Model-specific options for thinking, context management, and capabilities.
Thinking modes
Configured via /provider in the Claude tab:
| Mode | Behavior |
|---|
off | No thinking blocks |
auto / adaptive | Model decides when to think (Claude 4+ only) |
enabled | Fixed budget thinking with configurable token budget |
enabled): 1024, 2048, 5000, 10000, 20000.
Context management
Anthropic API context management features:
| Feature | What it does |
|---|
| Clear Thinking | Auto-clears old thinking blocks, keeps last 5 turns |
| Clear Tool Uses | Clears old tool use content when input exceeds 100k tokens |
| Compact | API-side compaction at 75% of context window |
Subagent options
Subagents inherit the parent’s provider options with one exception: contextManagement is stripped. Subagents are short-lived and token-budgeted, so context window management strategies don’t apply. Stripping also prevents errors when the subagent model doesn’t support the same features as the parent.
Graceful degradation
When provider options cause an API error, the system automatically retries with degraded options:
| Level | Options |
|---|
| Level 0 | Full options (thinking + context management) |
| Level 1 | Reduced (basic thinking, no context management) |
| Level 2 | No Anthropic-specific options |
Degradation is automatic and transparent. If a model doesn’t support thinking, SoulForge detects the error and retries without it — no user intervention needed.
