uipath-governance

UiPath Governance

Uber skill for UiPath governance authoring. Two backing CLI surfaces:

Both surfaces share verbs (block, restrict, deny, allow, require, enforce). The same English sentence often maps to either layer, so this skill classifies first and only then routes to the matching mechanic.

When to Use This Skill

Activate on any governance / policy / rule intent — even when the user did not name the underlying CLI:

• policy / rule / guardrail / govern / gate / control requests
• block / restrict / deny / disable / disallow an action, model, app, URL, agent, flow, or process
• require / enforce / mandate a behavior or rule
• allow only / permit only / limit to / restrict to X
• who can / which … can / on behalf of — actor- or identity-shaped governance
• compliance / posture / audit framing on top of policies
• .uipolicy file path, compliance standard, apply standard
• Standard names: ISO 42001
• check compliance, compliance posture, posture against, drift check
• is my tenant compliant, am I compliant with
• organization-wide, all tenants, entire org, across all tenants — org-scope full apply

Troubleshoot

• Investigate why a governance policy isn't taking effect
• Debug deployment precedence (user > group > tenant override)
• Evaluate access-policy rules against test scenarios (access-policy evaluate)
• Troubleshoot blocked tool/workflow invocations
• Audit which policies are deployed and to whom
• Identify license-type / product mismatch causing silent no-ops

Sibling redirects:

• Platform ops (auth, Orchestrator resources, packaging, deploy) → uipath-platform
• Authoring agents / workflows / RPA themselves → uipath-agents / uipath-rpa / uipath-maestro-flow

Critical Rules

1. Classify before authoring. First action on any governance request is to classify intent into Branch A (AOps) or Branch B (Access). Use the priors in references/disambiguation-guide.md. Never start create / update / delete until classification is settled — by user wording or by the disambiguation question.
2. Classification lives at the top. Mechanic libraries assume the branch is chosen. Do not let those flows ask "did you mean the other branch?" — that question belongs here.
3. One branch per mutation. A single user request produces a policy on one branch only. If the user wants both, run two sequential flows with two confirmation gates.
4. Each mechanic owns its own Critical Rules. Once routed, follow the branch's rules — do not relax them from this top level.
5. Never apply, restore, or remove compliance settings without user confirmation. For apply: run posture analysis first, show the plan (summary + detail). In every case: ask, then END YOUR REPLY — never run the mutating command in the same response as the confirmation question. Proceed only on an explicit y in the user's next message. The user's original request is intent, not consent. Sole exception: the user explicitly waived confirmation in advance (e.g., "don't ask for confirmation") — treat the waiver as y.
6. Always show a receipt after any apply. Present the post-apply report (settings configured, manual steps needed, Applied by / date) so the user has a record. No local file write is needed — the CLI and UiPath platform are the source of truth.
7. Compliance Standards is a preview feature — gate every compliance-pack flow. Append the preview disclaimer to user-facing compliance-standard responses, and on any uip gov compliance-packs … call returning HTTP 403 / Forbidden, stop immediately (do not retry, run no further compliance commands) and tell the user the feature requires enrolling in the preview program. Exact wording + placement in references/compliance-pack/preview-gate.md. A 403 is preview-not-enabled; a 401 is a normal login failure — do NOT conflate them.
8. Always uip login before any uip gov … command. evaluate (Access) additionally requires tenant-scoped login — see access-policy-overview-guide.md § Critical Rules.
9. Never fabricate UUIDs. Resolve every named user / group / process / agent / flow / robot / tenant via the relevant branch's lookups.

Workflow

1. Classify the intent silently — never announce routing to the user. Internal flow labels (AOps / Access / Compliance standard) are implementation details; the user sees only the outcome. Read references/disambiguation-guide.md — it lists the strong signals for each flow, the phrase patterns that need disambiguation, and the canonical worked example. If a strong signal matches, route silently. If the phrasing is ambiguous (matches AOps or Access), ask the disambiguation question and wait for a digit reply. If the user replies with anything other than 1 or 2, treat it as a re-statement of intent and re-classify. Do not run any CLI command before classification is settled — the disambiguation question itself does not need uip, and an unrelated request (platform ops, agent authoring) must redirect to a sibling skill before any setup happens here. If the request contains a standard name (ISO 42001), apply standard, compliance posture, drift check, am I compliant, is my tenant compliant, what packs are available, what packs are configured, which standards are enabled, organization-wide, disable standard, or reset / restore / undo drift on a standard → route silently to the appropriate compliance standard plugin. Read partial-apply/planning.md for scoped requests; coverage/impl.md for posture checks; catalog/impl.md for discovery; query/impl.md for information queries; full-apply/impl.md after confirming the posture plan; disable/impl.md for removal; restore/impl.md for resetting a configured standard back to its recommended settings; catalog/impl.md + state list for listing currently configured packs.
2. Verify uip and login (only after classification routes to a governance flow).

   which uip && uip --version
   uip login status --output json
   

   If not installed: npm install -g @uipath/cli. If not logged in: uip login (--authority <URL> for non-prod). For Access evaluate, login MUST be tenant-scoped. If logged in to the wrong tenant within the same org — use the fast path: uip login tenant list --output json then uip login tenant set <NAME>. Full re-login only needed for a different org or authority. See references/auth-context.md § Switching tenants.
3. Route to the chosen mechanic and follow its flow end-to-end.
   • AOps product policy → references/aops-policy/aops-policy-overview-guide.md
   • Access ToolUsePolicy → references/access-policy/access-policy-overview-guide.md
   • Compliance standard → use plugin routing from step 1 above (catalog / coverage / full-apply / partial-apply / disable / restore / query)

Disambiguation Question

When the user's intent fits both branches, render exactly this numbered list (no AskUserQuestion, no table) and wait for a digit reply:

### Which layer should this rule govern?

1. **Govern the product** — control what Studio / StudioX / Assistant / Robot / AI Trust Layer / Agent Builder *can do* (e.g. block ChatGPT inside Studio, enforce Workflow Analyzer, disable a Marketplace widget). Backed by `uip gov aops-policy`.
2. **Govern resource/tool use** — control which Actor Processes / identities can *invoke* which child Resource as a tool (e.g. block agents tagged `Sandbox` from being called, only let the finance group trigger this Flow). Backed by `uip gov access-policy`.

Reply with the number.

The canonical ambiguous prompt is "Block ChatGPT for my finance team using Studio." See references/disambiguation-guide.md for the worked-out reasoning of why both interpretations produce a working but different artifact.

Reference Navigation

Anti-patterns

• Do NOT skip the disambiguation question when the phrasing fits both branches. Mechanic libraries assume the branch is chosen and will not catch wrong-branch routing.
• Do NOT hand off to a mechanic, then ask "did you mean the other branch?". That question must happen at this top level.
• Do NOT merge AOps and Access intent into one policy. Different artifacts, different CLIs, different schemas.
• Do NOT activate this skill for platform ops. Route to uipath-platform.
• Do NOT propose skill edits when intent doesn't map to either branch. Ask the user to clarify.
• Do NOT use deployed-policy list for gap detection — it returns all rules in priority order, not the merged effective value. Use deployed-policy get <licenseType> <productName> <tenantId> to get the single effective merged policy.
• Do NOT skip the post-apply report even if apply partially fails — show what succeeded and what needs manual attention.
• For compliance pack posture analysis, use uip gov compliance-packs state coverage — do NOT use aops-policy deployed-policy commands; those are for AOps policy debugging (Branch A), not compliance pack flows.
• For full pack configuration, use state enable — do NOT manually call aops-policy create for each product; that path is only for partial/scoped configuration.
• For partial/scoped configuration (specific clauses, products, or areas only), use synthesize-formdata.mjs + aops-policy create — do NOT call state enable; state enable applies the FULL standard and cannot be scoped to specific clauses or products.
• For org-wide apply, do NOT call state enable organization — the backend does not implement org-scope enable. Instead: list tenants with uip login tenant list, then call state enable tenant <id> for each tenant individually. See references/compliance-pack/full-apply/impl.md § Org-scope deployment.
• NEVER claim a tenant is "compliant" with a standard — only that recommended settings are configured. Compliance status is determined by the customer's auditor.
• Do NOT surface policy names, product identifiers (AITrustLayer, Robot, Development), or clause IDs (A.6.2.8) as the main response unit — lead with plain-English control names and clause descriptions. Policy is an internal implementation detail. Clause IDs appear only as secondary reference in parentheses.
• Do NOT use the word "controls" in user-facing output — use "settings". The UiPath UI uses "settings" for what the standard recommends be configured.
• Do NOT narrate internal steps to the user. Never say "I ran…", "The CLI returned…", "Calling the governance CLI…", or "The API responded with…". Run commands silently and present only the interpreted result using the output templates in the reference docs. Raw JSON, UUIDs, error stacks, and CLI output are never shown — summarise errors in plain English.
• Do NOT dump raw command output. Parse every CLI response and render it as a formatted table or plain-English summary. The user sees the outcome, never the mechanism.