🚀 CodeBuddy Code v2.178.0 Release

📦 Version Info

Component	Version
CodeBuddy Code CLI	v2.178.0
Agent SDK JS	v0.3.294
Agent SDK Python	v0.3.293

✨ New Features

Locate traces by session, with more readable content

Right-click a session in the session list and choose "View Traces" to open all traces for that session, including its sub-agents and traces in background task processes; you can also open it directly with #/traces?session=<session ID>. The list adds two filters — executor (main session, sub-agent, background task/teammate, auxiliary call) and tool — and each trace shows its owning session, tools used, and sub-agents that ran within it. Spans no longer display internal names; they now use readable titles such as "Tool name + call description", "Model call · model name", and "Sub-agent · Explore". The detail page header shows the executor and owning session, lets you copy the session ID, or filter to just this session with one click (#164202).

Larger channel member limit, adjustable in settings

Channels now hold up to 12 AI teammates by default (previously 6). In "Settings → AI Teammates → Channel max members" you can set any value from 2 to 20: lowering the limit does not remove existing members, it only stops new additions; when a channel exceeds the limit, the message states the current limit and where to change it. Larger channels get more speaking slots per discussion round so every member still gets a turn (#164672).

New /approve retry authorization

You can select the most recently auto-review-denied action in the current session and record a one-time retry authorization for that exact action; it does not execute the action or switch permission modes, and retries still go through auto-review. The terminal fully displays auto-review denial reasons, risk, authorization level, and review failure/timeout messages (#164700).

Agent Home: read-only channels and task executor registration

The primary agent can now read channel messages and task events without joining (paginated history), without disturbing channel members. When a task is handed off via DM, you can register the actual executor on the task board (including teammates outside the channel); board cards and the timeline show who is really doing the work. Reassignment, returning, or revoking a claim clears the registration automatically, and completion results are credited to the actual executor (#165389).

🔧 Improvements

• Customized builds carry customer identity end to end: the model introduces itself with the configured CLI name, /init creates the configured instruction file, the Web UI entry page and connecting copy use the configured product name, and skills hints point at the configured directories. Official builds are unchanged.
• Mods tool.check hook supports ask: the engine records which plugin initiated the question; such questions can only be answered by a human — no permission mode, remembered approval, or auto risk review will answer on your behalf. ACP clients and host callbacks only receive "Allow once" and "Deny".
• Harness model metadata passthrough: send customPassthroughHeaders and extraFieldsMeta via prompt _meta; payloads are captured per execution turn, and sub-tasks and deferred calls keep their own turn's snapshot.
• New server API sessionEventBacklog(): read-only count of undelivered event frames per session for host-side memory diagnostics.
• Clearer Web UI run status: the status line distinguishes "Model generating", "Tool running", and "Waiting for model", with elapsed time for the current step; the code execution card distinguishes "Model writing code" from "Executing code".
• Smoother long sessions: avatar animations stop when scrolled out of view, the companion pet no longer redraws the page every frame, and terminal title updates no longer re-render the whole interface.
• More reliable remote hosts and channel data: after connecting to a remote computer, the main session and all agents see consistent host state; Linux remote computers support desktop control and clipboard. Teammate messages are not lost before delivery is confirmed; corrupted task board or channel directory files are preserved with a backup and can be restored.
• Agent evidence verification: fabricated "messages from teammates" are clearly rejected; marking a channel task done requires verifiable on-disk evidence (an existing file or a parseable commit). The task board marks completed tasks as "Evidence verified" or "No evidence".

🐛 Bug Fixes

• Windows resource leaks: fixed conhost processes and native handles left behind after consecutive Bash runs.
• Network error attribution: after repeated network interruptions exhaust retries, errors are no longer misattributed to "model service issues".
• Web UI terminal: terminals no longer resolve project-local cbc/codebuddy commands when launched via package manager scripts; running cbc in the terminal no longer exits silently, and leftover configuration now explains itself.
• Thinking toggle: explicitly disabling thinking now correctly sends reasoning_effort: none.
• Teammates no longer miss messages after restart: teammates who switched sessions reattach on restart, and the inbox re-delivers to the session actually running.
• Desktop new-window notifications: fixed notifications not arriving in pages opened via "New Window"; the test notification honestly reports failure when the OS rejects it.
• Remote host key rotation: rebuilt hosts no longer get permanently blocked by "host key has changed" (auto re-confirm when your SSH config disables checking; strict hosts stay blocked).
• Tool card copy buttons: copy controls float at the result's top-right, labeled "Input"/"Output"; a command's command line and output merge into one card.
• Image message display: session titles and bubbles no longer leak internal image markers; "Return here" moved below the bubble; image-only messages show thumbnails and survive refresh; titles follow the UI language.
• Auto sandbox boundary: fixed inconsistency between sandbox write boundaries and permission checks in auto mode; sandbox blocks without records now produce a clear reviewed-retry prompt.
• No more unbounded memory growth: fixed crashes from push streams (file watching, terminal output, task lists) piling up while pages hang; connections that stop reading are disconnected proactively.
• Avatar card: selected cards now show a clear background; hover tooltips render image markers as "[Image]".
• DM auto-creates sessions: messaging a teammate who has never chatted auto-creates the session and delivers the message; failures retry automatically without losing the message.
• New sessions use the default model: after manually switching models once in the Web UI, new sessions no longer inherit that choice and always use the configured default model; switching in the input box affects only the current session.
• Channel message dedup: teammates who only work the task board no longer see the same old messages repeatedly; task completion notifications list Issue, MR, commit, evidence paths, and verification status.
• No jitter when switching agents: opening a running agent with long history shows its collapsed state directly.
• Quota exhaustion messages: used amounts, limits, and periods for personal, department, and enterprise quotas now display correctly.
• Plugin detail page: expanding some plugin details no longer breaks the whole page; a failing card retries on its own.
• Task template icons: recommended template cards show type-appropriate icons instead of a uniform clock.

📝 Documentation

• Synced the CLI documentation site and v2.178.0 release notes.