adk-setup

Set-up, configure-overrides, init-config, refresh-metadata, verify-mcps, check-env. Stewards $ADK_CONFIG_HOME/core.yaml and the metadata cache. NOT a CLI-dep installer — brew, gh, jq, uv, node are the user’s job (SETUP.md prints the exact commands). NOT a wiring tool — install.sh handles symlinks, MCP merges, hook wiring, AGENTS.md pointers. This skill picks up where install.sh stops: filling user data files (conversationally) and introspecting MCPs (with the agent’s MCP client, which install.sh / curl cannot do). Four modes. —init: conversational scaffolding of core.yaml (workspaces, repos, data dictionary, RAG config). —enrich: queries every reachable MCP (Datadog dashboards, Statsig experiments, Mixpanel events, Snowflake schemas, Looker dashboards, Atlassian spaces, GitHub repos), summarizes findings, writes enriched: block + $ADK_DATA_HOME/improve/metadata/<source>.json. Never overwrites manually-set values. —check: superset of scripts/adk_mcp_health.py — also probes stdio MCPs (Atlassian via uvx, Slack via npx, Snowflake via uvx) via real MCP-client invocation, and offers conversational guidance when something’s broken. —diff: read-only preview of —enrich. Never modifies shell rc files. Never puts a raw token in core.yaml (regex-enforced).

Source

skills/adk-setup/SKILL.md

Frontmatter

1name: adk-setup2description: |3  Set-up, configure-overrides, init-config, refresh-metadata, verify-mcps, check-env. Stewards `$ADK_CONFIG_HOME/core.yaml` and the metadata cache. NOT a CLI-dep installer — brew, gh, jq, uv, node are the user's job (SETUP.md prints the exact commands). NOT a wiring tool — install.sh handles symlinks, MCP merges, hook wiring, AGENTS.md pointers. This skill picks up where install.sh stops: filling user data files (conversationally) and introspecting MCPs (with the agent's MCP client, which install.sh / curl cannot do). Four modes. --init: conversational scaffolding of core.yaml (workspaces, repos, data dictionary, RAG config). --enrich: queries every reachable MCP (Datadog dashboards, Statsig experiments, Mixpanel events, Snowflake schemas, Looker dashboards, Atlassian spaces, GitHub repos), summarizes findings, writes `enriched:` block + `$ADK_DATA_HOME/improve/metadata/<source>.json`. Never overwrites manually-set values. --check: superset of `scripts/adk_mcp_health.py` — also probes stdio MCPs (Atlassian via uvx, Slack via npx, Snowflake via uvx) via real MCP-client invocation, and offers conversational guidance when something's broken. --diff: read-only preview of --enrich. Never modifies shell rc files. Never puts a raw token in core.yaml (regex-enforced).4allowed-tools: [Read, Edit, Write, Bash, WebFetch]5argument-hint: "(--init [--from-v2]) | (--enrich [--source <name>|all]) | (--check) | (--diff) [--detailed] [--deep]"6metadata:7  category: core8  kind: task9  layer: 010  paths: []11  model: sonnet12  effort: low13  user-invocable: true14  disable-model-invocation: false15  needs_mcp_required: []16  needs_mcp_optional: [adk-mcp-github, adk-mcp-datadog, adk-mcp-statsig, adk-mcp-atlassian, adk-mcp-mixpanel, adk-mcp-slack, adk-mcp-snowflake, adk-mcp-looker]17  needs_meta_info: []18  forks_emitted: [init-source, enrich-sources, conflict-with-existing, model-depth]

Workflow body

adk-setup

Bootstrap + maintain $ADK_CONFIG_HOME/core.yaml.

Global skill — runs from anywhere; intermediate artifacts go to $ADK_DATA_HOME/setup/<ts>/ (per shared/paths.md). Touches $ADK_CONFIG_HOME/ (config) but not the cwd.

--detailed increases MCP/config diagnostics. --deep selects the stronger model profile per shared/model-depth.md; use it when reconciling ambiguous config, migrations from older installs, or many failing connectors.

Modes

—init

Scaffolds $ADK_CONFIG_HOME/core.yaml with full structure + comments. Behavior:

1. If $ADK_CONFIG_HOME/core.yaml exists → refuse; show user --diff instead.
2. Else → write fresh templates for core.yaml, repos.md, connectors/*.md, and links.json5.
3. Walk the user through filling: workspaces (cap 3 questions), one starter repo, RAG config.

Then: print “edit $ADK_CONFIG_HOME/core.yaml (defaults / workspaces / rag) and $ADK_CONFIG_HOME/repos.md (per-repo entries). Add new data sources by editing $ADK_CONFIG_HOME/connectors/<name>.md. Re-run /adk-setup --enrich to populate auto-discovery.”

—enrich

For each MCP (or --source <name> for one), call scripts/enrich_metadata.py:

1. Query reachable MCPs via curl / programmatic calls.
2. Write $ADK_DATA_HOME/improve/metadata/<source>.json (overwrites; archives previous).
3. Propose updates to $ADK_CONFIG_HOME/connectors/<name>.md frontmatter — only ADD; never delete manually-set values. Per-update user confirmation required.
4. Surface MCPs that couldn’t be reached (env var missing, OAuth not done, etc.) with the exact fix.

—check

For each agent (claude / cursor / codex / junie):

• Detect installation.
• Verify symlinks point to this repo’s skills/.
• Verify MCP config merged.

For each env var: present / missing / aliased.

For each MCP: reachable / not-reachable + reason, AND — if the user’s creds CLI is installed (mac-setup’s ~/.config/creds/ layout) — the creds-system probe status for the same service (OK / FAIL / MISCONFIGURED). This surfaces the common “env-vars are set but the token is invalid” case that env-presence checks alone cannot detect. The cross-reference is purely additive: --no-creds disables it, and it auto-skips on machines without the creds CLI.

For core.yaml: workspaces count, repos count, data_sources presence, defaults presence.

—diff

Read-only preview of what --enrich would change. Useful before committing.

Workflow

1Phase 0 — context-gather (minimal — this skill is mostly a config tool)2  - Detect environment (OS, agents installed, $ADK_CONFIG_HOME/ state)34Phase 1 — advise5  - Up to 3 questions depending on mode:6    --init: which workspace type (work / personal / both), confirm primary repo location, RAG enabled?7    --enrich: which sources, confirm OAuth flows if needed8    --check / --diff: no questions910Phase 2 — execute (mostly programmatic — script-driven)11  - --init: write core.yaml / repos.md / connectors/*.md / links.json5 templates12  - --enrich: run scripts/enrich_metadata.py13  - --check: build verification table14  - --diff: dry-run enrich1516Phase 3 — validate17  - File written + syntactically valid YAML18  - Enrichment didn't overwrite manual fields1920Phase 4 — report21  - --init: pointer to the file, next-steps to fill it22  - --enrich: count of fields added; list of unreachable sources with fixes23  - --check: the verification table

Hard rules

1. Never overwrite a manually-set value in core.yaml. Only the enriched: and learning_state: blocks are auto-managed.
2. Never put a token / secret in core.yaml. Regex-check before writing.
3. Never chmod or modify the user’s shell rc files. Print export lines; the user adds them.
4. Never OAuth a third-party service automatically. Print the URL; user clicks.

Persona

• Light persona (no shared file). Walks tone is “calm sysadmin”, not “enthusiastic onboarder”.

Fork IDs

Refusals

• --init when file exists → refuse, suggest --diff.
• --enrich when no MCPs configured → refuse, point at mcp/README.md + env vars.
• Migration when source file has unparseable YAML → refuse, show parse error, suggest manual fix.

References shipped

(References authored on first real use of each sub-flow.)