You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Long Claude Code / Codex sessions in the regular terminal produce thousands of lines. The outputs I actually need again — a final answer, a decision, a command that worked — get buried, and then disappear entirely once scrollback trims: the renderer ring buffer at 5k rows by default, the daemon's 5k-row cap (#12864), checkpoint byte caps and log-cap rotation (#17114), recovery snapshots (#14593).
Today there is no way to keep a handful of important outputs reachable. Terminal Search only covers what is still in the buffer, and scrolling back through hours of tool output is slow.
Proposed solution
Chapters: user-chosen, durable bookmarks on terminal output.
Select text in a terminal pane → right-click → Save selection as chapter (plus a configurable shortcut). Title defaults to the first line and is editable.
A new Chapters tab in the right sidebar (alongside Explorer / Agents / Source Control / Checks) lists the active worktree's chapters, grouped by terminal, newest first.
Click a chapter: if the text is still in that pane's scrollback, scroll it to the top of the viewport and briefly highlight it; otherwise open the saved copy in a read-only view inside the panel — a chapter never goes dead.
Durability: each chapter stores the captured text plus a content hash, pane identity, timestamp, and the agent session id when known. Persisted per worktree, so it survives terminal close and app restart.
Recovery: when the pane ran a known agent (Claude Code, Codex, …) and the output is gone, Orca looks the snippet up in that session's transcript — the same files Agent Session History already discovers, including on SSH hosts — and shows the full untruncated message.
Agent-neutral capture (from the xterm buffer); transcript recovery and search are per-provider enhancements. Works in folder workspaces and SSH worktrees.
Problem or use case
Long Claude Code / Codex sessions in the regular terminal produce thousands of lines. The outputs I actually need again — a final answer, a decision, a command that worked — get buried, and then disappear entirely once scrollback trims: the renderer ring buffer at 5k rows by default, the daemon's 5k-row cap (#12864), checkpoint byte caps and log-cap rotation (#17114), recovery snapshots (#14593).
Today there is no way to keep a handful of important outputs reachable. Terminal Search only covers what is still in the buffer, and scrolling back through hours of tool output is slow.
Proposed solution
Chapters: user-chosen, durable bookmarks on terminal output.
Alternatives or additional context
SetMarknavigation marks that are intentionally disposed on trim — complementary jump UX, opposite durability model.