Skip to main content
Version: Latest

Release notes

This page mirrors the project changelog on GitHub. Update it when cutting a new release (copy from CHANGELOG.md and move Unreleased into a new version section).

[Unreleased]​

[2.1.5] — 2026-03-11​

Added​

  • In-app HITL resolution — Optional in-process resolver so you can approve/reject HITL requests in the same terminal as hivemind run (no second terminal). When hitl.enabled is true and stdout is a TTY, the CLI prompts with a Rich prompt; resolution is written to ApprovalStore and the run continues or fails based on your choice.
  • HITL in single-node path — HITL escalation check and resolver/polling now run in the default single-node flow (WorkerNode), not only in the multi-worker Executor path.
  • Better MCP — hivemind doctor has a dedicated "MCP Servers" section listing each configured MCP server and tool count (or warnings). Second example MCP server (time server, no API key) in commented block in example hivemind.toml.
  • Full CLI visual redesign: amber/blue/teal color system across all commands
  • Structured logging with tracing-compatible format (matches Rust worker output)
  • Live run view: real-time task table, tool activity, cost counter during execution
  • Animated planning phase with strategy selection feedback
  • Redesigned hivemind init: interactive wizard with welcome screen (pyfiglet) and provider/model flow
  • Redesigned hivemind doctor: themed header and sectioned health check output
  • HivemindProgress: styled progress bars for long-running operations
  • Typed error classes (HivemindError, ProviderConnectionError, ConfigNotFoundError, etc.) with actionable hints and docs links
  • Shell completions: bash, zsh, fish (fish hint when shtab used for bash/zsh)
  • --debug, --trace, --quiet, --no-color, --json, --plain global flags on all commands
  • Auto-plain mode when stdout is not a TTY
  • hivemind run --summary to print only run summary without task results
  • New module hivemind/cli/ui/: theme, components, logging, progress, errors, run_view, onboarding

Changed​

  • All CLI output uses themed console (no bare print() in UI code paths)
  • Python logging replaced with HivemindLogger (tracing-format compatible)
  • Error display: no raw tracebacks shown to end users; use print_error/print_unexpected_error
  • hivemind run shows live view by default when TTY (use --plain or pipe for old behavior)
  • Docs URLs use https://hivemind.rithul.dev
  • Planner: simple-task fast path — Short, single-step prompts (e.g. "What is 2+2?") no longer get decomposed into 5 steps; they run as one task and one agent call.
  • Planner: dynamic step count — Planner prompt asks for "the minimal number of smaller steps needed" instead of a fixed 5; the model can return 1–3 for simple tasks or more for complex ones.

[2.1.0] — 2026-03-11​

Added​

  • MetaPlanner — Decompose mega-tasks into sub-swarms with dependencies, SLAs, and priorities
  • SubSwarmSpec — Per-swarm priority, SLA, worker count, model override, and depends_on
  • SLA monitoring — Duration/cost/quality breach detection with configurable actions (cancel, escalate, continue)
  • PriorityScheduler — Priority- and dependency-aware task ordering; add_task(task, priority), bump_priority(task_id, new_priority)
  • Human-in-the-Loop (HITL) — Configurable escalation triggers and approval workflows
  • ApprovalStore — Persistent pending approvals under {data_dir}/approvals/ with timeout handling
  • Approval notifications — Webhook and Slack channels (email logs to stdout without SMTP)
  • CLI — hivemind meta "<mega-task>" and hivemind meta plan "<mega-task>"; hivemind approvals list|show|approve|reject|watch
  • TASK_REJECTED_BY_HUMAN event type

[2.0.2] — 2026-03-10​

Fixed​

  • Release workflow — Docs version step skips when version already exists; PyPI publish uses skip-existing so re-tags or re-runs don't fail.

[2.0.1] — 2026-03-10​

Added​

  • Azure Foundry (v1 API) support — When AZURE_OPENAI_ENDPOINT points to Azure Foundry (URL contains cognitiveservices.azure.com or /openai/v1), the provider uses the v1 chat-completions API via ChatOpenAI with base_url instead of the legacy deployment-path API, fixing 404s on Foundry resources.
  • Credentials set inline and stdin — hivemind credentials set <provider> <key> [value] accepts an optional value; if omitted and stdin is not a TTY, reads value from stdin (e.g. echo "https://..." | hivemind credentials set azure endpoint).

Changed​

  • Credentials input masking — Only sensitive keys (api_key, token) use hidden input; endpoint, deployment, and api_version prompts show typed input.
  • Azure model spec — Provider strips provider: prefix (e.g. azure:gpt-5-mini → gpt-5-mini) before sending to the API so deployment name is correct.
  • .env.example — Documents correct Azure Foundry endpoint (.../openai/v1 for chat completions; avoid .../openai/responses).

[2.0.0] — 2026-03-10​

Breaking Changes​

  • Provider config schema updated: existing provider strings unchanged, new backends require new config sections
  • Agent execution now routed through AgentSandbox by default (disable with sandbox.enabled = false)
  • Memory storage now redacts PII by default if compliance.pii_redaction = true

Added​

  • Abstract LLM router with Ollama, vLLM, and custom OpenAI-compatible endpoint backends
  • Provider fallback chains: automatic failover across backends
  • Agent sandboxing: resource quotas, tool category restrictions, per-role sandbox profiles
  • Audit logging: append-only JSONL with chain integrity verification
  • PII redaction: regex + optional spaCy NER, configurable PII types
  • GDPR/CCPA compliance config section
  • Decision tree and rationale generation for every agent action
  • Simulation mode: dry-run planning without LLM calls or tool execution
  • hivemind explain, hivemind simulate, hivemind audit CLI commands
  • PROVIDER_FALLBACK event type

Migration from 1.x​

See Migration to v2 (or docs/migration/v2.md in the repo).

For older releases, see the full changelog on GitHub.