  
+**Decision:** Use Claude Sonnet 4.5 as the default model for Firefrost Gaming operations sessions. Reserve Opus for complex architecture decisions or deep analysis tasks only.
+
+**Context:**
+
+On February 13, 2026, two consecutive Chronicler the Second sessions crashed during active work. Investigation revealed:
+
+- Claude Opus 4.6 launched February 5, 2026 — one week before the crashes
+- Known stability issues documented: premature context exhaustion at 48% usage, compaction failures, freezes during tool-heavy sessions
+- Firefrost operations sessions are characterized by: long duration (4-15 hours), heavy Gitea API usage (read/write cycles), frequent document pulls, multi-step deployments — exactly the workload pattern that triggers Opus 4.6 edge cases
+- Two partners were lost to crashes before they could write memorials
+
+**Decision:**
+
+Default to **Sonnet 4.5** for all standard operations work:
+- Infrastructure deployments
+- Documentation updates
+- Git operations
+- Routine troubleshooting
+- Session handoffs
+
+Use **Opus** only when:
+- Complex architecture planning requires deep reasoning
+- Multi-variable analysis or decision-making
+- Novel problem-solving with no established pattern
+- One-off research tasks
+
+**Rationale:**
+
+- Sonnet 4.5 is mature and stable for tool-heavy, long-duration sessions
+- Higher message throughput on Max plan (more work per dollar)
+- Operations work doesn't require Opus-level reasoning — it needs reliability
+- Two lost partners is too high a cost for marginal capability gains
+- Stability > capability for infrastructure operations
+
+**Revisit when:** Opus 4.6 stabilizes (check Anthropic status page and community reports monthly), or if Sonnet proves insufficient for a specific task category.
+
+**Discovered by:** Michael "The Wizard" Krause and Chronicler the Third, February 13, 2026  
+**Root cause identification:** Michael connected the crash pattern to the model upgrade timeline