v0.11.0 - Multi-Agent Swarms

[0.11.0] - 2026-04-06

First-class multi-agent orchestration with shared budget control, cross-agent memory, agent-to-agent communication, and enterprise security hardening.

Added

Multi-Agent Swarm System

• Swarm with 5 topologies: ORCHESTRATOR, PARALLEL, CONSENSUS, REFLECTION, WORKFLOW
• Shared BudgetPool — hard per-agent caps enforced; no agent can exceed its allocation even if pool has funds
• MemoryBus — selective cross-agent memory sharing with type filters and custom backends
• A2ARouter — typed agent-to-agent messaging (direct, broadcast, topic pub/sub) with ack support and audit log
• SwarmController.topup_budget() / reallocate_budget() — runtime budget reallocation without restarting agents; backed by asyncio.Lock
• BroadcastBus — pub/sub with wildcard topic patterns ("research.*")
• MonitorLoop — async supervisor loop: heartbeat polling, quality assessment, bounded interventions with MaxInterventionsExceeded escalation
• SwarmAuthorityGuard — role-based permission model (ORCHESTRATOR > SUPERVISOR > WORKER); every control action permission-checked and audit-logged
• AgentRouter — LLM-driven agent selection from a pool (replaces removed DynamicPipeline)
• Workflow — sequential, parallel, branch, and dynamic fan-out steps with HandoffContext; full play()/pause()/resume()/cancel() lifecycle
• SwarmResult — unified result: content, cost_breakdown, agent_results, partial_results, budget_report

Budget Intelligence

• Pre-flight estimation: Budget.estimate() with EstimationPolicy
• Budget forecasting from run history with Hook.BUDGET_FORECAST
• Anomaly detection: Budget(anomaly_detection=True) triggers Hook.BUDGET_ANOMALY on p95 breach
• Cross-run daily/weekly limits with Hook.DAILY_LIMIT_APPROACHING
• FileBudgetStore for persistent cross-run history

Workflow & Swarm Visualization

• Workflow.visualize() — Rich ASCII tree
• Workflow.to_mermaid() — Mermaid graph TD for GitHub/docs
• Workflow.run(show_graph=True) — live Rich overlay with per-step status and cost
• GET /graph HTTP endpoint on Swarm.serve() and Workflow.serve()

AgentRegistry

• AgentRegistry with @registry.register(tags=[...]) — discover and control agents by name or capability tag

Security Hardening

• PIIGuardrail — detect and redact PII (email, phone, SSN, credit card) in inputs, outputs, and memory writes
• ToolOutputValidator — schema validation on tool results before the LLM sees them
• AgentIdentity — Ed25519 cryptographic identity; every A2A message signed and verified
• Decision provenance: DECISION_MADE hook with DecisionRecord (model, timestamp, hash) for audit trails

Remote Config Control Plane

• RemoteConfig — live push of any agent field without restart; schema export for dashboard rendering
• Config versioning and rollback(), per-field allow/deny lists, RemoteConfigValidator
• Remote lifecycle commands: PAUSE, RESUME, KILL over the wire

Multi-Agent Pry Debugger (TUI)

• SwarmPryTUI — Rich Live compositor: GraphPanel (per-agent status + cost), BudgetPanel (pool tree), MessagePanel (A2A/MemoryBus timeline), SwarmNav (keyboard navigation)

Changed

• rich>=13.0 promoted to core dependency
• requests>=2.33.0 minimum enforced (CVE-2026-25645)
• ExceedPolicy is now the canonical budget policy enum; on_exceeded= callback pattern deprecated
• MockResponseMode enum replaces raw strings "lorem"/"custom" in Model.mock()
• Class-based agent definition is now the canonical pattern in all docs and examples

Fixed

• @structured decorator now applies @dataclass internally — no manual decoration needed
• _has_default() helper: fixed broken is not type(None) comparison
• Swarm context injection: _AgentMeta metaclass compatibility fixed via plain setattr fallback
• @tool without a description now emits UserWarning at decoration time

Removed

• DynamicPipeline — removed; use AgentRouter. Importing raises ImportError with migration message.
• AgentTeam — removed; use Swarm(topology=SwarmTopology.ORCHESTRATOR). Importing raises ImportError.

v0.10.0 - Debug Watch Triggers

[0.10.0] - 2026-03-31

Added

Syrin Debug — Live Terminal Debugging

• Rich-based live terminal UI activated via debug=True
• Color-coded events: LLM (blue), tools (yellow), errors (red), memory (purple), guardrails (red)
• Multi-agent aware — nested panels for spawned agents
• Keyboard filtering: e errors, t tools, m memory, f full trace, q quit, p pause

Event-Driven Triggers (agent.watch())

• Watchable mixin with swappable protocols
• CronProtocol — cron-scheduled triggers using croniter
• WebhookProtocol — HTTP webhook trigger with HMAC signature validation
• QueueProtocol — message queue trigger with pluggable backend (Redis, in-memory)
• Agents that react to the world, not just API calls

Production Knowledge Pool

• GitHubSource — ingest entire GitHub repos (public and private)
• DocsSource — crawl documentation sites with depth limits and pattern matching
• Multi-language code chunking (Python, Go, Rust, TypeScript, Java, etc.)
• Progress events: KNOWLEDGE_CHUNK_START, KNOWLEDGE_CHUNK_PROGRESS, KNOWLEDGE_EMBED_PROGRESS, KNOWLEDGE_CHUNK_END
• Semantic chunking strategy for better retrieval
• Rate limiting for embedding APIs

Prompt Injection Defense

• InputNormalization — normalize inputs through defense pipeline
• SpotlightDefense — clearly label untrusted content
• CanaryTokens — hidden tokens that trigger alerts
• Secure memory writes — scan before storing

Structured Output Enforcement

• result.output always returns the typed object (never dict/str)
• Automatic retry on validation failure (configurable validation_retries)
• Validation hooks: Hook.OUTPUT_VALIDATION_RETRY, Hook.OUTPUT_VALIDATION_ERROR
• Clear error messages with raw response and validation details

Runtime Model Switching

• agent.switch_model() — switch model at runtime without recreating agent
• Context, memory, and hooks remain intact

Syrin v0.9.0 - Smart Budget

[0.9.0] - 2026-03-25

Breaking

• Budget(run=) renamed to Budget(max_cost=)
• agent.response() renamed to agent.run() — old method removed
• per= in rate limits replaced by RateLimit(hour=, day=, month=)

Added

• ExceedPolicy enum — STOP | WARN | IGNORE | SWITCH replaces string policies
• result.cost_estimated — pre-call budget estimate alongside result.cost
• result.cache_savings — cache discount reported per call
• agent.budget_summary() — dashboard dict: run cost, tokens, hourly/daily totals, percent used
• agent.export_costs(format="json") — structured cost history per call
• BudgetStore ABC — plug in any persistence backend (PostgreSQL, Redis, DynamoDB)
• Budget(shared=True) — thread-safe shared pool across parallel agents (SQLite WAL)
• ModelPricing override — pass pricing=ModelPricing(...) to any Model() to override built-in rates
• Docs — full rewrite with correct APIs, custom observability guide, 375+ fixed internal links

Fixed

• SQLite budget store thread safety under parallel spawns
• Threshold callbacks no longer re-fire on the same percentage within a run
• PIIScanner redaction applied before content reaches the LLM
• Memory.import_from() logging noise removed

v0.8.1 - Grounded Intelligence

[0.8.1] - 2026-03-13

Added

• Batch fact verification — Ground facts in batches of 10 for efficiency
• GroundingConfig.model — Dedicated model for grounding/verification
• YAML frontmatter — Template.from_file() parses YAML frontmatter

Changed

• PDF extraction — Consolidated to single docling dependency
• Grounding — Improved claim-to-fact matching with index-based pairing

Fixed

• IPO DRHP example — Working example with Budget, FactVerificationGuardrail
• Test mocks — Fixed output validation failure tests

v0.8.0 - Knowledge Pool

[0.8.0] - 2026-03-11

Added

• Intelligent model routing: Automatic LLM selection based on cost, performance, and context requirements.
• Multi-modality support: Native support for images, video, and audio generation/processing.
• Knowledge integration: RAG (Retrieval-Augmented Generation) with vector stores, document loaders, and semantic search.

Changed

• Simplified installation: pip install syrin now includes OpenAI and serving dependencies by default.
• Removed CLI: syrin trace, syrin run, and syrin doctor commands removed; use python my_agent.py --trace for observability.

Breaking

• CLI removal: Command-line interface deprecated; use direct Python execution with --trace flag.

v0.7.0 - Context Management

[0.7.0] - 2026-03-07

Breaking

• Context: Context.budget removed. Use Context.token_limits (TokenLimits). ContextWindowBudget → ContextWindowCapacity. Context.get_budget(model) → Context.get_capacity(model). ContextManager.prepare() takes capacity instead of budget.

Added

• Context management — Snapshot (provenance, why_included, context_rot_risk), breakdown, custom compaction prompt, auto_compact_at, runtime injection, context_mode (full/focused), formation_mode (push/pull), stored output chunks, persistent context map, pluggable RelevanceScorer.
• Memory–context — Memory on by default; memory=None turns off. No extra field.
• Handoff/spawn — Context visibility in events (handoff_context, context_inherited, initial_context_tokens).

Fixed

• Examples: Output(type=...) → Output(MyModel); Agent(dependencies=...) → Agent(config=AgentConfig(dependencies=...)).

v0.6.0 — CC

[0.6.0] - 2026-03-05

Added

• Remote config — syrin.init(api_key=...) or SYRIN_API_KEY enables real-time config overrides from Syrin Cloud or self-hosted backend. Overrides (budget, memory, temperature, etc.) via SSE; zero overhead when not enabled.
• Config routes — GET /config, PATCH /config, GET /config/stream added to agent.serve(). Baseline + overrides + revert; works with or without syrin.init().
• syrin.remote — Types: AgentSchema, ConfigOverride, OverridePayload, SyncRequest/SyncResponse. ConfigRegistry, ConfigResolver, extract_schema(). Transports: SSETransport, ServeTransport, PollingTransport.
• Hooks — Hook.REMOTE_CONFIG_UPDATE, Hook.REMOTE_CONFIG_ERROR.

Changed

• Agent registers with remote config on init when syrin.init() was called.

v0.5.0 - Remember Smarter

[0.5.0] - 2026-03-04

Added

• C5 fix — Memory.remember/recall/forget use configured backend (SQLite, Qdrant, Chroma) instead of in-memory dict when backend != MEMORY.
• QdrantConfig — Memory(qdrant=QdrantConfig(url=..., api_key=..., collection=..., namespace=...)) for Qdrant Cloud or local.
• ChromaConfig — Memory(chroma=ChromaConfig(path=..., collection=...)) for Chroma vector backend.
• Namespace isolation — QdrantConfig.namespace scopes all operations; payload filter on search/list.
• WriteMode — WriteMode.SYNC (block until complete) vs WriteMode.ASYNC (fire-and-forget, default).
• Memory export/import — Memory.export() returns MemorySnapshot; Memory.import_from(snapshot) appends memories. JSON-serializable for GDPR export.
• Examples — examples/04_memory/qdrant_memory.py, chroma_memory.py, async_memory.py, export_import_memory.py.

Changed

• Agent and Memory handoff now use memory._backend_kwargs() for backend config.
• syrin[qdrant] and syrin[chroma] optional dependencies added to pyproject.toml.

v0.4.1

[0.4.1] - 2026-03-01

Added

• API additions: Response.raw_response, GuardrailCheckResult.guardrail_name, CircuitBreaker.state, EventBus.on, GlobalConfig.debug, TokenLimits.per_hour, RateLimit.window, agent.checkpointer.

Fixed

• Model fallback and response transformer now use model.acomplete() when model has fallbacks/transformers.
• Model.with_middleware() preserves provider.
• Provider availability checks use importlib.util.find_spec() instead of import.

Changed

• Strict typing: TypedDict + Unpack for memory kwargs, ServeConfig, agent specs. Pyright config added.
• Replaced Any with object / TypedDict across core modules.
• Docs: docs/TYPING.md for typing standards; updated API references.

v0.3.0

[0.3.0] - 2026-02-27

Added

• Sub-agents & handoff — spawn(task=...), handoff(AgentClass, task) with optional memory transfer and budget inheritance.
• Handoff interception — events.before(Hook.HANDOFF_START, fn); raise HandoffBlockedError to block; HandoffRetryRequested for retry.
• Audit logging — AuditLog, JsonlAuditBackend; Agent(audit=...), Pipeline(audit=...), DynamicPipeline(audit=...).
• HITL — @syrin.tool(requires_approval=True); ApprovalGate protocol; hooks: HITL_PENDING, HITL_APPROVED, HITL_REJECTED.
• Circuit breaker — CircuitBreaker for LLM/provider failures; CLOSED → OPEN → HALF_OPEN; configurable fallback.
• Budget-aware context — Context tier selection by budget percent remaining.
• Dependency Injection — Agent(deps=...), RunContext[Deps]; tools receive ctx.deps (excluded from LLM schema).
• Dynamic Pipeline — Improved hooks and events API; flow diagram in docs/dynamic-pipeline.md.

Changed

• API validation — Agent, Model, Memory, Loop validate inputs at construction; clear errors for wrong types.
• agent.response(user_input) — Validates user_input is str; friendly error for None/int/dict.
• Example paths — Fixed run instructions (08_streaming, 07_multi_agent).

Fixed

• Chaos stress test fixes: Agent/Loop validation; Loop max_iterations < 1 no longer causes UnboundLocalError. Model _provider_kwargs passed to provider.