v0.0.11 — Effort-aware model pools and truthful routing

What’s new

• Added first-class reasoning-effort variants. Provider default, Low, Medium, High, Max, and other catalog-advertised efforts now behave as independent model instances with readable endpoint IDs, their own roles and readiness, variant-specific benchmark results, telemetry, and routing decisions.
• Updated pi-role-model to discover effort variants, route Pi CLI requests to the selected variant, serialize provider-compatible effort values, and preserve the selected identity through telemetry and downstream OpenAI-compatible requests.
• Unified the configured model pool across Overview, Models, Benchmark, Router, Observe, Connect, and the sidebar. Each configured eligible endpoint appears once, and model-pool views no longer silently omit candidates.
• Made benchmarking variant-exact from selection and progress through persisted scoring and routing profiles. Completed runs refresh candidate quality without a restart, while failed, cancelled, or stale results cannot overwrite valid profiles.
• Improved candidate profiles by combining current per-variant benchmark and live telemetry revisions, and attached those revisions to routing decisions for inspection.
• Added admission-time readiness checks. Newly added API-key models are verified before they can route or benchmark; a failure degrades only the affected instance and provides an actionable, secret-safe retry path.
• Made OAuth-backed admission follow the successful OAuth callback instead of starting a premature provider probe, preventing completed Codex subscriptions from appearing unavailable.
• Added explicit confirmation when ejecting the final controller and a truthful empty-pool/no-eligible-model state.
• Removed production fixture and mock fallbacks. Release packages are checked to ensure they contain no configured-model fixtures, benchmark fixtures, QA state, or credentials.
• Improved model-pool visualization with deterministic, distinct colors and complete scrollable candidate lists.
• Normalized provider presentation so managed adapters such as LiteLLM do not appear as duplicate user-configured providers.
• Improved adaptive recovery from upstream failures and preserved clear per-instance health state.
• Bundled the mandatory Track B runtime and all 13 extensions in every production package, with canonical variant identity available to supported routing, telemetry, message-lineage, contribution, recommendation, and storage projections.

Upgrade notes

• Existing legacy opaque effort IDs remain readable and migrate deterministically to the new readable identities.
• Runtime state and credentials remain outside the installation directory. Reusing the same state root during an upgrade preserves configured models, message history, and historical telemetry.