This release adds 3 notable features for engineering teams evaluating rollout.
✓ No known CVEs patched in this version
Affected surfaces
ReleasePort's take
Light signalThe Perplexity `approved_models` regression is fixed by adding fallback model lists in v0.6.0. Additionally, several enhancements and refactorings improve initialization workflows and routing.
Why it matters: Patch to v0.6.0 immediately if you use the Perplexity approved_models feature; benefits from new init guidance and router consolidation can be tested in dev environments.
Summary
AI summaryFixed Perplexity approved_models regression by adding fallback model lists.
Changes in this release
| Type | Severity | Summary | CVE |
|---|---|---|---|
| Feature | Medium |
Native Gemini provider routes through dedicated endpoint with retry and timeout handling. Native Gemini provider routes through dedicated endpoint with retry and timeout handling. Source: llm_adapter@2026-05-21 Confidence: high |
— |
| Feature | Medium |
`dragoman init` now installs per-vendor guidance files for each supported vendor. `dragoman init` now installs per-vendor guidance files for each supported vendor. Source: llm_adapter@2026-05-21 Confidence: high |
— |
| Feature | Medium |
Claude Code can delegate AgentTool calls to dragoman through agent surface. Claude Code can delegate AgentTool calls to dragoman through agent surface. Source: llm_adapter@2026-05-21 Confidence: high |
— |
| Bugfix | Medium |
Perplexity approved_models regression fixed; fallback models declared for empty live discovery. Perplexity approved_models regression fixed; fallback models declared for empty live discovery. Source: llm_adapter@2026-05-21 Confidence: high |
— |
| Refactor | Medium |
`dragoman init` rewritten as interactive wizard with provider discovery and secret storage. `dragoman init` rewritten as interactive wizard with provider discovery and secret storage. Source: llm_adapter@2026-05-21 Confidence: high |
— |
| Refactor | Medium |
Dynamic semantic model registry replaces hardcoded model catalogue. Dynamic semantic model registry replaces hardcoded model catalogue. Source: llm_adapter@2026-05-21 Confidence: low |
— |
| Refactor | Medium |
Router consolidation into single `routers/providers.py` module. Router consolidation into single `routers/providers.py` module. Source: llm_adapter@2026-05-21 Confidence: low |
— |
Full changelog
Supersedes v0.5.0, which has been yanked. 0.5.0 shipped with a
dragoman initregression that wrote noapproved_modelsfor Perplexity (Sonar API has no/modelsendpoint, so discovery returned empty and the wizard silently skipped it). Fixed here.
New
- Native Gemini provider — Gemini now routes through its own provider, not the OpenAI-compatible shim. Retry logic and timeout handling added alongside.
- Vendor agent-memory templates —
dragoman initnow installs per-vendor guidance files (vendor-anthropic.md,vendor-google.md,vendor-ollama.md,vendor-openai.md,vendor-perplexity.md,vendor-xai.md) so the subagent picks the right model per vendor's strengths. - AgentTool delegation — Claude Code can call dragoman through the agent surface, not just CLI.
Refactored
- Dynamic semantic model registry — replaces the hardcoded model catalogue. Models are discovered from each provider's live API where possible, with per-provider fallback lists for surfaces that don't expose one.
dragoman initrewrite — interactive wizard with provider discovery, secret storage, and clean install/uninstall paths.- Router consolidation — single
routers/providers.pyreplaces the per-provider modules.
Fixed
- Perplexity
approved_modelsregression —_PROVIDERSentries now declarefallback_models. When live discovery returns empty (as it does for Perplexity's Sonar surface), the fallback list is used. Verified againstdocs.perplexity.ai/docs/sonar/models.
Install
```bash
pipx install dragoman-ai
dragoman init
```
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
About Dragoman
All releases →Related context
Related tools
Beta — feedback welcome: [email protected]