context

Context efficiency from Claude Code transcripts, per session and over the window: context per main-thread turn (growth curve), peak context, compactions with context before and after, cache read / cache write / uncached input shares, estimated cache write cost, identical re-reads of a file, and tool result tokens per tool. Reads usage blocks, compact_boundary metadata and tool_use / tool_result sizes (plus Read/Edit file paths, in memory only to match re-reads). Writes counts only: no file path, prompt, command or tool output. Specs: METHOD.md (X-*). Doc: ../context.md.

• Input: ~/.claude/projects/**/*.jsonl (or config readers.paths.claude-code, or --root); prices from 02-measure/schemas/prices.json.
• Output: context.json in the data dir (or --out): readout (copied by the pipeline into readout.json section context, < 400 bytes), aggregate, curve_median, sessions (top 20 by peak).
• CLI: python3 04-auto-analysis/context/context.py [--days N] [--root DIR] [--config PATH] [--out DIR]
• Pipeline stage context (local, optional) in 03-analyze/pipeline.py. Tests: ../tests/test_context.py.

Reconfigure

Constants at the top of context.py: DAYS, ROOT, CHARS_PER_TOKEN, IMAGE_TOKENS, THRESHOLDS, CURVE_TURNS, CURVE_POINTS, TOP_TOOLS, TOP_SESSIONS, EDIT_TOOLS, MAX_READOUT_BYTES. Other harnesses are not read yet: add a scan returning the same per-file records.

Adapt, trade-offs, decisions for the human

• Adapt: THRESHOLDS to the model's window; CHARS_PER_TOKEN for code-heavy output.
• Trade-offs: costs are list-price estimates; tool result tokens are chars/4 and counted once although they stay in context for every later turn; a re-read only counts when path, offset and limit are identical and no edit to that file happened in between.
• Decisions for the human: whether to compact or start fresh sessions earlier, cap noisy tool output (Bash, MCP), or change cache TTL.