This release makes Conduct faster to start, more resilient during long-running work, and clearer when a turn is interrupted. It also adds support for more models and deeper grounding diagnostics for administrators.

Highlights

• ⚡️ Faster first responses: Conduct now prepares the sandbox and assistant when a chat opens or typing begins. Parallelized skill, memory, history, and client initialization removes several seconds from many cold starts.

• 🚡 More reliable long-running work: Extended reasoning, large tool results, context compaction, and brief gateway interruptions are less likely to stop a turn prematurely.

• 🩹 Clearer recovery: When a turn is stopped or interrupted, Conduct now explains what happened and identifies any answers, tool results, or files that were preserved.

• 🔊 More supported models: Conduct adds support for multiple new models per harness, see Expanded Model Catalog.

• 🦮 Personalized and transparent assistance: Context memory carries durable preferences across chats, while live plans, reasoning, and interactive questions make ongoing work easier to follow and steer.

New capabilities

Expanded model catalog

The following entries describe newly supported models. They appear as selectable options only in tenants where the corresponding deployment and configuration are present¹.

• Claude Code — strict enum of 9 Claude models (Haiku 4.5, Sonnet 4.5/4.6/5, Opus 4.5–4.8 and 5), most selectable via Anthropic-direct (anthropic-…) or Vertex AI (vertex-claude-…); default is Claude Opus 5 with Opus 4.8 as fallback.

• Codex — strict enum of Azure GPT deployments from GPT-4o and GPT-4.1 through the GPT-5 family up to GPT-5.6 Sol/Terra/Luna (default: 5.6 Sol).

• Pi — Pi-Agent is rated private-experimental and thus not further outlined.

Context memory

Conduct can retain durable preferences and personal context from conversations and apply them in future chats. The interface shows when memory is being updated, and saved context can be reviewed in Settings.

Note that there is limitations when combining Context Memory with Web Search.

Live plans and reasoning

Conduct now displays live task checklists that update as work progresses. Collapsible reasoning blocks provide additional visibility across Claude and Codex.

Grounding diagnostics

Environments with sandbox debug capture enabled now include a Grounding check tab in the message Debug panel which contains:

• the hallucination judge's exact input and verdict;

• the complete rendered LLM request, including base64 image data URLs;

• token usage;

• chat, assistant, and message identifiers; and

• the capture timestamp.

Downloaded records can be used to reproduce judge calls and build benchmark datasets.

Grounding checks now validate answers against MCP connector results even when the final response does not cite those results explicitly.

Performance improvements

Faster chat startup

• Sandbox preparation now begins when the chat opens and again on the first keystroke, improving response time for short prompts.

• The first message can adopt a prewarmed workspace and agent client created before the chat ID exists.

• Assistant clients now start as soon as the workspace, memory, and locally cached skills are ready. Conversation history and skill refreshes continue in parallel.

• Redundant background prewarms are skipped when a chat already has a warm agent.

• The pinned agent CLI no longer runs a redundant SDK version check on each connection, reducing cold client spawn time by approximately 0.3–2 seconds.

Faster skills and memory loading

• Knowledge Base skill files and skill layers are discovered and downloaded concurrently.

• Folder listings are batched, reducing node-chat API calls during a cold skill build by approximately 4×.

• Pods reuse shared discovery results, locally cached bundles, and per-layer bundles stored in sbx-storage instead of repeatedly traversing and downloading unchanged Knowledge Base content.

• Bundled skills are prepared while pods wait in the warm pool, allowing the first turn to link the cache instead of copying the complete tree.

• An expired pod-local skills cache no longer blocks a turn. Conduct starts with the last known skill tree while revalidation runs in the background.

• Personal skills download concurrently with shared skill layers.

• User memory uses the turn manifest and a pod-local read-through/write-through cache. Cold turns require at most one memory API call; warm turns require none.

Faster tools and connectors

• Connectors and connected-space tool servers now run in the sandbox's long-lived host process instead of being imported in a new Python process for each client spawn. This removes approximately 4–6 seconds from cold starts.

• If the in-process host is unavailable, the sandbox falls back automatically to the previous per-spawn process.

• Codex now uses a stable prompt-cache key across turns, reducing repeated processing of workspace instructions and lowering uncached input-token costs.

Faster response controls

• Stop requests now take effect within two seconds.

• New turns are accepted while the sandbox refreshes its skills cache instead of waiting for the download to finish.

Experience improvements

• Continuous reasoning can run for up to 15 minutes without visible output before the no-progress safeguard stops it. Turns with no reasoning or output still stop after five minutes.

• Responses now show clear notices when turn or cost limits are reached.

• Interactive questions now appear while work is running, retain their complete prompt, and allow Conduct to continue correctly after confirmation or denial.

• If work completes without a closing reply, the model gets one more opportunity to respond before Conduct displays a specific notice.

• Interrupted turns now explain terminal agent errors and list files saved before the interruption.

• Partial answers display an interruption warning directly below the response instead of falsely appearing to finish cleanly.

• Unresolved tool calls from a stopped, truncated, or stalled turn now say that the turn ended before a result arrived instead of incorrectly marking the call as failed.

• Resumed work continues from the interruption point instead of restarting and repeating completed steps.

• Questions raised through the Ask tool remain as one timeline step while Conduct waits for a response.

• Run statistics and debug information for stopped or interrupted turns remain attached to the visible answer.

• Uploads and file conversions now provide clearer progress feedback, and open file previews refresh after an edit.

• When Conduct rewrites a file you uploaded in the chat, the updated file is delivered as a downloadable card under the original filename. The original upload is left unchanged.

• Generated documents now include fonts required to render Chinese, Japanese, and Korean text correctly.

• Document templates are now loaded from a dedicated Knowledge Base folder.

Reliability fixes

Turns and streaming

• Stopping a response now reliably closes the active turn so a follow-up message can begin.

• Long silent periods during reasoning or large tool results no longer trigger a premature “connection was interrupted” error.

• A final render-flush error no longer discards an answer that has already streamed.

• Turns no longer report an unrecoverable error when the underlying work completed successfully.

• Stalled turns now identify completed tool calls and saved files so users know what can be reused.

• Repeated narration across tool calls is now detected by the no-progress safeguard instead of allowing the turn to end silently.

• Short or repeated correct answers are no longer mistaken for duplicate output and blocked.

• Skill invocations from the chat input no longer intermittently return a generic empty-response fallback.

Long conversations and context

• Large conversations, including threads with rendered charts, PDFs, or images, no longer exceed stricter request limits on the openai-proxy, anthropic-proxy, and cursor-proxy routes.

• Chats affected by an automatic context-compaction failure now return a retryable error instead of hanging indefinitely.

• Long-running reasoning survives context compaction in progress. If a turn still ends without answer text, run statistics and the Debug and Report Issue actions remain available.

• Claude models with a one-million-token context window, including Opus and Sonnet 4.6 or later, are now detected correctly and no longer compact repeatedly as if limited to 200,000 tokens.

Skills, tools, and files

• Skills removed from the Knowledge Base are now removed from cached chat workspaces.

• A transient refresh error no longer causes cached skills to disappear from a conversation.

• Codex chats now use mounted connectors and retain large connector results so answers can include their sources.

• Web search and other agent tools now work correctly in chats served by a pre-started sandbox client.

• Large document downloads no longer exhaust sandbox memory and interrupt the turn.

• Reused agent clients, connectors, confirmations, citations, and generated files now use the current turn's identity instead of retaining or falling back to an earlier message.

Documents and presentations

• Conduct now inspects charts and figures visually when answering document questions instead of relying only on extracted text.

• Presentation quality checks now detect additional layout problems, including title overflow, detached chart labels, and mid-word wrapping.

Sandbox startup and connectivity

• A message sent while prewarm is still starting now preempts the prewarm instead of failing with “Your previous message is still being processed.”

• Agent preboot now completes in seconds instead of occasionally blocking the first message for up to a minute.

• The internal LLM proxy retries brief gateway connection failures with backoff, including interruptions caused by a gateway pod restart.

Security and privacy

• Chat-owned files, transcripts, and workspace checkpoints now verify chat access before retrieval or restoration, preventing another same-company user from accessing them by supplying a different chat ID.

• Conduct sandboxes can access only security-approved public node-chat endpoints through the sandbox gateway. Authentication, user and company identity injection, header filtering, and restrictions on non-public APIs remain enforced.

Administrator and operator notes

Cost management

• LLM usage accounting now applies provider-specific cache rates and normalizes OpenAI and Anthropic token semantics.

• Conduct usage is attributed to the originating chat and assistant, improving spend-limit enforcement and usage reporting.

• Historical usage records without cache-token splits remain indicative rather than directly comparable with newly calculated costs.

Skill assets larger than 10 MB

Conduct no longer copies individual skill files larger than 10 MB into the agent workspace. It creates a small pointer containing the Knowledge Base content ID, and the agent retrieves the asset when needed.

Skills that depend on large bundled assets must fetch those assets from the Knowledge Base at runtime. Configure the threshold per cluster with CONDUCT_SKILL_ASSET_MAX_MB; set it to 0 to disable the limit.

Claude workspace rules

Claude Code now receives platform workspace rules in the system prompt rather than through the full CLAUDE.md workspace marker file. This places approximately 14,000 tokens in the chat-invariant Anthropic prompt-cache prefix. The marker file is now a small stub. Codex, Cursor, and Pi continue to use file-based rule delivery.

Observability

• Slow workspace boot alerts now fire after two minutes of elevated p90 boot times.

• Boot duration is recorded when warmup completes instead of at the end of the turn.

• Claude Code errors now record SDK status code, stop reason, and error count in sandbox logs.

• Conduct now emits structured sandbox logs, distributed traces across the sandbox boundary, and turn-level metrics for token usage, cost, latency, retries, and interactive questions.

• Codex run statistics now include turn count, tool time, LLM wait time, and max-turn termination.

• An auditable QA-only feature flag can enable verbose payload logging for debugging. It is disabled by default and cannot be enabled in UAT or production.

Infrastructure and capacity

• Sandbox services now support ARM-based nodes.

• Operators can configure off-hours warm-pool scale-to-zero schedules. Active sessions remain running; sessions started outside warm-pool hours may incur a cold start.

Action required

• Pi model migration: Kimi K2.6 is no longer supported by the Pi agent. Select Kimi K3 where it is deployed and enabled for the tenant; otherwise, configure another supported replacement.

• Large skill assets: Review skills containing files larger than 10 MB and confirm they retrieve those assets from the Knowledge Base at runtime, or adjust CONDUCT_SKILL_ASSET_MAX_MB for the cluster.

¹ Model availability varies by tenant and environment. A model mentioned in these notes is available only when it has been deployed, configured, and enabled for the tenant. Platform support does not make a model automatically available to every customer.

Known limitations

• If you press Stop while the agent is still starting the turn — for example while it shows “Setting up your workspace…” or initial “Thinking…” — or if the agent stops responding, the chat can stay blocked and later messages in that same chat may not start. Stop already works once the agent is running. A follow-up patch to this release will cancel those stuck starts automatically and unblock the chat. Until then, start a new chat if one stays (b-)locked.