Lataa TermCanvas 0.20.0

Added
- SuperHydra P0/P1 agent runtime with filesystem tools, model registry, cost tracking, compaction, o-series adaptation, streaming, fallback model, tool hooks, context injection, and session persistence
- API key encryption at rest via Electron safeStorage
- Approval bridge with policy-based auto-approve and escalation
- JSONL session persistence with append-only transcript and resume
- Composable pre/post tool hooks with permission decisions
- Fallback model and max_output_tokens recovery with progressive retry
- OpenAI o-series adaptation with max_completion_tokens, developer role, reasoning_effort
- Auto-compaction with circuit breaker
- Cost/budget tracking with per-turn USD calculation
- Error categorization and loop-level retry
- Static model capability registry
- ReadFile, GlobFile, GrepFile read-only filesystem tools for agent
- Async/background tool execution with pending task registry
- Structured coordinator system prompt builder
- Challenge skill
- Hydra telemetry enrichment and retry for agent watch
- Hydra watch support for spawned agents
Fixed
- Network errors classified as retryable; compaction same-size loop guarded
- 5 agent review-round bugs: cost model, approval crash, callback imbalance, stale context, dead code
- currentModel passed to provider.stream() so fallback model switch takes effect
- Consecutive same-role messages merged in Anthropic conversion
- Synchronous settled flag for pending task detection
- Provider-level retry removed to eliminate double retry
- HTTP 413 categorized as prompt_too_long
- termcanvas-hook.mjs moved to skills/scripts/ to match registered path
- Hydra spawn uses hydra list instead of hydra watch