Resolve the canonical deploy target for a file/path and cross-check it against repo config.

Unified open items across GitHub PRs/issues and local git. includeVault (default false) opts into a bounded, most-recent-first Obsidian vault marker-sample ; off by default because even the bounded scan adds iCloud latency.

List tables in the configured local SQLite database, or (with table) inspect one table's columns/indexes/row count. Read-only via the system sqlite3 CLI; table names are validated to prevent injection.

List on-disk ML model artifact files across configured model directories. Presence on disk does NOT mean a model is loaded or active.

Row counts and recency for known prediction/feedback tables in the local dev SQLite database. Honestly labeled as local/dev, not production; flags feedback famine.

Curated architecture docs (gateway skills). No area: list doc names/sections/sizes. With area: return one doc's content (truncated to 8000 chars), matched by name/basename substring.

Read-only static scan for fabricated-core leaves (Math.random()-fed confidence/accuracy/score/decision values) across configured engine directories. A signal, not proof ; verify each finding against stub-detection-audit before certifying REAL/ABSENT. dirs overrides the configured SYSTEM_BRAIN_FABRICATION_DIRS list for this call.

List available reasoning lenses (mental models / thinker-inspired frames) for framing a decision. kind optionally filters to one lens kind. Output includes an applyGuidance string the calling model is expected to execute: pick the 1-2 lenses that bite hardest on the live problem and apply them, rather than treating the list as reference material.

Scan configured Obsidian vaults for plan/roadmap/next-steps notes (filename-matched, most-recently-modified first) and count their open - [ ] checklist items. Content reads are bounded by limit (default 6, max 12) so an iCloud-hosted vault cannot stall the call.

Rank next steps from the brain's own read-only evidence (analytics, backlog, models; deep: true adds the roadmap scan). A deterministic rule table, NOT a model: every candidate cites the tool and number it came from, and per-source health is reported so degraded evidence cannot read as confirmed support. Output includes a reasoning contract the calling model is expected to execute: rank the candidates against live session context and state what would change its mind.

Reframe a real brain_recommend recommendation (looked up by recommendationKey), or a bounded task-only subject when omitted, through a reasoning lens. Priority/evidence/title/why/action/systems/phases are always derived server-side from the real recommendation and its evidence health ; never accepted from the caller. Output includes a reasoning contract (static, in-session instructions) the calling model is expected to execute: actually answer the lens questions against the cited evidence and emit a proceed/reframe/reject verdict.