Tags

0.9.5

0.9.5 — Phase 65j: chat/thread storm gates, visible-close etiquette, breaker log sanity

0.9.4

0.9.4 — stability & UX fix batch

- hive-client: composer/sidebar selection fixes, exception-safe UI jobs, first unit tests, UiApi select-agent hook
- routine orphan self-heal (thread delete → rebind, scheduler no longer disables on error)
- thread-etiquette system prompt (cross-agent noise + ack-storms)
- browser freeze fixes: zombie-aware x11vnc watchdog, idle watcher respects busy agents, pkill  fixes
- workspace read-only flip fixed: screen recreate now passes --userns=keep-id
- deploy skill v1.3.0

0.9.3

Phase 62-64: amnesia fix, per-message turn stats (ctx/cache/iterations), deploy-skill GPU gate fix

0.9.2

0.9.2 — secret broker agent tools (secrets_list/secrets_get, Phase 61); spawn-frame secret delivery (Phase 60/60b/60c)

0.9.1

0.9.1 — client polish: windowed history, context menus, per-agent stop, scheduled-tasks UX

- Phase 55 (7f95d1f): bounded thread history — GET /api/threads/:id/messages
  returns the latest message_window + has_more; ?before= pages back;
  ServerConfig.message_window (default 200, 10..5000) via GET/PUT
  /api/config + webUI Settings → Chat history; client renders the window
  with a 'load earlier…' pager (no more whole-thread loads)
- Unread dot clears instantly on thread select (5c30a4c)
- Phase 56 (4333342): thread-list context menu — Clear history (new
  POST /api/threads/:id/clear) + Delete per-row; title-bar trash removed
- Phase 57 (566d4fe): Stop moved to the agent context menu as per-agent
  cancel (ChatService::cancel_agent_turns + POST /api/agents/:id/cancel);
  title-bar Stop (thread-wide) removed; ui_stop UiApi kept
- Phase 58 (a3ecb30, 99ab8fc, 8daff35): Scheduled Tasks consolidated —
  agent menu '⏰ Schedule Task…' (Add for that agent), title bar opens
  the manager (list/pause/run/edit/delete + '➕ Add task…'), list
  self-reloads after Add/Edit, pause-switch tooltips
- test_thread_pool: 3 load-sensitive races fixed (the -j4 flake, 0ddeb80)
- hive-worker.service template with KillMode=mixed + install docs (13d0035)

0.9.0

0.9.0 — live agent output, agent-managed routines, screen-share repair

Phase 52 (47c1f89):
- hive-client renders the live partial as an ephemeral streaming bubble
  while an agent works (activity endpoint was wired but only 'busy' read)
- hub flushes a dying turn's partial to its thread with *(interrupted)*
  on cancel + watchdog retire (LiveTurn.partial_flushed)
- hive-agent captures the stop tool's summary into the reply

Phase 53 (ed0f5ce):
- agent routine tools restored: schedule_task / list_schedules /
  update_schedule / delete_schedule — request/response over
  routine_req/routine_resp riding the NDJSON/op plumbing
- hub HubRouter applies them against RoutineStore with per-agent
  ownership checks, mirroring the Scheduled page semantics

Phase 54 (0e37391, 852c41b, a94b400, c566753, 381726b):
- noVNC screen share repaired — six layered bugs: stale port-mapping
  status lie, http:// redirect downgrade, uncaught 500 on offline
  workers, 3s bring-up race in the client, fleet drift-check fighting
  screen sessions (screen-image pin + recreate mutex), and pasta port
  forwarding dying with the worker service (port_accepts probe +
  KillMode=mixed on both worker hosts)

0.8.1

0.8.1 — real-world test pass fixes + live disabled-skills gate  - browser_driver.py installs to /usr/local/bin (/home/hive is bind-mount masked)  - Chrome launches with --no-sandbox (rootless podman can't run SUID sandbox)  - browser driver self-provisions the Xvfb display (worker screens are on-demand)  - Phase 50: disabled-skills gate wired hub → worker → agent, live via task frames  - Phase 50b: LoadSkillTool refuses disabled skills before loading (no content leak)  All verified live against the fleet; 101/101 ctest green.

0.8.0

0.8.0 — web tools restored, browser_* computer-use tools, engine test suite, per-app test layout

- Phase 46: web_search + web_fetch restored in hive-agent
- Phase 47: browser_* computer-use tools recreated
- Phase 48: engine unit tests restored (83 ported), daedalus→hive-agent
  rename sweep, SkillManager is_skill_active() Phase 15 fix
- Phase 49: per-app tests/ (hive-server, hive-worker, hive-client)
- deploy-hivement skill: full deployment runbook
Deployed: droplet hub @5d82d93, gpu-rig images 05577659/dd93b65c/4041bc83

0.7.0

0.7.0 — engine merged in-tree + hub-owned LLM registry (Phases 42-43)

- hive-agent: daedalus engine merged into hive-agent/src (namespace hive);
  no more fetched hive-brain — one repo, one build
- Phase 43: hub-owned provider/model registry (llm.db) — hub injects full
  llm config into every task frame; brains upsert before each turn
- webUI Models page: Providers table (Add/Edit/Delete/Discover with
  provider /models probing, masked keys) + Models per provider + hub default
- Full daedalus → hive-agent rename: code, runtime paths
  (.hive-agent/<Name>/, hive-agent.db, HIVE_AGENT_* envs), migration from
  both old layouts, docs rewrite
- Worker hardening: task frames queued during container boot (race fix),
  llm api_keys redacted from journal logs
- Hub hardening: watchdog retires cancels the worker never confirms
  (HIVEMIND_CANCEL_GRACE_S, def 90s) — fixes cancel-after-dropped-frame
  busy wedge; dispatch failures post a visible thread message
- hive-client: agent click recreates a deleted DM (DM-focus guard fix)

0.6.0

0.6.0 — Phase 33: worker-native skills, MCP tunnel, file locality

- Skill discovery tools (search_skills/install_skill) + registry (19 skills)
- Skill sync server→worker over the channel (reconcile on reconnect)
- /skills:ro container mounts + home symlinks both sides
- MCP tunnel: loopback streamable-http bridge → mcp_rpc channel op →
  worker spawns MCP servers inside agent containers (podman exec -i)
- Per-agent mcp_servers roster config + daedalus_home fix (per-agent
  SettingsStore) + skills_dirs opt-in fix
- file_push: server→worker file transfer, files stay on the rig
- Live-tested on dev-rig: remotion 3s mp4 render, Blender 4.5 render
  via MCP through the agent tool loop

0.5.3

0.5.3 — worker collaboration + cross-machine file delivery

- hive-worker: shared per-rig /workspace (sibling agents on one rig can collaborate)
- OmniNet wss idle-close fix rolled out everywhere (worker channel stable)
- Phase 31: send_file agent tool (worker → company workspace, confined, 50MB cap, notify)
- resolve_confined_create confinement helper + tests

0.5.1

0.5.1 — OmniNet 1.0.0 (connection-leak fix), container network self-heal, client selection/message-drop fix, app icon

0.5.0

HiveMind 0.5.0 — distributed workers milestone

184 commits since 0.2.0 (2026-08-08 → 2026-08-26):

- Phase 30: hive-worker — remote machine executor (enroll/revoke CLI,
  hmwk_ keys), persistent WS worker channel (auto-reconnect, chunked
  exec streams), remote container exec via WorkerRouter, screen proxy
  (noVNC/RFB over the channel), Workers page in web UI + 'Runs on'
  agent chip
- Phases 13/14: HiveMind owns LLM provider/model config (LLMStore),
  credential injection into daedalus agents
- Phases 24/25, 27/28/29: GTK client polish, routines, triggers,
  approvals auto-review
- OmniNet WS hardening: keepAliveTimeout on upgraded sockets, readLoop
  coalesced-frame handling

Validated end-to-end: hive-client → hive-server → hive-worker (dev-rig)
→ podman container agent + live screen share.

0.2.0