v0.1.1 — honest denominator

Measurement-integrity release. knitbrain profile previously excluded tool results under 400 chars (which pass through uncompressed) from its denominator, which made the headline read 55.8% when the all-inclusive number is 49.3%.

We found this ourselves, before anyone asked — and since the entire point of this project is publishing the number nobody else does, it's fixed at the product level, not the wording level:

• knitbrain profile now reports both lines: sizable blocks (≥400 chars: 55.8%) and TOTAL over every tool-result token including untouched small outputs: 49.3% on 3.49M tokens from 70 real sessions.
• Per-session distribution published: median 56.2%, middle half of sessions between 48% and 64%, best 82.6%, worst 0% (nothing compressible).
• README, package description, and all claims realigned to ~50%.

No optimizer behavior changed — same compression, same losslessness, same never-expand guard. Only the measurement got stricter.

Run it on your own transcripts: npx knitbrain profile

v0.1.0 — first public release

The local-first brain for coding agents: per-project memory, task-tier workflow routing, and lossless context compression.

Note: the savings figure originally quoted here (55.8%) measured only sizable blocks (≥400 chars). v0.1.1 corrects the measurement to count every tool-result token: 49.3% all-inclusive on 3.49M real tokens. Use v0.1.1+.

Highlights

• 25 MCP tools — learnings + session handoffs, knowledge graph (imports/exports/blast-radius), tier classifier with a self-healing false-positive loop, on-demand skills engine, guardrailed agent generation, shared team board, context-window meter, explicit optimize/retrieve.
• LLM proxy (knitbrain-proxy) — compresses requests on the wire: rolling window, cross-turn dedup, directive kept verbatim, cache-stable prefix. Anthropic + OpenAI protocols.
• Structure-preserving compression — JSON keeps its schema, code keeps its signatures (tree-sitter AST), logs collapse to templates, prose keeps its opening and conclusion.
• Lossless by construction — every elision carries a content-addressed ⟨ccr:hash⟩; byte-exact original one retrieve away. Never-expand guard enforced by tests.
• Self-healing — TOIN backs off over-retrieved compression kinds; classifier recalibrates after 3 wrong-verdict votes.
• Live dashboard at 127.0.0.1:8790. Local-first: nothing leaves your machine.