CCCC v0.4.15 Release Notes

v0.4.15 is a group portability, capability lifecycle, and browser-runtime reliability release.

The previous releases made ChatGPT Web Model setup clearer and continued stabilizing the website-hosted runtime path. This release broadens that work into the surrounding product system: copying full working groups instead of maintaining fragile templates, managing capabilities as a lifecycle instead of isolated buttons, and making projected browser surfaces less dependent on screenshot polling.

The main theme is operational maturity. CCCC now treats groups, capabilities, and browser-backed runtimes as durable systems with explicit state, safer copy boundaries, better UI feedback, and stronger failure semantics.

Copy Groups Replaces Blueprints

v0.4.15 introduces Copy Groups as the supported path for duplicating, migrating, and backing up working groups.

The old blueprint/template model copied a curated whitelist of fields. That approach became harder to maintain as CCCC groups gained ledgers, memory, attachments, assistants, automation, browser state, capability state, and richer settings. Copy Groups flips the model: export durable group state by default, then exclude unsafe or runtime-only material.

Copy packages include the state users actually expect when moving a group:

• ledger history,
• actor definitions,
• memory and context state,
• attachments and blobs,
• assistant configuration,
• automation configuration, and
• durable group settings.

They intentionally do not include workspace repository files, live credentials, browser sessions, provider auth state, transient runtime state, caches, locks, or generated indexes. Imported groups start idle with actors stopped, so copying a group does not immediately execute work in the destination environment.

The import path also validates archives defensively: traversal paths, unsafe names, symlinks, duplicate entries, oversized archives, zip-bomb patterns, unsupported schema versions, secret-containing packages, and workspace-including packages are rejected before import.

Capability Center Becomes the Capability Control Plane

v0.4.15 turns capability management into a more complete lifecycle.

The new Capability Center workflow covers:

• discovering skills, MCP toolpacks, and capability packs,
• importing from supported sources,
• enabling capabilities for group usage,
• hiding capabilities from slash surfaces without disabling them,
• removing installed or imported sources,
• inspecting source instances, policy, readiness, and usage state, and
• keeping capability state visible in one Web workspace.

This is a practical shift from ad hoc capability actions to a registry-backed control plane. Installed capabilities are now easier to audit, easier to remove, and less likely to drift between "installed somewhere" and "available to the group".

Capability overview requests were also made lighter. UI surfaces that only need items can skip source-instance aggregation, and Capability Center receives kind counts from a single overview response instead of issuing multiple stats-only requests. This reduces unnecessary work as capability catalogs grow.

Slash Commands Route Through CCCC

v0.4.15 adds a stronger slash-command path for capability workflows.

Most importantly, /install is routed through the CCCC capability registry instead of leaving the model to decide whether to install a local Codex skill package, a CCCC skill, an MCP toolpack, or something else. The default install path now tells the agent to use cccc_capability_install with group scope, so installed skills appear in the group's slash command surface and remain manageable through Capability Center.

Chat also gained capability-backed slash command discovery and dispatch. This makes frequently used capability actions easier to invoke while preserving CCCC's central lifecycle and policy checks.

Projected Browser Viewing Moves to VNC When Safe

v0.4.15 replaces the most fragile part of projected browser viewing on supported Linux/Xvfb sessions.

Instead of relying primarily on repeated screenshot capture or CDP screencast frames, CCCC can now expose CCCC-owned Xvfb browser sessions through a localhost x11vnc server and a noVNC Web viewer. Browser automation and delivery still use the daemon-owned CDP/Playwright session; VNC is only the visual and interactive viewer transport.

This separation matters. The embedded viewer no longer has to compete with the same CDP path used for browser automation and ChatGPT message delivery. On Linux/container setups with a CCCC-owned Xvfb display and x11vnc available, the viewer should feel closer to a remote desktop surface. When VNC is disabled, unsupported, or unsafe, CCCC falls back to the built-in CDP screencast viewer.

For safety, CCCC does not attach VNC to an inherited host desktop display. VNC is only enabled when the runtime knows the display is dedicated to the projected browser session. This avoids exposing unrelated local windows through the embedded browser panel.

ChatGPT Web Model Delivery Is More Honest

This release continues the ChatGPT Web Model hardening line.

Delivery now distinguishes stronger submitted evidence from weaker browser symptoms. A visible user-message echo is treated as strong submitted evidence. Weaker signals, such as "ChatGPT appears to be running" or "the composer was cleared", are classified as ambiguous rather than definitive success.

That changes the failure model in an important way:

• confirmed delivery is marked as submitted,
• unconfirmed but plausible delivery is marked as ambiguous,
• true failures remain failed, and
• failed or ambiguous batches do not automatically get bundled into later messages.

This is intentionally conservative. Repeating an old failed batch into a later ChatGPT message can corrupt the conversation more severely than asking the user to inspect or retry explicitly. The Web UI now has better delivery-state projection so users can see when CCCC is submitting, submitted, pending, ambiguous, or failed instead of guessing from a silent browser automation attempt.

The default browser delivery path also gained more stable send-control waiting, better verification-budget handling, and clearer runtime state transitions.

Web UI Reliability and Polish

v0.4.15 includes several Web UI improvements that support the larger runtime work.

Markdown code blocks now render through CCCC's own code block UI without redundant nested <pre><code> containers, improving density and readability for long technical replies. The persistent composer and group store were tightened so selected group state and composer ownership stay aligned earlier in app startup. Settings initialization and Voice Secretary composer hooks were adjusted to remove stale hook warnings and keep lint clean.

Runtime control and browser readiness paths were also made less brittle for slower machines. The Web child readiness timeout is now less likely to fail on low-end hardware, while browser status and viewer state provide clearer fallback information.

Capability and Event Consistency Fixes

Several smaller fixes close consistency gaps around capability state:

• capability reinstall now clears hidden slash visibility when the capability becomes usable again,
• GitHub source-instance grouping keeps root-level skills from different refs distinct,
• source deletion can target a source instance instead of only broad source IDs,
• capability overview can return kind counts without extra UI requests, and
• failed capability.changed event publication no longer makes a persisted capability mutation look like a failed API operation.

That last item is important for operational correctness: a notification failure should not make a successfully saved enable/visibility change appear to have failed.

Why Upgrade

Upgrade to v0.4.15 if you want to:

• duplicate, migrate, or back up full CCCC groups with Copy Groups instead of legacy blueprints,
• manage capability discovery, installation, visibility, source instances, and removal from a more complete Capability Center,
• use /install through the CCCC capability lifecycle rather than an ambiguous model-chosen install path,
• get a smoother embedded browser viewer on Linux/Xvfb deployments through VNC/noVNC,
• reduce false ChatGPT Web Model delivery certainty and avoid automatic re-sending of old failed batches,
• improve Web UI readability for code-heavy conversations, and
• pick up stronger capability-state consistency and CI coverage.

In short, v0.4.15 is the release where CCCC replaces fragile group templates with durable group copying, turns capability management into a real lifecycle, and makes browser-backed Web Model operation more observable, safer, and easier to recover.