pruner

!"${CLAUDE_SKILL_DIR}/bin/pruner" mode --banner

Agents that do not expand the line above (e.g. Codex): run <this-skill-dir>/bin/pruner mode --banner first and follow it.

Offline indexer + lexical/structural relevance scorer (keyword overlap, call graph, error-stack membership, test relationships; no embeddings, no LLM). How scores work: references/scoring.md.

pruner is not on PATH: run <this-skill-dir>/bin/pruner.

1. Select a slice:

   <this-skill-dir>/bin/pruner select --task "<what you're doing>" --budget <N> \
     [--error "<traceback>" | --error-file <path>] [--changed <files...> | --auto-changed]

   If you have an error, always pass --error; error-stack membership is the strongest relevance signal. First run builds a SQLite index under .pruner/; later runs re-parse only changed files. Output lists required_context, supporting_context, relevant_tests, relevant_config as exact file:start-end ranges with a reason each. Read those ranges, not whole files.
2. The slice is a starting point, not a fence. If confidence is low/medium (a hint: line is printed) or the slice doesn't explain the behavior, fall back to targeted grep and reads.
3. Missing a dependency? Don't re-run select from scratch or fall back to open-ended exploration: check omitted_candidates, then bin/pruner expand <slice_id> --add <id or file:start-end> (or --file <path> --lines A:B). The token budget is hard-enforced: select never returns more than --budget tokens and expand refuses to exceed it unless you pass --budget-extra N.
4. bin/pruner show <slice_id> / list re-print saved slices without re-scoring.
5. With the plugin, an error pasted into the prompt is sliced automatically (context starting "pruner (ran automatically ..."): start from that slice, expand it rather than re-running select.

Full CLI (--graph, --parser, tree-sitter backend, language scope): references/schema.md.