Syncing a source delta into the Brain (additive, gated, no regression)

A git pull of the immutable raw repo (ramco-erp-brain/) periodically lands new files (a new artefact class, more of an existing source, a whole new BPC) or refreshes existing ones. This skill turns that raw delta into Brain content safely and repeatably. It is the orchestration layer; the mechanics live in reindex-brain (how to add one thing) and the proof lives in eval-brain (how to grade without regressing).

Prime directive (from reindex-brain): positive delta, never regress. Raw is read-only; registers are run-id'd and never clobbered; the accepted pointer advances only on a full gate pass; a new BPC adds coverage scope but never a new wiki ontology.

The 7-step loop (run in order; each step gates the next)

1 · Snapshot the delta (git facts)

• Record HEAD before/after, commit list, and git diff --stat HEAD@{before}..HEAD (added / modified / deleted counts; disk footprint). Deletions in raw are rare and must be handled as provenance-tagged supersedence, never as a Brain deletion.
• Flag any new top-level berp-* folder — that is a new BPC/module tier (new coverage scope + a domain-model note in CLAUDE.md and git_walkthrough.md), not a new page type.

2 · Document it (human-readable inventory) — before touching the Brain

• Update git_walkthrough.md (§3 top-level map, §6 berp-general table, §7 A/B/C cross-ref, §9 gaps) with the new folder-by-folder counts.
• Write a dated git_new_addion.md-style delta addendum: commit-by-commit attribution, per-folder counts grounded in the working tree, and the gap-registry impact.
• (These two artefacts are the audit trail; produce them even for small deltas.)

3 · Classify every changed path → change shape × artefact class

For each new/changed path decide (a) the reindex-brain change shape and (b) whether it hits an existing parser glob or needs a new class:

More of an existing source (same shape, new dirs) ....... §A GLOB WIDENING (disjoint recount)
A brand-new structured artefact class ................... §A NEW PARSER  (parser_plan_v2 class row)
A binary/prose the LLM must read, not a parser .......... §A HYBRID / LLM-DIRECT (container shim only)
A new relationship the data now supports ................ §B CROSS-CONNECTION (extend an edge generator)
The Brain already has it — chatbot just can't surface it. §C PROMPT / TOOL (cheapest — try first)

Dispatch by magic bytes, not extension. A new BPC = many §A classes (per-component loop) + a domain-model note. Never invent an orphan page type.

4 · Route each class through reindex-brain

Invoke the reindex-brain skill per class, honoring its section:

• §A — parser mode decision → inherit the output contract (run-id'd JSONL, header counts.in/out/skipped_named, named residuals, golden-file suite) → acceptance discipline (golden reproduction → LLM spot-audit ≥5 files → gate pass → then the accepted pointer) → build_runmaps.py → re-run generators (idempotent). Glob widening: prove the new glob is disjoint, recount per BPC so old_total + new_disjoint = new_total with the delta enumerated, tag new sub-classes, re-run only the newly-included dirs.
• §B — extend one of the five edge generators (gen_cim_cardinality / gen_sp_callgraph / gen_reverse_indexes + gen_cross_fk / gen_journey_index / [[wikilinks]]); resolve by basename; every emitted link must resolve.
• §C — append (never contradict) a rule/tool in brain_chatbot/src/prompt.js or tools.js; rebuild the BM25 index.

5 · Sequence the work (run-anytime vs gated vs barrier)

• Run-anytime (P8.0): glob widenings, named-skip enumeration, prose refresh — they only ADD and break nothing. Do these first.
• Gated (P8.2–P8.9): classes that consume V1 registers; fan out per gap.
• Barrier (P9): after all classes land, re-run dependent overlays (journeys / a2ui / migration / rules-semantics), rebuild the BM25 index, then the Regression Gate.

6 · Regression Gate (mandatory — use eval-brain) before advancing any pointer

• check_links.py → 0 dangling [[links]], 0 unreconciled A2UI packs.
• verify_journeys.py → 0 schema-invalid, 0 dangling refs.
• Every touched batch's gates.json → all pass; count sums reconcile.
• Re-run the Layer-1 eval via eval-brain vs the last accepted baseline: FULL% and answerable% must not drop; no previously-FULL question may regress; new coverage only adds FULL answers. Apply eval-brain's grading rules (drop malformed/two-word questions; a superset of the gold is CORRECT, not a hallucination). Diff per-question verdicts, not just headline %.
• Spot-check 3–5 previously-good questions → unchanged or better, same citations.

7 · Record & institutionalize

• status.json — open/close gaps[] with the run that changed them; bump parser / generator versions; a closed gap keeps its record (never overwrite).
• log.md — append the run history.
• git_walkthrough.md / CLAUDE.md — apply the domain-model deltas (e.g. an 11th BPC).
• Only on all-green: advance the accepted pointer. If any prior answer regressed, do not advance — the prior baseline is your rollback until the regression is fixed.

What this skill will NOT do (guardrails)

• Never write into ramco-erp-brain/ (raw is immutable).
• Never delete a register run, overwrite a good run with an ungated one, or merge two sources so a prior row disappears — all forbidden (they regress the Brain).
• Never advance the accepted pointer without a green Regression Gate.
• Never simplify or skip the acceptance discipline "to save time" — a bad run must stay .FAILED and invisible, not be patched into the wiki.

Related

• reindex-brain — the per-change mechanics (§A/§B/§C, the 8 non-regression devices).
• eval-brain — the grading harness the Regression Gate calls.
• ramco-journey — regenerate journeys for components a delta touched.
• Canonical process docs: brain_plan/master_plan.md, brain_plan/V2/master_plan_v2.md, brain_plan/parser_plan.md (+ V2/parser_plan_v2.md), brain_plan/status.json (gaps).