v0.18.3

What's Changed

• feat(memory): chunk large messages into multi-vector embeddings (#2551, #2552, #2543) by @bug-ops in #2564
• fix(tools): add PII NER circuit breaker for paginated reads by @bug-ops in #2572
• refactor: remove 9 feature flags, eliminate ~320 cfg gates by @bug-ops in #2583
• arch: DRY improvements — consolidate hash, truncate, config, and event conversion utilities by @bug-ops in #2584
• fix(subagent): prevent exit after 1 turn on text-only LLM response by @bug-ops in #2585
• fix(memory): apply multi-vector chunking in real-time embed paths (#2570, #2571) by @bug-ops in #2586
• fix(sanitizer,clippy): skip ML classification for policy_blocked outputs; raise hard threshold; fix doc_markdown lints by @bug-ops in #2588
• arch: trust taxonomy rename and ProviderName newtype by @bug-ops in #2587
• feat(subagent): context propagation Phase 1 — history, cancellation, model inherit by @bug-ops in #2589
• fix(memory): resolve SQLite pool starvation causing multi-turn session hangs by @bug-ops in #2592
• feat(memory,index): GAAMA episode nodes + IndexMcpServer (#2503, #2508) by @bug-ops in #2594
• feat(subagent): MCP tool annotation in sub-agent system prompt (#2581) by @bug-ops in #2598
• security(skills,memory): OOM cap, hint truncation, D2Skill/SkillOrchestra tests by @bug-ops in #2601
• feat(skills): D2Skill step-level error correction + SkillOrchestra RL routing head by @bug-ops in #2600
• test(config): add TOML float deserialization regression tests (#2599) by @bug-ops in #2602
• feat(security): cross-tool injection correlation + AgentRFC protocol audit by @bug-ops in #2603
• feat(memory): contextual tool embeddings, Focus compression, density-aware budgets by @bug-ops in #2604
• feat(skills): NL skill generation and GitHub repo mining (#2418, #1889) by @bug-ops in #2605
• security: SKILL.md injection sanitization, URL domain allowlist, trust fallback fix (#2500, #2506) by @bug-ops in #2611
• feat(agent): BATS budget hint + utility 5-way action policy (#2267, #2477) by @bug-ops in #2613
• feat(memory): SleepGate forgetting pass + performance-floor compression predictor by @bug-ops in #2614
• fix(agent): defer utility gate System hint injection until after tool results (#2615) by @bug-ops in #2616
• fix(skills): RL cold start, injection hard block, /skill create dedup by @bug-ops in #2618
• feat(skills): wire channel allowlist, cap skill-create description, wrap github_token in Secret by @bug-ops in #2617
• refactor: consolidate duplicated utilities into zeph-common by @bug-ops in #2619
• fix(agent): exclude [skipped] utility messages from semantic memory (#2620) by @bug-ops in #2623
• fix(skills): scan input for injection patterns, fix Qdrant dedup (#2621, #2622) by @bug-ops in #2624
• refactor: split large modules into focused submodules by @bug-ops in #2626
• docs: update book, specs, and READMEs for ~100 merged commits by @bug-ops in #2625
• fix(memory): ensure conversation row exists before graph episode insert by @bug-ops in #2629
• refactor: code quality — split god files, extract long functions, fix DRY violations by @bug-ops in #2631
• fix(persistence): flush orphaned tool_use on shutdown by @bug-ops in #2632
• refactor: split agent/learning.rs into focused submodules by @bug-ops in #2633
• release: v0.18.3 by @bug-ops in #2634
• docs: update book, READMEs, and specs for v0.18.3 by @bug-ops in #2640

Full Changelog: v0.18.2...v0.18.3