Skip to content

CHANGELOG.md

Latest commit

 

History

History
199 lines (152 loc) · 8.91 KB

CHANGELOG.md

File metadata and controls

199 lines (152 loc) · 8.91 KB

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[1.10.0] - 2026-04-22

Added

HTTP Input LLM Invocation Models (Symbiont v1.10.0)

  • WebhookInvocationRequest — Caller-supplied payload accepted by the HTTP Input endpoint: prompt or message (user text), optional system_prompt (capped at 4096 bytes by the runtime, truncated on a UTF-8 character boundary), plus pass-through of additional fields.
  • WebhookToolRun — Per-tool ORGA execution preview (tool, input, output_preview). output_preview is UTF-8 safe and truncated to at most 500 bytes by the runtime.
  • WebhookExecutionStartedResponse — Returned when the target agent is in the Running state and the request is dispatched on the communication bus (status: "execution_started" with agent_id, message_id, latency_ms, timestamp).
  • WebhookCompletedResponse — Returned when the agent is not running and the request is served by the on-demand LLM invocation path, which runs an ORGA tool-calling loop against ToolClad manifests (status: "completed" with response, tool_runs, model, provider, latency_ms, timestamp).
  • WebhookInvocationResponse — Discriminated union alias (Union[WebhookExecutionStartedResponse, WebhookCompletedResponse]).
  • WebhookInvocationStatus — String enum with EXECUTION_STARTED and COMPLETED.

Changed

  • Version alignment with Symbiont runtime v1.10.0.
  • Package version bumped to 1.10.0 in setup.py and symbiont.__version__.

Notes

  • Symbiont v1.9.0 and v1.9.1 introduced ToolClad v0.4.0 runtime features (session / browser execution modes, HTTP and MCP proxy backends, output parsers, custom types, secrets injection, W3C traceparent propagation, bounded channels, rate limiting, optional LanceDB feature flag). These changes are runtime-side; existing SDK models (ToolManifestInfo, ToolExecutionResult, etc.) remain backward compatible. The backend string on ToolManifestInfo now accepts "http", "mcp", "session", and "browser" in addition to previously documented values.

[1.8.1] - 2026-03-23

Added

ToolClad Manifest Management

  • ToolCladClient — Full client for ToolClad manifest operations via client.toolclad
    • list_tools() — List all discovered .clad.toml manifests
    • validate_manifest() — Validate a manifest file
    • test_tool() — Dry-run a tool (no execution)
    • get_schema() — Get MCP-compatible JSON schema for a tool
    • execute_tool() — Execute a tool with validated arguments, returns evidence envelope
    • get_tool_info() — Get detailed tool manifest information
    • reload_tools() — Trigger hot-reload of tool manifests
  • Pydantic models: ToolManifestInfo, ToolValidationResult, ToolTestResult, ToolExecutionResult

Inter-Agent Communication Policy Gate

  • list_communication_rules() — List all communication policy rules
  • add_communication_rule() — Add a policy rule (allow/deny between agents)
  • remove_communication_rule() — Remove a rule by ID
  • evaluate_communication() — Evaluate whether a communication is allowed
  • Pydantic models: CommunicationRule, CommunicationEvaluation

Agent Lifecycle

  • delete_agent() — Delete an agent and its metadata
  • re_execute_agent() — Re-execute an agent with optional new input (resets ORGA loop)

ORGA-Adaptive Features

  • reasoning.get_tool_profiles() — Tool execution timing statistics and success rates
  • reasoning.get_loop_diagnostics() — Stuck-loop detection, iteration history, adaptive parameters

Changed

  • Version alignment with Symbiont runtime v1.8.1

[1.6.1] - 2026-02-28

Added

Reasoning Loop APIs

  • ReasoningClient — Full ORGA (Observe–Reason–Ground–Act) cycle client with 15 methods
  • Cedar policy evaluation, journal system, circuit breakers, knowledge bridge
  • Pydantic models with enums for all reasoning types

Changed

  • Version alignment with Symbiont runtime v1.6.1

[0.6.0] - 2026-02-15

Added

Markdown Memory Persistence

  • MarkdownMemoryStore — File-based agent context persistence using markdown format
    • save_context() / load_context() — Atomic save with daily log files
    • delete_context() / list_agent_contexts() — Context lifecycle management
    • compact() — Remove log files older than retention period
    • get_storage_stats() — Storage statistics across all agents

Webhook Verification

  • HmacVerifier — HMAC-SHA256 webhook signature verification with prefix stripping
  • JwtVerifier — JWT-based webhook verification with optional issuer validation
  • WebhookProvider — Pre-configured providers (GitHub, Stripe, Slack, Custom) with factory method
  • SignatureVerifier ABC for custom verifier implementations

Agent Skills (ClawHavoc Scanning + Loading)

  • SkillScanner — Security scanning with 10 built-in ClawHavoc rules and custom rule support
    • Detects pipe-to-shell, wget-pipe-to-shell, env file references, SOUL.md/memory.md tampering, eval+fetch, base64-decode-exec, rm-rf, chmod-777
  • SkillLoader — Skill discovery and loading from configured paths
    • YAML frontmatter parsing for skill metadata
    • Optional SchemaPin signature verification (soft dependency)
    • Configurable scan-on-load behavior

Metrics Collection & Export

  • MetricsClient — Sub-client for runtime metrics API (GET /metrics/snapshot, etc.)
  • FileMetricsExporter — Atomic JSON file export with compact mode
  • OtlpExporter — OTLP export stub (requires opentelemetry-api)
  • CompositeExporter — Fan-out to multiple export backends
  • MetricsCollector — Background thread for periodic metrics export
  • MetricsSnapshot — Serializable snapshot with scheduler, task manager, load balancer, and system metrics

New Exceptions

  • WebhookVerificationError, SkillLoadError, SkillScanError, MetricsExportError, MetricsConfigError

New Pydantic Models

  • Webhook: WebhookProviderType, WebhookVerificationConfig
  • Skills: SignatureStatusType, ScanSeverityType, ScanFindingModel, ScanResultModel, SkillMetadataModel, LoadedSkillModel, SkillsConfig
  • Metrics: OtlpProtocol, OtlpConfig, FileMetricsConfig, MetricsConfig, SchedulerMetricsSnapshot, TaskManagerMetricsSnapshot, LoadBalancerMetricsSnapshot, SystemResourceMetricsSnapshot, MetricsSnapshot

Optional Dependencies

  • skills extra: schemapin>=0.2.0
  • metrics extra: opentelemetry-api, opentelemetry-sdk, opentelemetry-exporter-otlp

Changed

  • Aligned with Symbiont Runtime v1.4.0
  • Client.metrics_client property — Lazy-loaded MetricsClient sub-client
  • All new types exported from symbiont package

[0.5.0] - 2026-02-11

Added

AgentPin Integration

  • AgentPinClient — Client-side credential verification, discovery, and trust bundle support via agentpin PyPI package (v0.2.0)
    • client.agentpin.verify_credential() — Full 12-step online verification
    • client.agentpin.verify_credential_offline() — Offline verification with pre-fetched documents
    • client.agentpin.verify_credential_with_bundle() — Trust bundle-based verification (no network)
    • client.agentpin.fetch_discovery_document() — Fetch .well-known/agent-identity.json
    • client.agentpin.issue_credential() — Issue ES256 JWT credentials
    • client.agentpin.generate_key_pair() — P-256 key generation
    • Key pinning (TOFU) and JWK utilities
  • Client.agentpin property — Lazy-loaded AgentPinClient accessible directly from the main Client instance
  • AgentPinClient export in symbiont package

Changed

  • Aligned with Symbiont v1.0.1 release

Fixed

  • Fixed setup.py version drift (was stuck at 0.2.0, now matches __version__)

[0.4.0] - 2026-02-07

Added

Scheduling Parity & API Alignment

  • SchedulerHealthResponse — New dataclass for the 13-field GET /health/scheduler endpoint
  • get_scheduler_health() — New method on ScheduleClient to query scheduler health
  • Client.schedules property — Lazy-loaded ScheduleClient accessible directly from the main Client instance
  • Schedule type exports — All schedule dataclasses now exported from symbiont package (CreateScheduleRequest, ScheduleSummary, ScheduleDetail, SchedulerHealthResponse, etc.)

Test Coverage

  • ScheduleClient tests — 18 test cases covering all 11 methods (list, create, get, update, delete, pause, resume, trigger, history, next-runs, scheduler-health) plus edge cases

[0.3.1] - 2025-01-16

Added

  • ScheduleClient for cron schedule management API
  • Test compatibility fixes for enhanced Client class

[0.3.0] and earlier

See previous releases for historical changes.