v0.9.0

AgentRunbook memory + a visible, honest LLM backend.

AgentRunbook multi-granularity memory (opt-in)

Distillation layers above raw session findings — Event (per-transition), Runbook (reusable procedures), Gotcha (recurring failure modes) — plus query-decomposed multi-pool retrieval. Configurable via tesserae compile --distill, tesserae context --multi-pool, distillation.enabled in config, and the MCP compile_context multi_pool arg. All byte-idempotent and default-off.

A visible LLM backend

• tesserae config status — resolved provider + codex-home/claude-dirs with the source of each setting, plus a live ping. A rate-limited/mis-authed backend is now obvious.
• Loud session-extraction failures — distinguishes "model returned nothing" from "model never answered"; a failed chunk is no longer cached as empty, and compile warns loudly.
• Codex reasoning effort pinned to medium for Tesserae's calls (configurable), so a global xhigh doesn't make every compile many times slower.

Determinism hardening

Byte-idempotence guards scoped to the deterministic compile; no-LLM compiles re-emit cached community summaries instead of dropping them.

Drop-in upgrade from 0.8.x: pip install --upgrade tesserae==0.9.0. Release note (7 languages): docs/release-notes/v0.9.0.md.