aix CLI 2.26.0

• Shared sessions show every subagent (/share; Aix Code v0.24.0: the Share button). The
  viewer used to collapse the helpers into one line ("3 agents working") behind the phone's task
  bar. Now:
  • Agent chips under the header on phones (the side list on wider screens): each agent with
    what it's doing now and its own task progress.
  • Agent cards in the conversation: the call that started a helper opens it; calls that wait
    on one or write to one link to it too.
  • An agent page for each helper, live: the task it was given, its own task list, the
    messages it sent and got, and its activity — reasoning, replies, tool calls with their output
    on tap — plus model, tool calls, tokens, context and cost. Switch between agents from the page;
    the phone's back button closes it. Finished agents stay openable, with how they ended.
• Light on the network: an agent page is fetched only while it's open (about once a second; an
  unchanged agent answers with an empty reply), never pushed to every viewer.
• Cost and cache-savings estimates are unchanged; nothing about what the agents do changes.

aix CLI 2.25.4

• Fix Gemini API-key chats failing on tool calls with missing thought_signature. Google's
  Gemini connection now uses the native Gemini transport, preserving signed tool calls and
  reasoning state across requests and saved sessions. Existing models/gemini-* selections,
  API keys and older unsigned/mixed-model history continue to work without resetting the chat.
• Report content_filter: OTHER as a provider refusal with its reason, not an empty response.
  Native Google block/finish explanations are retained in errors and diagnostics. This does not
  disable Google's safety controls or guarantee that a provider-blocked request will be accepted.
• Shared with Aix Code v0.23.4; restart after updating to apply the transport fix to existing chats.

Aix Code 0.24.0

• Shared sessions show every subagent (aix v2.26.0). On the phone, the shared view used to fold
  the helpers into one "3 agents working" line. Now each agent has a chip under the header (the
  side list on wider screens) with what it's doing and its task progress; the calls that start or
  message an agent show as cards that open it; and every agent has its own live page — the task it
  was given, its task list, the messages it sent and got, and its activity: reasoning, replies and
  tool calls with their output on tap, plus model, tokens, context and cost. Switch between agents
  from the page; the phone's back button closes it. Finished agents stay openable.
• An agent page loads only while it's open, so sharing stays light for everyone else watching.

aix CLI 2.25.3

• Restore GLM 5.3 in Z.ai's fallback and incomplete catalogs, with 1M context and low/high/max
  effort. Shared with Aix Code v0.23.3.
• Add Gemini 4 Carbon (gemini-4-carbon), including OpenCode Zen's native Gemini endpoint,
  reasoning, tool calls and thought-signature replay. Zen requires a paid API key for Carbon.
• Antigravity model listing no longer waits for project onboarding, retries Google's alternate
  endpoint, and falls back before the picker's timeout. Live context/output limits are honored;
  disabled/internal entries are hidden. Carbon appears there only when Google enables it for
  the connected account; gemini-4-carbon-disabled is not a usable model.
• Aix Code's provider brand icons are desktop-only; terminal rendering is unchanged.

Aix Code 0.23.4

• Fix Gemini API-key chats failing on tool calls with missing thought_signature. Google's
  Gemini connection now uses the native Gemini transport, preserving signed tool calls and
  reasoning state across requests and saved sessions. Existing models/gemini-* selections,
  API keys and older unsigned/mixed-model history continue to work without resetting the chat.
• Report content_filter: OTHER as a provider refusal with its reason, not an empty response.
  Native Google block/finish explanations are retained in errors and diagnostics. This does not
  disable Google's safety controls or guarantee that a provider-blocked request will be accepted.
• Shared with aix v2.25.4; restart after updating to apply the transport fix to existing chats.

aix CLI 2.25.2

• Fix Claude subscription requests rejected because the advertised Claude Code version was
  2.1.280. Both the user-agent and billing header now advertise at least the current published
  version, 2.1.321, even when an older Claude Code is installed locally. Newer installed versions
  remain honored. Shared with Aix Code v0.23.2; API-key requests are unchanged.

aix CLI 2.25.1

• Claude Fable 5.2 (claude-fable-5-2) is available in the shared model catalog and Anthropic's
  offline fallback list, including Aix Code v0.23.1. Select it with /model claude-fable-5-2.
  Supports a 1M context window, 128K output, images, low through max effort, and Fast mode.
• Requests keep adaptive thinking on even when the thinking toggle is off, with thinking-binding
  controls so edited history can continue. Standard pricing is $8/$40 per million input/output
  tokens, cached reads are $0.20, and Fast pricing is $16/$80. Cost and cache-savings estimates use
  these rates; older Fable models keep their existing rates. Existing model defaults are unchanged.
• Anthropic requires 30-day data retention for Fable 5.2 unless it expressly authorizes otherwise.
  Model details: https://platform.claude.com/docs/en/models/fable-5-2/overview

Aix Code 0.23.3

• Replace provider monograms with bundled brand icons in the model pill, picker, settings and
  usage views. Icons adapt to light/dark themes; unknown custom providers keep a monogram.
• Restore GLM 5.3 in Z.ai's fallback and incomplete catalogs, with 1M context and low/high/max
  effort in the shared aix v2.25.3 core.
• Add Gemini 4 Carbon (gemini-4-carbon), including OpenCode Zen's native Gemini endpoint,
  reasoning, tool calls and thought-signature replay. Zen requires a paid API key for Carbon.
• Antigravity model listing no longer waits for project onboarding, retries Google's alternate
  endpoint, and falls back before the picker's timeout. Live context/output limits are honored;
  disabled/internal entries are hidden. Carbon appears there only when Google enables it for
  the connected account; gemini-4-carbon-disabled is not a usable model.

Aix Code 0.23.2

• Fix Claude subscription requests rejected with “Claude Code 2.1.280 does not support this
  model.” The shared aix v2.25.2 core advertises at least Claude Code 2.1.321 in both the
  user-agent and billing header, without letting an older local installation lower that version.
  Newer installed versions remain honored. API-key requests are unchanged.

Aix Code 0.23.1

• Claude Fable 5.2 (claude-fable-5-2, aix v2.25.1) is available in the model picker and with
  /model claude-fable-5-2: 1M context, 128K output, images, low through max effort, and Fast mode.
  Adaptive thinking stays on as required by Anthropic, with thinking-binding controls for edited
  history. Pricing and cache savings use $8/$40 per million input/output tokens, $0.20 cached
  reads, and $16/$80 in Fast mode. Existing model defaults are unchanged.
• Anthropic requires 30-day data retention for this model unless it expressly authorizes otherwise.
  Model details: https://platform.claude.com/docs/en/models/fable-5-2/overview