token-reducer

token-reducer

Cut context size without cutting answer quality.

Why token use still spikes

Claude Code often answers code questions with native Read / Grep on whole files, which loads raw text into the model and bypasses this pipeline. Long chat history is re-sent every turn, so costs compound even when code is compressed.

Workflow (prefer this order)

1. Do not paste large code or logs into chat — that bypasses reduction and burns tokens.

2. Run the slash command first so the pipeline runs before reasoning, for example: use /token-reducer with a short objective and paths (defaults come from plugin settings.json: small chunks, low --top-k, word budget, relevanceFloor).

3. CLI (same pipeline) when you want a packet on disk or in a script:

   python "${CLAUDE_PLUGIN_ROOT}/scripts/context_pipeline.py" run --inputs ./src --query "Locate JWT validation" --top-k 3

   Use a specific query (not “auth stuff”) so low-scoring chunks are dropped by the relevance floor before summarization.

4. Session hygiene: around 10 turns the hook suggests /compact; by 40–50 turns start a new chat for coding after planning.

Pipeline (what the tool does)

1. Preprocess large/noisy corpus into overlap-aware chunks (size/overlap from settings.json → chunkSizeWords / chunkOverlapWords).
2. Index chunks into SQLite FTS5 and local embeddings.
3. Retrieve with BM25-first policy; vector fallback when configured.
4. Merge + rerank; keep top K (default 3 from defaultTopK in settings.json).
5. Compress with TextRank/word budget; drop chunks below relevanceFloor.
6. Emit citation-rich packet + savings telemetry.

Commands

• End-to-end run (defaults from plugin settings.json; override flags as needed):

  python "${CLAUDE_PLUGIN_ROOT}/scripts/context_pipeline.py" run --inputs . --query "${ARGUMENTS}" --hybrid-mode fallback --top-k 3

• Self-test:

  python "${CLAUDE_PLUGIN_ROOT}/scripts/context_pipeline.py" self-test

Tuning (plugin settings.json under tokenReducer)

• compressionWordBudget — lower for shorter summaries (e.g. 150).
• chunkSizeWords / chunkOverlapWords — smaller chunks before compression (e.g. 100 / 20).
• defaultTopK — fewer final chunks (e.g. 3).
• relevanceFloor — higher values drop more weak chunks before summarization (e.g. 0.18).

Session reminders: top-level promptGuard (autoCompactTurn, autoResetTurn, criticalResetTurn, reminderTurns) plus historyCompactReminderTurns inside tokenReducer.

Deep References

• ./references/implementation-guide.md
• ./references/context7-integration.md