CCCC v0.4.13 Release Notes

v0.4.13 is the ChatGPT Web Model and local voice release.

The headline change is simple: CCCC can now bring a GPT-5.x ChatGPT web session into a working group as a real actor, backed by the same CCCC messaging, MCP, workspace, and local execution model used by other runtimes. This release also makes Voice Secretary substantially more local-first through built-in ASR, diarization, model management, and cleaner transcript routing.

Compared with v0.4.12, this is less about adding another panel and more about turning two experimental surfaces into usable product workflows:

1. ChatGPT Web Model becomes a practical CCCC runtime for GPT-5.x.
2. CCCC MCP grows into a local-development tool surface that ChatGPT can actually use.
3. Voice Secretary gains local speech recognition and a cleaner document/ask/prompt split.
4. The browser, Docker, Node, CI, and setup paths are hardened enough for wider testing.

ChatGPT Web Model

v0.4.13 promotes ChatGPT Web Model from an integration experiment into a first-class runtime path.

A user can expose CCCC through an HTTPS tunnel, create a ChatGPT Web Model actor, copy one MCP URL into ChatGPT, and let a GPT-5.x ChatGPT web conversation participate in the group. The web model can receive CCCC work, call CCCC MCP tools, read and edit the local workspace, run validation, send visible replies through CCCC, and hand work back to peers.

The setup model was intentionally simplified around one ChatGPT MCP app and one CCCC ChatGPT Web Model actor. That keeps the mental model closer to how ChatGPT custom MCP apps actually behave: one remote connector, one bound CCCC actor, one target ChatGPT conversation at a time.

Important product changes:

• ChatGPT Web Model is shown as its own runtime in actor creation, settings, and runtime details.
• The Web Model setup flow now centers on the few things users need to understand: public HTTPS access, the MCP URL, ChatGPT sign-in, and the target chat.
• Settings, runtime inspection, browser delivery, and tool-confirm handling all use the same daemon-managed ChatGPT browser profile.
• Changing or creating a target ChatGPT conversation reseeds the actor so the web model understands its CCCC identity and tool contract.

Local Development Through ChatGPT

The main reason to use ChatGPT Web Model is not just to chat with another model. It is to let a GPT-5.x ChatGPT session do useful local work inside CCCC.

v0.4.13 adds the missing local-power layer: scoped repository reads and edits, Codex-style patching, shell and git wrappers, file attachment transfer, and a higher-throughput cccc_code_exec loop for multi-step engineering work.

For day-to-day development, this means ChatGPT can now handle workflows such as:

• inspect the current diff
• read relevant files with hashes
• apply exact edits or Codex-style patches
• run focused tests or build checks
• send generated files back as CCCC attachments
• reply with evidence through the normal CCCC message channel
• complete the browser-delivered turn when finished

cccc_code_exec is the most important ergonomics upgrade in this area. It gives the web model a controlled JavaScript loop that can call nested CCCC MCP tools, so review/patch/test/report work does not require dozens of separate ChatGPT tool turns. The code mode remains scoped: it does not expose Node host APIs, and real work still goes through CCCC MCP tools.

The help surface was also tightened. Web Model actors now get compact common work loops, tool-name lookup, focused tool_help(...), attachment guidance, and clearer recommended actions for common errors such as stale file hashes, missing reply targets, unsupported code-mode APIs, and truncated output.

Browser Delivery Reliability

The hardest part of ChatGPT Web Model is that the runtime lives inside a website. v0.4.13 therefore treats browser delivery as an operational system, not a best-effort click script.

The release adds stronger evidence and diagnostics around delivery:

• CCCC now records whether the last browser delivery was submitted, pending a new-chat bind, or failed.
• Prompt submission requires stronger evidence than "the composer changed".
• Pending new-chat delivery no longer resends the first message while waiting for ChatGPT to expose the final /c/... URL.
• The runtime and settings panels show a derived health snapshot with browser state, target-chat state, delivery state, and the next recommended action.
• ChatGPT tool-confirm prompts can be auto-approved so model work is less likely to stall behind repetitive UI confirmation cards.

This does not pretend ChatGPT's website is as deterministic as a local CLI. It does make failures visible and recoverable instead of silent.

Voice Secretary Local ASR

Voice Secretary also takes a major step forward in this release.

v0.4.13 adds local ASR infrastructure around Sherpa ONNX, final transcript processing, speaker diarization, model manifests, and model/runtime cache controls. The Web settings surface now makes the local runtime, ASR model, and diarization model explicit rather than hiding them behind a single vague "recognition" switch.

The Voice Secretary workflow was also cleaned up after the local ASR expansion:

• Document, Ask, and Prompt modes are routed through their own output channels.
• Document transcripts are kept with the target document instead of mixing with Ask or Prompt activity.
• Ask replies are returned as visible assistant results.
• Prompt refinement submits composer drafts instead of producing chat noise.
• Browser ASR and local ASR settings now make clearer distinctions where browser APIs cannot support the same capabilities.

The net effect is that Voice Secretary is closer to a local working assistant for notes, documents, and quick voice commands, rather than just a streaming transcript display.

Operational Polish

This release also includes a set of important reliability and packaging improvements that support the larger runtime surface.

Notable improvements:

• Docker image builds now install Node.js in a way that includes npm.
• Managed Node deprecation warnings no longer spam the service logs.
• CI and pre-commit checks are bounded and faster on impacted paths.
• The default voice model manifest is packaged with the Python distribution.
• The release docs and README surfaces now describe ChatGPT Web Model and the GPT-5.x / GPT-5.x Pro boundary more clearly.

GPT-5.x Pro Boundary

v0.4.13 deliberately distinguishes regular GPT-5.x ChatGPT web sessions from GPT-5.x Pro sessions.

Regular GPT-5.x ChatGPT sessions can use the CCCC MCP connector and therefore can participate as local-development actors. GPT-5.x Pro remains useful for architecture review, strategy, analysis, and high-level feedback inside ChatGPT, but current ChatGPT Pro behavior does not provide the same third-party MCP local-access surface. CCCC therefore does not market GPT-5.x Pro as a full local runtime in this release.

Why Upgrade

Upgrade to v0.4.13 if you want to:

• use GPT-5.x ChatGPT as a CCCC actor with local development tools
• run ChatGPT-assisted repo review, patching, testing, and file handoff through CCCC MCP
• use Voice Secretary with local ASR and diarization instead of only browser recognition
• get clearer Web Model setup, delivery status, and recovery guidance
• avoid Docker/Node and browser-delivery issues fixed since v0.4.12

In short, v0.4.13 is the release where CCCC starts treating website-hosted models and local voice input as serious runtime surfaces, not side experiments.