Skip to content Apps, Subagents, GPT-6, and Claude Opus 5.5
Features
- Apps — Build and run custom workspace Apps inside Consoles. Apps can create focused tools, dashboards, and workflows that connect to your Sources without exposing raw credentials. Apps get a local runtime manager, a popout window, and hooks in the right panel and composer, and the agent can create and edit them from chat.
- Overview App — A built-in App gives you a high-level view of your workspace activity, sessions, and usage.
- Subagents — Delegate work to specialized agents that can run alongside your main conversation and report progress back into the session.
- Orchestrator mode — Coordinate larger tasks across agents with clearer progress and handoffs.
- iOS beta — Early iOS support is available for testing the mobile Consoles experience.
- GPT-6 family on ChatGPT/Codex — GPT-6 Astra, GPT-6 Sol, and GPT-6 Luna join the ChatGPT/Codex model picker as “6 Astra”, “6 Sol”, and “6 Luna”, ahead of the GPT-5.6 family and GPT-5.5. All three run on the Responses Lite path with Code Mode and keep native
max effort instead of saturating at xhigh. Zero-config Codex subagents now use GPT-6 Sol @ max for execution and GPT-6 Luna @ medium for Explore/reading, with GPT-6 Astra @ max directing and merging. - Claude Opus 5.5 and Claude Fable 5.1 — Claude Opus 5.5 is available on direct Anthropic connections and leads the model list, with its always-on adaptive thinking handled automatically; connections that follow the catalog pick it up without changing the selected default model, and curated model lists are left untouched. Claude Fable 5.1 is selectable on Anthropic and Bedrock (us/eu/global) connections, and the Pi Anthropic lane sends Anthropic’s thinking-binding controls so mismatched replayed thinking blocks are dropped instead of failing the request. (#23)
- Claude Sonnet 5 metadata and routing — Sonnet 5 is normalized in the model picker as “Sonnet 5” with the updated speed/intelligence description, becomes the default Sonnet alias for subagents, keeps Sonnet 4.6 selectable as the previous generation, and has Bedrock US/EU/Global routing wired without changing the existing Opus 4.5–4.8 options.
- Muse (Meta) provider — Meta’s Muse Spark models (1.3, 1.2, and 1.1, including Contributor variants) are available as a first-class provider. Sign in with a Muse membership or an API key during onboarding or in Settings → AI; Muse gets its own icon and branding in the picker and chat, and runs through a native Meta Responses provider on the Pi lane.
- Goals keep their plan — Goal mode snapshots the latest to-do plan onto the goal and carries it into the next continuation, surviving turn boundaries and compaction. A new resumable “usage limited” state covers turns that end on a provider rate or billing limit — it never auto-continues and resumes from the goal chip — and the goal popover shows the latest steps and stop reason in one place. (#22)
- In-session subagents on ChatGPT/Codex — Subagent fan-out on the Pi/Codex lane now matches the Claude lane:
task subagents stay inside the parent session as hidden side conversations instead of appearing as separate sessions, and the session list no longer shows placeholder entries leaked by streaming races. - Codex-style right panel — The right panel now has retained workspace Files tabs, rich code/Markdown/image/PDF previews, open-outside handling for unsupported media and Office files, expandable panel width, and hidden ephemeral Side chats that inherit the parent conversation without disrupting it.
Improvements
- Safer Source access — Apps request “Use this Source” permissions and call through the Source layer instead of receiving raw API keys or OAuth credentials.
- App runtime isolation — Local Apps run in a sandboxed environment with explicit environment variables only.
- Claude Agent SDK uplifted to 0.3.280 — The bundled Claude SDK now matches Claude Code v2.1.280 behavior (up from 0.3.197 earlier in this cycle), including native Sonnet 5 awareness for aliases, context-window handling, effort defaults, first-class Claude Fable 5.1 support, and recognition of Claude Opus 5.5, which older bundles rejected as an invalid request.
- Native Codex auto-compaction and batching — ChatGPT/Codex sessions now trigger auto-compaction exactly where native Codex does — 244,800 tokens against the 258,400-token usable window — instead of 12,240 tokens early on live-discovered models, and the context meter marker matches. Responses Lite turns are guided to batch independent reads into one Code Mode
exec call with Promise.allSettled. - True usage ledger for Dashboard — Console now writes exact usage rows to each session’s
usage.jsonl for main turns plus instrumented internal model calls like call_llm, title generation, and summaries, and Dashboard totals are ledger-only so provider/model/connection switches are tracked without legacy header or SDK-log inference. - Review parity and reliability — Review now uses the active paired Craft code theme, supports Color or +/− diff markers, keeps hunk actions stable after context expansion, preserves staged files in Last turn, persists per-file rich-preview fallback, and rolls failed Git actions back instead of leaving optimistic UI behind.
- Live Review status — Review now refreshes from a server push whenever git status changes, with a 15-second fallback poll.
- Smoother streaming under load — Bursts of thinking and task-progress events are coalesced before they reach React, and the model picker is memoized, so long reasoning turns and large fan-outs keep the UI responsive.
- Shared command icons — Notebook command rows reuse the desktop tool-icon registry and command parser. Blender, Unreal Engine, uv, and FFmpeg join Python, npm, Bun, GitHub, and Cloudflare in the shared defaults.
Bug Fixes
- Blocked tools no longer end the turn — When Plan/safe mode blocks or denies a tool call, the model now receives the reason and continues, instead of the turn ending on a generic “No response” card that hid why.
- Background sessions no longer park forever — Three stale liveness flags could leave a session waiting on a background wake indefinitely (seen at 100+ minutes, recoverable only by restarting the app); recovery paths now check them against reality instead of trusting them.
- No phantom turns after long commands — Long-running foreground tool calls (like a slow Bash) no longer queue a background update when they finish, which used to trigger an extra continuation turn after the real turn ended.
- Idle CPU usage fixed — A re-render loop in the chat composer (for sessions without labels) kept a CPU core near 100% even with nothing running; it’s fixed, along with a Radix ScrollArea animation-frame leak.
- MCP sign-in no longer routes through upstream’s relay — Signing in to an MCP source (like Cloudflare), or to an API source that registers its OAuth client automatically, now redirects straight to the app’s local callback instead of bouncing through Craft’s
agents.craft.do relay. The provider’s consent screen shows localhost, and the authorization code never leaves your machine. Built-in Google, Slack, and Microsoft sign-in is unchanged. - Pi deferred session tools are discoverable again — ChatGPT/Codex sessions can now load every registered cold session control through
tool_search, including native MCP OAuth and credential prompts, instead of registering those tools in the subprocess while leaving them invisible to the model. - Pi SDK manual compaction usage reset — Manual
/compact in Pi SDK sessions now clears stale context/cache counters, carries the native compaction summary metadata through the event pipeline, and waits for the next provider-reported response to show exact post-compaction usage so the context badge no longer jumps back to the pre-compaction percentage when the turn completes or after reload. - Markdown code-fence paste handling — Long text pasted inside an explicit Markdown code fence now stays inline instead of being converted into a
pasted-text attachment, and empty code fences no longer render as undefined in chat.
Breaking Changes
- Apple Silicon only on macOS — macOS releases are Apple Silicon only, and downloads now focus on the supported Mac and Windows builds.