v2026.5.5

Added

• Native extended-thinking effort. New model.thinking field (false | minimal | low | medium | high | xhigh) populates PydanticAI's ModelSettings["thinking"] at agent-build time, validated at load against reasoning-model support, with per-flow-agent and per-persona overrides. Thinking-token usage is read from the model usage details and recorded on RunResult and in the audit trail. See docs/core/reasoning.md.
• Per-role output modes. New output.mode (auto | tool | native | prompted | text, default auto) wraps the resolved output type in the matching PydanticAI wrapper (ToolOutput / NativeOutput / PromptedOutput / TextOutput); auto defers to PydanticAI's native model-profile detection. See docs/core/structured-output.md.
• Verified reflexion. ReflexionStrategy can bind the LLM judge as a verification gate: after each critique round it scores output against the role's success criteria, short-circuits once all criteria pass, and otherwise feeds the per-criterion failure reasons into the next continuation prompt. Verdicts are audited, turning "reflect N times" into "reflect until verified, max N". See docs/core/reasoning.md.
• Hybrid retrieval. New retrieval strategy (vector | hybrid | hybrid_rerank, default vector) on the document store. hybrid adds a BM25 full-text index alongside the vector column and fuses results with reciprocal rank fusion; hybrid_rerank adds an optional cross-encoder stage that degrades to plain RRF when sentence-transformers is absent. New hybrid_search method on the DocumentStore ABC. See docs/core/rag-guide.md.
• Durable, resumable flows. New durability block on FlowSpec with an audit-backed checkpoint journal (flow/checkpoint.py) keyed by flow_run_id. Completed delegation envelopes are recorded through the HMAC audit chain and replayed on initrunner flow resume <id> instead of re-running. backend is none | journal (default none), so single-shot and REPL runs are unaffected. See docs/orchestration/durability.md.
• Ensemble voting. New ensemble delegate-sink and team strategy plus a SpawnPool-backed fan-out that runs the same prompt K times. A reducer selects the winner by majority, weighted (per-target YAML weight), or judge (the LLM judge), and emits the winning envelope and full vote trace to the audit chain. See docs/orchestration/flow.md.
• First-class evaluation on pydantic-evals. The YAML TestSuite schema is unchanged, but cases now run through a pydantic-evals Dataset of evaluators producing an EvaluationReport. Adds span-based assertions over OpenTelemetry spans and the structured run-event timeline, and reuses the existing LLM judge as a custom evaluator so local, no-Logfire setups still work. See docs/operations/testing.md.
• Cyclic loop-back routing. New bounded loop_back edge ({target, max_iterations, until}) on the delegate sink enables writer/critic refine loops. Cycle detection in initrunner/_graph.py now permits only explicitly-marked loop-back edges; every other cycle is still rejected. See docs/orchestration/flow.md.
• Structured blackboard shared-state. A typed Blackboard is now the flow graph's run state, exposed to agents through a run-scoped blackboard tool (post / read / claim / list). Fan-in joins read attributed entries instead of only concatenating prompt strings, and the final board is persisted to the audit chain. See docs/orchestration/blackboard.md.
• In-process local embeddings. New provider: local embeds documents in-process via fastembed (the local-embeddings extra), distinct from the HTTP-backed ollama provider, so no endpoint is required and no document text leaves the process. See docs/configuration/providers.md.
• Dashboard surfacing for the new capabilities. The audit table and cost views now show thinking and reasoning token counts; the audit detail drawer renders a per-run event timeline (tool calls, tool results, thinking) and verified-reflexion judge verdicts via a new GET /api/audit/{run_id} endpoint; and the agent builder exposes a thinking-effort selector backed by the valid config enums from the builder options endpoint.

Changed

• Full streaming-event consumption. The streaming executor now branches on ThinkingPartDelta, FunctionToolCallEvent, and FunctionToolResultEvent in addition to text deltas, persists a redacted structured event timeline to the append-only audit log, attributes reasoning tokens separately, and treats the native typed events as the source of truth over the legacy ContextVar tool-event path. Events are forwarded to the display and SSE consumers.
• Audit schema. AuditRecord gains thinking_tokens, a structured run-event timeline, and judge-verdict fields; the HMAC chain covers the new fields and an idempotent migration backfills existing databases (legacy rows verify and default to 0 or empty).
• Store context managers. DocumentStore.__enter__ and MemoryStoreBase.__enter__ now return Self, preserving the concrete store type inside with blocks.

Fixed

• Durable daemon flows no longer prune checkpoints after a failed run. The trigger-driven flow daemon (initrunner flow up) pruned the checkpoint journal after any graph.run() that did not raise, so a run where a sub-agent returned RunResult(success=False) still had its checkpoints deleted. initrunner flow resume <id> then found zero completed services and re-ran the flow from the start, defeating resume-after-failure. The daemon now gates the prune on overall-run success (a per-run flag set false when any completed sub-agent reports success=False), mirroring the orchestrator's run_once path. Failed runs keep their checkpoints so resume replays the completed services and re-runs the failed one; fully successful runs still prune.
• initrunner run now persists the structured run-event timeline. The run-event timeline (tool calls, tool results, thinking deltas) was only built and stored when a live on_event consumer was wired, so every CLI run (buffered and streaming) wrote event_timeline_json = NULL to the audit database. The timeline is now reconstructed from the run's final message history whenever the run is audited, independent of any live UI consumer, reusing the existing redacted entry shapes and length bounds. Capture is gated on audit being enabled, so --no-audit runs add no overhead, and the live on_event path is unchanged.
• ensure_private_dir no longer crashes on a parent directory it does not own. Pointing --audit-db (or a store path) at a file inside a shared directory such as /tmp previously raised PermissionError because the helper unconditionally chmoded the parent to 0o700. It now tightens permissions only on a directory it just created or already owns, and otherwise leaves the directory untouched. Data privacy is unaffected: the audit database and store files are still locked to 0o600.

Dependencies

• Add pydantic-evals to the observability optional-dependencies extra.
• Wire the local-embeddings extra (fastembed) into the embedding factory.
• Bump dashboard vite 8.0.13 to 8.0.14 (#132).
• Bump dashboard svelte group with 2 updates (#131).