Changelog

All notable changes to confused-ai are documented here.
Format: Keep a Changelog · Semantic Versioning

Full changelog

The authoritative CHANGELOG.md lives in the repository root.
View on GitHub →

v2.3.0 — Current

Added

• 9 Extended Multi-Agent Orchestration Patterns (@confused-ai/orchestration) — Mixture-of-Agents (MoA), Actor-Critic loops, Socratic tutor guiding, Prompt Chaining pipelines, Program-of-Thought code sandbox runtimes, Skeleton-of-Thought parallel generation, Step-Back conceptual abstraction solvers, Rejection Sampling (Best-of-N) evaluations, and validation-driven Self-Correction.
• createGSDCoordinator() (Get Shit Done) — spec-driven workflow coordinator that executes project goals inPlan-Execute-Verify phases, using a workspace .planning folder to isolate contexts.
• createRalphLoop() (RALF) — autonomous cycle executor that leverages fresh session isolation to prevent context bloat while propagating iteration summaries.
• Mastermind Context Compression (@confused-ai/compression) — a multi-stage intelligent context compression suite featuring:
  • CacheAligner (KV-cache prefix alignment).
  • Specialized crushers (smart-crusher for JSON, minifiers for Code, log timestamp/duplicate aggregators, XML, CSV).
  • Sliding-window group budget enforcers to prevent orphaned tool call/result pairs.
  • Code & Context Reduction (CCR) annotations with retrieveTool for on-demand details recall.

---

v1.1.7

Added

• DbScheduleStore (@confused-ai/scheduler) — bridges ScheduleManager with any AgentDb backend. Persist schedules to SQLite, Postgres, MySQL, MongoDB, Redis, DynamoDB, or Turso with no custom glue code. See Scheduler → Production persistence.
• DB health in /health endpoint — createHttpService now accepts a db?: AgentDb option. When provided, GET /health (and /v1/health) runs a live db.health() probe. Returns HTTP 503 with { status: 'degraded' } when the database is unreachable.

Fixed

• @confused-ai/db — uuid() security — all 8 backends (InMemory, SQLite, Postgres, MongoDB, Redis, JSON, MySQL, DynamoDB, Turso) now generate IDs with crypto.randomUUID() instead of the previous Math.random()-based implementation.
• @confused-ai/db — init() race condition — concurrent callers no longer double-initialize the connection. All async backends (Postgres, MongoDB, MySQL, DynamoDB, Turso) now share a single _initPromise guard.
• PostgresAgentDb — getKnowledgeItems(), getTrace(), and getTraces() now correctly re-serialize JSONB content and metadata columns to string (the pg driver returns these as parsed objects).
• MongoAgentDb — all findOne and find calls now include { projection: { _id: 0 } } so MongoDB's internal _id field is never included in returned rows.
• DynamoDbAgentDb — constructor now calls validateTableNames() to catch invalid table names at construction time instead of silently failing at runtime.
• DbSessionStore — now() helper now returns Unix epoch seconds (Math.floor(Date.now() / 1000)) to match the AgentDb timestamp contract (was returning milliseconds, causing created_at/updated_at to be off by ×1000).
• TursoAgentDb — single-row casts (LibSqlRow → SessionRow, MemoryRow, etc.) now use the as unknown as T double-cast pattern, fixing TypeScript strict-mode errors.
• PostgresAgentDb — close() method was accidentally stripped during a refactor; restored.
• Unified class API — SimpleAgent and LegacyAgent were removed from public exports; Agent is now the only class-based API and includes both legacy defaults and modern fluent methods.
• Durable runtime lifecycle correctness — resume now rejects terminal workflows consistently; terminal-state handling no longer allows invalid resume paths.
• CQRS error propagation — event-bus handler failures are now surfaced via AggregateError after handlers run.
• State machine lifecycle hardening — start() is idempotent; transition commits in send() and jumpTo() are atomic (state changes apply only after target onEntry succeeds).
• Snapshot restore semantics — snapshots persist startup status (started); restore defaults legacy snapshots to started to avoid duplicate initial onEntry execution.

---

v1.1.6

Changed

Monorepo restructure — packages fully independent

• All source code now lives in independently-built workspace packages under packages/. The src/ directory is retained as a backward-compatible re-export barrel — no breaking API changes.
• packages/tools rewritten with clean functional defineTool implementations; removed all class-based files with broken relative imports.
• packages/test-utils is now a fully standalone package: createMockLLM, createMockAgent, runScenario with zero cross-package dependencies.
• CI pipeline updated to 4 sequential jobs: typecheck → lint → test (Node 18 / 20 / 22) → build all packages.

Fixed

• router/selectForBudget — removed incorrect × 1,000,000 scaling; budget comparison is now a direct dollar-per-million comparison.
• adapter-redis/session-store — removed unnecessary optional chain on non-null hGetAll result; fixed template literal number type.
• tools/types.ts — migrated from deprecated ZodTypeAny → z.ZodType; _def private field access replaced with .def.
• Removed 33 broken package copies that had relative src/-path imports causing circular resolution failures.
• Documentation URLs updated from rvuyyuru2.github.io/agent-framework to confused-ai.github.io/confused-ai throughout all docs.
• Version consistency: ARCHITECTURE.md and SECURITY.md now match the package.json version.

Security

• SECURITY.md: added ShellTool sandbox requirements section documenting blocked command patterns.
• SECURITY.md: documented RedisRateLimiter as the required solution for multi-instance distributed rate limiting.
• README.md: qualified audit-logging claim — removed unqualified SOC 2 / HIPAA label; added compliance footnote.

---

v1.1.0

Added

agent.stream() — async iterable streaming on every agent

• Every CreateAgentResult now has a built-in stream(prompt, options?) method
• Returns an AsyncIterable<string> — chunks arrive as the LLM generates
• Works with for await loops; no extra setup; accepts all run() options except onChunk

for await (const chunk of agent.stream('Explain quantum computing')) {
  process.stdout.write(chunk);
}

defineAgent() builder — budget(), checkpoint(), adapters()

• .budget(config) — USD spend caps without dropping down to createAgent()
• .checkpoint(store) — durable crash recovery wired in one line
• .adapters(registry) — plug in adapter registry or explicit bindings
• Full builder method table now documented in Creating Agents

Performance

AgenticRunner — Zod→JSON Schema cached per agent (not per run)

• Tool definitions (Zod → JSON Schema) are computed once in the constructor and reused
• Zero toolToLLMDef() overhead on hot-path run() calls after initial agent creation

Tool execution — timer leak fixed

• Promise.race timeout timer is now always cleared via .finally() on every tool call
• Prevents 30-second timer handles accumulating in long-running processes
• Timing now uses performance.now() for sub-millisecond accuracy

AuditPlugin — O(1) event queries

• Internal Map indexes maintained on every onEvent() call
• getEventsByType(), getEventsForNode(), getEventsForExecution() are all O(1) index lookups
• Previously O(n) full scans — eliminates bottleneck on high-event-volume workflows

OpenTelemetryPlugin — OTel module imported once and cached

• The @opentelemetry/api dynamic import is now cached after the first successful load
• Previously re-imported on every onNodeStart() call

---

v1.0.0

Added

Reasoning Module

• ReasoningManager — chain-of-thought and self-critique loops over any generate function
• ReasoningEventType discriminated union: step, action, complete, error
• NextAction typed decision point: continue | finish | backtrack | escalate
• ReasoningStore — pluggable trace persistence (audit, replay, fine-tuning)

Scheduler Module

• ScheduleManager — CRUD for cron-based job schedules with pluggable ScheduleStore + ScheduleRunStore
• InMemoryScheduleStore (dev) / SqliteScheduleStore (prod)
• In-process handler registry — no HTTP endpoint required
• Full lifecycle: create / update / delete / enable / disable / triggerNow / listRuns

CompressionManager

• Transparent context-window compression before LLM calls; truncate | summarise | rolling strategies

ContextProvider

• Retrieves and injects grounding documents into the system prompt or user message at run time
• Pluggable ContextBackend: InMemoryContextBackend, SqliteContextBackend

Freedom Layer — bare / compose / pipe

• bare(opts) — zero-defaults agent; caller owns LLM, tools, hooks, everything
• compose(...agents, opts?) — sequential pipeline; output of each agent → input of next
• compose options: when (conditional routing) + transform (reshape data between steps)
• pipe(agent).then(agent).run(prompt) — builder-style equivalent to compose()

Eval Regression Suite

• runEvalSuite — labeled dataset, per-sample scoring, baseline comparison, CI exit code
• InMemoryEvalStore / SqliteEvalStore — durable baseline persistence across CI jobs
• setBaseline: true — promote current run as new reference; regressionThreshold — allowable score drop

Real-World Example Library (4 new runnable examples)

• examples/reasoning-agent.ts — Incident Triage Bot (no API key needed)
• examples/scheduled-agent.ts — Nightly Market Digest (no API key needed)
• examples/code-review-pipeline.ts — PR Code Review Pipeline (no API key needed)
• examples/eval-regression.ts — CI Eval Regression Guard (no API key needed)

Documentation

• 19 · Incident Triage Bot — ReasoningManager, event streaming
• 20 · Scheduled Agent Jobs — cron scheduling, run history
• 21 · Code Review Pipeline — bare(), compose(), pipe()
• 22 · Eval Regression Guard — runEvalSuite, CI baseline guard

---

v0.7.0

Added

Budget Enforcement

• budget?: BudgetConfig added to CreateAgentOptions — hard USD caps per run and per user
• BudgetEnforcer instantiated in factory, recordAndCheck(userId) called after the run loop
• userId?: string added to AgenticRunConfig for per-user cap enforcement

HITL Approval HTTP Endpoints

• GET /v1/approvals — list pending approvals
• POST /v1/approvals/:id — submit decision { approved, comment, decidedBy }
• approvalStore?: ApprovalStore added to CreateHttpServiceOptions

Distributed Trace Context

• W3C traceparent / tracestate extraction from incoming HTTP request headers
• traceId from incoming trace propagated in JSON and SSE responses
• src/observability/trace-context.ts — extractTraceContext(), injectTraceContext()

---

v0.6.0

Added

Testing Module (confused-ai/testing)

• MockToolRegistry — records all invocations; calls(), lastCall(), reset()
• createTestAgent() — zero-config test harness with MockLLMProvider + MockSessionStore
• createTestHttpService() — integration test helper on a random port

HTTP Runtime

• X-Request-ID correlation header on every response
• rateLimit middleware option in CreateHttpServiceOptions
• auditStore option — SQLite-backed persistent audit log
• WebSocket transport (websocket: true) — attaches to existing http.Server
• Admin API (adminApi: true) — /admin/health, /admin/agents, /admin/audit, /admin/stats, /admin/checkpoints

Adapter System (confused-ai/adapters)

• 20-category adapter system covering SQL, NoSQL, vector, cache, object storage, message queues, observability, embedding, session, memory, guardrail, RAG, tool registry, auth, rate limiting, audit log
• createProductionSetup() — opinionated full-stack wiring with progressive upgrade path

LLM Router (confused-ai/llm)

• LLMRouter — intelligent routing by task type, complexity, and strategy
• Four strategies: balanced, cost, quality, speed
• Factories: createBalancedRouter, createCostOptimizedRouter, createQualityFirstRouter, createSpeedOptimizedRouter

Deployment Templates (/templates)

• Dockerfile, docker-compose.yml, fly.toml, render.yaml, k8s.yaml
• Grafana dashboard JSON (grafana-dashboard.json)

DX Improvements

• defineTool() helper — AI SDK-style fluent builder with Zod schemas, needsApproval, streaming hooks
• createWorkflow().then(step).commit() — Mastra-style typed step workflows
• createStepWorkflow, StepWorkflow, StepWorkflowBuilder, StepWorkflowStep exports

Resilience

• withResilience() — circuit breaker + rate limiter + retry + health check wrapper
• RedisRateLimiter — distributed rate limiting via Redis

Secret Manager (confused-ai/config)

• createSecretManager() with adapters: EnvSecretManagerAdapter, AwsSecretsManagerAdapter, AzureKeyVaultAdapter, VaultAdapter, GcpSecretManagerAdapter

Orchestration Extensions

• AgentRouter — capability-based, round-robin, least-loaded routing
• HandoffProtocol — structured agent-to-agent task handoff with tracing
• ConsensusProtocol — multi-agent voting (majority, unanimous, weighted, best-of-n)

---

v0.5.0

Added

• Checkpoint/resume for long-running agents — checkpointStore? in AgenticRunnerConfig
• createSqliteSessionStoreSync — sync init, safe for factory use
• Persistent user profiles and learning modes
• Eval dataset persistence — EvalStore, InMemoryEvalStore, SqliteEvalStore, runEvalSuite
• Plugin system — confused-ai/plugins with built-in logging, rate-limit, telemetry plugins
• Contracts layer — confused-ai/contracts for shared interfaces without runtime code

---

v0.4.0

Added

• Full adapter system for all infrastructure categories
• Multi-tenancy with createTenantContext()
• JWT RBAC on HTTP routes
• SOC 2 / HIPAA audit trail

---

v0.3.0

Added

• ReAct agentic loop with createAgent
• createHttpService HTTP runtime with OpenAPI
• 50+ built-in tools
• RAG / KnowledgeEngine

---

v0.1.0

Initial release.