Skip to content

v0.14.0

Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 21 Jul 01:20
Immutable release. Only release title and notes can be modified.
v0.14.0
26e179e
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Notable changes

Vekil v0.14.0 is a major routing release. It adds explicit multi-provider failover and semantic model selection, introduces managed coding-agent launchers, and lets Responses-native models serve Chat-compatible clients.

  • Provider-agnostic model routing and safe failover — Schema v2 adds explicit model_routes, global ownership of public model IDs, ordered targets, replay-safe failover, provider-state binding, and per-attempt observability. Existing schema-v1 configurations remain supported.

    routing:
  mode: priority_failover
  max_target_attempts: 2
  max_upstream_sends: 2

    Validate configurations before deployment:

    vekil config validate --providers-config providers.yaml

  • Semantic policy routing — Opt-in policy profiles classify eligible Chat Completions requests and select a configured lightweight or powerful route. The initial release includes off, asynchronous observe, and synchronous enforce modes, along with privacy controls, readiness checks, decision provenance, and policy metrics.

    Start safely in observe mode:

    vekil \
  --providers-config policy-routing.yaml \
  --policy-routing=observe

  • Chat compatibility for Responses-native models — Models exposing only /responses can now serve OpenAI Chat Completions, translated Anthropic Messages, Gemini generation and token-counting requests, and dashboard insights. The adapter includes strict validation, typed streaming conversion, parallel function calls, and bounded tool-call replay.

    Configured native endpoint:  /responses
Client-facing compatibility: /chat/completions
                             /v1/messages
                             Gemini generateContent

  • Managed coding-agent launchers — New vekil launch commands run Claude Code, Codex CLI, and GitHub Copilot CLI through an ephemeral authenticated proxy with credential isolation, child-process cleanup, and per-session usage reporting. Claude and Codex can use their configured defaults when --model is omitted; Copilot CLI continues to require an explicit model.

    vekil launch claude
vekil launch codex --model gpt-5.4-mini
vekil launch copilot --model gpt-5.4-mini

    Agent arguments can be forwarded after --:

    vekil launch codex -- \
  exec --ephemeral "Review this workspace"

  • Improved protocol fidelity and failure handling — Anthropic requests are forwarded natively to compatible Copilot models, token-limit and incomplete-response semantics are preserved more accurately, and uncoded Azure/Foundry streamed quota failures are classified as rate limits with retry metadata where possible.

    HTTP/1.1 429 Too Many Requests
Retry-After: 30
X-Vekil-Request-ID: ...

Upgrade notes

  • Existing zero-config and schema-v1 configurations do not require migration.
  • Schema v2 validation is intentionally strict; use vekil config validate.
  • priority_failover is ordered failover—not load balancing—and only switches while replay remains safe.
  • Semantic policy routing is off by default and initially supports text and standard function tools on native POST /v1/chat/completions.
  • Stateful Responses traffic and Chat-over-Responses tool replay remain process-local; use sticky routing for stateful traffic across multiple replicas.
Assets 12
Loading