claude-engineering-plugin/plugins/compound-engineering/skills/ce-session-extract/SKILL.md

name, description, user-invocable, context

name	description	user-invocable	context
ce-session-extract	Extract conversation skeleton or error signals from a single session file at a given path. Invoked by session-research agents after they have selected which sessions to deep-dive — not intended for direct user queries.	false	fork

Session extract

Agent-facing primitive. Extract filtered content from a single Claude Code, Codex, or Cursor session file — either a conversation skeleton or error signals.

This skill exists so that agents do not read multi-megabyte session files into context. The scripts under scripts/ own the JSONL shape knowledge and emit a narrative-readable digest.

Arguments

Space-separated positional args:

1. <file> — absolute path to a session JSONL file (typically a file value returned by ce-session-inventory).
2. <mode> — skeleton or errors.
3. <limit> (optional) — head:N or tail:N to cap output at N lines (e.g., head:200). Omit to return full extraction.

Execution

Skeleton mode — narrative of user messages, assistant text, and collapsed tool-call summaries:

cat <file> | python3 scripts/extract-skeleton.py

Errors mode — just error signals:

cat <file> | python3 scripts/extract-errors.py

If <limit> is head:N, pipe through head -n N. If tail:N, pipe through tail -n N. Apply the limit after the Python script, never before — the _meta line is emitted last and a head cap may drop it; that is acceptable when the caller asks for a head cap.

Return the raw stdout verbatim. Do not paraphrase, annotate, or synthesize — the caller does synthesis across multiple sessions.

What each mode returns

Skeleton

Narrative output with one logical event per block, separated by ---:

• User messages (text only, no tool results, framework wrapper tags stripped)
• Assistant text (no thinking/reasoning blocks — those are internal or encrypted)
• Tool call summaries; 3+ consecutive same-name calls are collapsed (e.g., [tools] 5x Read (file1, file2, +3 more) -> all ok)

Ends with a _meta line: {"_meta": true, "lines": N, "parse_errors": N, "user": N, "assistant": N, "tool": N}.

Errors

One line per error, separated by ---:

• Claude Code: tool results with is_error: true
• Codex: exec_command_end events with non-zero exit or non-empty stderr
• Cursor: always empty — Cursor agent transcripts do not log tool results

Ends with a _meta line: {"_meta": true, "lines": N, "parse_errors": N, "errors_found": N}.

Error handling

If the file cannot be read, let the error surface to the caller. If _meta reports parse_errors > 0, return the output as-is — partial extraction is still useful and the caller decides whether to widen the search or deep-dive further.