termdeck
Claude Code
Claude Code writes every session to a JSONL file under the projects directory. Here is the exact path, the slug rule, and how to find a chat again.
Guides
Engine guides
Three coding agents, four questions each, answered from the behaviour rather than from the marketing. Every page here states what one engine actually does and puts it beside the other two.
Session storage
Where each engine writes a session
Codex CLI
Codex CLI writes rollout files into a dated directory tree and compresses the cold ones. Here is the exact layout and why paths resolve by id.
Grok CLI
Grok CLI gives every session its own directory with a live summary index. Here is what each file holds and which one is the authority on the project path.
Permission modes
One vocabulary, three engines
Claude Code
Plan, supervised, accept edits, auto and full access: what each Claude Code permission mode actually allows, and which one new sessions start in.
Codex CLI
Codex pairs an approval policy with a sandbox level. Here is what read only, plan, auto and full access each permit, and why accept edits is missing.
Grok CLI
Grok denies rather than asks in read only mode and has no plan mode of its own. Here is what each Grok permission mode really does before you rely on it.
Resuming a session
How each engine picks a conversation back up
Claude Code
Resume picks up a Claude Code conversation that already exists on disk. Here is what it restores, what it does not, and how to find the session id.
Codex CLI
Codex resumes a thread before every turn, so it always sees terminal edits. Here is what that costs on a long thread and why it needs the working directory.
Grok CLI
Grok re-reads session state on every turn rather than resuming explicitly. Here is what that means for terminal work and for finding a session again.
Models and effort
Model and reasoning controls by engine
Claude Code
Claude Code has a model, a reasoning effort and a separate thinking budget. Here is how the three interact and why the order they are set in matters.
Codex CLI
Codex advertises a larger context window than you can use and lists effort levels per model. Here are the real numbers and where the levels come from.
Grok CLI
Grok reports a 500,000 token window and compacts at 85 percent of it. Here is where the model list comes from and why the window needs no correction.
Claude Code, Codex and Grok in one browser console