Skip to content

v0.7.0

Choose a tag to compare

@sherifkozman sherifkozman released this 28 Mar 01:16
· 25 commits to main since this release
5ccd38c

[0.7.0] - 2026-03-27

Added

  • Mode-aware runtime wiring across Council and Orchestrator
    • runtime mode, runtime_profile, reasoning_profile, temperature, max_tokens, and output-schema overrides now flow into execution
    • execution plans now expose effective mode, schema, model pack, providers, phase budgets, and degradation details
  • Routed handoff support
    • council run router ... --route can now continue into the router-selected subagent and mode
    • routed runs can carry model-pack, execution-profile, budget-class, and capability recommendations into the follow-up run
  • Capability-aware execution baseline
    • capability planning and execution-profile selection added to the runtime
    • initial packs include repo-analysis, docs-research, planning-assess, diff-review, security-code-audit, and red-team-recon
  • Deterministic evaluation tooling
    • new council eval, council eval-compare, and council eval-import-pr commands
    • public runtime baseline dataset under evals/runtime-baseline.yaml
  • Deep doctor support
    • council doctor --deep now distinguishes installed/configured providers from providers that can answer a trivial non-interactive prompt

Changed

  • Canonical CLI provider names are now codex, gemini, and claude, with legacy aliases preserved for compatibility
  • Public docs now describe the current runtime surface, local-only eval import boundary, and routed execution behavior
  • Bounded runtime request caps were recalibrated to better match real review workloads

Fixed

  • Explicit user-selected providers and configured models now take precedence over subagent defaults
  • Auto-fallback now stays limited to the default unhealthy path instead of overriding explicit user provider choices
  • CLI provider timeout settlement now kills the full subprocess tree instead of hanging on timed-out children
  • Gemini CLI approval-mode handling updated to current CLI behavior
  • OpenAI timeout and provider error reporting now preserve the underlying cause chain for easier diagnosis
  • Bounded review prompts are slimmer and no longer force schema-heavy draft/critique instructions