Code Command Center went deep this cycle — a cloud backbone and billing layer, video generation that ships into the pipeline canvas, a backlog that researches and tracks its own stories, voice control on smart glasses, and guardrails that watch every session so you don't have to.
From a cloud identity layer that spans orgs to a backlog that tracks its own stories, every piece landed behind test, wired into the ledger, and ready for production.
Three increments, one week. Start with Linear-style stories that auto-research, track session history, and roll up costs. Add purpose-sessions, Ralph and BigRalph orchestration, and a sign-off-to-PR pipeline. Land at v3: planning models, multi-repo research, story priority, threaded comments, budget tracking, sub-intents, and instant PR-merge detection. Your backlog writes itself, executes itself, and closes its own tickets.
Video generation ships as a first-class lane alongside images. A vid- model prefix routes to the video path; the Images page becomes /media (308 redirect from /images); gallery cards are playable <video> elements; and an image-to-video upload flow lets you animate anything already in your gallery. Pipeline nodes accept vid- models with the same cap and n=1 guarantees — outputs land in the blackboard served as video/mp4. 2,127 server tests pass. The cloud provider seam is documented; local-only for v1.
Missions landed across four increments this cycle. P1 ships the orchestrator-driven goal executor. P2 adds a judge layer, spend budgets, heartbeat checks, and session handoff. P3 tightens the steering surface — in-flight chat, live plan edits, pause/resume, and concurrent agent dispatch. P4 wires Missions into triggers and pipelines. A fifth PR rounds it off with a conversational mission-start dialog backed by a planning-mode orchestrator, so the goal drives the plan rather than the other way around.
The multi-tenant cloud layer that will power codecc.ai shipped across eight increments: Supabase identity schema and RLS, control-plane config and DEK escrow, entitlements and billing seed data, Edge Functions for device-link and session-handoff flows, org management with seat-gated invites and last-owner invariant, a JWKS verifier with offline disk cache, five cloud-sync repositories, and the full signup / device-link / invite / OAuth web flow. The seam is drawn; the schema is locked; the auth layer is live.
The Even Realities G2 HUD gains a full voice stack in three increments. Increment 2a streams any CCC session to the lens display. 2b adds cloud Whisper transcription + live translation — continuous mic to English captions on-lens, offline-clamped and metered. 2c ships a wake-word Voice Command mode: speak a command, the intent grammar resolves it (or Claude prompts for it), and CCC executes.
Eight PRs remade the chat surface from the ground up. The composer gains drag-drop and paste. Queued messages echo optimistically. Thinking cards, live todo checklists, edit-diff cards, and a context/cost chip render Claude's work as it happens. The slash-command menu lands in the composer. /btw opens a transcript-grounded side channel for questions without breaking the session. /resume shows a conversation picker. Bash commands surface risk annotations. Tables and headings render correctly. Drafts persist.
Two guardrail layers ship together. Stuck and loop detection watches every session in real time — when a session spins, a banner and badge fire immediately without auto-killing. Opt-in spend caps let operators set a per-session token budget at spawn time; automations that would exceed the org rate limit are held back and rescheduled automatically. Both layers compose: a stuck session that's also over budget gets both signals.
The local FLUX / ERNIE image backend gains a job queue with live progress, cancel, and a blank-image guard that retries rather than serving a white card. Per-user visibility lets each user hide or permanently delete their own images. The size dropdown now populates from the selected model's declared sizes instead of a hardcoded list. The IMAGEGEN_BASE_URL env gate keeps the page hidden until a backend is actually configured.
Four improvements landed on the pipeline canvas. Connectors now snap to 90°-locked orthogonal routing by default. Draggable bend points let you reshape any connector by grabbing it — persisted to the backend. BigRalph plan and impl models are now configurable per node via the inspector. And pipeline variables ship with editor autocomplete, inline warnings, run-dialog prompting, and e2e coverage — so the same pipeline can run against dev, staging, and prod without edits.
Screens pulled straight from this cycle's pull requests — captured against live, isolated instances.
Intents v3 ships a full story-creation surface with planning model selection, budget caps, priority levels, sub-intent scaffolding, and multi-repo research scope — all before the first session starts. CCC researches the story, assigns it to agents, and closes the ticket when the PR merges.
The session chat surface now renders Claude's internal reasoning as collapsible thinking cards, live todo checklists that check off as work completes, edit-diff cards with a drill-down view, and a context/cost chip in the composer. Every signal your operator needs, without touching the terminal.
One platform now grows its own cloud backbone, extends to the lens you wear, tracks every intent from idea to merged PR, and watches every session so you don't have to — each increment landing under test, each PR reviewed and closed before the next one starts.
Nine more capabilities that landed quietly this cycle.
Label a GitHub issue with a configured label and CCC auto-resolves a branch, an agent, and a pull request — no manual dispatch.
Export any session transcript as structured data, archive it for later, or generate a tokenized read-only share link for stakeholders.
Slack and Discord bots now support per-user chat access policies — wiki-only, chat-only, or a custom tool subset, set per project.
Bots no longer thread on every message in a channel — they require an @-mention before opening a conversation, reducing noise.
The development server switches to Turbopack, cutting dev-mode RAM by ~55% and making edit-reload cycles noticeably faster.
Right-click any session tile in the grid view for an instant context menu: rename, disconnect, kill, or jump to chat — no toolbar required.
The ccc CLI gains three new command groups — images, themes, and secrets — and injects a full command inventory into every session's context automatically.
The Assistant Gateway now exposes a standards-compliant MCP endpoint with per-tool lockdown, so external MCP clients can connect with controlled tool access.
When the local Spark GPU is busy on a StoryTranslate job, workloads are automatically rerouted to cloud — CPU-light tasks get their own tier.
Behind the headline launches, the cycle poured itself into correctness: PTY session spawning fixed to use the user-writable Claude install so self-updates take effect; control tokens minted for every trigger-spawned session; systemd-handed restart so the web process never runs on a stale build; two chat AskUserQuestion fixes so prompts survive view swaps; Discord Enterprise Grid workspace IDs accepted; and a self-heal for stale .git/config.lock in project auto-commit. Power you can trust is power that keeps working on the tenth run.