Session Summary CLI

Restores bounded working context from a Claude Code session transcript: finds sessions on disk (list) and prints a digest of verbatim quotes from the tail of a chosen session (load). The digest is never an agent-written summary — every line comes straight from the transcript: user prompts, assistant text, tool calls with their key argument, and system errors.

Bounded, not schema-strict

The Claude Code transcript format keeps growing new root event types and new field shapes on existing ones. This parser treats that as normal: an unrecognized root type is skipped silently (never a parse failure for the rest of the line's siblings — see [claude_session_types::events::SessionEvent::Unknown]), and a handful of known fields (message.content, toolUseResult) accept more than one wire shape rather than dropping the whole event.

Reads are byte-budgeted, not full-file scans

Both list and load read a bounded window from the end of the file (read_tail_lines) with a single seek plus one read — cost is proportional to the window, not to file size, so a multi-gigabyte transcript loads in well under a second. Session titles can lag behind a large tail window on a very bursty session, so title detection falls back to a small bounded read from the start of the file (read_head_lines) when the tail window doesn't carry one.

Topic selection

In priority order: the session's own custom-title (Claude Code's UI title), then ai-title (model-generated), then last-prompt (latest verbatim human prompt), then the first genuine human-typed prompt in the session (skipping harness-injected meta turns, command-palette injections, and tool-result-only turns) — never a generic placeholder unless none of the above exist at all.