stci-dataops

STCI Data Operations

Overview

Turn a proposed pricing source into a reproducible evidence packet and, when requested, a tested STCI collector change. Treat published prices as time-sensitive evidence and legal conclusions as a human-review boundary; see the source evaluation rubric.

Prerequisites

Work from the STCI repository root with Python 3.11 or later.
Read the current observation schema, collector base class, methodology, source profile template, and legal-risk register before proposing changes.
Use existing dependencies and test fixtures. Ask before installing packages, changing authentication, enabling schedules, deploying, or publishing data.
Confirm whether the request is research-only, profile, implementation, or review. Research-only mode must not modify repository files.

Workflow

Establish repository truth with Glob and Read: locate the current schema, collector modules, tests, methodology, source-profile convention, provider IDs, and mappings.
Define the requested source and collection boundary: official provider, aggregator, manual submission, or fixture; public page or API; authentication; expected units; intended refresh cadence.
Research with WebSearch and WebFetch, preferring the provider's official pricing, API, terms, robots.txt, rate-limit, and changelog pages. Record the exact URL and retrieval time for every claim.
Separate observations from judgments. A reachable endpoint or permissive robots.txt file is not proof that automated collection or redistribution is legally authorized. Mark ambiguous terms LEGAL REVIEW REQUIRED.
Reconcile the proposed source tier against the repository's current written methodology and implementation. Report contradictions instead of silently choosing one; in particular, aggregator classification must be consistent across docs, schema usage, and collector code.
Produce the evidence and acceptance matrix in the evaluation reference. Stop in profile mode after presenting it.
In implementation mode, obtain approval for the proposed file set, then use Write or Edit to add the smallest collector, fixture, profile, and test changes that satisfy the collector contract.
Run the narrow test first, then the repository suite. Use fixture or dry-run modes before any live collection. Inspect git diff and report all generated or data-bearing files; do not commit, push, enable CI schedules, deploy, or publish unless explicitly requested.

Approval boundaries

Proceed: read public documentation; inspect local code and fixtures; draft an evidence packet; run existing local tests.
Ask first: install dependencies, use credentials, call a metered/private API, write repository files, run a live collection that stores data, or alter source-tier policy.
Explicit separate authorization: deploy, publish observations, enable a schedule, change cloud/IAM configuration, commit, push, or open a PR.
Never: scrape around access controls, expose tokens, invent permission from silence, copy restricted content, or label uncertain evidence as verified.

Validation

python -m pytest tests/test_collector.py -v
python -m pytest tests/ -v --tb=short
python -m services.collector.pipeline --fixtures --dry-run
git status --short
git diff --check

Validate every candidate observation against schemas/observation.schema.json and confirm normalization with fixed input/output unit tests. A live endpoint response is evidence, not a golden fixture; redact secrets and minimize retained payloads according to approved policy.

Output

Return:

mode, source identity, retrieval timestamp, and official evidence URLs;
auth, rate-limit, automation, and redistribution status, with unknowns clear;
unit mapping and required observation fields;
source-tier recommendation plus any policy contradiction;
proposed or changed files and approval received;
tests run, exit results, observation counts, and validation failures;
unresolved legal, operational, provenance, or data-quality risks;
next human decision, without claiming deployment or publication occurred.

Error Handling

Terms or permissions unclear: stop before automation and request legal or owner review.
Authentication required: document the mechanism; never request secrets in chat or place them in code, fixtures, URLs, or logs.
Rate units ambiguous: retain the raw field and stop normalization until an official definition resolves the unit.
Schema drift: preserve a minimal redacted sample, report changed fields, and fail closed rather than dropping records silently.
Rate limit or outage: honor retry guidance; do not retry indefinitely or substitute stale data without the methodology's explicit carry-forward rule.
Cross-source disagreement: report both values and provenance; do not average or choose a source outside the approved methodology.
Partial write or failed test: show the diff and failure, preserve evidence, and do not publish.

Examples

Research-only request:

Evaluate Provider X's official pricing API for STCI. Produce evidence and a
source-tier recommendation, but do not edit files or call authenticated APIs.

Implementation request:

Using the approved source profile, add the provider collector and fixed fixtures.
Run collector tests and the fixture dry-run; do not deploy or enable schedules.

About this skill