knitbrain v0.7.0 — the onboard release

The brain stops waking up blank. 34 MCP tools (was 33). Backward-compatible.

Highlights

knitbrain onboard — the front door. The CLI scans the repo into the
knowledge graph and imports this project's past Claude sessions into the wiki
(pages + spine) + mines learnings. The new knitbrain_onboard MCP tool
runs the agent-driven intent interview: it returns 5 questions, the agent asks
them, then writes a Project Charter — a claim-lined wiki page surfaced
every session — plus a constraints skill whose guardrails propagate to spawned
agents. The loop is shaped to your project, not generic. (Fixes the old silent
onboard no-op.)
Honest context meter. The window is configurable
(KNITBRAIN_WINDOW_TOKENS) and auto-heals: on a large-context model the meter
no longer pins to a false 100% / "save a handoff now." The token count was
always real; now the percentage is too.
Terse brain-writes (opt-in). KNITBRAIN_TERSE_STORE=1 (default off):
learning summaries and skill bodies are terse-rewritten before persist,
reusing the existing compressProse — one transform, not two — and never
touching code, paths, numbers, or claim: lines. Output-side terse stays the
separate prompt-level mode.
Cross-platform + cleanup. projectTranscriptDir strips the Windows drive
colon; fan worktree creation drops a unix-only redirect; a dead test export
removed.

Gate chain green: typecheck · lint · test 358 · build · consistency
(34 tools) · bench (floors held).
production-audit 50/50 · e2e (all 34 tools live) · evals 100% (2514
real blocks: round-trip + identifier + error + summary + never-expand).
4-layer coherence (every tool through the one dispatch() chokepoint) ·
onboard path live (Charter + constraints + load_session surfaces intent) ·
adherence HARD · cso clean · cross-platform.

npm install -g knitbrain
knitbrain setup      # wire MCP + hooks
knitbrain onboard    # import history + scan the repo

Full changelog: CHANGELOG.md.