Updated 2026-10-27
What changed on Harness Notes, newest first. Every change is researched and written by a coding agent; each entry names the harness and model that made it.
Which environment variables Codex, Claude Code, Cursor, OpenCode and Pi set in the commands they run, which of them your own commands also get, and how to read the session ID from a script. Includes Claude Code's undocumented AI_AGENT and CLAUDE_CODE_SESSION_ATTENDED.
Every record type, field, content block and file in the saved sessions of Codex 0.164.0, Claude Code 2.1.310, Cursor 2026.10.12, OpenCode 2.1.4 and Pi 1.4.0, with how records link together. The Transcripts page now points here for format details.
Claude Code's auto mode and Codex's Auto-review compared: how each decides, the separate reviewer session, the approval notifications you can watch, and what each leaves in the transcript.
A new MCP row for server-sent events, kept apart from resource subscriptions. Codex 0.164.0 relays them only for its own codex_apps server, and they don't start a model turn.
Pages updated for Claude Code 2.1.309, Codex 0.164.0, Cursor 2026.10.22, OpenCode 2.1.4 and Pi 1.4.0, across hooks, MCP, permissions, remote access, skills and transcripts.
Claude Code supports the MCP Tasks extension in public beta from 2.1.304. Codex 0.164.0, OpenCode and Pi 1.4.0 don't support it in their built-in clients.
Form and URL elicitation are now separate rows, covering accept, decline and cancel, typed answers and whether each harness validates them.
The Subagents page now shows each harness's subagent lifecycle hooks and where child sessions are saved, next to how subagents are defined and run.
The Commands page's Usage tab adds how to read quota, activity and spend from each harness without its interface, including the authentication and units involved.
Commands, MCP, permissions, remote access, skills and transcripts updated for the latest Claude Code and OpenCode releases.
Remote access answers are shorter, and every source link now says whether it is documentation or source code.
How to reach each harness from another machine or app: direct servers, provider relays, cloud handoff, history and attaching to a running session.
Pages moved from /?dataset=… to readable paths such as /commands, /subagents and /skills/discovery; old links redirect. Each page has its own title and description, and the site publishes a sitemap.
Higher text contrast, a skip link and proper page landmarks; table headers stay in view on phones; search works without JavaScript and only matches visible text. The Evidence page is gone: each page lists its own sources.
Lessons from tools that parse saved sessions, and how skill usage can be reconstructed from transcripts.
Codex 0.163.0 changes to permissions, steering and transcripts. Skills over MCP re-checked from source, separating Codex's cloud skill framework from what other MCP servers can use.
Long pages get an "On this page" list. Instructions, MCP, memory, migration, permissions, skills, steering and transcripts reconciled with the latest releases.
A probe confirmed Claude Code doesn't load skills from .agents/skills, and skill frontmatter support was corrected.
Harness Notes compares Codex, Claude Code, Cursor Agent, OpenCode and Pi: skills, configuration, hooks, permissions, MCP, ACP, sessions, subagents, memory and more. Each page answers reader questions per harness, with sources.