Transmission — Cycle 18 · Oct 5 → Oct 12, 2026

Plugins. Clef. Run isolation. CuttleCode's platform layer deepens.

CuttleCode gets its Plugin Catalog — 13 plugins ship built-in, each scoped to exactly the permissions it declares. Clef, the ambient AI assist layer, lands on main and auto-enables. FX-5 moves every agent spawn under a dedicated OS run user, keeping secrets isolated and spawn revocable at the kernel level. A Proposal Inbox surfaces PR merges, flag flips, and deploy holds as swipeable cards with one-tap execution. A four-phase Analytics Platform lands with self-service dashboards, session provenance capture, and fleet telemetry. And Ralph gains a stuck-run detector and frontier escalation — it pauses when it stops making progress and escalates when the local model isn't enough.
180 Commits merged 80 New features 169 Pull requests 7 Days elapsed
0
Commits merged
0
New features
0
Pull requests
0
Days elapsed

01 / New this cycle

Nine systems that extend what CuttleCode can run.

A plugin catalog that keeps the server surface fixed, ambient AI assistance, OS-level agent isolation, a structured action inbox, a full analytics stack, a loop definition catalog, orchestrator improvements, and a Ralph that knows when to stop and escalate.

01

Platform

Cuttle Plugin Catalog — 13 Plugins Built-in

CuttleCode gains a first-class plugin system. Import a .cuttle file or browse the built-in catalog — Automation Suite (nightly triage, weekly dependency audit, agent code review), ComfyUI (image generation via a local server), and eleven more ship at launch. Each plugin declares its exact permission scopes — automations.contribute, sessions.spawn, network:${config.url} — and the operator approves each one before installation. Plugins are isolated from the server binary; extending CuttleCode never means editing a config file or restarting the daemon.

02

AI Assist

Clef — Ambient AI Assist, Auto-Enabled

The Clef stack lands on main and turns on automatically when the backend signals readiness. Clef is CuttleCode's always-on AI intelligence layer: it watches session context, surfaces relevant wiki pages, and can trigger actions before you ask. Where System One audits every ACP tool call for safety, Clef operates in the suggestion layer — it recommends rather than enforces. The stack ships with a Clef assist toggle in the session header and a dedicated settings panel.

03

Security

FX-5 — Dedicated Run User, Sandboxed Spawn

Seven slices of FX-5 landed this week. Every agent spawn now runs under a dedicated OS-level run user (ccc-run), separate from the CuttleCode server process. The PTY daemon shifts to external mode with SO_PEERCRED peer-credential verification. A revocable spawn latch lets the host admin cut off new spawns without restarting the server. Secrets in /etc/credstore are readable only by the run user; fx5-migrate --reverse round-trips byte-for-byte if rollback is needed. Every attack surface reduction ships with a revert-failing test proving the gap it closed.

04

Inbox

Proposal Inbox — PR Merges, Flag Flips, Deploy Holds

A structured Proposal Inbox surfaces high-stakes actions as typed proposals — pr_merge, flag_flip, deploy_hold, cleanup, bug_triage. Each proposal carries a reasoning-provider score; swiping right executes the action behind CCC_INBOX_EXECUTE. Producer tokens (inbox.produce scope) let external digests and providers write proposals without broader API access. The inbox entry appears in the sidebar with an open-item badge and an opt-in high-stakes push notification.

05

Analytics

Analytics Platform — Four Phases in One Cycle

Phase 1 ships a metrics engine and /api/analytics rollups. Phase 2 adds a self-service Analytics dashboard. Phase 3 captures spawn source, human turns, device class, and PR origin per session. Phase 4 enables opt-in fleet telemetry with an owner fleet dashboard — model usage by node, token velocity, and GPU utilisation. All four phases landed in a single cycle; analytics is now a first-class platform surface.

06

Loops

Loop Catalog Phase A

The first installment of a structured loop definition catalog: a definition schema with cycle records and readiness keys, fix-with-proof runs that open and close only when the proof passes, a pytest runner for fail-before/pass-after proofs, and per-environment configurations. Phase A covers the core lifecycle; subsequent phases will add specialised loop types.

07

Ralph

Stuck-Run Detector + Frontier Escalation

Ralph now detects when a run stops making progress — N identical unmet constraint sets without forward motion — and pauses automatically at a named checkpoint. From the run view, operators see the pause reason. Separately, frontier escalation decisions are now recorded (Phase 1) and applied with revert and spending caps (Phase 2). Ralph can escalate to a stronger model when the local model isn't resolving the loop, and roll back if the decision turns out to be wrong.

08

Orchestrator

Adopt Existing Sessions into an Orchestration

The orchestrator can now adopt an already-running session into an active orchestration — via the CLI (ccc orchestrator adopt) and the REST API. The session gets the orchestration's parentage, appears in the session tree, and can be awaited or cancelled by the orchestration's control flow. Previously, only sessions spawned by the orchestrator could participate; now any live session can join.

09

Sessions

Orchestrator Sessions — Parentage + Sidebar Folding

Orchestrator sessions gain first-class representation in the sessions pane: a collapsible tree shows parent and child sessions with depth indicators, a Sessions Wait chip counts how many children are blocked on external events, and the sidebar folds sub-trees when the orchestrator is not the active session. Session-level parentage is written to the database and exposed in the REST API.

This cycle CuttleCode became a platform you extend with plugins, isolated every agent behind a dedicated OS identity, shipped ambient AI assistance with Clef, gave operators a structured inbox for high-stakes actions, and launched a full analytics platform alongside a loop definition catalog and a Ralph that knows when to stop, escalate, and revert.

— CuttleCode · orchestration + inference substrate

03 / Also aboard

Smaller payloads, real reach.

Twelve more capabilities that landed quietly this cycle.

Browser Extension

Firefox Support

The CuttleCode browser companion now ships for Firefox. The same pairing flow, the same tab-sharing bridge, and the same agent-controlled browsing — now on a second major engine alongside Chrome.

Browser Companion

Companion Popup Redesign

The browser extension popup is rebuilt at a real 520px canvas — a clean pairing flow, a step-numbered guide, and an Appearance picker. It opens faster and stays out of the way.

Mobile

Session Header ⋯ Sheet

The mobile session header's trash icon is replaced by a ⋯ button that slides up a sheet: Hide tool calls, Hide thinking, Show timestamps, and End session. Less accidental termination; more control.

Agent Sessions

ccc-control MCP in Agent Sessions

Agent sessions can now include the ccc-control MCP server, letting an agent drive CuttleCode's own API from within a session. Opt-in behind CCC_SESSION_CONTROL_MCP.

Triggers

Typed Trigger Parameters

Trigger definitions now declare typed parameter schemas validated before substitution on every execution. Bad parameter types are caught at trigger time, not buried in a failed session.

Wiki

Recommended Pages as Hidden Context

Wiki pages the knowledge advisor marks as relevant to the active session are now delivered as hidden system context — agents start each session already aware of what the wiki says about the work.

Intents

Model Picker in Intents

Intents can now select the model and provider to use when launching a session — from the connected provider menus. The choice persists and re-routes correctly after a retry.

Desktop App

Update Claude Code from the Notice

When Claude Code's in-app update notice appears, a single click now installs the update directly — no terminal, no copy-pasted upgrade command.

Sessions

Left-Rail Hover Cards

Project rows in the left rail now surface a hover card with session counts, active pipelines, and a pinned-project toggle. Session rows show model, branch, and duration — no click required.

Worker Nodes

Offline Notifications

Workspace admins receive a notification when a worker node goes offline after its 2-minute heartbeat window expires. No more silent capacity loss.

Fleet

Model Capability Cards

The fleet dashboard now shows model capability cards generated from actual runs, cycles, and canaries — not from static metadata. Coding models route to worker nodes that match their GPU footprint.

Ralph

Per-Run Learning Record

Ralph writes a per-run learning record table after each cycle, capturing what worked, what got reversed, and why. Stored and queryable across runs under CCC_LEARNING_RECORD_ENABLED.

04 / Under the hull

The unglamorous 80%.

Behind the headline launches, this cycle closed 44 bugs. The automation result detector now catches AUTOMATION_RESULT on every runner lane, not just the primary. Git reads in bwrap are hardened and fail closed against paths that escape the work tree. The watchdog identifies Linux processes by /proc starttime, not ps lstart, closing a false-positive kill on process-ID reuse. Local Ollama and paid cloud PTY lanes authenticate correctly with ANTHROPIC_AUTH_TOKEN. The sessions pane keeps past-session rows visible under a crowded live list. Session sharing surfaces readable conflicts, revocation paths, and invite flow in plain UI. Project auto-commit is skipped for sandboxed workspace dirs. A long local-model reply now streams a continuation notice rather than truncating. Every fix removes a class of failure, not a single instance.

44
Bug fixes
80
Features
169
Pull requests