v0.6.0 — Framework Hardening

Framework-hardening release. The agent runtime is more robust under streaming interruptions, hung tools, and corrupt state, and the package is friendlier to import and embed. No breaking changes — existing databases upgrade automatically.

Added

• Per-tool execution timeout — Agent(tool_timeout=...) / daemon.tool_timeout (default 60s). A hung tool becomes a tool-error fed back to the model instead of stalling the cycle. 0 disables it.
• Install extras — hive-agent[anthropic], [openai], [mcp], [cli]. SDKs still ship in core deps; a missing one raises a clear MissingDependencyError naming the extra.
• Context managers + lazy init on Hive — with Hive() as h: / async with Hive() as h:, and Hive(path).spawn(...) scaffolds .hive/ lazily so init() is optional. ensure_hive_dirs() exported.
• StructuredParseError and MissingDependencyError typed exceptions.

Changed

• Streaming falls back instead of failing — a stream ending without a terminal DONE event (or erroring mid-stream) keeps the text already shown or falls back to a retried non-streaming call; the stream is closed on cancellation.
• StructuredTaskResult.parsed is now T | None — failed structured runs return parsed=None instead of an unvalidated object.
• Provider error detection is structured-first for unsupported response_format/json_schema.

Fixed

• Previously-silent failures surfaced — memory recall, corrupt suffering/persona restore, and plugin-init failures now log with context; corrupt suffering snapshots fall back to a fresh state.
• Robust structured-output parsing — a brace-matching scanner replaces the naive first-{/last-} slice.
• One active generated goal per agent — the existence loop re-checks before saving (subgoals/delegation still create multiple by design).
• Malformed tool-call arguments are logged instead of silently dropped.
• Budget: the run_once wrap-up generation is now budget-checked; a hard budget can overshoot by at most one generation (documented).

Full changelog: https://github.com/chiruu12/Hive/blob/main/CHANGELOG.md

v0.5.2 — Hardened No-Tools Recovery

Hive v0.5.2 Release Notes

Headline: Hardened No-Tools Recovery

A follow-up to v0.5.1 that incorporates review feedback on the no-tools recovery
path, plus the Phase 3 simulation core that landed on main.

What's Fixed

• Streaming recovery is scoped to pre-content failures: the OpenAI-compatible
  adapter now recovers a tool_use_failed only when it happens before any text has
  reached the caller, so a mid-stream error can never produce duplicated output. If a
  recovery stream errors after emitting text, the terminal result preserves the text
  already streamed.
• Recovery no longer masks real errors: a tool_use_failed on a request that
  did offer tools (e.g. a malformed tool schema) is surfaced instead of being
  swallowed by the recovery retry.
• No mid-thread system messages: both the agent-layer wrap-up nudge and the
  adapter's text-only recovery nudge are sent as user-role messages, which strict
  providers accept mid-conversation (some reject mid-thread system messages). The
  agent nudge is sent only for the wrap-up call and is not written to the logged
  conversation.

Also Included

• Phase 3 simulation core: registry-driven world catalogs (events & jobs) and
  wired simulation feedback loops.

Upgrade

pip install --upgrade hive-agent

No code changes required. The happy path (tools provided, or no error) is unchanged.

v0.5.1 — Resilient No-Tools Wrap-Up (Groq tool_use_failed recovery)

Hive v0.5.1 Release Notes

Headline: Resilient No-Tools Wrap-Up Across Strict Providers

A focused reliability fix. When Agent.run_once finishes its tool loop it makes a
final wrap-up call with no tools offered. If the model still emits a tool call on
that call -- common on multi-action requests -- strict OpenAI-compatible providers
(notably Groq) reject it with a tool_use_failed 400, failing the whole turn
even though the tools that ran during the loop already persisted.

What's Fixed

• Provider-agnostic recovery in the OpenAI-compatible adapter: a tool_use_failed
  rejection on a no-tools request is detected by error code/message (no provider
  hardcoding) and recovered with a single bounded retry carrying a strong text-only
  instruction. If the retry still fails, the turn completes with clean text instead
  of raising -- the tools already ran. Covers both generate_with_metadata and
  generate_stream.
• Belt-and-suspenders nudge at the agent layer: Agent.run_once now appends a
  "tool budget exhausted, reply in plain text" system message before the final
  wrap-up call, reducing the chance of hitting the error at all.

A multi-action turn (e.g. "make three notes") now completes on Groq without surfacing
a 400.

Upgrade

pip install --upgrade hive-agent

No code changes required. The happy path (tools provided, or no error) is unchanged.

v0.5.0 — Composable Core + Daemon Scalability

Hive v0.5.0 Release Notes

Headline: A Composable Core That Scales

This release hardens and decouples Hive's shared core. The runtime Agent is now
a clean, standalone, streaming-capable building block, and the daemon runs many
agents concurrently. Most changes are additive -- the public Agent, provider,
and toolkit APIs stay backward compatible.

What's New

Streaming

• BaseProvider.generate_stream() + StreamEvent -- token streaming with a base
  fallback so every provider works out of the box.
• Native streaming for Anthropic and OpenAI-compatible providers (Groq, Fireworks,
  Ollama, LM Studio, OpenRouter).
• Agent(on_text=...) streams assistant text token-by-token during run().

Provider capabilities & availability

• Capability + supports(...) -- branch on what a provider can do, not its class.
• Availability + availability() -- distinguishes "no API key" from "unreachable";
  hive models now shows the reason for unavailable local servers.

Daemon scalability

• Agent cycles run concurrently with bounded concurrency
  (daemon.max_concurrent_agents, default 8). Each cycle is isolated, so one slow,
  timed-out, or failing agent never blocks or breaks the others.
• Per-agent provider and profile are cached across cycles.

Framework polish

• Concurrent tool execution -- multiple tool calls in one model turn run in
  parallel, with per-call error isolation and ordered results.
• Typed errors -- HiveError, AgentNotFoundError, ProfileNotFoundError.
• Standalone Agent is first-class: 2-line usage, no daemon (see example 23).
• InstructionLike protocol unifies Instructions / Persona / custom objects.
• Structured output works on every provider (prompt-based fallback default).

Persistence

• SQLite store gained indexes on hot columns and versioned migrations
  (PRAGMA user_version) that upgrade older databases in place.

Tools

• ClipboardToolkit.read_clipboard -- read the system clipboard, complementing the
  existing copy tools.

Upgrade

pip install --upgrade hive-agent

No code changes required for existing agents. To stream, pass on_text=... to
Agent; to bound daemon throughput, set daemon.max_concurrent_agents.

Stats

• Anthropic, OpenAI, Groq, Fireworks, Ollama, LM Studio, OpenRouter providers
• Full test suite green on Python 3.11 / 3.12 / 3.13

v0.4.2 — Missing Toolkit APIs

What's New

Fills in missing CRUD operations that downstream consumers (Nudge) were forced to hack around using private attributes and raw SQL.

TaskToolkit

• uncomplete_task -- reopen a completed task, setting it back to pending
• update_task -- modify description, priority, or due date on an existing task
• Priority filtering -- list_tasks(priority="high") filters by priority level

KnowledgeToolkit

• delete_note -- delete a note by ID (replaces ._memory.forget() hack)
• update_note -- edit content or tags in-place, preserving ID and creation timestamp

AlarmToolkit

• set_alarm_at -- set alarms for absolute times like "3pm", "15:00", "tomorrow 9am"
• Uses python-dateutil for parsing, interprets naive times as local timezone
• Time-only inputs auto-roll to tomorrow if already past; explicit past dates are rejected

Memory Layer

• TFIDFBackend.update() -- modify notes in-place without losing ID or timestamp
• SemanticMemory.update() -- facade with hasattr guard for backend compatibility

Other

• Groq test fix -- test_groq_unavailable_without_key no longer leaks OPENAI_API_KEY from .env
• Docs -- TaskToolkit, KnowledgeToolkit, and AlarmToolkit added to toolkit guide
• Changelog -- v0.4.0, v0.4.1, v0.4.2 entries added

New Dependency

• python-dateutil>=2.8 (main), types-python-dateutil>=2.8 (dev)

Stats

• 828 tests passing (32 new)
• 14 files changed, ~500 lines added
• No database schema changes -- all new methods use existing columns

Full Changelog: v0.4.1...v0.4.2

v0.4.1 — Shared Toolkit Race Fix

What's New

Bugfix release -- fixes a race condition when sharing toolkits across concurrent agent requests.

Fixed

• Clone-on-rebind -- when an Agent rebinds a toolkit already bound to a different agent, it now clones the toolkit instance instead of mutating the shared one. Prevents state corruption under concurrent usage (e.g. Nudge handling multiple voice requests simultaneously).

Stats

• 815 tests passing
• 2 files changed

Full Changelog: v0.4.0...v0.4.1

v0.4.0 — Production Hardening + ClipboardToolkit

What's New

Bug fixes found while building Mutter/Hum, new toolkit APIs for host applications, and the ClipboardToolkit.

Bug Fixes (Critical)

• tc.arguments None crash — tool.call(**tc.arguments) crashed when LLM called tools with no required params (e.g. list_alarms()). Fixed in both run() and run_once().
• Integer params break Groq — Groq's API rejects "type": "integer" in JSON schema when the LLM sends strings. All LLM-facing numeric params changed to str with int(float()) parsing. Affected: AlarmToolkit, KnowledgeToolkit, LinkToolkit, A2AToolkit.
• WhisperLocal blocking inference — ML inference now runs in run_in_executor() to avoid blocking the event loop.
• MLX whisper model map — Fixed HuggingFace repo names (base/small/large-v3 repos renamed with -mlx suffix).

New Features

• ClipboardToolkit — copy_to_clipboard, copy_note, copy_task, copy_link. Uses pbcopy (macOS) / xclip (Linux).
• Public query methods on toolkits for host applications:
  • TaskToolkit.query_tasks(status) / query_all_tasks(status)
  • AlarmToolkit.query_pending_alarms() / query_all_pending_alarms()
  • KnowledgeToolkit.query_recent(limit) → list[dict]
• Configurable notification title — AlarmChecker(notification_title="Hum")
• Agent auto-rebinds toolkits — Per-request Agent creation with shared toolkits just works.
• Metadata search in TFIDF — Tags, URLs, and metadata are now indexed alongside content.

Stats

• 796 tests passing
• CI: 5/5 green (lint, mypy, tests 3.11/3.12/3.13, build wheel)
• Verified with 5 end-to-end test scenarios including live API calls

Full Changelog: v0.3.2...v0.4.0

v0.3.2 — Voice/Assistant Infrastructure

What's New

Voice/assistant infrastructure layer — everything needed to build voice apps like Mutter on top of Hive.

STT Providers

• WhisperLocal — mlx-whisper (Apple Silicon) or faster-whisper (Linux/CPU), auto-detected
• GroqSTT — Groq Whisper API with connection pooling and exponential backoff retry
• DeepgramSTT — Deepgram Nova-2 API with Pydantic response validation
• Factory with auto-detection: create_stt_provider() picks the best available

AudioRecorder

• Microphone capture via sounddevice with thread-safe callbacks
• WAV output using stdlib struct (no scipy dependency)
• Device channel validation, proper memory cleanup

IntentRouter

• LLM-based text classification using generate_structured() with Pydantic models
• Falls back to text parsing if provider doesn't support structured output
• User-defined intents, configurable fallback

Trigger Systems

• HotkeyTrigger — global hotkeys via pynput with per-trigger in-flight guards
• WebhookTrigger — lightweight HTTP server using stdlib asyncio (no extra deps)
• Content-Length validation and 10MB cap for security

LinkToolkit

• Save, search, list, and scrape web links using SemanticMemory
• Async HTTP throughout (no event loop blocking)
• Wired into daemon's _build_toolkits()

Hardening

• All data types are Pydantic BaseModel with validation
• Connection pooling with shared httpx.AsyncClient
• Exponential backoff retry on 429/500/502/503
• transcribe_bytes() properly wraps raw PCM in WAV headers

New Optional Dependencies

pip install hive-agent[audio]    # sounddevice, scipy, numpy
pip install hive-agent[hotkeys]  # pynput
pip install hive-agent[voice]    # all of the above

Whisper backends (mlx-whisper, faster-whisper) are manual install.

Stats

• 787 tests passing (127 new)
• 31 files changed, 3200+ lines added
• Stress tested: 50 concurrent webhooks, 5 concurrent STT calls, real mic recording, real hotkey listener

Full Changelog: v0.3.1...v0.3.2