From galeharness-cli
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.
npx claudepluginhub wangrenzhu-ola/galeharnesscodingcli --plugin galeharness-cliThis skill uses the workspace's default tool permissions.
Agent-facing primitive. Extract filtered content from a single Claude Code, Codex, or Cursor session file — either a conversation skeleton or error signals.
Creates isolated Git worktrees for feature branches with prioritized directory selection, gitignore safety checks, auto project setup for Node/Python/Rust/Go, and baseline verification.
Executes implementation plans in current session by dispatching fresh subagents per independent task, with two-stage reviews: spec compliance then code quality.
Dispatches parallel agents to independently tackle 2+ tasks like separate test failures or subsystems without shared state or dependencies.
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.
Space-separated positional args:
<file> — absolute path to a session JSONL file (typically a file value returned by gh-session-inventory).<mode> — skeleton or errors.<limit> (optional) — head:N or tail:N to cap output at N lines (e.g., head:200). Omit to return full extraction.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.
Narrative output with one logical event per block, separated by ---:
[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}.
One line per error, separated by ---:
is_error: trueexec_command_end events with non-zero exit or non-empty stderrEnds with a _meta line: {"_meta": true, "lines": N, "parse_errors": N, "errors_found": N}.
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.