v2026.4.17

Added

• spec.model.fallback accepts a list of provider:model strings (or aliases). When set, _build_model wraps the primary and fallbacks in pydantic-ai's FallbackModel so runs survive single-provider outages (5xx, 429, auth, connection resets) without call-site changes. Each entry is resolved and validated at load time; Ollama and custom-base_url providers are rejected in the fallback chain because aliases cannot carry a base_url. The executor catches FallbackExceptionGroup on both sync and async paths and classifies by the last inner exception, emitting an error string that lists every provider's failure. See examples/roles/fallback-demo.yaml and docs/configuration/providers.md.
• Human-in-the-loop approval. Mark any tool config with approval: required and runs pause whenever the model wants to call it, using PydanticAI's native DeferredToolRequests / DeferredToolResults contract so each runner mode resumes without re-prompting or losing message history.
  • RunResult gains status and pending_approvals; executor exposes execute_run_resume and execute_run_resume_async.
  • New pending_approvals audit table persists the message history so another process can resume.
  • REPL prompts y/N inline; single-shot exits 2 with a CLI resume hint; daemon persists and keeps serving other triggers; API returns finish_reason: tool_calls_pending_approval and exposes POST /v1/approvals/{run_id}.
  • CLI: initrunner pending, initrunner approve RUN_ID [--all|--tool-call-id] [--deny]. Docs: docs/security/approvals.md.
• Dashboard approval queue. /approvals queue route with keyboard navigation and bulk actions; inline ApprovalCardGroup in RunPanel; ApprovalDrawer for multi-call detail; ShortcutOverlay (?). Sidebar Approvals entry under Operate with a count badge (tabular-nums, polling every 20s + SSE bump). New router routers/approvals.py (GET /api/approvals/pending, GET /api/approvals/{run_id}, POST /api/approvals/{run_id}) calls services.execution.resume_run_sync in-process. Streaming layer fires a new approval_required SSE event when the executor pauses and carries status + pending_approvals in the result payload.
• Streaming structured output. Structured-output roles now stream progressively-validated partials via StreamedRunResultSync.stream_output() (the previous hard forbid on output.type != "text" is gone). Sync and async streaming paths accept an on_partial callback; the async path additionally accepts an on_event callback that yields typed AgentStreamEvent instances via run_stream_events(). Dashboard emits partial_output SSE frames for structured roles instead of falling back to non-streaming.
• spec.execution block for agent-level execution semantics (retries, output_retries, end_strategy, tool_timeout_seconds), distinct from guardrails budgets.
• spec.execution.max_concurrency ({max_running, max_queued}) wires pydantic_ai.ConcurrencyLimit on Agent for per-agent backpressure. max_running is required when the block is present.
• spec.deps_schema with {{var}} templating. Flat-scalar interpolation into spec.role and related fields, rendered through a local hook (no pydantic-handlebars dep). initrunner run --var KEY=VALUE threads values into single-shot runs. See docs/getting-started/agent-spec-import.md.
• Agent Spec import/export. initrunner run --agent-spec <file> imports a PydanticAI Agent Spec (JSON or YAML) as a transient role; initrunner export agent-spec <role.yaml> emits the reverse mapping. retries, output_retries, end_strategy, and tool_timeout round-trip through spec.execution.
• Run identifiers in PydanticAI metadata=. agent.run(metadata=...) now carries initrunner.run_id, initrunner.agent_name, and initrunner.trigger_type (when set) alongside the existing input_validated flag, so Logfire and equivalent backends surface them on PydanticAI's emitted spans without scraping InitRunner's parent span attributes.

Changed

• Agents now receive their static directive via PydanticAI's instructions= kwarg instead of system_prompt=. The composed spec.role + skill prompts + auto-skill catalog + tool-search hint flow to instructions=; the two dynamic @agent.system_prompt decorators (procedural memory, resume context) stay as-is. Matches PydanticAI 1.71+ guidance. No role YAML change.
• stores/_helpers._filter_system_prompts rewritten. Preserves ModelRequest.timestamp, run_id, and metadata through filtering (previously dropped by bare ModelRequest(parts=...) reconstruction). Also normalizes ModelRequest.instructions by retaining it only on the newest two retained requests, matching PydanticAI's _get_instructions resolver fallback.
• Five internal helper agents (eval/judge, agent/policies classifier, three services/agent_builder wizard/import agents) migrated to instructions= for consistency.
• _validate_provider iterates the full model + fallback chain so a missing provider SDK fails fast at load time, not at failover time.

Fixed

• Race on shared run-scoped state. todo, think, spawn, and clarify toolsets now build with FunctionToolset(sequential=True). PydanticAI's default tool-execution mode is parallel; if the model emitted two calls to the same toolset in one turn, they could race on ReflectionState.todo or SpawnPool internals.

Removed

• FlowOrchestrator.max_agent_workers parameter. Non-functional since the pydantic-graph async migration; no callers referenced it outside its declaration line.