OtoDock 1.6.0

Added

• GPT-6 Astra on the Codex engine. OpenAI's new flagship is a predefined
  Codex model. "Auto" keeps choosing GPT-5.6 Sol: Astra costs 2.5 times more.
• Local models, everywhere. A Codex agent on a paired machine can run on a
  local endpoint; one endpoint serves both the Direct LLM API and Codex engines
  and can carry an API key; chats show a local model's thinking and pass effort.
• The Direct LLM API engine catches up. OpenAI models use the Responses API
  (effort with tools, reasoning shown, web search with sources, nothing stored).
  Agents get file tools and the sidecar MCPs. Tools and skills load on demand
  (DIRECT_LLM_TOOL_SEARCH=off restores the old way). Needs the current relay.
• Phone routes know who is calling. Each caller gets a private folder and
  memory (Per caller, the default), or the call runs as a chosen platform user
  (role capped at manager); callers are viewers with no platform or shell tools.
• Caller data retention. Callers' folders, conversations and the call log
  age out together (Setup → Phone → Caller Data, 90 days, "Forget all").
  Callers' folders count against a new per-agent quota (1 GB).
• Browser control can use your own browser. A remote machine can opt in to
  "Use my own browser": the agent works in the Chrome, Edge or Brave you are
  signed into, through the Playwright Extension (save its token for tasks).
• The video-tools MCP, 0.4.0 to 0.4.2. Colour-managed grading (HLG and PQ
  to Rec.709, LUT chains, clarity, sharpness), a colour-managed edit_video,
  an align_audio tool, Rec.709 renders at 48 kHz. Update it from the catalog.
• Sandbox limits and a stall watchdog. Each local sandbox's /tmp is capped
  (SANDBOX_TMP_SIZE_MB, 4096 MB, at most half the host's memory). A watchdog
  logs the blocking stack on event-loop stalls (LOOP_WATCHDOG_THRESHOLD_S, 2 s).

Changed

• Codex CLI 0.153.4 and Claude Code 2.1.263. Paired machines update at
  their next reconnect. Codex's plan tool stays on so its checklist keeps working.
• Cost shown only when the turn cost money. API-key and hosted relay chats
  keep the badge, subscription and local-model chats hide it; usage is unchanged.
• Phone routes on upgrade. Existing routes come up in Per caller mode: a
  caller who worked in the shared space gets a private one and no longer sees
  the shared memory (tie the route to a platform user to keep the old way).
• Phone daemon (rebuild it with the proxy): call-log retention follows the
  caller-data window; outbound calls wait for their pre-generated opening; the
  PHONE_AGENT_ROLES setting is gone; admin-only agents no longer run as admin.
• "Primary" is gone from AI Engines. The model decides the provider. System
  Default prefers a provider with a subscription, so a local-only install no
  longer defaults to a hosted model. Stored primary flags are cleared on upgrade.
• Claude models on Direct LLM search the web with the basic search tool again.
• Satellite 0.5.118 (fleet auto-update): unattended Codex sessions on paired
  machines run the same permission floor as on the server; lean skill cards.

Fixed

• Dictation no longer repeats words. The 1.5.0 fix for vanishing words could
  print the last words of a phrase twice. It now commits exactly what was heard.
• Codex on local models. Discover finds an Ollama server's models again. On
  Ollama 0.34 the MCP tools load on demand (0.34.0-rc1 cannot call them yet,
  reported upstream). A slow first prompt no longer dies at five minutes.
• Codex sessions start clean. The first turn waits for every MCP server. A
  stale ChatGPT login no longer shadows an API key or a local model. llama.cpp
  and LM Studio are told they drop MCP tools. Meeting tools no longer prompt.
• Direct LLM chats. A chat on an agent pinned to a paired machine no longer
  believes it runs there. Switching provider mid-chat works. Web search, fetch
  and code rows finish. A cut-off tool call no longer breaks the next request.
• Phone calls. No two-second dead air after each sentence with ElevenLabs
  voices (close_at_flush). The ambience plays through gaps. A call ends at the
  agent's end marker. A timed-out audio read no longer leaves the call deaf.
• The proxy no longer stalls on a slow database disk. Status, turn saves,
  chat actions and log writes run off the event loop. Tunneled MCP streams are
  no longer cut every 15 minutes. send_files sees same-turn remote files.
• Server sandbox. Bash commands that start with a variable assignment run
  again. The permission gate looks past do, then, if and ( … ) wrappers.
  The session scratchpad under /tmp is writable.
• Company map. Zooming out of a department lands on the same whole-map
  view that paging between departments uses, with no second zoom-out step.
• Meetings end each turn at the routing call. A participant that kept
  working after direct_to stalled the meeting, and a summary written after
  end_meeting was lost. Later tool calls are refused, the summary is kept.
• Smaller fixes. New models show in the pickers at once. No "Default Session
  Mode" on a Direct LLM agent. write_xlsx charts and conditional formatting
  come out as asked and survive later edits.

Security

• Sessions without a human user (tasks, triggers, meetings, phone) are tighter:
  permission state clears when a local session closes, agent configuration and
  warmup need the master key, the memory API refuses foreign-agent sessions,
  phone-minted tokens die at hangup, delegation refuses external principals.
• Uploads, context files, recover-bin restores and previews re-check their
  destination on the resolved path (a planted symlink cannot redirect a read
  or write); the file tools check every path with the platform, including
  container-absolute ones. Phone numbers no longer appear in logs.