Skip to content

MIRA v0.340.1

Choose a tag to compare

View all tags
@tarekedOz tarekedOz released this 20 Aug 05:38
· 5 commits to main since this release
v0.340.1
3a719f2

MIRA v0.340.1 — public beta. Separate liveness/readiness health probes, dependency security patches, and automation improvements. One breaking HTTP-API change (see below); safe to upgrade the binary in place.

Self-hostable single binary, AGPL-3.0-or-later, signed with the vexillon.ai release key.

Install

# Linux / macOS
curl -fsSL https://get.vexillon.ai/install.sh | sh
# Windows (PowerShell)
irm https://get.vexillon.ai/install.ps1 | iex

Health probes: /livez + /readyz (breaking)

The overloaded /health route is split into two purpose-built probes:

  • /livez — liveness. Dependency-free: returns 200 ok without touching a provider, the database, or any subsystem. Use it to answer "is the process up?".
  • /readyz — readiness. Returns 200/503 based on provider reachability, and is cached (serve-stale-while-revalidate + single-flight) so a fleet of load-balancer probes triggers at most one provider round-trip per window and never blocks. The body reports cached_age_secs. Tune with the new server.readiness_cache_ttl_secs (default 30; 0 disables caching).
  • /health now falls through to the in-app System Health page (SPA), not a machine endpoint.

Action for self-hosters: point external monitors and container HEALTHCHECKs at /livez (or /readyz for provider readiness). MIRA's own supervisor probe auto-migrates a configured …/health URL to …/livez on load; the bundled Dockerfile HEALTHCHECK already uses /livez.

Dependency security patches

Cleared the outstanding Dependabot advisories: react-router 6 → 7 (declarative APIs only, React 19-compatible), plus postcss, js-yaml, brace-expansion, quinn-proto, and rpassword 7.5.4 / rtoolbox 0.0.5 (the latter also fixes the macOS build).

Automations

  • Create-on-behalf-of. An admin can create a schedule for another user (POST /api/schedules accepts an optional user_id; non-admins get 403). The owner's quota applies; the acting admin is audited separately.
  • Editor budgets. Per-run max_iterations and max_action_secs are exposed under an Advanced disclosure. Sub-daily schedules now default to a fresh conversation per run (with a warning if you point a repeating trigger at a named conversation).
  • Skills forward user context. Built-ins invoked through a skill wrapper now receive the caller's context (user, conversation, channel, timezone), with trusted server-side values overriding anything the model supplies — fixing user-scoped built-ins (wiki, memory, calendar) that previously failed when reached through a skill.

Chat

  • The conversation auto-titler no longer persists a model's reasoning scaffolding (markdown-bold or numbered-list lines like 1. **Analyze User Input:**) as the title — it skips them and reaches the real title.

Verify (optional)

Every archive is signed with the vexillon.ai minisign key committed at verification/release-pubkey.minisign:

minisign -Vm mira-0.340.1-<target>.tar.gz -p verification/release-pubkey.minisign
Assets 14
Loading