Add embedded Claude Code Skills + AGENTS.md for AI agents

[Snodgrass-Wilkerschnoz]

What & why

Extends the CLI's agent integration with the operational monitoring lifecycle, not just monitor setup. It adds eight current-format Claude Code Skills embedded in the binary, plus a repo-root AGENTS.md so non-Claude agents (OpenAI Codex, Cursor, Gemini CLI, …) get the same guidance from one source of truth.

This is purely additive. The existing setup-* skills, skills install --claude-code, --codex, --dir, and onboard behave exactly as before.

Skills added

Embedded under skills/claude-skills/<name>/SKILL.md:

Skill	Purpose	Mutates config?
apimetrics-project-bootstrap	Select a project and stand up verified scheduled monitoring	Yes; explicit invocation
apimetrics-weekly-health-review	Analyze the previous seven complete days	No
apimetrics-failure-investigation	Root-cause one result or a failure series	No
apimetrics-incident-triage	Blast radius and common cause across monitors	No
apimetrics-monitoring-estate-audit	Coverage, hygiene, stale/duplicate monitors	No
apimetrics-slo-review	SLO design, attainment, error-budget risk	No
apimetrics-config-as-code	Safe bulk checkout/diff/push workflow	Yes; explicit invocation
apimetrics-performance-analytics	Latency percentiles, component timing, DNS analytics	No

CLI change (additive)

skills/skills.go:

• New embed //go:embed claude-skills (separate from the untouched embed/*.md).
• New flag apimetrics skills install --claude-skills → installs .claude/skills/<name>/SKILL.md (current Claude Code Skills format; preserves disable-model-invocation/argument-hint, so read-only skills auto-load and mutation skills stay gated).
• Existing --claude-code (flat .claude/commands/), --codex, --dir, and onboard are unchanged.

AGENTS.md (non-Claude agents)

New repo-root AGENTS.md — the cross-agent standard read by OpenAI Codex, Cursor, Gemini CLI, and others. It carries the core CLI facts (flat commands, JSON-on-stdin, no --api-key, output/filtering, the non-uniform list envelopes, discovery-first) and a workflow index that points at the skill files. Any agent can also run apimetrics onboard to pull the full workflows.

Grounding

Every command name and JSON shape is grounded in the built CLI's actual command surface (branch develop, commit 9697405; 213-command tree). Notable corrections baked into the skills: get-result is a summary (forensics come from get-result-content/query-*/conformance-results); percentiles are server-computed via query-*-performance; API assertions use set-call-conditions; SLOs are CRUD-only (no attainment endpoint); MCP uses run-monitor; and there is no --api-key flag.

Testing

• go build ./... and go build -tags prod — pass.
• go test ./... — pass.
• apimetrics skills install --claude-skills writes the eight .claude/skills/<name>/SKILL.md files.
• apimetrics skills install --claude-code still writes the three flat .claude/commands/*.md (unchanged).

Follow-ups (not in this PR)

• The three existing setup-* embedded skills still reference a non-existent --api-key flag in their 401/403 recovery notes; worth a one-line fix each (left out here to keep this PR additive).
• No tests currently guard the skills package (frontmatter validity, name==dir slug, embedded inventory, no stale names) — recommended as a small follow-up.

🤖 Generated with Claude Code

[rujames]

❗ Much better to derive attainment from the get-incidents operation (which seems to have been missed, or misunderstood) than to start from scratch.