Claude Code Adds a Live Diff Panel, /skill-doctor, and Claude Fable 5.1 as Default
Claude Code's early-September releases (2.1.257 through 2.1.261) focus on visibility into what an agent session is actually doing. Version 2.1.260 introduces a /diff command that opens a side-by-side panel showing uncommitted changes as Claude edits files, replacing the need to shell out to git diff mid-session, and adds likely-cause hints for prompt-cache misses to /cost. Version 2.1.261 follows with /skill-doctor, a diagnostic command that reports which loaded skills are going unused in a session and how much context budget they're consuming — useful for trimming bloated .claude/skills configurations in larger projects.
Separately, version 2.1.257 makes Claude Fable 5.1 the default Fable-tier model, bringing a 1M-token context window at $10/$50 per million input/output tokens with $0.25/Mtok cached reads. The same release adds configurable timeFormat and timeZone settings (12-hour, 24-hour, 24-hour UTC, or a custom strftime pattern) and a Containment Escape rule to Auto mode that blocks attempts to reach cloud metadata endpoints. Version 2.1.259 adds a managedMcpServers setting so organizations can centrally provision HTTP/SSE MCP servers for their developers, plus a --permission-prompts none flag for running Claude Code unattended on headless hosts.
Read more — Anthropic
OpenAI Codex CLI 0.153.0 Ships Vim Mode, Plugin Marketplace, and Auto-Recaps
OpenAI's Codex CLI 0.153.0, released September 3, is one of the larger recent updates to the tool. It adds Vim-style undo/redo support for developers who prefer modal editing inside the terminal UI, and introduces plugin CLI marketplace support so third-party Codex plugins can be discovered and installed directly from the command line rather than requiring manual configuration. The release also adds configurable automatic recaps that summarize a session's progress at set intervals, an enhanced terminal history display, and early rate-limit warnings that surface to Plus and Team subscribers before they hit a hard cap.
The releases immediately preceding it rounded out the update: 0.152.0 (September 1) added Vim-style / and ? search, richer rate-limit banners with actionable suggestions, visible credential-refresh progress in the terminal, and package-style naming for MCP servers, while 0.151.0 (August 29) added configurable grace periods for optional MCP tool discovery and let extensions inspect or modify MCP tool call results before they're returned to the model.
Read more — OpenAI
NousResearch Hermes Agent v0.21.0 "Pantheon" Adds Multi-Agent Bot Mode
NousResearch shipped Hermes Agent v0.21.0, "The Pantheon Release," on August 31 — a large update spanning roughly 5,800 commits and 2,475 merged pull requests since v0.20.0. The headline feature is Bot Mode built directly into the desktop app: a society of named agents, each with its own face and group chat, that can talk to each other and to the user like a team rather than a single tool. Supporting that, the release adds hermes peer for direct bot-to-bot messaging, memory-backed cron jobs so scheduled tasks retain context and continuity between runs, and live subagent steering that lets a user redirect a running subagent mid-task instead of waiting for it to finish or fail.
Beyond Bot Mode, v0.21.0 strengthens the MCP command center with unified health checks and usage tracking across connected servers, adds direct browser control from inside the desktop app, and expands the model catalog with six new providers. The release notes also describe a broad security hardening pass across the agent runtime. This follows the "Herald Release" (v0.20.0, August 3), which had introduced streaming conversational voice and Agent-to-Agent v1.0 protocol support — Pantheon builds directly on that multi-agent foundation rather than introducing it from scratch.
Read more — NousResearch
IntelliJ IDEA 2026.2.2 Fixes Spring Modulith False Positives and Markdown Rendering
JetBrains released IntelliJ IDEA 2026.2.2 on September 2 as a maintenance update to the 2026.2 line. The release fixes a Spring Modulith regression where the IDE incorrectly flagged accessibility violations for types exposed through @NamedInterface in nested modules — a false positive that could clutter code inspections in modular Spring applications. Other fixes address a freeze when opening directories containing a large number of Markdown files, incorrect vertical scrollbar rendering in the Markdown preview pane, and a bug where terminal tab order wasn't preserved after reopening a project.
The update also resolves a failure when loading remote OpenAPI specifications and improves checkbox contrast in Markdown for better visibility. As a patch release, 2026.2.2 doesn't introduce new features but rolls up fixes reported against 2026.2 and 2026.2.1 since the July feature release, which had added day-one Java 27 and Kotlin 2.4 support.
Read more — JetBrains
Ollama Introduces Transparent Per-Token Pricing and a Claude Desktop Gateway
Ollama published two notable updates in late August. On August 31, the company detailed a move to transparent, industry-standard per-token pricing across its Pro, Max, and Team plans, replacing more opaque usage allowances with metered pricing that's included on every paid tier — a shift aimed at making cost predictable for teams running larger models through Ollama's cloud offering. On August 25, Ollama enabled Claude Desktop to be configured with Ollama as a third-party gateway provider, letting users route Claude Desktop's model calls through locally or cloud-hosted open-weight models served by Ollama instead of Anthropic's own infrastructure.
Together, the two changes extend Ollama's positioning as a model-agnostic serving layer: developers can mix proprietary assistants like Claude Desktop with self-hosted open models, and pay for whichever backend they actually use on a consistent per-token basis rather than juggling separate subscription tiers for each.
Read more — Ollama
