infra-context

infra-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 infra-context (or --section backend, --section environments, ...). NEVER read references/infra-map.html raw: its SVG is most of the bytes and none of the facts.
• DO: treat a placeholder map as "no infra map". Say so and hand the user project-discovery Phases 2-3; never guess a run command, an environment URL or an auth flow.
• DO: take environment URLs from .agents/project.yaml and credentials from .env. The map says which environments exist and how they differ; it never holds a secret or a value that has its own source.
• WHEN a session observes something that contradicts a section (a command that no longer runs, a pipeline step, an auth request that changed): 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 CI files, no other skill, 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 diagrams (redrawing the architecture figure), run the point-of-use check in agentic-qa-core/references/preflight-gate.md §8. A table is fine where it says it better.

Read full SKILL.md when: building a briefing for framework adaptation, an environment or CI change, an auth-in-tests decision or a non-functional check; deciding whether a section is stale; or proposing an edit to the map.

What this skill knows

One aspect of the system under test: how it is built and where it runs. Its containers and external services, the stacks and commands of each side, how a user authenticates, which environments exist, how code reaches them, and the budgets that make a slow or fragile build a defect. 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/infra-map.html	the synthesis: overview, architecture, external-services, nfr-<slug>, backend, frontend, environments, ci-cd, discovery-gaps	this skill's map (read via bun run context:map)
the product repos ({{BACKEND_REPO}}, {{FRONTEND_REPO}}), their manifests, compose files and CI workflows	the real stack, commands and pipelines	wins over the map on any conflict
.agents/project.yaml environments	environment URLs	the only source for a URL
.env / .env.example	credentials and their key names	the only source for a secret
.context/ADR/	the hard-to-reverse test-architecture decisions	cited from architecture
project-discovery Phases 2-3	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

• Business vocabulary and the business model → business-domain-context.
• Entities, triggers and state machines → business-data-context.
• Endpoint groups and error semantics → business-api-context.
• This framework's own KATA wiring → test-framework-adaptation, and the code under tests/.

References

• references/infra-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 stack.

Refinements

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