v1.0.1

Security & Reliability Hardening

• Strict Multi-Tenant Resource Isolation:
  • Enforced project-level authorization across runs, results, traces, plans, and asynchronous jobs.
  • Cross-tenant resource queries return 404 Not Found fail-closed to prevent resource enumeration.
  • SQL-level tenant scoping applied before loading spans and trace records.
• Durable Asynchronous Worker Leases & Recovery:
  • Implemented lease expiration and heartbeat tracking (worker_id, lease_expires_at, heartbeat_at) in run_jobs table.
  • Background worker thread renews active leases every 15 seconds.
  • Automatic stale job recovery sweeps abandoned worker reservations and marks dead executions failed.
  • Cooperative cancellation checked at case execution boundaries.
• Financial Concurrency & Execution Integrity:
  • Additive SQLite Migration 8: added idempotency_key column on run_jobs with unique partial index, and project_id on budget_reservations.
  • Atomically reserved budget per case during concurrent execution, releasing held reservations immediately upon case completion or failure.
  • Deterministic financial transaction IDs (tx-{run_id}-{case_id}) with INSERT OR IGNORE semantics preventing duplicate spend booking.
• Control-Plane MCP Parity:
  • Added get_budget_status, list_jobs, and cancel_job MCP tools.
  • Strictly forbade agent self-approvals, preserving human confirmation requirements.
• SDK Multi-Tenant & Idempotency Support:
  • Extended Python SDK plan() and benchmark() with project_id.
  • Added X-Project-Id and Idempotency-Key headers to TypeScript SDK RunportClient.

v1.0.0

Added

• Universal Developer Instrumentation API (runport.instrumentation):
  • @trace decorator for synchronous and asynchronous Python functions, agents, and tool calls.
  • Context managers start_trace() and start_span() with contextvars hierarchical span tracking.
  • LLM client wrappers wrap_openai() and wrap_anthropic() extracting exact token metrics and outputs as canonical child spans.
• Framework Interoperability:
  • OpenTelemetry (OTel) GenAI Semantic Conventions and OpenInference trace ingestion & export (POST /api/v1/traces/otlp, runport import-otlp, runport export-otlp).
  • Bidirectional Langfuse v2 batch export & import (POST /api/v1/traces/langfuse/export, POST /api/v1/traces/langfuse/import, runport export-langfuse, runport import-langfuse).
  • Native LangGraph SSE streaming adapter (LangGraphStreamingConnector) and in-process callback tracer (RunportLangGraphTracer).
• Production-Grade Model Context Protocol (MCP):
  • Streamable HTTP (SSE) transport (GET /mcp/sse, POST /mcp/messages) supporting protocol revisions up to 2026-07-28.
  • Bearer token authentication and strict self-approval prevention.
  • Ready-to-use agent IDE blueprints for Claude Desktop, Cursor, and VS Code.
• Hosted Multi-Tenancy & RBAC (docs/MULTI_USER_SECURITY_MODEL.md):
  • Additive SQLite Migration 6: organizations, projects, memberships, and audit_log tables.
  • Project-scoped filtering across API keys, runs, traces, and ledger records.
  • Granular RBAC permissions for VIEWER, MEMBER, APPROVER, ADMIN.
  • Fail-closed self-approval guard blocking principals from approving their own billable plans.
• Versioned REST API v1 & SDKs:
  • OpenAPI 3.1 REST API v1 (/api/v1/) with typed Pydantic v2 schemas and RFC-7807 error envelopes.
  • Asynchronous background job worker (JobManager) with non-blocking execution and cancellation.
  • Remote Python client (runport.Client) with retry handling and typed exceptions.
  • Standalone TypeScript SDK package (@runport/sdk).
• Canonical Tracing & Trajectory Evaluation:
  • Pydantic v2 Trace and Span models enforcing Invariant #1 ("Null is not 0").
  • SQLite WAL persistence with indexes for trace hierarchy and parent-child span trees.
  • Interactive web waterfall visualization (TraceView.tsx) and collapsible TUI trace tree.
• Automated GA Release Gates (tests/test_release_gates_v1.py):
  • 6 automated release gates verifying package cleanliness, offline hermeticity, fail-closed spend caps under concurrency, cryptographic tamper detection, null integrity, and secret redaction.

v0.1.0-alpha

Added

• Universal Target Execution: Unified Target abstraction with connectors for openai_compat, openrouter, dify, http_workflow, python_target, and ollama.
• Preflight Plan Fingerprint: Cryptographic SHA-256 fingerprint binding models, endpoint routes, dataset hash, prompt template hash, and budget caps prior to live execution.
• Fail-Closed Spend Ledger: Append-only SQLite WAL ledger tracking spend with strict provenance vocabulary (ACTUAL, ESTIMATED, UNRESOLVED, UNAVAILABLE). Unresolved spend halts subsequent runs fail-closed.
• Strict Ground Truth Verification: Supervised evaluation metrics (Accuracy, Balanced Accuracy, Precision, Recall, Macro F1, Wilson 95% confidence intervals, Cohen's Kappa, and Confusion Matrices) strictly evaluated against human ground truth.
• Local Read-Only Interfaces:
  • Full-screen Rich Terminal User Interface (runport ui) with live histograms and monitor panes.
  • Interactive launcher and workflow menu (runport).
  • Local web dashboard (runport web) built with FastAPI, React, TypeScript, and Vite, served strictly on localhost.
• Sanitized Trace Exports: Zip/JSON trace exporters with automatic scrubbing of API keys, bearer tokens, and sensitive security identifiers (runport export-traces).
• Offline Replay Runner: Zero-cost, isolated replay harness (REPLAYED provenance, $0.00 spend) for deterministic quickstart demos and contract testing.