soria-stack

Data pipeline skills for Soria Analytics. Cognitive modes for upstream pipeline work (scrape → extract → value-map → publish) and for building dives (dbt marts + manifest + React component + DivesPage registration + rows in the shared verifications seed + methodology wired into the component). All skills drive the Soria platform through the `mcp__soria__*` MCP tool namespace. SQL and React dives are authored locally in `frontend/src/dives/` and shipped via `git push` + PR — there is no `soria` CLI. Soria skills: /tools (verify MCP + local stack), /env (preflight; no isolated envs), /status (what exists), /plan (ETVLR orchestrator), /ingest (scrape + extract + publish), /map (value mapping), /parent-map (centralized parent company resolution), /dive (build a dive end-to-end), /preview (render a dive in chat), /verify (prove data correct), /dashboard-review (adversarial browser QA against dev.soriaanalytics.com), /diagnose (failure triage), /ticket (file structured tickets mid-session), /promote (warehouse_diff + warehouse_promote + PR + CI), /newsroom (news pipeline ops), /lessons (retrospective), /browse (persistent headless Chromium for verification), plus engineering skills /dev-env, /dev-dives, /test, and /code-review for branch-local app work, dev HTTPS dive frontend repair, test evidence, and Soria-specific code review. Suggest the right skill by stage: starting a session → /tools; investigating what exists → /status; planning work → /plan; building a pipeline → /ingest; normalizing values → /map; resolving parent companies → /parent-map; building a dive or reviewing its SQL → /dive; proving data correct → /verify; testing live dive UI → /dashboard-review; something broke → /diagnose; filing a bug/feature ticket → /ticket; promoting to prod → /promote; news pipeline → /newsroom; reviewing recent work → /lessons.

Soria-Inc 0 1 Updated 4w ago

Resources

GitHub

Install

npx skillscat add soria-inc/soria-stack

Install via the SkillsCat registry.

SKILL.md

Preamble (run first)

mkdir -p ~/.soria-stack/artifacts
echo "SoriaStack v6 loaded"
echo "---"
echo "Recent artifacts:"
ls -t ~/.soria-stack/artifacts/*.md 2>/dev/null | head -5 || echo "  (none)"

Reversibility, not isolation. There are no isolated dev environments.
Every MCP write lands on shared Postgres + soria_duckdb_staging — but
every write is soft-delete reversible via deleted_at + the
PipelineEvent audit trail. Promotion to soria_duckdb_main is PR-gated
through .github/workflows/promote.yml and dbt-deploy.yml.

Read ETHOS.md before any data pipeline work. See MCP_TOOL_MAP.md for
the tool surface.

SoriaStack — Data Pipeline And Engineering Skills

SoriaStack contains data pipeline skills and engineering skills. Data pipeline
skills drive the platform through the mcp__soria__* MCP tool namespace.
Engineering skills operate inside app repo worktrees and follow that repo's
AGENTS.md, scripts, and local dev environment rules.

Skill routing

If the user is...	Suggest
Starting a new session	`/tools`
Sanity-checking the dev stack	`/env`
Asking "what do we have for X?"	`/status`
Saying "let's work on X" or "come up with a plan"	`/plan`
Ready to scrape, extract, or publish	`/ingest`
Normalizing values across eras	`/map`
Resolving company names to parent companies	`/parent-map`
Building a dive, writing dbt SQL, or reviewing a dive	`/dive`
Checking if data is correct, proving it	`/verify`
Testing a live dive in a browser	`/dashboard-review`
Wanting to see a dive rendered in chat	`/preview`
Something broke or isn't working	`/diagnose`
Hit a bug, need to file a ticket	`/ticket`
Promoting to production (`git push` + PR)	`/promote`
Working with the news pipeline	`/newsroom`
Reviewing recent work for lessons	`/lessons`
Needing a persistent browser (repro, recon)	`/browse`
Entering or repairing a branch-local app dev environment	`/dev-env`
Repairing `dev.soriaanalytics.com` dive frontend/runtime alignment	`/dev-dives`
Testing Soria engineering changes or choosing E2E proof	`/test`
Reviewing a Soria code diff or PR	`/code-review`

The sequence

/tools (verify MCP reachable + local stack installed — always first)
   ↓
/status → /plan → /ingest → /map → /dive → /verify → /promote
                          ↑              ↑
                   /parent-map      (verify rows live in
                   (parallel to /map) the shared seed,
                                     authored inside /dive)
   + /dashboard-review (browser QA — after dive is in git)
   + /preview (render dive output in chat — any time)
   + /diagnose (enters from any phase when something breaks)
   + /ticket (side-quest — file a bug/feature ticket from any phase)
   + /newsroom (separate domain)
   + /lessons (periodic)

Each skill produces an artifact the next skill consumes.
Don't skip steps — every pipeline that went poorly started with the AI
building before looking.

ETVLR Framework

Every data concept follows this lifecycle:

E (Extract)    → /ingest Gate 1: scrape files         (mcp__soria__scraper_run)
T (Transform)  → /ingest Gates 2-4: group, schema,    (mcp__soria__detection_run /
                 extract, validate                     extraction_run / validation_run /
                                                       schema_manage / schema_mappings)
V (Value Map)  → /map: normalize values to canonicals (mcp__soria__value_manage)
L (Load)       → /ingest Gate 5: publish to staging   (mcp__soria__warehouse_manage
                                                       action="publish" → soria_duckdb_staging)
R (Represent)  → /dive: dbt marts model + manifest +  (dbt run locally →
                 TSX component + DivesPage entry +     mcp__soria__warehouse_query to
                 verify seed rows + methodology        validate, then git push + PR)

/plan orchestrates the phases. /verify runs after each.

Reversibility model

All write-path skills (/ingest, /map, /parent-map, /dive,
/promote) write directly to shared state — there are no isolated envs
to "switch to." The safety net is:

Soft-delete. Every Base ORM model has deleted_at / deleted_by.
The do_orm_execute listener filters soft-deleted rows from normal
reads. SOFT_DELETE_CASCADES propagates the delete down required
relationships. Undo via mcp__soria__database_mutate setting
deleted_at = NULL.
Audit trail. PipelineEvent captures every create/update/delete
with actor + timestamp. Read via mcp__soria__pipeline_activity /
pipeline_history / pipeline_cascade.
PR-gated warehouse promotion. Prod MotherDuck
(soria_duckdb_main) is written to only by CI on PR merge. Staging
(soria_duckdb_staging) is the working surface.

Read-only skills (/status, /verify, /plan, /newsroom, /lessons,
/preview, /ticket) don't change state but should surface the current
shared-state position in their output.

Quick reference

Principles in ETHOS.md — the source of truth
Tool surface in MCP_TOOL_MAP.md — what every mcp__soria__* call does
Artifacts in ~/.soria-stack/artifacts/ — state passed between skills
Gates in every skill — hard stops where the human must review
Completion status — every skill ends with DONE / DONE_WITH_CONCERNS / BLOCKED / NEEDS_CONTEXT
Local dev — use /dev-dives for frontend-only dive work at https://dev.soriaanalytics.com against the prod DBOS API + Clerk. No local backend is needed for dive frontend iteration.

soria-stack

Resources

Install

Preamble (run first)

SoriaStack — Data Pipeline And Engineering Skills

Skill routing

The sequence

ETVLR Framework

Reversibility model

Quick reference

Categories

Install

Recommended Skills