Intent, trend / coverage flags, verdict and fact extraction use structured calls. A single reader accepts valid calls, safely recovers simple malformed calls by parsing, and rejects foreign ones. The shared client adds reasoning headroom: LOW 1,500, MEDIUM 4,000 and HIGH 8,000 tokens (override capped at 30,000).

Model-written code and production Python are parsed, never executed for call recovery or lineage. SQL is read statement-by-statement; execution allows SELECT / WITH only, enforces table allowlists and row / time limits, and blocks file, network, system and scheduler access from the execution environment.

Checks are isolated and time-bounded. If one fails, the best available draft survives and the degraded path can be counted. Typed trace signals feed proposals for knowledge growth; a human must approve changes.

Tool content is bounded, cleaned and labelled as data; credentials are redacted. The interface streams real status, supports cancellation and structured chart / filter actions, while analysis detail stays out of browser responses unless debugging is enabled. Optional basic authentication and a liveness endpoint support operation.

Flow-sensitive parsing follows definitions in order, resolves assignments, formatted paths, helper returns and SQL-built path declarations, tags comments / dead code as not live, and links scripts through shared storage locations.

Stored values are indexed across columns derived from graph metadata and live scans. A typed name is matched case-insensitively to the exact stored form; a miss states what was searched rather than claiming non-existence.

Facts are re-cleaned on every request, reduced to plain tokens and one-line values, with paths shortened and invisible characters removed. The encoding marks arguments as data.

Actual work-in-progress steps, source provenance and typed signals are traced. A proposals-only learning loop can suggest knowledge growth, but humans gate the changes.