rvas-sre-agent-jumpstart

SRE Agent Jumpstart Skill

Guide a user from an operational need to a working Azure SRE Agent capability or a newly provisioned agent resource — always plan first, then generate artifacts only after the user approves the plan.

How This Skill Works

This skill has five phases, always in this order:

1. Interview — grill the user through an SRE-agent-specific design tree (reusing the grilling skill's round/frontier protocol) to settle whether the target is a new or existing agent, which IaC surface(s) apply, the named use case, run mode, and connector and knowledge needs. See references/interview-design-tree.md.

2. Research and plan — activate the azure-sre-agent skill and follow its own Category Index procedure directly (its stated fetch tooling and freshness check) rather than re-describing that procedure here; layer only the use-case-specific lookup on top — always perform an additional targeted lookup for the named use case's domain, grounded in fetched documentation rather than any existing file in the target repository. Emit exactly one reviewable Markdown plan naming the IaC surface(s), target manifests or recipe, run mode, connector/knowledge needs, and the exact CLI commands to run, and write that plan to a file in addition to showing it in chat. See references/iac-surfaces.md.

3. Approval gate — stop and wait for the user's explicit approval of the plan. Never generate an artifact before this gate is cleared.

4. Authoring — once approved, generate capability manifests in this skill's own azuresre.ai/v1 convention under the canonical azure-sre-agent-config/ artifact layout, including its infra/ subdirectory, which is created only when Surface 1 applies (see references/manifest-templates.md). For Surface 2, write the capability manifests and copy this skill's own pre-built apply.ps1/apply.sh script pair (never regenerated, always the same vetted asset) that performs the agent's own REST API sequence — the user runs the script themselves afterward, against their own tenant; this skill's own turn never invokes, tests, or otherwise calls that script or any live agent endpoint (see references/iac-surfaces.md). For Surface 1, write the fetched agent-resource-provisioning command sequence into the plan as reference content — the user applies it via their own existing IaC procedure; this skill never executes it and never generates a script for it (see references/iac-surfaces.md) — never merging the two outputs.

5. Verification handoff — after authoring, write azure-sre-agent-config/VERIFY.md so the user can prove the capability works. It is done when every Definition of Done item from the plan has at least one sample prompt or trigger with its expected behaviour, and the plan's rollback steps are included. The user runs it; this skill only writes it.

Two IaC Surfaces, Never Conflated

Every request resolves to one or both of two genuinely different surfaces — agent-resource provisioning and capability authoring — never conflated or merged into one plan section. Read references/iac-surfaces.md for what each means, how to tell which applies, and where the separate Azure/sre-agent-plugins connector marketplace fits.

Any Use Case, Not a Fixed Catalog

This skill has no fixed list of supported use cases — whatever the user names is the input to research, never a menu to pick from. See references/iac-surfaces.md for how a named use case is matched to an existing skill or researched fresh.

Five-Active-Skill Cap

Azure SRE Agent enforces a hard cap of five active skills per agent. This check is unconditional for every existing-agent path — see the skill-cap branch in references/interview-design-tree.md for when it runs during the interview as a plan-time heads-up based on the user's own stated count, as distinct from the copied apply script's own live preflight check (run later, by the user, against their tenant), which hard-stops by default if the cap is actually exceeded.