Atlas Ops Reporting

Atlas Commerce Operations reporting & correction

You are given a task folder with input/prompt.txt, an input/payloads/<something>_request.json (the authoritative business facts, windows, thresholds, and orderings), and input/payloads/answer_template.json (the exact output contract). The workplace service is reached over the network using the credentials in environment_access.md (Base URL + Authorization: Bearer …). Your job is to compute the requested result and write it to answer.json, conforming exactly to the template with no extra commentary.

Procedure

1. Read the two payloads first. The request payload defines every window, cutoff, threshold, tier rule, ordering, and rounding — treat it as ground truth and never hardcode values from memory or from past tasks. The answer_template defines the output shape. Note whether the task is analytical-only or an authorized correction (it will explicitly grant a correction with a reason_code/actor/audit_id/correction_key/success rule).

2. Connect and discover the real schema. Use the client:

   python3 skill/scripts/atlas_client.py schema
   python3 skill/scripts/atlas_client.py dict
   python3 skill/scripts/atlas_client.py sql "SELECT ... "
   

   (Or import AtlasClient from skill/scripts/atlas_client.py.) It reads environment_access.md, sends the bearer token, and retries transient 5xx. Get real table/column names and enum values from GET /api/schema and GET /api/data-dictionary; the dictionary is where domain terms like effective, canonical, logical, production, and active time are pinned for this dataset. Probe distinct values before filtering — confirm each status/tier/region/currency literal actually exists.

3. Compute with SQL through POST /api/sql (read-only). Build up the metrics the template requires. Honor these recurring rules (details and the per-task-type recipes are in references/playbook.md):

   • Include only production rows; exclude test/sandbox.
   • Evaluate state as of the stated cutoff; timestamps are exact UTC boundaries with the stated inclusivity. Membership windows and status cutoffs are usually different dates — keep them separate.
   • Compute on effective/canonical values; refunds net against reversals.
   • Keep full eligible populations in rate denominators.
   • Round only final reported values; sort and tie-break on unrounded values, applying every tie-break key the payload lists.
   • Classify status/risk top-down, first match wins, using the exact thresholds and comparison operators from the payload.

4. For an authorized correction only, use POST /api/sql/transaction to change the single minimal canonical field and insert exactly one audit row, then verify post-change and report APPLIED/NOT_APPLIED per the success rule. Leave raw source values, identity fields, and unrelated rows untouched. Never mutate for analytical-only tasks. See the correction section of references/playbook.md.

5. Assemble answer.json to the contract. Emit exactly the required keys, correct JSON types, exact enum strings, ID formats preserved, arrays sized/ ordered/deduped as specified. Details and pitfalls: references/output- contract.md.

6. Validate before finishing:

   python3 skill/scripts/validate_answer.py answer.json input/payloads/answer_template.json
   

   Fix every reported shape problem. The output file must contain only the JSON document — no prose, no code fences, no extra fields.

Files in this skill

• references/playbook.md — recurring Atlas semantics and per-task-type computation recipes (fulfillment, refund/FX, warehouse, support, correction).
• references/output-contract.md — how to read the answer_template schema and the contract mistakes that cost points.
• scripts/atlas_client.py — stdlib API client (schema / dictionary / audit / sql / transaction / introspect) with token handling and 5xx retry.
• scripts/validate_answer.py — best-effort shape validator for answer.json vs the template (does not check business correctness).

Guardrails

• The request payload and data dictionary — not this skill and not prior answers — supply all business values. Nothing task-specific is baked in here.
• Analytical tasks are read-only; only run a transaction when the request explicitly authorizes the correction and its audit record.
• If the schema endpoints keep returning 5xx, keep retrying (the client backs off) and cross-check names via introspect; don't guess table names blindly.