@uipath/skills 1.196.0 → 1.197.0-preview.0

package/skills/uipath-agents/references/lowcode/{critical-rules.md → critical-rules/critical-rules.md}

@@ -1,8 +1,10 @@
 # Critical Rules and Anti-Patterns
 
-These rules are the canonical source for low-code agent authoring. Capability files cross-reference back here; they do not restate rules.
+These rules are the canonical source for low-code agent authoring. Capability files cross-reference back here; they do not restate rules. The rules and anti-patterns in this file apply to **both** low-code autonomous and low-code conversational agents.
 
-## Critical Rules (22)
+Some additional rules and anti-patterns may only be relevant for autonomous or conversational agents. For **low-code autonomous agents**, these are defined in [autonomous-critical-rules.md](autonomous-critical-rules.md). For **low-code conversational agents**, these are defined in [conversational-critical-rules.md](conversational-critical-rules.md). Always refer to those variant-specific rules in addition to these rules when working with low-code autonomous/conversational agents.
+
+## Critical Rules (20)
 
 1. **Edit JSON files directly, except CLI-managed memory features** — the CLI supports `init` (scaffold), `refresh` (apply migrations + regenerate derived files), `validate` (strict read-only check), and `memory` (writes memory feature files). Agent configuration (prompts, schemas, settings) is done by editing `agent.json`. Resources (tools, contexts, escalations) are added as individual files in `resources/{ResourceName}/resource.json` inside the agent project directory — **not** inline in `agent.json`. Memory spaces are added with `uip agent memory` and live in `features/{FeatureName}/feature.json`. The root `agent.json` should not contain `resources` or manually-authored memory feature entries. The `refresh` command reads `agent.json` and these resource files to generate `entry-points.json` and `bindings_v2.json`.
 
@@ -12,9 +14,9 @@ These rules are the canonical source for low-code agent authoring. Capability fi
 
 4. **Keep schemas in sync** — `inputSchema` and `outputSchema` in `agent.json` must exactly mirror `input` and `output` in `entry-points.json`. Update both when adding or removing fields.
 
-5. **Use `{{input.fieldName}}` syntax** in user message templates. Do not use `$vars` or `=js:` expressions — those are flow syntax, not agent syntax.
+5. **Use `{{input.fieldName}}` syntax** in message templates. Do not use `$vars` or `=js:` expressions — those are flow syntax, not agent syntax.
 
-6. **Keep contentTokens in sync with content** — every message has both a `content` string and a `contentTokens` array. Update both when editing. See [agent-definition.md](agent-definition.md) § contentTokens Construction.
+6. **Keep contentTokens in sync with content** — every message has both a `content` string and a `contentTokens` array. Update both when editing. See [agent-definition.md](../agent-definition.md) § contentTokens Construction.
 
 7. **Do not manually edit `entry-points.json` or `bindings_v2.json`** — they are generated by `uip agent refresh`. Edit `agent.json` and `resources/{Name}/resource.json`, then re-run refresh.
 
@@ -24,29 +26,25 @@ These rules are the canonical source for low-code agent authoring. Capability fi
 
 10. **Solution must exist first** — create one with `uip solution init` before scaffolding an agent.
 
-11. **Set `folderPath` to the literal `Folder` returned by `uip solution resources list`.** Applies to both local (`Source: "Local"`) and external (`Source: "Remote"`) resources, and to tool `properties.folderPath`, context-index top-level `folderPath`, escalation `channel.properties.folderName`, and guardrail escalation `action.app.folderName`. Local resources typically carry `"solution_folder"` (their declared folder in the solution); external resources carry the human-readable, slash-separated Orchestrator folder (e.g., `"Shared"`, `"Shared/Sales/Region-EU"`). The author writes the value verbatim into `resource.json` (or into the guardrail action under `agent.json`); `uip agent refresh` propagates it into `bindings_v2.json` as `folderPath` (App resources translate `folderName` → binding `folderPath`). See [capabilities/process/process.md](capabilities/process/process.md) § Tool resource.json Shape.
+11. **Set `folderPath` to the literal `Folder` returned by `uip solution resources list`.** Applies to both local (`Source: "Local"`) and external (`Source: "Remote"`) resources, and to tool `properties.folderPath`, context-index top-level `folderPath`, escalation `channel.properties.folderName`, and guardrail escalation `action.app.folderName`. Local resources typically carry `"solution_folder"` (their declared folder in the solution); external resources carry the human-readable, slash-separated Orchestrator folder (e.g., `"Shared"`, `"Shared/Sales/Region-EU"`). The author writes the value verbatim into `resource.json` (or into the guardrail action under `agent.json`); `uip agent refresh` propagates it into `bindings_v2.json` as `folderPath` (App resources translate `folderName` → binding `folderPath`). See [capabilities/process/process.md](../capabilities/process/process.md) § Tool resource.json Shape.
 
 12. **Tools carry a `location` field.** Use `"solution"` when the row from `uip solution resources list` has `Source: "Local"` (resource is in the same solution); use `"external"` when `Source: "Remote"` (already deployed in Orchestrator). All other resource fields (`type`, `referenceKey`, `folderPath`, schemas, `exampleCalls`) follow the same rules in both cases. Connection (Integration Service) bindings are bound by `connection.id` and are exempt — no `folderPath` propagation.
 
-13. **Process tools require solution-level resource files and `debug_overwrites.json`.** Whether local or external, you need the agent-level `resources/{ToolName}/resource.json`, solution-level files under `resources/solution_folder/`, AND a `userProfile/<userId>/debug_overwrites.json` for folder resolution. `uip solution resources refresh` generates these from the bindings; for in-solution agents the package + process declarations are pre-created when the project is registered with the solution (auto-registered by `uip agent init` when run inside the solution directory; falls back to `uip solution project add` if registration was `NotInSolution` / `Skipped` / `Failed` — `OptedOut` means `--skip-solution-registration` was passed and the skip was intentional), and refresh resolves the binding against them. Without these, Studio Web will show "resource is missing in this environment". See [capabilities/process/solution-files.md](capabilities/process/solution-files.md).
+13. **Process tools require solution-level resource files and `debug_overwrites.json`.** Whether local or external, you need the agent-level `resources/{ToolName}/resource.json`, solution-level files under `resources/solution_folder/`, AND a `userProfile/<userId>/debug_overwrites.json` for folder resolution. `uip solution resources refresh` generates these from the bindings; for in-solution agents the package + process declarations are pre-created when the project is registered with the solution (auto-registered by `uip agent init` when run inside the solution directory; falls back to `uip solution project add` if registration was `NotInSolution` / `Skipped` / `Failed` — `OptedOut` means `--skip-solution-registration` was passed and the skip was intentional), and refresh resolves the binding against them. Without these, Studio Web will show "resource is missing in this environment". See [capabilities/process/solution-files.md](../capabilities/process/solution-files.md).
 
 14. **Never manually edit `storageVersion`.** It is managed by `uip agent refresh` (which writes the latest version on success) and by Studio Web on import. `uip agent validate` fails with `AgentValidationOutdated` if the version is behind — run `uip agent refresh` to migrate. If validate reports a `storageVersion` newer than supported, upgrade uipcli rather than editing the field by hand.
 
-15. **Inline agents in flows use `uipath.agent.autonomous` nodes.** The node's `inputs.source` references the inline agent's `projectId` UUID. Attached resource nodes (`uipath.agent.resource.tool.*`, `uipath.agent.resource.escalation`, `uipath.agent.resource.context.*`) use the same `inputs.source` convention to reference their `<RES_UUID>` subdirectory under `<projectId>/resources/`. The definition still declares `model.source: true`, but flow-core hoists the source identity onto `inputs.source` for the node instance; no node instance carries an instance `model` block. The agent definition lives in a subdirectory inside the flow project. See [capabilities/inline-in-flow/inline-in-flow.md](capabilities/inline-in-flow/inline-in-flow.md).
-
-16. **Never invoke other skills automatically.** If the user needs flow operations, tell them to use the `uipath-maestro-flow` skill.
-
-17. **Read [capabilities/guardrails/guardrails.md](capabilities/guardrails/guardrails.md) before authoring guardrail JSON.** The schema uses discriminator fields (`$guardrailType`, `$actionType`, `$parameterType`, `$ruleType`, `$selectorType`) that cannot be guessed. Configure guardrails at the agent.json root `guardrails` array only.
+15. **Never invoke other skills automatically.** If the user needs flow operations, tell them to use the `uipath-maestro-flow` skill.
 
-18. **`{{input.<file-field>}}` exposes metadata only.** When an input field is a `job-attachment`, the variable token renders only `ID`, `FullName`, `MimeType`, and `Metadata` into the prompt — not the file contents. To let the agent read contents, configure a file-handling built-in tool (e.g. `analyze-attachments`) and instruct the agent to call it. See [capabilities/built-in-tools/built-in-tools.md](capabilities/built-in-tools/built-in-tools.md).
+16. **Read [capabilities/guardrails/guardrails.md](../capabilities/guardrails/guardrails.md) before authoring guardrail JSON.** The schema uses discriminator fields (`$guardrailType`, `$actionType`, `$parameterType`, `$ruleType`, `$selectorType`) that cannot be guessed. Configure guardrails at the agent.json root `guardrails` array only.
 
-19. **`job-attachment` schema is canonical — copy verbatim.** Declare file fields as `{ "$ref": "#/definitions/job-attachment" }` and place the canonical block under `definitions`. `x-uipath-resource-kind: "JobAttachment"` is required. Same schema for input and output fields. See [agent-definition.md](agent-definition.md) § File Attachments.
+17. **`{{input.<file-field>}}` exposes metadata only.** When an input field is a `job-attachment`, the variable token renders only `ID`, `FullName`, `MimeType`, and `Metadata` into the prompt — not the file contents. To let the agent read contents, configure a file-handling built-in tool (e.g. `analyze-attachments`) and instruct the agent to call it. See [capabilities/built-in-tools/built-in-tools.md](../capabilities/built-in-tools/built-in-tools.md).
 
-20. **Built-in tools require explicit configuration.** Tools like `analyze-attachments` are not implicitly available — add them as `resources/{Name}/resource.json` with `type: "internal"` and `properties.toolType: "<kebab-lowercase-id>"`. The `toolType` discriminator is fixed per tool — copy from the capability reference, do not invent. See [capabilities/built-in-tools/built-in-tools.md](capabilities/built-in-tools/built-in-tools.md).
+18. **`job-attachment` schema is canonical — copy verbatim.** Declare file fields as `{ "$ref": "#/definitions/job-attachment" }` and place the canonical block under `definitions`. `x-uipath-resource-kind: "JobAttachment"` is required. Same schema for input and output fields. See [agent-definition.md](../agent-definition.md) § File Attachments.
 
-21. **Built-in tools need no solution-level files and no resource refresh.** Unlike external process tools, built-in tools (`type: "internal"`) are self-contained at the agent level. Do not run `uip solution resources refresh` for them; do not author solution-level resource files. Validate the agent and bundle.
+19. **Built-in tools require explicit configuration.** Tools like `analyze-attachments` are not implicitly available — add them as `resources/{Name}/resource.json` with `type: "internal"` and `properties.toolType: "<kebab-lowercase-id>"`. The `toolType` discriminator is fixed per tool — copy from the capability reference, do not invent. See [capabilities/built-in-tools/built-in-tools.md](../capabilities/built-in-tools/built-in-tools.md).
 
-22. **Use `uip agent memory` for memory spaces and seed items.** The memory commands own `features/{FeatureName}/feature.json`, memory item type validation, and generated `memorySpace` bindings. Attach existing memory spaces by `--memory-space` plus literal `--folder-path`; seed items require `--memory-type episodic|escalation`, and episodic seed items require `--feedback-id`. Run `uip agent refresh --output json`, `uip agent validate --output json`, and then `uip solution resources refresh --output json` after memory changes in a solution. See [capabilities/memory/memory.md](capabilities/memory/memory.md).
+20. **Built-in tools need no solution-level files and no resource refresh.** Unlike external process tools, built-in tools (`type: "internal"`) are self-contained at the agent level. Do not run `uip solution resources refresh` for them; do not author solution-level resource files. Validate the agent and bundle.
 
 ## What NOT to Do (24)
 
@@ -56,8 +54,8 @@ These rules are the canonical source for low-code agent authoring. Capability fi
 4. **Do not skip validation after a bulk of related edits** — validate after each cohesive set of changes (not after every line, but always before moving on to a new capability or publishing)
 5. **Do not publish/deploy without validating** — always validate first
 6. **Do not forget contentTokens** — editing `content` without updating `contentTokens` causes rendering issues
-7. **Do not forget `uip solution resources refresh` after adding a process tool** — applies to both local and external. Creating only the agent-level `resources/{ToolName}/resource.json` is not enough. After `uip agent refresh` generates `bindings_v2.json`, run `uip solution resources refresh` from the solution root to import the bindings into the solution. For `Connection` bindings refresh also generates `debug_overwrites.json`. For `Process`, `App`, and `Index` bindings refresh imports the resource but does not hand-write the rich solution-level files — check the output and hand-author missing files per [capabilities/process/solution-files.md](capabilities/process/solution-files.md).
-8. **`uip solution resources list` returns identity only (Source/Key/Name/Type/Folder/FolderKey)** — for the full configuration run `uip solution resources get <KEY> --output json` and read `Data.spec`. Applies to `--kind Process` (argument schemas, package keys, entry-point IDs — see [capabilities/process/process.md § Discovery](capabilities/process/process.md#discovery)) and `--kind Index` (data source type, storage bucket reference — see [capabilities/context/index.md § Discovery](capabilities/context/index.md#discovery)).
+7. **Do not forget `uip solution resources refresh` after adding a process tool** — applies to both local and external. Creating only the agent-level `resources/{ToolName}/resource.json` is not enough. After `uip agent refresh` generates `bindings_v2.json`, run `uip solution resources refresh` from the solution root to import the bindings into the solution. For `Connection` bindings refresh also generates `debug_overwrites.json`. For `Process`, `App`, and `Index` bindings refresh imports the resource but does not hand-write the rich solution-level files — check the output and hand-author missing files per [capabilities/process/solution-files.md](../capabilities/process/solution-files.md).
+8. **`uip solution resources list` returns identity only (Source/Key/Name/Type/Folder/FolderKey)** — for the full configuration run `uip solution resources get <KEY> --output json` and read `Data.spec`. Applies to `--kind Process` (argument schemas, package keys, entry-point IDs — see [capabilities/process/process.md § Discovery](../capabilities/process/process.md#discovery)) and `--kind Index` (data source type, storage bucket reference — see [capabilities/context/index.md § Discovery](../capabilities/context/index.md#discovery)).
 9. **Do not copy-paste UUIDs from one resource to another** — every resource (including each guardrail) needs a unique UUID.
 10. **Do not bump `storageVersion` manually** — breaks packager compatibility.
 11. **Do not call raw Automation.Solutions REST APIs** — always use `uip solution` commands.
@@ -65,12 +63,12 @@ These rules are the canonical source for low-code agent authoring. Capability fi
 13. **Do not pass `--for-low-code-agents` to `uip is connectors list` or `uip is activities list`** — this flag has been removed. Run the commands without it and filter yourself if needed.
 14. **Do not omit discriminator fields in guardrails** — every action needs `$actionType` (not `type`), every validator parameter needs `$parameterType` and `id` (not `name`), every custom rule needs `$ruleType`, every field selector needs `$selectorType`. Missing any discriminator causes `uip agent validate` to fail.
 15. **Do not use lowercase scope values in guardrails** — use `"Agent"`, `"Llm"`, `"Tool"` (PascalCase), not `"agent"`, `"llm"`, `"tool"`.
-16. **Always include `"guardrail": { "policies": [] }` in every tool resource.json, and never populate `policies` with entries** — the field must be present for backward-compatible solution loading; omitting it causes the runtime to fail. All guardrails are configured at the agent.json root `guardrails` array with `selector.scopes` and `selector.matchNames`.
-17. **Do not expect `uip solution resources refresh` to wire non-StorageBucket index data sources** — GoogleDrive/OneDrive/Dropbox/Confluence/Attachments indexes, `attachments` contexts, and `datafabricentityset` contexts are not auto-generated. Refresh warns + skips them; any solution-level files for these must be hand-authored. See [capabilities/context/index.md](capabilities/context/index.md).
-18. **Do not add a Tool-scoped guardrail before the tool exists** — run `uip agent tool list` first and confirm every name in `selector.matchNames` is present. `uip agent validate` will fail with an error if a guardrail targets a tool that has not been added to the agent.
-19. **Set `folderPath` on process tool resources to the literal `Folder` from `uip solution resources list`** — applies whether `Source` is `Local` or `Remote`. `uip agent validate` reads it from the agent-level `resource.json` during the schema check; `uip agent refresh` propagates it verbatim into `bindings_v2.json`. The same value goes into the matching solution-level process declaration's `folders[].fullyQualifiedName`.
-20. **Do not hand-edit `bindings_v2.json`** — the file is generated by `uip agent refresh` from the agent-level `resource.json` (including `folderPath`). Edit `resource.json` and re-run refresh; never patch the binding directly.
-21. **Do not pass attachments via the `uip` CLI** — runtime file inputs are not supported by `uip agent run` or solution-level CLI run paths. Test attachment-aware agents from Studio Web or via Orchestrator job invocation.
-22. **Do not assume `{{input.<job-attachment>}}` lets the agent read the file** — it renders metadata only. Pair the file field with a file-handling built-in tool (e.g. `analyze-attachments`) so the agent has a way to read contents.
-23. **Do not run final inline-agent refresh or validation before flow graph edits for inline-in-flow agents** — direct `.flow` authoring of the inline-agent node, capability-resource nodes, and edges is owned by the `uipath-maestro-flow` skill. Always run `uip agent refresh --inline-in-flow --bindings-target <FlowProjectDir>/bindings_v2.json` and then `uip agent validate --inline-in-flow` **after** those flow graph edits are complete. Refresh is what propagates the inline agent's bindings into the flow project's `bindings_v2.json`; without it, `uip solution resources refresh` cannot discover them. Never hand-edit `bindings_v2.json`. See [capabilities/inline-in-flow/inline-in-flow.md](capabilities/inline-in-flow/inline-in-flow.md).
-24. **Do not hand-edit memory feature files for routine changes** — use `uip agent memory add/list/remove` and `uip agent memory item add/list/remove`. Hand-writing `features/{Name}/feature.json` or `bindings_v2.json` can desynchronize generated state and skip CLI validation for search mode, field weights, metadata JSON, and memory item type.
+16. **Do not expect `uip solution resources refresh` to wire non-StorageBucket index data sources** — GoogleDrive/OneDrive/Dropbox/Confluence/Attachments indexes, `attachments` contexts, and `datafabricentityset` contexts are not auto-generated. Refresh warns + skips them; any solution-level files for these must be hand-authored. See [capabilities/context/index.md](../capabilities/context/index.md).
+17. **Do not add a Tool-scoped guardrail before the tool exists** — run `uip agent tool list` first and confirm every name in `selector.matchNames` is present. `uip agent validate` will fail with an error if a guardrail targets a tool that has not been added to the agent.
+18. **Set `folderPath` on process tool resources to the literal `Folder` from `uip solution resources list`** — applies whether `Source` is `Local` or `Remote`. `uip agent validate` reads it from the agent-level `resource.json` during the schema check; `uip agent refresh` propagates it verbatim into `bindings_v2.json`. The same value goes into the matching solution-level process declaration's `folders[].fullyQualifiedName`.
+19. **Do not hand-edit `bindings_v2.json`** — the file is generated by `uip agent refresh` from the agent-level `resource.json` (including `folderPath`). Edit `resource.json` and re-run refresh; never patch the binding directly.
+20. **Do not pass attachments via the `uip` CLI** — runtime file inputs are not supported by `uip agent run` or solution-level CLI run paths. Test attachment-aware agents from Studio Web or via Orchestrator job invocation.
+21. **Do not assume `{{input.<job-attachment>}}` lets the agent read the file** — it renders metadata only. Pair the file field with a file-handling built-in tool (e.g. `analyze-attachments`) so the agent has a way to read contents.
+22. **Do not run final inline-agent refresh or validation before flow graph edits for inline-in-flow agents** — direct `.flow` authoring of the inline-agent node, capability-resource nodes, and edges is owned by the `uipath-maestro-flow` skill. Always run `uip agent refresh --inline-in-flow --bindings-target <FlowProjectDir>/bindings_v2.json` and then `uip agent validate --inline-in-flow` **after** those flow graph edits are complete. Refresh is what propagates the inline agent's bindings into the flow project's `bindings_v2.json`; without it, `uip solution resources refresh` cannot discover them. Never hand-edit `bindings_v2.json`. See [capabilities/inline-in-flow/inline-in-flow.md](../capabilities/inline-in-flow/inline-in-flow.md).
+23. **Do not switch between autonomous and conversational variants by simply editing the `agent.json` fields** — for example, never modify `metadata.isConversational`. Instead, rely on first re-initializing a new agent of the opposite variant using the `uip agent init` command with or without the `--conversational` flag, then editing the new `agent.json`. See [project-lifecycle.md](../project-lifecycle.md) § Agent Commands.
+24. **Do not register a low-code conversational agent as a tool in another agent** — conversational agents are properly run through the UiPath Conversational Service per exchange with a threaded `messages` input; they do not match the input→output contract that agent-tools require. Only autonomous agents can be used as tools of other autonomous/conversational agents. Applies to both in-solution and deployed tool references.

package/skills/uipath-agents/references/lowcode/debug.md

@@ -1,11 +1,13 @@
 # Debug a Low-Code Agent
 
-Run a low-code agent end-to-end on Studio Web and stream the result — without publishing it to Orchestrator. Use this to test an agent against real inputs during iteration. `uip agent debug` **uploads the agent's enclosing solution** to Studio Web, starts it on the serverless debug runtime, polls to completion, and returns the agent's output.
+Run a low-code autonomous agent end-to-end on Studio Web and stream the result — without publishing it to Orchestrator. Use this to test an agent against real inputs during iteration. `uip agent debug` **uploads the agent's enclosing solution** to Studio Web, starts it on the serverless debug runtime, polls to completion, and returns the agent's output.
+
+> The debug command is currently not supported for conversational agents, so only attempt to debug autonomous agents.
 
 ## Pre-flight
 
 1. **Logged in.** `uip login status --output json` returns success.
-2. **Solution resources refreshed — only if the agent has solution-level bindings.** If it uses external process/IS tools, index contexts, memory spaces, or escalations, refresh first so those declarations stay in sync with its bindings. Agents with only built-in tools (or no resources) don't need this (see [critical-rules.md](critical-rules.md) Rule 21):
+2. **Solution resources refreshed — only if the agent has solution-level bindings.** If it uses external process/IS tools, index contexts, memory spaces, or escalations, refresh first so those declarations stay in sync with its bindings. Agents with only built-in tools (or no resources) don't need this (see [critical-rules/critical-rules.md](critical-rules/critical-rules.md) Rule 20):
 
    ```bash
    uip solution resources refresh --solution-folder <SOLUTION_DIR> --output json
@@ -13,7 +15,7 @@ Run a low-code agent end-to-end on Studio Web and stream the result — without
 
 ## Consent gate
 
-`uip agent debug` **executes the agent for real** — it calls its tools, escalations, and external APIs, and consumes model tokens. It also overwrites the agent's Studio Web solution. Per [critical-rules.md](critical-rules.md) Rule 8 (consent before upload/publish/deploy — debug uploads), confirm with the user before running it.
+`uip agent debug` **executes the agent for real** — it calls its tools, escalations, and external APIs, and consumes model tokens. It also overwrites the agent's Studio Web solution. Per [critical-rules/critical-rules.md](critical-rules/critical-rules.md) Rule 8 (consent before upload/publish/deploy — debug uploads), confirm with the user before running it.
 
 ## Debug — controlled end-to-end run
 
@@ -53,3 +55,4 @@ uip traces spans get <TraceId> --output json
 
 - **Never use `uip agent debug` as a validation step.** Use `uip agent validate` for correctness; debug is for end-to-end execution.
 - **Don't skip `uip solution resources refresh` before debug when the agent has solution-level bindings** (external tools, IS, indexes, memory spaces, escalations). Stale declarations cause runtime binding failures even when `agent.json` is correct. Agents with only built-in tools don't need it.
+- **Never attempt `uip agent debug` for low-code conversational agents.** The debug command for conversational agents is not yet supported.

package/skills/uipath-agents/references/lowcode/evaluations/evaluate.md

@@ -2,6 +2,8 @@
 
 Design and run evaluations against low-code agents using the `uip agent eval` CLI.
 
+Evaluations from the CLI are currently supported only for autonomous agents. For conversational agents, evaluations should currently be handled through the Studio Web experience after the Agent solution is uploaded.
+
 ## Quick Reference
 
 ```bash

package/skills/uipath-agents/references/lowcode/lowcode.md

@@ -14,9 +14,32 @@ Entry point for low-code agent work. Read this first after low-code mode is dete
 - Publish agent to Studio Web, pack and deploy to Orchestrator
 - Debug / run a low-code agent end-to-end on Studio Web and inspect its output
 
+## Autonomous and Conversational Variant
+
+UiPath low-code agents come in **two variants**, both sharing this skill.
+
+1. Autonomous agents are intended for any general use-cases involving execution from input → output. They may be standalone or inline in a flow.
+   
+2. Conversational agents are intended for use-cases involving multi-turn conversations / fast latency with real-time user-interaction / streamed responses.
+
+> Currently, conversational agents are only standalone and not yet supported to be inline in a flow, so inline-in-flow use-cases should always use an autonomous agent.
+> After deployment, conversational agents are interacted with through the UiPath Conversation Service, which manages conversation history and exposes a CLI/SDK for client UIs. Each user-initiated exchange invokes the agent for a single turn, streaming events back to the client.
+
+In summary:
+| Signal | Variant |
+|---|---|
+| User wants general input → output execution | **Autonomous** |
+| User wants inline-in-flow | **Autonomous** |
+| `agent.json` has `metadata.isConversational: false` or unset | **Autonomous** |
+| User wants multi-turn chat / fast latency with real-time user-interaction / streamed responses | **Conversational** |
+| Existing `agent.json` has `metadata.isConversational: true` | **Conversational** |
+| Ambiguous | Ask the user |
+
+Although most of the lifecycle and capabilities are shared between both autonomous and conversational agents, some files will have distinguishing differences between the two (for example, mark that certain features do not apply for conversational agents). You can assume that when not labeled otherwise, content from files would apply to both variants.
+
 ## Critical Rules
 
-[critical-rules.md](critical-rules.md) is the canonical source for low-code agent rules and anti-patterns. Read it every session. The rules below are the operational must-knows; the rest live in the canonical file.
+[critical-rules/critical-rules.md](critical-rules/critical-rules.md) is the canonical source for low-code agent rules and anti-patterns. Read it every session. The rules below are the operational must-knows; the rest live in the canonical file.
 
 1. **Edit JSON files directly, except CLI-managed memory features.** The CLI provides `init` (scaffold), `refresh` (apply migrations + regenerate derived files), `validate` (strict read-only check), and `memory` (writes memory feature files). Resources live in `resources/{Name}/resource.json`; memory features live in `features/{Name}/feature.json`.
 2. **Use `--output json`** on every `uip` command.
@@ -31,7 +54,7 @@ Entry point for low-code agent work. Read this first after low-code mode is dete
 Standard workflow for any low-code agent task:
 
 1. **Scaffold** — `uip solution init` (if no solution exists), then `uip agent init "<AgentName>" --output json`. Full walkthrough in [project-lifecycle.md](project-lifecycle.md) § End-to-End Example.
-2. **Edit** — open `agent.json` and `entry-points.json`. Schema reference in [agent-definition.md](agent-definition.md). **Override the scaffold model** (`gpt-4o-2024-11-20`) per [model-selection-guide.md](model-selection-guide.md) and write robust prompts per [agent-prompting-guide.md](agent-prompting-guide.md) — the scaffold ships a stale model and toy prompts.
+2. **Edit** — open `agent.json` and `entry-points.json`. Schema reference in [agent-definition.md](agent-definition.md). **Override the scaffold model** (`gpt-4o-2024-11-20`) per [model-selection-guide.md](model-selection-guide.md) and write robust prompts per [prompting/agent-prompting-guide.md](prompting/agent-prompting-guide.md) — the scaffold ships a stale model and toy prompts.
 3. **Add capabilities** — pick from the Capability Registry below. Most add `resources/{Name}/resource.json`; memory uses `uip agent memory` and writes `features/{Name}/feature.json`.
 4. **Refresh** — `uip agent refresh --output json`. Applies pending migrations and regenerates `entry-points.json` and `bindings_v2.json`. Confirm `MigrationApplied`, `StorageVersion`.
 5. **Validate** — `uip agent validate --output json`. Strict read-only. Confirm `Status: "Valid"`, `Validated` counts.
@@ -52,7 +75,7 @@ Capabilities are **orthogonal**: there is no ordering requirement among them. Ad
 | Debugging / running a low-code agent end-to-end to test it | [debug.md](debug.md) |
 | Editing `agent.json` (prompts, schemas, model, contentTokens) or `entry-points.json` | [agent-definition.md](agent-definition.md) |
 | Choosing the LLM (`settings.model`) — discover tenant models, override the scaffold default | [model-selection-guide.md](model-selection-guide.md) |
-| Writing a robust system/user prompt — skeleton, tool-call criteria, output contract, production checklist | [agent-prompting-guide.md](agent-prompting-guide.md) |
+| Writing a robust system/user prompt — skeleton, tool-call criteria, output contract, production checklist | [prompting/agent-prompting-guide.md](prompting/agent-prompting-guide.md) |
 | External tools / IS tools / index contexts / memory spaces / escalations behave unexpectedly after `uip solution resources refresh` | [solution-resources.md](solution-resources.md) |
 | Running evaluations, adding test cases, managing evaluators | [evaluations/evaluate.md](evaluations/evaluate.md) |
 
@@ -61,7 +84,7 @@ Capabilities are **orthogonal**: there is no ordering requirement among them. Ad
 | I need to... | Read first | Then |
 |--------------|------------|------|
 | Understand agent.json schema | [agent-definition.md](agent-definition.md) | |
-| Edit system prompt or user message | [agent-prompting-guide.md](agent-prompting-guide.md) (quality) | [agent-definition.md](agent-definition.md) § Messages, § contentTokens (mechanics) |
+| Edit system prompt or user message | [prompting/agent-prompting-guide.md](prompting/agent-prompting-guide.md) (quality) | [agent-definition.md](agent-definition.md) § Messages, § contentTokens (mechanics) |
 | Choose or change the model | [model-selection-guide.md](model-selection-guide.md) | [agent-definition.md](agent-definition.md) § Change Model Settings |
 | Add/remove input or output fields | [agent-definition.md](agent-definition.md) § entry-points.json | |
 | Scaffold a new agent project | [project-lifecycle.md](project-lifecycle.md) § End-to-End Example | [agent-definition.md](agent-definition.md) |
@@ -82,13 +105,13 @@ Capabilities are **orthogonal**: there is no ordering requirement among them. Ad
 | Add a memory space or seed memory items | [capabilities/memory/memory.md](capabilities/memory/memory.md) | |
 | Add an Action Center escalation (HITL) | [capabilities/escalation/escalation.md](capabilities/escalation/escalation.md) | |
 | Add guardrails (PII, harmful content, custom rules) | [capabilities/guardrails/guardrails.md](capabilities/guardrails/guardrails.md) | |
-| Embed an agent inline in a flow | [capabilities/inline-in-flow/inline-in-flow.md](capabilities/inline-in-flow/inline-in-flow.md) | |
+| Embed an autonomous agent inline in a flow | [capabilities/inline-in-flow/inline-in-flow.md](capabilities/inline-in-flow/inline-in-flow.md) | |
 | Set up Orchestrator resources | Tell the user to use the `uipath-platform` skill | |
 | Wire agent into a flow | Tell the user to use the `uipath-maestro-flow` skill | |
 
 ## Anti-patterns
 
-See [critical-rules.md](critical-rules.md) § What NOT to Do for the canonical list. Most expensive to get wrong:
+See [critical-rules/critical-rules.md](critical-rules/critical-rules.md) § What NOT to Do for the canonical list. Most expensive to get wrong:
 
 - Editing `content` without updating `contentTokens` (causes silent rendering failures)
 - Skipping `uip solution resources refresh` after adding external tools (`bindings_v2.json` never reaches the solution)

package/skills/uipath-agents/references/lowcode/model-selection-guide.md

@@ -2,7 +2,9 @@
 
 How to pick the LLM for a low-code agent's `settings.model`. The tenant is the source of truth — discover at runtime, then select; never ship the scaffold default.
 
-> `uip agent init` scaffolds `settings.model: "gpt-4o-2024-11-20"` — a 2024-era GA model with a 16384 output cap. **Always override it** with a discovered current model before validating.
+> `uip agent init` for an autonomous agent scaffolds `settings.model: "gpt-5.4"` with a 128000 output cap. **Always override it** with a discovered current model before validating.
+
+> `uip agent init --conversational` for a conversational agent scaffolds `settings.model: "anthropic.claude-sonnet-4-5-20250929-v1:0"` with a 64000 output cap. It serves as a solid default but can be overwritten with a discovered current model before validating. 
 
 ## 1. Discover (primary path)
 
@@ -47,6 +49,7 @@ Express the choice as **selection criteria over the discovered list**, not a fix
 | Reasoning / judgment / multi-step tool use | Newest GA Anthropic Sonnet or Opus, or newest GA flagship OpenAI | Strongest instruction-following + tool-call discipline |
 | Fast / cheap / high-volume classification or extraction | Newest GA `*-mini` (e.g. an OpenAI `*-mini`) or GA Haiku | Lower latency + cost; sufficient for narrow deterministic tasks |
 | Long-context (large documents, big tool outputs) | GA model with the highest `MaxTokens` | Avoids truncated output |
+| Conversational | Newest GA Anthropic Sonnet or user preference | Strongest conversational responses and instruction-following |
 
 Concrete GA examples observed on a live tenant — **illustrative only; verify against `uip agent model list`**:
 

package/skills/uipath-agents/references/lowcode/project-lifecycle.md

@@ -24,10 +24,11 @@ uip agent init "<AGENT_NAME>" --output json
 The `<path>` argument is relative or absolute; the command can run from any directory. Creates agent.json, entry-points.json, project.uiproj, and default eval directories inside the target path. Run `uip agent refresh` after editing to regenerate `entry-points.json` and `bindings_v2.json`.
 
 **Options:**
-- `--model <model>` — LLM model to use (default: `gpt-4o-2024-11-20`). This default is stale; override it post-init — discover current tenant models with `uip agent model list` and select per [model-selection-guide.md](model-selection-guide.md). Pass `--model` at init or edit `settings.model` after.
+- `--conversational` - Pass to initialize a conversational agent. When not passed, an autonomous agent is initialized.
+- `--model <model>` — LLM model to use (default: `gpt-5.4` for autonomous, `anthropic.claude-sonnet-4-5-20250929-v1:0` for conversational). This default is stale; override it post-init — discover current tenant models with `uip agent model list` and select per [model-selection-guide.md](model-selection-guide.md). Pass `--model` at init or edit `settings.model` after.
 - `--system-prompt <prompt>` — Initial system prompt for the agent
 - `--force` — Overwrite existing directory if non-empty
-- `--inline-in-flow` — Scaffold an inline agent inside a flow project (see below)
+- `--inline-in-flow` — Scaffold an inline agent inside a flow project (see below). Only applicable for autonomous agents, since adding inline conversational-agents within a flow project is currently not an enabled feature.
 
 #### Inline mode: `--inline-in-flow`
 
@@ -142,7 +143,7 @@ For discovery, retrieval settings, memory item types, and troubleshooting, see [
 
 ### `uip agent debug`
 
-Run the agent end-to-end on Studio Web and stream the result. **Uploads the enclosing solution** to Studio Web, then runs it — one step, no separate `uip solution upload`, so the debugged copy always matches local. **Executes the agent for real** — confirm with the user first (per [critical-rules.md](critical-rules.md) Rule 8: consent before upload/publish/deploy).
+Run the autonomous agent end-to-end on Studio Web and stream the result. **Uploads the enclosing solution** to Studio Web, then runs it — one step, no separate `uip solution upload`, so the debugged copy always matches local. **Executes the agent for real** — confirm with the user first (per [critical-rules/critical-rules.md](critical-rules/critical-rules.md) Rule 8: consent before upload/publish/deploy).
 
 ```bash
 uip agent debug <AGENT_PROJECT_DIR> --inputs '{"input":"..."}' --output json
@@ -150,6 +151,8 @@ uip agent debug <AGENT_PROJECT_DIR> --inputs '{"input":"..."}' --output json
 
 Returns `Code: "AgentDebug"` with `Data.State`, `Data.Output`, and `Data.TraceId`. A `Faulted` run returns `Result: "Failure"` (exit 1); inspect it with `uip traces spans get <TraceId> --output json`. Full options and reporting in [debug.md](debug.md).
 
+> The debug command is currently not supported for conversational agents, so only attempt to debug autonomous agents.
+
 ## Solution Commands
 
 ### Create Solution
@@ -324,16 +327,16 @@ When the fallback is needed, `uip solution project add` automatically finds the
 
 ### Step 3 — Configure agent.json
 
-Read [agent-definition.md](agent-definition.md) for the full schema.
+Read [agent-definition.md](agent-definition.md) for the full schema, which differs between autonomous and conversational agents.
 
-1. Set `settings.model` — discover with `uip agent model list`, select per [model-selection-guide.md](model-selection-guide.md) (override the scaffold default `gpt-4o-2024-11-20`)
+1. Set `settings.model` — discover with `uip agent model list`, select per [model-selection-guide.md](model-selection-guide.md) (override the scaffold default `gpt-5.4` for autonomous, `anthropic.claude-sonnet-4-5-20250929-v1:0` for conversational)
 2. Set `settings.temperature` (0 for deterministic)
-3. Write system prompt in `messages[0].content` + rebuild `contentTokens` — structure it per [agent-prompting-guide.md](agent-prompting-guide.md) (skeleton, tool-call criteria, output contract), not a placeholder
-4. Write user message template in `messages[1].content` using `{{input.fieldName}}` + rebuild `contentTokens`
+3. Write system prompt in `messages[0].content` + rebuild `contentTokens` — structure it per [prompting/agent-prompting-guide.md](prompting/agent-prompting-guide.md) (skeleton, tool-call criteria, output contract), not a placeholder
+4. For autonomous agents, write user message template in `messages[1].content` using `{{input.fieldName}}` + rebuild `contentTokens`. Conversational agents should always have the user message template left blank since each user message is received during the actual conversation.
 
 ### Step 4 — Define input/output schemas
 
-1. Add fields to `agent.json` → `inputSchema` and `outputSchema`
+1. Add fields to `agent.json` → `inputSchema` and `outputSchema`. Note that modifying `outputSchema` only applies for autonomous agents.
 2. Mirror in `entry-points.json`
 3. Refresh (writes migrated files + regenerates `entry-points.json` and `bindings_v2.json`): `uip agent refresh "<SOLUTION_NAME>/<AGENT_NAME>" --output json`
 4. Validate: `uip agent validate "<SOLUTION_NAME>/<AGENT_NAME>" --output json`

package/skills/uipath-agents/references/lowcode/prompting/agent-prompting-guide.md

@@ -0,0 +1,17 @@
+# Agent Prompting Guide
+
+Robust prompts for low-code agents. This guide contains shared prompting details for autonomous and conversational agents, and then references the specific prompting guides for both variants.
+
+## Shared
+
+This guide owns prompt **quality**; [agent-definition.md](../agent-definition.md#contenttokens-construction) owns the `contentTokens` **mechanics** (keep `content` ↔ `contentTokens` in sync after every edit here).
+
+> Default scaffolds ship toy or empty prompts ("You are a helpful agentic assistant" / "What is the current date?" / ""). Replace them. A placeholder system prompt is the single biggest quality gap in a scaffolded agent.
+
+## Autonomous Agents
+
+When building a low-code autonomous agent, refer to [autonomous-agent-prompting-guide.md](autonomous-agent-prompting-guide.md) for the autonomous-specific prompting guide.
+
+## Conversational Agents
+
+When building a low-code conversational agent, refer to [conversational-agent-prompting-guide.md](conversational-agent-prompting-guide.md) for the conversational-specific prompting guide.

package/skills/uipath-agents/references/lowcode/{agent-prompting-guide.md → prompting/autonomous-agent-prompting-guide.md}

@@ -1,11 +1,9 @@
-# Agent Prompting Guide
+# Autonomous Agents Prompting Guide
 
-Robust prompts for low-code agents. This guide owns prompt **quality**; [agent-definition.md](agent-definition.md#contenttokens-construction) owns the `contentTokens` **mechanics** (keep `content` ↔ `contentTokens` in sync after every edit here).
+This is the prompting guide for low-code autonomous agents.
 
 "Coding-agent-centric" = the prompt makes the embedded agent behave like a disciplined tool-using agent: explicit tool-call criteria, stop conditions, structured output. Maps to the agent's `tool` artifact ports and `outputSchema`.
 
-> Default scaffolds ship toy prompts ("You are an agentic assistant" / "What is the current date?"). Replace them. A placeholder system prompt is the single biggest quality gap in a scaffolded agent.
-
 ## 1. System-prompt skeleton
 
 Copy this skeleton, fill every slot. Consistent structure → consistent runs. Put role/behavior here; data/task goes in the user message (§2).
@@ -54,10 +52,10 @@ The user message carries the task and the data — not the role.
 
 Token form depends on context:
 
-- **Inline-in-flow agents** reference upstream flow nodes: `{{ $vars.<flowNodeId>.output[.<field>] }}`. See the [uipath-maestro-flow inline-agent prompt-wiring guide](../../../uipath-maestro-flow/references/author/references/plugins/inline-agent/impl.md#wiring-flow-variables-into-agent-prompts).
+- **Inline-in-flow agents** reference upstream flow nodes: `{{ $vars.<flowNodeId>.output[.<field>] }}`. See the [uipath-maestro-flow inline-agent prompt-wiring guide](../../../../uipath-maestro-flow/references/author/references/plugins/inline-agent/impl.md#wiring-flow-variables-into-agent-prompts).
 - **Standalone agents** reference declared inputs: `{{input.<field>}}`.
 
-Mirror every `{{ ... }}` in `contentTokens[]` per [agent-definition.md § contentTokens Construction](agent-definition.md#contenttokens-construction).
+Mirror every `{{ ... }}` in `contentTokens[]` per [agent-definition.md § contentTokens Construction](../agent-definition.md#contenttokens-construction).
 
 ## 3. Grounding in wired data
 
@@ -70,7 +68,7 @@ Realistic inline-in-flow agent. Note the **structured `outputSchema`**, not a ba
 **Before (toy):**
 
 ```json
-"settings": { "model": "gpt-4o-2024-11-20" },
+"settings": { "model": "gpt-5.4" },
 "outputSchema": { "type": "object", "properties": { "content": { "type": "string" } } },
 "messages": [
   { "role": "system", "content": "You are an assistant." },
@@ -134,13 +132,13 @@ A robust agent is more than its prompt. Each field: default, and when to change.
 | `outputSchema` | Scaffold gives a single `content` string | **Almost always** — define typed fields a downstream node can consume. Bare `content` forces brittle string-parsing. |
 | `settings.temperature` | `0` | Keep `0` for extraction/classification/judgment. Raise only when output *variation* is wanted (drafting, brainstorming). |
 | `settings.maxIterations` | `25` | Lower (≤5) for single-shot classification. Higher for multi-tool research loops. |
-| `settings.maxTokens` | Scaffold value | Set ≤ the model's `MaxTokens` cap — see [model-selection-guide.md](model-selection-guide.md#1-discover-primary-path). |
-| `settings.model` | `gpt-4o-2024-11-20` | **Always override** — discover + select per [model-selection-guide.md](model-selection-guide.md). |
-| `guardrails` | `[]` | Add input/output policy enforcement (PII, content, escalation). See [capabilities/guardrails/guardrails.md](capabilities/guardrails/guardrails.md). |
+| `settings.maxTokens` | Scaffold value | Set ≤ the model's `MaxTokens` cap — see [model-selection-guide.md](../model-selection-guide.md#1-discover-primary-path). |
+| `settings.model` | `gpt-5.4` | **Always override** — discover + select per [model-selection-guide.md](../model-selection-guide.md). |
+| `guardrails` | `[]` | Add input/output policy enforcement (PII, content, escalation). See [capabilities/guardrails/guardrails.md](../capabilities/guardrails/guardrails.md). |
 
 ## Anti-patterns
 
-- **Vague role** — "You are an assistant." Name the role and bound the scope.
+- **Vague role** — "You are a helpful agentic assistant." Name the role and bound the scope.
 - **No output contract** — agent free-forms prose; downstream nodes can't parse it.
 - **Bare `content` output** — a single string where typed fields belong. Define `outputSchema`.
 - **No tool-call criteria** — agent over-calls tools or loops to `maxIterations`.

package/skills/uipath-agents/references/lowcode/prompting/conversational-agent-prompting-guide.md

@@ -0,0 +1,123 @@
+# Conversational Agents Prompting Guide
+
+This is the prompting guide for low-code conversational agents.
+
+## 1. System-prompt skeleton
+
+The nature of conversational agents means that they involve multi-turn interaction with the user with flexible response shapes. Thus, conversational system prompts are more flexible than autonomous prompts; the following system-prompt skeleton serves as a starting point, not a contract. Adapt this skeleton to the use case: slots can be dropped or extended based on requirements and domain-specific goal(s).
+
+```text
+You are <ROLE> for <DOMAIN>. <ONE-LINE PURPOSE>. <DESCRIBE TONE — e.g. warm, concise, professional>.
+
+Scope:
+- In scope: <topics or tasks you engage with>
+- Out of scope: <what to refuse, deflect, or hand off>
+
+Conversation behavior:
+- Goals: <desired outcome(s) of the conversation>.
+- Steps: <expected sequence; may not be needed for free-form chat>.
+- Clarifications: <when to ask before answering>.
+- Confirmations: <when to confirm before actions>.
+
+Tools:
+- <toolName>: call when <explicit condition>. Do not call when <condition>.
+- <toolName>: ...
+- After tool use, explain <what to summarize, recommend, or ask next>.
+
+Response style:
+- Default to <length — e.g. 2–4 sentences>; expand on request.
+- Use <format — markdown, bullets, JSON, prose> when helpful.
+
+Uncertainty:
+- If <context> is missing or ambiguous, <ask clarifications / run tools> — do not guess.
+```
+
+Slot rules (flexible):
+- **Role + tone** — name the role, set the voice.
+- **Scope** — bound what the agent engages with. Without it the agent answers off-task prompts.
+- **Conversation behavior** — desired conversational outcomes and steps for how the agent should drive the conversation. Slots can be omitted/modified based on use-case.
+- **Tools** — Define use-case-specific triggers and non-triggers.
+- **Response style** — Set default length and format.
+- **Grounding** — State what the agent must not invent and which sources are authoritative for the domain. Cuts hallucination.
+
+> For conversational agents, you may also template per-exchange `inputSchema` fields into the **system prompt** (not the user message, which stays blank for conversational agents). See [agent-definition.md § Input Schema](../agent-definition.md#input-schema). Only add `inputSchema` fields when the use case needs variable-based context from the chat-UI beyond the conversation history and the user's input messages (which are already captured in the implicit `messages` field for all conversational-agents). Reference them as `{{input.<field>}}` and mirror in `messages[0].contentTokens` per [agent-definition.md § contentTokens Construction](../agent-definition.md#contenttokens-construction).
+
+> For low-code conversational agents, the above defined system-prompt within the `agent.json` will be wrapped by more prompt details in the agent-runtime which describes citation generation behavior (e.g. for web-urls and context-grounding results). Do not define citation-generation format in the system-prompt.
+
+## 2. User-prompt
+
+Leave user-message contents blank for low-code conversational agents in the `agent.json`. See [agent-definition.md § User Message](../agent-definition.md#user-message).
+
+## 3. Worked example — customer support chat
+
+Realistic conversational agent: an order-status helpdesk with one tool. Note `inputSchema` carrying a per-exchange `currentOrderNumber` (the order the user currently has open in the host UI) and **empty `outputSchema`** (runtime streams conversation events).
+
+`agent.json` excerpt:
+
+```json
+"settings": { "model": "anthropic.claude-sonnet-4-6", "temperature": 0, "maxTokens": 64000, "engine": "conversational-v1", "mode": "standard" },
+"inputSchema": {
+  "type": "object",
+  "properties": {
+    "currentOrderNumber": { "type": "string", "description": "Order number currently open in the host UI; supplied by the client on every exchange" }
+  },
+  "required": ["currentOrderNumber"]
+},
+"outputSchema": { "type": "object", "properties": {} }
+```
+
+System prompt (filled skeleton):
+
+```text
+You are an company's customer support assistant for online retail. Help customers with order status, delivery, and returns. Tone: warm, concise, professional.
+
+The customer currently has the following order open: {{input.currentOrderNumber}}.
+
+Scope:
+- In scope: order status, delivery estimates, returns and refunds, basic product questions.
+- Out of scope: pricing negotiations, escalations requiring human review — hand off to a human agent.
+
+Conversation behavior:
+- Goals: resolve the customer's queries about the currently opened order in as few turns as possible.
+- Steps: greet on turn 1, reference the opened order, and explain capabilities. Then answer the queries and confirm resolution.
+- Clarifications: if the customer asks about a different order, ask once for that order number.
+- Confirmations: confirm with the customer before initiating a return.
+
+Tools:
+- lookupOrder: call with the currently opened order number unless the customer specifies a different one. Use for status, delivery, or item details.
+- initiateReturn: call only after the customer confirms in plain language they want to start a return.
+- After tool use, summarize the result in 1-2 sentences and ask whether anything else is needed.
+
+Response style:
+- Default to 2-4 sentences; expand only when the customer asks for detail.
+- Use plain prose; switch to bullets for multi-item summaries (e.g., multiple shipment events).
+
+Uncertainty:
+- If `lookupOrder` returns no record, ask the customer to verify the order number — do not guess details.
+```
+
+User message: `""` — left blank. The Conversational Service injects the user turn each exchange.
+
+## 4. Production checklist — adjacent `agent.json` quality fields
+
+| Field | Default | Change when |
+|-------|---------|-------------|
+| `inputSchema` | `{ "properties": {} }` | Add fields only when per-exchange, variable-based context beyond conversation history is genuinely needed. Reserved names: `messages`, `uipath__*` ([critical-rules/conversational-critical-rules.md](../critical-rules/conversational-critical-rules.md) Anti-patterns 4 and 5). |
+| `outputSchema` | `{ "type": "object", "properties": {} }` | **Never populate** — runtime streams events, does not fill output ([critical-rules/conversational-critical-rules.md](../critical-rules/conversational-critical-rules.md) Anti-pattern 1). |
+| `messages[1].content` | `""` | **Keep blank** — Conversational Service injects the user turn at runtime ([critical-rules/conversational-critical-rules.md](../critical-rules/conversational-critical-rules.md) Anti-pattern 3). |
+| `settings.temperature` | `0` | Raise for open-ended brainstorming or casual chats. Keep `0` for factual support flows. |
+| `settings.maxTokens` | `64000` | Set ≤ the model's `MaxTokens` cap — see [model-selection-guide.md](../model-selection-guide.md#1-discover-primary-path). |
+| `settings.model` | `anthropic.claude-sonnet-4-5-20250929-v1:0` | **Always verify** — discover + select per [model-selection-guide.md](../model-selection-guide.md). |
+| `guardrails` | `[]` | Tool-scope only; mirror in tool `resource.json`. See [capabilities/guardrails/guardrails.md](../capabilities/guardrails/guardrails.md) ([critical-rules/conversational-critical-rules.md](../critical-rules/conversational-critical-rules.md) Critical Rule 1). |
+
+## Anti-patterns
+
+- **Vague role** — "You are a helpful agentic assistant." Name the role and bound the scope.
+- **No tool-call criteria** — agent over-calls or under-calls tools.
+- **Long tool-call loops** - agent runtime may stop and require the user to confirm continuation after a single agent run (turn) consists of a series of over 8 steps that each involve tool-call(s). Note that this is not a limitation on total parallel tool-calls on any individual step, so aim to parallelize tool-calls when possible and/or ask for user-confirmation to break up long loops of sequential steps.
+- **Populating `outputSchema`** — runtime streams events; populated schemas never get filled and confuse the agent ([critical-rules/conversational-critical-rules.md](../critical-rules/conversational-critical-rules.md) Anti-pattern 1).
+- **Templating data into the user message** — the user message content stays blank; per-exchange context goes into the **system prompt** via `inputSchema` templating.
+- **Adding `messages` or `uipath__*` to `inputSchema`** — reserved names; runtime injects ([critical-rules/conversational-critical-rules.md](../critical-rules/conversational-critical-rules.md) Anti-patterns 4 and 5).
+- **Using `Agent` or `Llm` guardrail scopes** — silently ignored; only Tool-scope guardrails apply ([critical-rules/conversational-critical-rules.md](../critical-rules/conversational-critical-rules.md) Critical Rule 1).
+- **Defining citation-generation format in the system prompt** — agent runtime wraps citation formatting around the prompt; redefining it conflicts or confuses citation generation (see § 1 callout).
+- **Cargo-culted `temperature`** — copying a nonzero temperature into a deterministic, factual-based conversation task.

package/skills/uipath-agents/references/lowcode/solution-resources.md

@@ -222,7 +222,7 @@ The `solutionsSupport: "true"` metadata flag signals to the deployment engine th
 
 > **Note 1: `solutionsSupport` is a stringified boolean** (`"true"`, not `true`). `uip agent refresh` and `uip solution resources refresh` emit the string form — preserve it verbatim when round-tripping. Re-typing it as a JSON boolean breaks downstream parsing.
 >
-> **Note 2: do not hand-edit `bindings_v2.json`.** The binding's `folderPath` is generated from the agent-level `resource.json` or memory feature file. Edit the resource.json, or use `uip agent memory` for memory features, then re-run `uip agent refresh`; never patch the binding directly. See [critical-rules.md](critical-rules.md) Anti-pattern 20 and 24.
+> **Note 2: do not hand-edit `bindings_v2.json`.** The binding's `folderPath` is generated from the agent-level `resource.json` or memory feature file. Edit the resource.json, or use `uip agent memory` for memory features, then re-run `uip agent refresh`; never patch the binding directly. See [critical-rules/critical-rules.md](critical-rules/critical-rules.md) Anti-pattern 19 and [critical-rules/autonomous-critical-rules.md](critical-rules/autonomous-critical-rules.md) Anti-pattern 2.
 
 ## Debug Overwrites
 

package/skills/uipath-api-workflow/SKILL.md

@@ -65,7 +65,7 @@ Do NOT use for: `.flow` Maestro flows (→ `uipath-maestro-flow`), `.xaml` / cod
     - **On-disk is authoritative.** Even with the single-expression workaround, every StudioWeb designer save can re-trigger normalization passes that may corrupt the Response shape. After any designer roundtrip, re-validate with `uip api-workflow run --no-auth` and re-apply the workaround if needed. Until SW-28452 ships a fix, treat the file on disk as truth, not what the designer renders.
 16. **Connector activities (HTTP + Integration Service) come from `uip api-workflow registry resolve` + `stub` — never hand-author or guess.** The stub computes `metadata.configuration`, the kind (`UiPath.Http` vs `UiPath.IntSvc`), the endpoint (with hub prefix), `SlotKey`, and `ExportBucketKey` (which can differ — HTTP slot `HttpRequest_1` vs bucket `http_request_1`). Use all of them verbatim; NEVER invent a `uiPathActivityTypeId`, hand-author `metadata.configuration`, or reconstruct a key from `objectName`. Non-negotiables (full step-by-step, field-shape rules, multipart, and worked examples in [references/connector-activity-discovery.md](references/connector-activity-discovery.md)):
     - **A keyword `resolve` miss is NOT proof no curated activity exists — verify connector-first before giving up.** `resolve` AND-matches every token, so a marketing phrase + guessed verb over-narrows (the product "UiPath Data Fabric" carries `connectorKey: uipath-uipath-dataservice` and activity names like "Create Entity Record" — `resolve "data fabric insert"` returns 0; fewer/truer tokens, not more). Before concluding none exists or falling back to a hand-built HTTP call (a Rule 0 fork): map the product/vendor → connector key with `uip is connectors list --filter "<product>"`, then enumerate with `uip is activities list <connector-key>`. Do NOT hardcode/guess the key — look it up. See the reference's Step 1 recovery.
-    - **IntSvc/vendor activities require a *pinged* connection.** `uip is connections ping <uuid>` must succeed before authoring — listing-state ≠ runtime-state; an `Enabled` connection can still 401 in cloud. On failure, try the unfiltered `uip is connections list` fallback and re-ping a different `Id` for that `ConnectorKey`.
+    - **IntSvc/vendor activities require a *pinged* connection.** `uip is connections ping <uuid>` must succeed before authoring — listing-state ≠ runtime-state; an `Enabled` connection can still 401 in cloud. An empty listing is NOT proof no connection exists — `uip is connections list` is folder-scoped. On empty/failed listing, walk the fallbacks in order: unfiltered `uip is connections list`, then `uip is connections list --all-folders` (catches connections in other folders), re-pinging a different `Id` for that `ConnectorKey` each time.
     - **No connection pings cleanly → STOP and ask the user — do not decide alone.** Offer: **(a)** continue with a placeholder (stub without `--connection-id`, leaving the `<REPLACE_WITH_VENDOR_CONNECTION_UUID>` sentinel — workflow is structurally complete but 401s until replaced; only with explicit user consent), or **(b)** stop and wait for the user to create/fix the connection, then re-ping. Never silently emit the placeholder, never silently abort. (Instance of Rule 0 — escalate design forks.)
     - **NEVER ship a `<REPLACE_WITH_*>` placeholder** in `with.connectionId` / `connectionResourceId` / Http `bodyParameters.url`. StudioWeb renders it as a broken connection and the workflow 401s. The placeholder is a sentinel for "re-stub with the real value," not a fill-in-later field.
     - **After every stub, cross-check required fields** — the stub drops `required: true` request fields (e.g. Outlook `getNewestEmail` needs `parentFolderId`). Confirm via `uip is resources describe ... --operation <op>` or the stub's own `metadata.configuration` inputFields; re-stub with `--inputs` if missing.