Adding context-os under Developer Productivity.
What it is: a stdlib-Python Claude Code plugin (MIT) that gives Claude a static-analysis repo graph (symbols + line ranges + imports + git-hot files) injected pre-prompt, so the first turn opens the right file instead of grepping for it. Plus PreToolUse hooks (smart_read, smart_bash) that intercept whole-file Reads — and pure cat/less dumps via Bash — and hand Claude the file's structural outline so it reads only the slice it needs. Plus /savings Receipts that measure both sources causally from the transcript.
Why distinct from existing entries:
context-mode: that processes large tool outputs in subprocesses to keep summaries in context; context-os attacks earlier — which file Claude opens first + whether a whole-file dump enters context at all.codebase-graph: that's an MCP server with tree-sitter + FalkorDB; context-os is a stdlib-only hook (no MCP server, no DB, no embeddings), trading some recall for zero-deps + ~50ms hook latency.Honest evidence (CI-gated, reproducible):
claude --print A/B: −40.9% aggregate tokens (N=36, p=5e-7) — cold-cache one-shots; in a warm interactive session the dollar delta is smaller due to caching, so the durable win is fewer first-turn tool calls + later compaction.python3 python/evals/runners/replay_history.py.The plugin passes the contributing checks: addresses a real use case (first-turn exploration + compounding context bloat), doesn't duplicate the existing entries (different mechanism), follows the plugin structure (.claude-plugin/plugin.json, commands/, etc.), tested (CI badge green).
Happy to adjust the description length or category if you'd prefer it elsewhere.