Skip to Content

Changelog

All notable changes to the isA platform.

v1.1.0 — July 21, 2026

isA_Docs — AEP Control Plane

  • [New] Canonical AEP control-plane guide covering ownership boundaries, roles, and the complete project-to-release operating pipeline.
  • [New] English and Chinese Platform navigation for the guide, with architecture cross-links into the AEP operations portal.
  • [Updated] Added docs contract coverage for localized navigation, pipeline vocabulary, and portal action links.

v0.9.0 — May 17, 2026

Platform Shell and Connectors

  • [New] Per-connector policies and evaluator support in the main isA app.
  • [New] Conversation tool-request plumbing and interactive connector surfaces.
  • [Updated] Claude parity certification flow for connector behavior.

isA_App_SDK

  • [New] @isa/service-agents package for embeddable vertical service agents.
  • [New] Web widget bundle (service-agents-widget.js) with window.IsAServiceAgents initialization and Shadow DOM isolation.
  • [New] Service-agent gateway endpoints for listing agents and sending chat messages.
  • [New] A2UI interaction contracts for knobs, color pickers, 2D sliders, draggable cards, drop zones, comparison grids, annotated diffs, and flow diagrams.

isA_Console

  • [New] Customer service control plane under /dashboard/customer-service.
  • [New] Agent prompt, knowledge, conversation, and deployment management pages.
  • [Updated] Developer console pages for API keys, webhooks, SDKs, code examples, settings, and personalized quickstart snippets.

isA_Model

  • [New] Sleep intelligence API mounted under /api/v1/sleep.
  • [New] Sleep evaluation and recommendation routes: /api/v1/sleep/evaluate and /api/v1/sleep/recommend.
  • [New] Commercial sleep model products: com_sleep_quality, com_sleep_stage, and com_health_risk.
  • [New] Commercial model serving evidence gate script for validating model products before release.

isA_Data

  • [New] Framework data-product marketplace search, product detail, and subscribe flows.
  • [New] Certification report APIs and policy-evaluation support for data products.
  • [New] Support and trust cards for marketplace listings.
  • [Updated] Canonical product identifiers and framework schema bundle/versioning.

isA_Cloud

  • [New] Big-data foundation stack for local KIND validation: Kafka, PostgreSQL big-data metadata, Hive Metastore, MinIO, Iceberg tools, StarRocks, Apicurio Registry, Flink, Fluss, and Flink CDC jobs.
  • [New] Validation targets: setup-datalake-kind, verify-bigdata-kind, verify-bigdata-kind-readiness, verify-bigdata-kind-smoke, and teardown-bigdata-kind.
  • [Updated] Chart coverage now includes cert-manager, External Secrets Operator, and Prometheus Operator hardening.

isA_MCP

  • [New] /live liveness endpoint for Kubernetes probes.
  • [New] Runtime configurable branding for white-label deployments.
  • [New] Backup/restore scripts and disaster-recovery runbook scaffolding.
  • [Updated] Liveness probes and chart hardening documentation.

isA_Mate

  • [New] OpenTelemetry tracing propagation across the gateway and channel senders.
  • [New] A2A delegation facade, event subscriptions, background jobs, and declarative DAG workflow configuration.
  • [Updated] Local development readiness now gates on /health; risky browser actions support human-in-the-loop review.

isA_OS

  • [New] Streaming execution surfaces for Python REPL, Code REPL, and Cloud OS.
  • [New] Web Automation A2A service.
  • [New] Pool Manager Prometheus metrics for active, warm, idle, max, healthy, utilization, and queue-depth resource state.

isA_user

  • [New] Memory service stats endpoint: GET /api/v1/memories/stats?user_id=....

isA_Agent_SDK

  • [New] Model usage counters for model calls, input tokens, output tokens, and cost.
  • [New] MemoryStack list and stats helpers.
  • [Updated] Human-in-the-loop query APIs and interrupt checkpoint support.

isA_Vibe

  • [Fixed] isa-vibe --list-workflows now fills legacy defaults instead of crashing when older configuration fields are absent.

v0.8.0 — April 11, 2026

isA_Console — Developer Platform Parity

Closing feature gaps identified in a competitive analysis against OpenAI and Anthropic developer platforms. 33 stories across 6 epics.

  • [New] Prompt Management: /dashboard/prompts — create, version, and test reusable prompt templates with {{variable}} syntax
  • [New] AI Prompt Generator: generate and improve prompts from natural language descriptions
  • [New] Evaluation Suite: /dashboard/evaluations — test cases, grading (exact match, LLM-as-judge, regex), side-by-side model comparison
  • [New] Batch Processing: /dashboard/batch — submit JSONL jobs at 50% reduced cost, progress tracking, download results
  • [New] Request Logs: /dashboard/logs — inspect individual API requests with filters, expandable payloads, live tail
  • [New] Audit Log: /dashboard/audit-log — immutable org event trail (user, key, agent, auth events)
  • [New] Conversations: /dashboard/conversations — manage hosted conversation state
  • [New] Projects: /dashboard/projects — project-scoped isolation for keys, usage, and budgets
  • [New] Members & RBAC: /dashboard/members — role management (Owner, Admin, Developer, Billing, Viewer), 30 granular permissions
  • [New] Webhook Management: Settings > Webhooks tab — CRUD, event subscriptions, delivery logs, test endpoint
  • [New] Vector Stores: /dashboard/vector-stores — managed RAG with file upload, semantic search
  • [New] Service Health Dashboard: /dashboard/health — real-time service status cards, auto-refresh
  • [New] Pricing Page: /pricing — public page with tier comparison, per-model pricing table, cost calculator
  • [New] Structured Outputs: playground toggle with JSON Schema editor and validation feedback
  • [New] Sidebar Navigation: added Conversations, Prompts, Evaluations, Logs, Vector Stores, Batch, Health, Projects, Members, Audit Log

isA_Console — Admin Services

  • [New] PromptAdminService — CRUD + versioning for prompt templates
  • [New] EvalAdminService — evaluation suites, test cases, runs, results, comparison
  • [New] BatchAdminService — batch job lifecycle with multipart upload
  • [New] ObservabilityAdminService — request logs and audit log queries
  • [New] WebhookAdminService — webhook CRUD, test, delivery logs
  • [New] usePermissions hook — client-side RBAC with 30 granular permissions
  • [New] useServiceHealth hook — SWR-based health polling across 6 services

isA_Docs — Documentation Gaps

  • [New] Prompt Engineering Guide: comprehensive guide covering zero-shot, few-shot, CoT, tool use, agent prompts, model-specific tips
  • [New] Error Reference: complete HTTP error code reference with causes, resolution, retry examples
  • [New] Migration Guides: Chat API, Agent Config, Auth, MCP version migration paths
  • [New] SDK Quickstarts: Python, TypeScript, cURL — copy-paste ready, under 5 minutes
  • [Updated] Changelog: structured format with dated entries, categorized by area

v0.7.0 — March 27, 2026

isA_Agent_SDK — Long-Running Agent Hardening (Epic #510)

  • [New] Tiered memory hierarchy: replaced flat summary with 3-tier system (working → episodic → semantic memory)
  • [New] BudgetGuard: max_credits/max_tokens enforcement with pause/warn/stop actions
  • [New] SessionWatchdog: stall detection with NATS heartbeat publishing and Redis liveness polling
  • [New] Checkpoint compaction: CheckpointCompactor with configurable limits, preserves tagged snapshots
  • [New] Session state garbage collection: cleanup service for expired checkpoints, stale Redis keys, old task results; supports dry-run + configurable TTLs
  • [New] Warm restart: MCP tool schemas cached in Redis to reduce cold-start penalty on resume
  • [New] Idempotency tokens: SHA-256 keys + Redis cache prevent double-execution on checkpoint replay
  • [New] CommandGuard: 30 regex patterns blocking destructive filesystem/git/k8s/DB operations pre-execution

isA_MCP — OAuth 2.0 & Scalability (Epics #411, #417)

  • [New] OAuth 2.0 resource server with RFC 8707 metadata endpoint (/.well-known/oauth-protected-resource)
  • [New] Bearer token validation middleware (RS256/ES256) for browser POST tools
  • [New] Connection pooling with LRU eviction: max 30 concurrent connections, lazy-connect, pool stats API
  • [New] Server priority tiers: CRITICAL/STANDARD/OPTIONAL with tiered health check intervals; critical servers exempt from eviction
  • [New] O(1) tool routing: _tool_index dict in ToolAggregator replaces O(n) scans
  • [New] Parallel health checks with bounded concurrency (env: HEALTH_CHECK_CONCURRENCY, default: 10)
  • [New] Configurable asyncpg pool size via DB_POOL_MIN_SIZE/DB_POOL_MAX_SIZE
  • [Breaking] Scope migration: a2a.* scopes renamed to mcp:*. Auto-normalization middleware preserves backward compat during transition window.

isA_Model — Platform Billing & Metering (Epic #495)

  • [New] Complete billing infrastructure: atomic Redis Lua-script reservation, NATS billing events, Dead Letter Queue stream
  • [New] Per-meter tracking: MeterType enum for tokens, tool calls, storage ops, pipeline runs
  • [New] /api/v1/usage endpoint: credit balance, consumption history, cost attribution by project/api_key
  • [New] Billing hooks for MCP tools (/api/v1/tools/usage), storage (/api/v1/storage/usage), pipelines (/api/v1/pipelines/usage)
  • [New] Credit alerts via NATS when balance approaches threshold
  • [New] Per-resource RBAC: resource:action model, 11 resource types, feature-flagged rollout
  • [New] Immutable audit logging via NATS JetStream: non-blocking AuditPublisher, admin query endpoint
  • [New] GDPR compliance: cascading tenant deletion across PG/Redis/Qdrant/NATS; K8s CronJob cleanup; admin endpoints

isA_user — OAuth 2.0 Authorization Server (Epic #161)

  • [New] Authorization Code + PKCE flow: GET /oauth/authorize, POST /oauth/consent-approval, token exchange
  • [New] Client ID Metadata Document (CIMD) fetcher and cache per RFC 8707 Dynamic Clients
  • [New] /.well-known/oauth-authorization-server metadata: advertises authorization_endpoint, code_challenge_methods_supported, CIMD support
  • [New] New DB tables: oauth_authorization_codes, oauth_client_metadata
  • [Breaking] Default OAuth scopes changed from a2a.* to mcp:*. Legacy scopes auto-mapped.

isA_Orch — Platform Code Index (Epic #1)

  • [New] isa-orch CLI: index, search, impact, stats, projects, project-context, test-scaffold commands
  • [New] Python AST + TypeScript tree-sitter parsing across all 19 isA projects
  • [New] Semantic code search: local sentence-transformers embedding index, --semantic flag
  • [New] Cross-project edge resolution with same-project preference (3-pass algorithm)
  • [New] Test directory scaffolding: detect_context() + scaffold_test_dirs() CLI commands
  • [New] Incremental indexing with checksum skipping and stale file pruning

isA_Vibe — Multi-Turn Dispatch (Epic #199)

  • [New] Phased CDD/TDD/Deploy execution: Phase enum, per-phase graph executions with session resume
  • [New] Session manager: phase 1 uses execute(), phases 2+ use query() for context continuity (1800+ turns/run)
  • [New] Manifest chunker: FileManifest/FileSpec + chunk_manifest() for batch-aware prompts
  • [New] Phase completion detection: scans for CDD/TDD/Deploy outputs, generates continuation prompts
  • [New] Sandbox isolation: git worktree per run — create/validate/merge/cleanup lifecycle
  • [New] Auto-revert on post-push test failure
  • [New] BudgetGuard and CheckpointStore wired into orchestrator

isA_Data — Vector & AI Features

  • [New] Milvus vector DB backend: MilvusVectorDB implementing BaseVectorDB, registered as VectorDBType.MILVUS
  • [New] Hybrid GraphRAG with RRF fusion: parallel graph + Qdrant retrieval fused via Reciprocal Rank Fusion (k=60)
  • [New] AI data dev copilot: POST /api/v1/data/copilot — NL → ETL/model/quality rule/indicator/SQL artifact generation
  • [New] Official Dataphin SDK integration: all 4 backends migrated to SDK methods
  • [New] Master data backend with AI entity resolution and merge candidates

isA_Agent — Agent Management

  • [New] Template management API: CRUD, versioning, categories, tags, 8 builtin templates (undeletable), version history + rollback
  • [New] Deployment automation: canary/blue-green/rollback with configure/promote/rollback/status endpoints
  • [New] Config versioning: immutable snapshots on every create/update, GET /versions, POST /rollback
  • [New] Audit trail: field-level diffs, bounded deque store, persists after config deletion

isA_Console — UI & Observability

  • [New] OAuth consent screen: /oauth/consent, /oauth/callback, /oauth/error pages with scope descriptions
  • [New] Marketplace: /dashboard/marketplace — 24 built-in skills across 9 categories with search + filters
  • [New] Agent lifecycle badges: StatusBadge component, version history page, Grafana monitoring embeds
  • [New] Loki log shipping: batched via sendBeacon, global error handlers, global-error.tsx boundary
  • [New] OpenTelemetry + Web Vitals instrumentation
  • [New] Responsive sidebar with mobile drawer + hamburger

isA_Cloud — Infrastructure

  • [New] 4 Grafana monitoring dashboards: agent-execution, mcp-tools, resource-pools, service-mesh
  • [New] APISIX routes for OAuth /.well-known discovery endpoints (all 3 environments)
  • [New] mcp-worker ArgoCD app manifests (staging auto-sync, production manual)
  • [New] Custom HPA metrics rendering in Helm template

[Unreleased]

Added

Agent SDK

  • Subscription proxy system with account pool rotation, cooldown, and usage alerts (#119-#126)
  • Inter-graph messaging for agent-to-agent communication (#109)
  • AgentLifecycle for long-running proactive agents (#108)
  • GenerationClient and Vibe generation pipeline (#107)
  • AgentTemplateStore with versioning and CRUD (#106)
  • Writer/Reviewer pattern with independent review context (#105)
  • Loop detection for stuck agents (#104)
  • Runtime observability with Prometheus metrics and structured logging (#103)

Model Service

  • One-click deploy fine-tuned model to Ollama (#198, #266)
  • Fine-tuned model checkpoint registry (#197, #265)
  • Fine-tune job endpoint with async execution (#196, #264)
  • Dataset upload and management API (#195, #263)
  • LoRA SFT algorithm using peft/trl (#194, #262)
  • Latency-aware model ranking in router (#261)
  • Model selection stats collector and telemetry events to NATS (#256, #260)
  • Per-provider latency tracking with percentiles (#257)
  • Local model benchmarking on first pull (#254)
  • Cache warming for high-frequency queries (#253)
  • Semantic cache migration from Redis brute-force to Qdrant vector search (#252)
  • Runtime adapter abstraction for pluggable deployment (#234, #251)
  • Training GPU isolation with preemptible allocations (#232, #250)
  • GPU monitoring service for status, allocations, and plan preview (#233, #249)
  • CLOUD provider consolidation with sub-provider routing (#231, #248)
  • Thinking mode control for reasoning models (#235, #247)
  • Triton multi-model co-location for CV workloads (#229, #246)
  • Deployment Planner for auto engine/GPU/TP decisions (#230, #245)
  • ISA backend auto-routing with fallback chain (#227, #244)
  • CapabilityRegistry for multi-capability model sharing (#243)
  • SGLang adapter and engine instance management with TP support (#242)
  • Model Lifecycle Manager with tiered loading and LRU eviction (#241)
  • GPU Allocator with VRAM-aware placement and MoE support (#225, #240)
  • GPU Discovery Service for auto-detecting GPUs and topology (#224, #239)
  • LOCAL_GPU backend type for on-prem GPU inference (#223, #238)
  • Provider consolidation to ISA + CLOUD (#222, #237)

MCP

  • list_user_files MCP tool wrapping storage service (#280, #281)
  • Health failure persistence and reconnect_server tool (#273, #274, #277)
  • Connection event tracking and history MCP tools (#271, #272, #276)
  • 5 MCP tools for aggregator observability (#270, #275)
  • SHA256 integrity, script gating, lock file, and per-skill config (#261-#265, #268)
  • Skill upgrade with automatic rollback (#260, #267)
  • Skill enable/disable toggle and dry-run install preview (#259, #264, #266)
  • InternalTool data class with format translators (#235-#237, #238)
  • Cursor-based pagination for resource, prompt, session, and tool discovery (#231-#233)
  • Resource update notifications on catalog changes (#215, #222)
  • Weak reference cleanup for session disconnect (#216, #221)
  • Resource subscription registry with subscribe/unsubscribe handlers (#212-#214, #218-#220)
  • Category-based prompt injection detection (#210)
  • Security scanning engine for skills (#183, #209)
  • Trust level system with security scoring (#202-#207)
  • Skill publisher profiles and download counter (#198-#200)
  • Error handler, rate limiting, health checks, API version header (#245-#251, #254)
  • Request timeout middleware, graceful shutdown, and L5 smoke tests (#253)

Trade

  • Locust load test and performance baseline (#171, #188)
  • Kustomize overlays for staging and production (#189)
  • Redis AOF persistence for production (#176, #191)
  • Prometheus alerting rules and Trivy container scanning (#167, #168, #186)
  • Versioned GHCR push and live trading approval gate (#183)
  • Database connection resilience with retry (#182)
  • AGENT_PROVIDER env var for subscription proxy support (#195)

Cloud

  • Backup/restore scripts for NATS, Consul KV, and MQTT (#116)
  • PodDisruptionBudget template for isa-service Helm chart (#115)
  • Observability env vars in Helm chart and scrape targets (#102, #113)
  • PostgreSQL backup CronJob for production (#69, #105)
  • Vault + External Secrets Operator for production secret management (#67, #100)
  • Consul watch for near-instant APISIX route sync (#99)
  • Unit tests for tracing and observability modules (#97, #103, #104)
  • Unit tests for shared Prometheus metrics client (#101)

Mate

  • A2A delegation team for dev workflows via isa-vibe (#176)
  • Unified AI access into single onboarding step (#174)
  • Multi-user credential routing with BYOK and isA Subscription (#171, #172)
  • Credential validation, CLI, onboarding, mode switching, and audit logging (#172)
  • OAuth transport for codex-sub and anthropic-sub providers
  • Mate persona, skill injection, and expanded tool list (#146, #150)
  • Skill Registry API with CRUD endpoints for community skills (#91, #115)
  • Skill publishing pipeline with security scanning (#116)
  • Skill install, list, remove, search, browse, update, and connect CLI commands (#117-#123)
  • Signal, Microsoft Teams, and Matrix channel adapters (#104)
  • Docker quickstart and onboarding docs refresh (#89, #90, #108)

Console

  • SWR migration, console cleanup, and loading states (#62)
  • Vitest + RTL test framework with L1/L2 tests (#61)
  • Production Dockerfile (#58)
  • Security headers in Next.js config (#57)
  • Error boundaries and error UI for dashboard (#56)
  • Next.js middleware for server-side auth protection (#55)
  • SkillCategory/GuidanceSkill types and useAgentResources hook (#36)

User Services

  • Prometheus /metrics endpoints for all 35 services (#107)
  • Graceful SIGTERM shutdown handlers for all services (#103)
  • Rate limiting middleware and circuit breaker pattern (#101)
  • Inter-service clients for inventory, tax, fulfillment (#105)
  • Missing billing endpoints for quota and records (#104)
  • Alembic database migration framework (#99)
  • Service layers for inventory, tax, fulfillment (#83)
  • Metrics migration to isa_common shared observability clients (#110)

Creative (new project)

  • DAM Manager with bulk operations and expiration policies (#45, #60)
  • Malware scanning in asset ingestion pipeline (#46, #61)
  • Real brand compliance checking in Visual Designer (#44, #59)
  • Multi-variation content generation with scoring and selection (#28, #34)
  • Real tool integrations for generate_image and apply_layout (#25, #33)
  • Checkpoint/restore for workflow execution loop (#32)
  • WorkflowStore abstraction for persistent workflow storage (#31)
  • Design System Agent veto power through shared state (#5, #14)
  • DAG-based wavefront execution with parallel steps (#13)
  • Shared state model with field ownership (#12)
  • SwarmOrchestrator with in-process agents (#11)
  • Circuit breaker pattern and JWT authentication (#42, #43, #49)
  • Per-client rate limiting with 429 responses (#30, #38)
  • Prometheus metrics instrumentation and /metrics endpoint (#35)
  • SDK multi-agent workflow primitives (#23, #36)

Marketing (new project)

  • 8-agent architecture with Content, SEO, Campaign, and Compliance agents (#18, #43)
  • A2A endpoints for campaign analyst and SEO agents (#39)
  • Shared state through pipeline with STATE_FIELD_OWNERS pattern (#4, #35)
  • In-process SwarmOrchestrator pipeline replacing HTTP microservices (#36)
  • Single-container architecture consolidating 8 agent containers (#5, #7, #40)
  • Orchestrator multi-agent scenario and integration tests

Orchestrator (new project)

  • Platform CLI with index, search, impact, and stats commands (#15)
  • Cross-project edge resolution with same-project preference (#13)
  • Multi-project indexing loop (#12)
  • Project registry loader from config/projects.yaml (#11)
  • Tree-sitter parsing for TypeScript and JavaScript (#17)
  • Tier 2 local embedding index for semantic code search (#16)
  • Project-scoped query API with get_symbols method (#14)
  • TDD test-layer routing and upgrade skills for full pyramid (#26)
  • Project context detection and test directory scaffolding (#27)

Vibe

  • Provider-aware model defaults for codex-sub support (#146)
  • Global preference persistence and memory CLI subcommand (#138)
  • Lazy skill loading and ops-to-product feedback loop (#137)
  • Codebase intelligence tools and handoff notes (#136)
  • Human-in-the-loop pause wired into streaming with —quiet flag (#135)
  • Inter-agent handoff notes, requests, and streaming UX (#134)
  • Session listing via workflow list command (#132)
  • —resume CLI flag for session resume (#114, #131)
  • SharedState injection into sub-agent prompts for cross-team context (#129)
  • Session metadata persistence and memory injection ordering (#123)

OS Services

  • Pool Manager executor state persistence across restarts (#95-#100)
  • Code deployment subsystem merged into Pool Manager (#101, #106)
  • Tiered compute pricing plans (#74)
  • Volume mount and network_mode support for Cloud OS backends (#94)
  • WebSocket streaming and security API routes for cloud_os (#116)
  • Dockerfile, Helm values, and env configs for desktop_os (#141)
  • GitHub Actions CI/CD pipeline for all services (#136)
  • Metrics migration to isa_common shared observability clients (#138)

Fixed

Agent SDK

  • PoolManagerClient endpoint and payload inconsistency (#114)
  • Stripped OPENAI_API_KEY from subprocess environments (#125)

Model Service

  • Clarified embedding service stubs as rerank-only (#259)
  • Cleaned up test artifacts and moved misplaced test files (#258)

MCP

  • Skill search AttributeError and cache reload after install (#257)

Trade

  • Require actual credentials before starting streaming feeds (#199)
  • Async test helpers using new_event_loop() to prevent suite-wide failures (#193)
  • A2A agent card URLs using 127.0.0.1 instead of localhost (#192)
  • K8s image tag updates and README/SECURITY/CHANGELOG additions (#173-#175, #184)

Cloud

  • Client bugs in MinIO typo, path traversal, deprecated datetime, thread cleanup (#125)
  • CI security scans made blocking by removing continue-on-error (#114)
  • State mutation fix with observability tests and scrape targets (#113)
  • Default registry fallback when custom registry is None in metrics client

Mate

  • Local user bypass removed so all BYOK users can use credentials (#173)
  • Unified proxy port mismatch and GPT model upgrade (#145)
  • Route default chat through runtime agent for local tool access
  • Pre-warm SDK context at gateway startup to eliminate first-query delay (#144)
  • A2A team URLs using 127.0.0.1 and isa_vibe timeout increase (#178)

Console

  • Auth token propagation to all SDK service instances (#60)
  • Consistent getCurrentToken() across all admin services (#59)
  • P3 backlog fixes for errors, chart, auth context, and URLs (#35)
  • Non-functional settings controls disabled with coming-soon indicators (#34)
  • Error states on analytics, MCP prompts, and main dashboard (#33)
  • Service error propagation to UI instead of swallowing (#32)

User Services

  • Consul deregistration before rejecting requests during shutdown (#109)
  • Production readiness fixes for async-safe singletons, Helm tags, and service CORS (#100)
  • Stub health check DB probe replaced with real SELECT 1 query (#97)
  • Commerce services added to Helm manifests and deploy scripts (#96)

Creative

  • Stale imports, API test mismatches, and Pydantic V2 warnings (#19)
  • A2A port standardization to 18791 with 127.0.0.1 (#55)
  • Visual designer test imports to match actual function names (#56)
  • Missing Pillow dependency (#62)

Marketing

  • Stale port references and docs update for single-container architecture (#58)
  • Dead specialist main.py files and deprecated agent_client.py removal (#57)
  • Missing logging module (#55)
  • Stale unit tests for non-existent base_agent and Config (#46, #52)
  • A2A URLs using 127.0.0.1 instead of localhost (#51)
  • Empty contracts and report generator tool name alignment (#22)

Vibe

  • Proxy reuse, model tier resolution, and streaming deduplication (#148)
  • Streaming output duplication and Redis TTL persistence (#141)

OS Services

  • Bare except blocks audited and fixed across all services (#143)
  • Missing vm_id field and watchdog restart callbacks wired (#142)
  • All service URLs made env-configurable in desktop_os (#140)
  • cloud_os production code importing from test contracts (#139)
  • Missing prometheus_client dependency for desktop_os (#137)
  • Regional warm pool methods wired into PoolState (#135)
  • 64 test failures and 13 collection errors resolved in desktop_os (#134)

[0.6.0] - 2026-02-14

Added

Agent SDK (isa_agent_sdk)

  • Updated docs for stabilized Swarm + DAG orchestration workflows.
  • Added current guidance for A2A integration (Agent Card + JSON-RPC adapter flow).
  • Synced reliability/testing status with latest suite (234 passing tests including audit coverage).

Model Service (isa_model)

  • Added documentation for generic ML prediction APIs:
    • training lifecycle routes
    • prediction inference routes
    • paginated listing (limit / offset) for models and runs
  • Added current Lightning training API route coverage.
  • Updated cache API examples to current HTTP methods (POST invalidation/clear).

MCP (isa_mcp)

  • Added new guidance pages for:
    • Security (/mcp/security)
    • Multi-Tenant (/mcp/multi-tenant)
    • Reliability (/mcp/reliability)
  • Updated MCP overview to reflect hybrid multi-tenant architecture and operational hardening.

isA Vibe (isa_vibe)

  • Updated Vibe docs for generic multi-project orchestration support.
  • Added backlog command coverage (ready, dispatch, status, --watch).
  • Documented new security/infrastructure tooling:
    • SAST + dependency scanning
    • cross-service log query
    • migration runner helpers

Fixed

  • Corrected stale Model cache endpoint examples to match current service behavior.

[0.5.5] - 2026-02-06

Added

Agent SDK

  • Complete voice feature suite with audio buffer, dictation rewriter, and voice orchestrator (#56-#61)
  • Team delegation module for multi-agent coordination
  • TLS/mTLS support for inter-service communication (#101)
  • Per-user tier-based rate limiting (#99)
  • Docker CI/CD pipeline with GitHub Actions (#97)

MCP

  • Hybrid multi-tenant platform with org-scoped isolation (#1)
  • Marketplace HTTP API with skill installation endpoints (#51-#53)
  • Aggregator REST endpoints for external MCP server federation
  • Comprehensive security, multi-tenant, and reliability documentation

Model Service

  • Realtime session gateway with WebSocket streaming (#73)
  • Voice privacy filter and retention modes (#77)
  • Role-based model resolution wired into /invoke (#63-#66)
  • Capability-based filtering and cost-aware ranking (#67-#72)
  • Model selection explain endpoint for observability (#69)
  • Latency tracking with soft constraints and feedback loop (#70-#72)
  • OpenRouter as LLM provider (#59, #62)
  • Anthropic and subscription proxy providers
  • Generic ML prediction service with 2-tier Ray backend

User Services

  • X-API-Key authentication for core services (#4)
  • User-managed API key lifecycle (#1)
  • Dependency injection infrastructure across 9 services
  • Loki centralized logging integration
  • Event-driven architecture with async service clients

Cloud

  • Loki + Grafana observability stack with AsyncLokiClient (#47)
  • Terraform modules and environment configs (#40)
  • NetworkPolicies for namespace isolation (#8)
  • SAST and dependency scanning CI workflow (#12)
  • Claude skills for cluster ops, backup/restore, and service management

OS Services

  • Desktop OS transformed into local-first AI agent with isA_Agent_SDK (#88)
  • Telegram channel integration with webhook mode and media support (#31-#36)
  • Channels runtime with ingest/dispatch/retry/dead-letter workers (#28-#29)
  • DDoS protection and network security (#78)
  • Rate limiting and API security for Pool Manager (#78)
  • CLI renamed from isa-desktop to isamate (#93)

Data

  • Initial release with data lake, RAG pipelines, and file processing

Changed

Cloud

  • Removed all Go code — platform is now pure Python
  • Replaced gRPC layer with direct native Python clients
  • Rewrote Makefile with docker-compose for local dev (#38)
  • Removed Consul Discovery from APISIX in favor of direct routing

Model Service

  • Major refactor to standalone OpenAI-compatible client (v0.5.0)
  • Increased default timeouts from 30s to 300s
  • Removed duplicate CORS middleware, deferred to APISIX gateway (#111)

Fixed

Agent SDK

  • Concurrency and thread-safety issues in voice feature (3 critical fixes)
  • Empty responses after tool execution (#12)
  • Cascade failures optimized from O(n²) to O(1) (#53)
  • @tool pipeline crash, tools= param, and name resolution (#74)
  • Pydantic V1 to V2 migration across all models

MCP

  • Auth bypass vulnerabilities (#3, #6, #12, #21)
  • Tool execution injection and path traversal (#4, #13, #14)
  • Bash command injection hardening (#34)
  • SSE injection prevention (#64-#70)
  • Qdrant retry logic and vector ID overflow protection (#16, #17)
  • Session leak fixes and ExceptionGroup unwrapping (#9, #11, #15)

Model Service

  • Critical, high, medium, and low severity security audit fixes (~30 issues)
  • Restricted pickle unpickler and sensitive header redaction (#110)
  • HMAC integrity checks on model save/load (#10)
  • DeepSeek-R1 streaming tool call implementation

User Services

  • P0 critical auth vulnerabilities (#18-#22)
  • P1 high-severity security vulnerabilities (#23-#29)
  • P2/P3 medium and low severity issues (#30-#43)
  • SDK-backend mismatches and billing resilience (#7 critical fixes)

Cloud

  • Hardcoded passwords removed from Kubernetes secrets
  • SQL injection prevention with identifier validation (#4)
  • Pod security context added to Helm chart (#5)
  • CORS wildcard replaced with explicit origin allowlist (#7)
  • NATS consumer resilience and JetStream message acknowledgment (#18, #26-#27)

[0.5.1] - 2025-06-15

Added

  • MCP tool server initial release with LangGraph agent integration
  • Docker deployment and Railway hosting support
  • Event sourcing and memory services
  • Web tools, data graph, and digital graph services
  • Model service Triton integration and Replicate/Gemma/Whisper providers
  • OS web automation with Browserbase element extraction
  • Python REPL and Cloud OS services with Kubernetes deployment
  • MQTT and Device Management services for IoT

Fixed

  • MCP server crash and exposed secrets remediation
  • Model service tool calling optimization and adapter fixes

[0.5.0] - 2025-01-29

Added

Agent SDK

  • New Workflow class for multi-step agent tasks
  • Support for parallel tool execution
  • Memory persistence with Redis and PostgreSQL backends
  • Streaming responses with async for syntax

MCP

  • 15 new tools: browser_*, database_*, api_*
  • Tool composition for complex operations
  • Automatic retry with exponential backoff

Model Service

  • Support for Claude claude-sonnet-4-20250514 and Opus models
  • Local model support via Ollama integration
  • Response caching with configurable TTL
  • Cost tracking and budgets

User Services

  • Cognitive memory service with 6 memory types
  • Enhanced payment processing with Stripe
  • Organization hierarchy and family sharing
  • IoT device management with OTA updates

Cloud

  • ArgoCD GitOps deployment
  • Multi-environment support (dev/staging/production)
  • Automated CI/CD with GitHub Actions
  • Horizontal pod autoscaling

Changed

  • Improved error messages across all services
  • Better type hints in Python SDK
  • Faster startup time for agents

Fixed

  • Memory leak in long-running agents
  • Race condition in concurrent tool calls
  • Token refresh issues in auth service

[0.4.0] - 2025-01-15

Added

  • MCP tool framework with 35 initial tools
  • Data lake with DuckDB integration
  • Vector search with Qdrant
  • Document processing pipeline

Changed

  • Migrated from REST to gRPC for internal services
  • Unified logging format across services

Fixed

  • Connection pooling issues in PostgreSQL client
  • Rate limiting accuracy improvements

[0.3.0] - 2025-01-01

Added

  • Agent SDK initial release
  • Model service with OpenAI and Anthropic support
  • Basic authentication service
  • File storage with MinIO

Changed

  • Restructured repository for monorepo

[0.2.0] - 2024-12-15

Added

  • Consul service discovery
  • APISIX API gateway
  • Redis caching layer
  • Health check endpoints

[0.1.0] - 2024-12-01

Added

  • Initial platform architecture
  • Kubernetes deployment manifests
  • Basic CI/CD pipeline
  • Documentation site setup

Versioning

This project follows Semantic Versioning :

  • MAJOR: Incompatible API changes
  • MINOR: New functionality, backwards compatible
  • PATCH: Bug fixes, backwards compatible

Release Schedule

  • Patch releases: As needed for bug fixes
  • Minor releases: Every 2-4 weeks
  • Major releases: As needed for breaking changes

Deprecation Policy

  • Deprecated features are marked in documentation
  • Minimum 2 minor versions before removal
  • Migration guides provided for breaking changes
Home