SuperAPI AI Directory

business-api-context

Testing & quality

View repository

Updated 10-16 07:16

About this skill

business-api-context

Kind context with a declared write scope (agentic-qa-core/references/skill-composition-strategy.md §2b). Delivered once by upstream as a placeholder, then owned by the project: the map inside is this project's synthesis. Procedure: agentic-qa-core/references/business-context-maps.md.

Compact Rules

  • DO: read the map through bun run context:map business-api-context (or --section <id> for one endpoint group). NEVER read references/business-api-map.html raw: its SVG is most of the bytes and none of the facts.
  • DO: treat a placeholder map as "no map". Say so and hand the user project-context mode api; never answer API questions as if the API were empty.
  • DO: take field names, types and required flags from the schema ([API_TOOL] schema read, api/schemas/), and the MEANING from the map. On a conflict the schema wins for shape, the running API wins for behaviour.
  • WHEN a session observes something that contradicts a section (a status, a field, an auth rule): PROPOSE the one-section edit with its evidence to the user (or to the conductor when you are a supervised worker), apply it only on approval. Procedure: references/refresh.md.
  • DO NOT: write anywhere but this skill's own references/. No Jira, no .context/, no other skill, no api/schemas/, no product code.
  • DO NOT: copy map content into this SKILL.md. Judgment goes in ## Rules or references/gotchas.md, dated and measured.
  • Before a step that uses api-schema (verifying a section) or diagrams (redrawing a figure), run the point-of-use check in agentic-qa-core/references/preflight-gate.md §8.

Read full SKILL.md when: building a briefing for an API-level dispatch, deciding whether a section is stale, or proposing an edit to the map.

What this skill knows

One aspect of the system under test: its API as a business surface. Which endpoint groups exist and what each one is for, who is allowed to call them, what they change, how a multi-call flow chains them, what each error means to a user, and which third parties sit behind them. Loading it changes what the agent KNOWS; the only thing it ever does is propose an edit to its own map.

Sources of truth (cited, never copied)

Source What lives there Role
references/business-api-map.html the synthesis, one <section> per endpoint group or cross-endpoint flow, plus discovery-gaps this skill's map (read via bun run context:map)
the OpenAPI contract ([API_TOOL] schema read, api/schemas/ after bun run api:sync) paths, methods, payload shapes wins for shape
the running API of the active environment (curl, agentic-qa-core/references/api-testing-doctrine.md) real statuses and bodies wins for behaviour
project-context mode api the generator that CREATEs the map and UPDATEs its stale sections owns regeneration

Rules (judgment, dated)

(none yet: each rule carries YYYY-MM-DD · rule · measured: how)

Not here

  • Payload schemas and types → the OpenAPI contract and api/schemas/.
  • Entities, state machines and triggers → business-data-context.
  • User journeys and the feature catalog → business-e2e-context.
  • Tokens and credentials → .env and bun run api:login, never a map.

References

  • references/business-api-map.html — the map (generated; read through the reader).
  • references/refresh.md — the self-update procedure and this aspect's staleness signals.
  • references/gotchas.md — measured traps in reading this project's API.

Refinements

Lessons land as proposals per agentic-qa-core/references/skill-refinement-protocol.md, never as direct edits.