v0.6.0

Summary

• Added guided onboarding and Control UI onboarding/status support for more provider paths, including Anthropic setup-token auth profiles, Bedrock validation, and Vertex setup guidance.
• Standardized model routing on explicit provider:model syntax and expanded Vertex AI support to Anthropic, Meta, Mistral, and Nvidia third-party publishers via streamRawPredict.
• Added migration/import flows for OpenClaw, OpenCode, Aider, and NemoClaw so existing provider configuration can be brought into Carapace with previewed mappings.
• Hardened secret/auth-profile/plugin state handling with Argon2id new writes, versioned auth-profile store envelopes, deferred auth-profile metadata persistence, runtime auth-profile caching for tool calls, and transactional managed-plugin writes.
• Added a Claude CLI backend/provider for guarded local CLI-backed use cases.

Breaking Changes

• Explicit provider:model routing is now required across config and request surfaces. Bare model names and implicit/default provider inference are no longer accepted.
• Existing configs or automation that relied on unprefixed model IDs must be updated before upgrade. After upgrade, bare or unrecognized model IDs fail config validation and startup until they are corrected.

Migration Steps

• Before upgrading, create a backup:
  • cara backup --output ./carapace-backup.tar.gz
• Audit configured model IDs and convert them to canonical provider:model form.
  • Examples:
    • anthropic:claude-sonnet-4-20250514
    • gemini:gemini-2.0-flash
    • codex:default
    • vertex:gemini-2.5-flash
    • vertex:publishers/anthropic/models/claude-sonnet-4-20250514
• If you are migrating from another tool, prefer an import flow instead of hand-copying secrets:
  • cara import openclaw
  • cara import opencode
  • cara import aider
  • cara import nemoclaw
• After upgrade, verify the install:
  • cara verify --outcome auto
  • cara verify --outcome autonomy

Rollback Steps

• Reinstall the previous known-good binary, for example v0.5.0.
• If v0.6.0 performed any writes to encrypted config secrets, auth profiles, or backups, restore the backup you created before upgrade:
  • cara restore --path ./carapace-backup.tar.gz
• Re-run:
  • cara status --port 18789
  • cara verify --outcome auto --port 18789
  • cara verify --outcome autonomy --port 18789

Security

• New encrypted secret writes now use Argon2id-backed enc:v2 envelopes, while legacy PBKDF2 material remains decryptable for compatibility on the new release.
• Auth-profile store state is versioned and rollback-aware, with request hot paths moved off synchronous last_used persistence and tool-side auth-profile resolution moved onto a bounded runtime cache.
• Control onboarding/status responses now use typed, fail-closed projection instead of ad hoc browser-facing detail forwarding.
• Managed plugin install/update writes are now transactional and hardened against local symlink-swap races.
• No new public advisories are introduced by this release.

Verification

• Verify published artifacts and Sigstore bundles:
  • RELEASE_TAG=v0.6.0 ./scripts/smoke/verify-release-artifacts.sh
• After upgrading, verify runtime behavior:
  • cara verify --outcome auto
  • cara verify --outcome autonomy
• If you use Vertex AI third-party publishers, validate one real target after enabling the model in Vertex Model Garden and authenticating with gcloud auth application-default login.

Known Caveats

• Rolling back after v0.6.0 has rewritten encrypted secrets, auth profiles, or backup artifacts should be done with the pre-upgrade backup, not by downgrading the binary alone.
• Control UI onboarding/status coverage is substantially broader than v0.5.0, but not every provider has full browser-native write parity yet.
• Vertex setup validation remains strongest for Google-native Gemini paths; third-party publisher setup still depends on per-model enablement and provider-specific availability inside Vertex AI.
• Claude CLI provider support depends on a locally installed and authenticated Claude CLI.