cAppDeveloper docsUse Agent to inspect the selected application, search evidence, propose changes and run granted workflows. Agent's language model proposes work; cApp's scoped tools execute it and record the result. This guide shows how to start a useful session and verify its effects.
Before you begin
Open a project and establish the intended application component. You need cApp's connected local service for brokered operations. For local model-driven chat, include Local AI, load a compatible model and check its readiness. Saved sequence workflows and knowledge tools can remain available without Local AI; that is not the same as having a model answering chat.
Choose a small, explicit first task. For inspection, use an Observe session and a read-only workflow. Review Agent approvals before granting source writes, builds or application actions: new built-in chats can automatically approve actions already within their grants.
Open the right project first
The Agent page names the project and component. A component called main is not a project selector. Keep the names visible and open another project explicitly before asking Agent to work on it; a matching folder name in a prompt is not authority to switch workspaces.
Choose the workflow that fits the result: review, knowledge search, proposed changes or compilation. Its grants define which tools the session may use. A model, imported prompt or saved workflow cannot grant itself more capabilities.
Use the left-side Inference, Models, Workflows and Activity menus. Models opens the Local AI library. The Studio prompt box is a separate manual model conversation, not the Agent workflow executor. In compact layouts, Chat options contains workflow, theme and new-chat controls.
Press Enter to send and Shift+Enter for a new line; Ctrl/Command+Enter also sends. Leaving the page keeps the same conversation in floating chat. Minimising, navigating or choosing a chat theme does not cancel the turn. Latest resumes following incoming replies after you scroll back.
Start with a verifiable request
- Confirm the selected project, component, workflow and approval policy.
- Ask for a bounded result, such as “Review this application's entry point without changing files” or “Compile the current project.” Name Windows or Android for an installer request.
- Watch the reported state: waiting for the model, tool execution, compilation or review. Use Debug → Work plan and recorded steps to inspect progress when needed.
- Handle any review on the existing operation. Do not send the same mutation again merely because preparation or a reply was slow.
- Open View result and the relevant Activity receipt. Check the project, component and actual result before requesting the next change.
You can type the next draft while work runs, but Send remains disabled until the current work finishes or stops. A draft is not sent automatically. A compile/package request without a target may ask you to choose one before execution.
Choose the right evidence
Use Guide for how cApp works, project knowledge for the application's documents, and scoped source tools for its implementation. Ask explicitly when you need both product guidance and project evidence. Read the cited source when a claim affects a build or write.
Keyword retrieval needs no embedding model. Hybrid retrieval needs the separately configured embedding model and reports fallback when unavailable. Include historical project documents only for historical investigations. More retrieved text is not automatically better evidence, and no match is not permission to invent an API.
Approvals and operations
Open Inference settings → Agent approvals. Approve all granted actions starts enabled for new built-in chats/workflows. It can allow linked-source edits, compilation, installer builds and declared app tasks without a separate Agent prompt, but only inside the session's existing grants. It saves for the selected project without reloading the model.
Turn it off to use the selected autonomy mode's review behaviour. Use Observe for work that should remain read-only. The switch applies to new built-in sessions, not a request already waiting for approval or an external agent. Check the current session shown in the panel rather than assuming its grants changed.
| Control | Effect | Not an effect |
|---|---|---|
| Allow once & continue | Authorises and resumes the reviewed recorded operation | A new blanket grant for unrelated actions |
| Deny & stop turn | Denies the operation and stops that turn | Permission to try a substitute mutation |
| Pause | Stops new progression | An undo of completed work |
| Resume | Checks recorded state before continuing | A promise to repeat an uncertain mutation |
| Stop & revoke | Blocks further requests from the session | Rollback of effects already performed |
Review cards name the application, component and effect, with arguments, hashes and diffs under technical details. Changed files, stale approvals, a different project, expired sessions and revoked grants still block execution. Sensitive Sidecar settings retain explicit review. Guide tasks and operating-system dialogs keep their own controls.
Package builds must prepare and verify the exact release/plan in either approval mode. Retry preparation retries that preparation for the same recorded operation; it does not mean an installer already exists. Rollback is available only for supported file patches whose applied bytes remain unchanged, not arbitrary external effects.
Interactive Sidecar controls
Open Sidecars in chat to browse, inspect and configure through the same project-scoped broker. The form uses real Library controls; it does not require a model to reinterpret each click. Changes have operation receipts and apply to the next package, not automatically to an already running provider.
These questions request different evidence:
| Ask | What the result describes |
|---|---|
| “What Sidecars are available?” | The actual catalogue, not just general documentation |
| “Which Sidecars are selected?” | The selected component's next-package selection |
| “Which Sidecars are running?” | Observed host-backed tool availability, not complete provider health |
| “Configure Guide” | Opens its admitted settings without saving a change |
| “Add Guide Sidecar for Windows” | Requests a scoped selection change for that explicit platform |
Inspect a component for its full description, and use Find a setting for a label or field name. Resizing, filtering and paging do not save your form; use its explicit save. Imported packages and protected fields stay in their owning editors. After a lost save acknowledgement or conflict, inspect Library before repeating the write.
Save and update workflows
The Workflows menu contains workflow JSON, history and knowledge options. Export a workflow to keep a copy. Import loads configuration into the editor for review; Save applies it to future runs. It does not execute the imported definition.
Older saved workflows may not include newer tools. Use Load updated starter definitions, review and save, then start a new session. Existing sessions retain their original grants and running workflows retain their saved definitions. A granted workflow-save operation cannot invent additional authority or recursively dispatch workflows merely by storing JSON.
External agents use the credentials and scoped sessions created through Agent settings/Activity, not the owner's ambient service access. Treat exported access credentials as private and revoke the relevant session when it is no longer needed.
Model settings and resource limits
In Inference settings → Behaviour & limits, distinguish the input budget, reply budget and the loaded model's context. Instructions, history, tool descriptions and retrieved evidence all consume input. Both input and reply must fit the actual context.
Saving Agent budgets does not reload the model. Apply inference settings applies the corresponding runtime/model settings. Use larger chat defaults is a convenience, not evidence that your hardware or model can support the larger context. Check the reported effective generation budget and memory/fallback state; narrow the task or reduce context before assuming more memory allocation is the solution.
A model being loaded does not certify its tool-use ability. Start with narrow tasks and inspect operations. When Local AI reports a missing engine, model or service connection, fix that prerequisite in Local AI rather than repeatedly resubmitting the Agent request. See the Local AI reference.
Check the result, not the wording
A compile result has an actual compile operation and an exact verified release link. A Windows or Android installer additionally needs a package operation with artifact results. Existing historical artifacts are not proof of a new build. View result opens the recorded release, not one guessed from the model's prose.
Agent compile does not yet carry every saved Build-page version, backend and planning choice. Open the result in Build and check the configuration before distribution. “Compiled” does not mean installed, tested or published.
Stop in chat prevents further chat work; it does not cancel an independently started Build-page compile. Navigation also does not stop a turn. For a failed or lost mutation response, inspect Activity and the recorded operation rather than attempting to force a second execution.
Sharing diagnostics
Use Debug → Execution diagnostics to refresh, copy or download the bounded report. Model-decision telemetry is separate from recorded operations. The export omits prompts, source contents and credentials, but review project/release identifiers and any other contextual details before sharing it.
Report the requested result, selected workflow, observed failing stage and relevant operation/result. Do not publish a full private conversation, source tree or access credential to demonstrate that a model failed.
Troubleshooting
| Symptom | Next check |
|---|---|
| Send is disabled | Finish, resume or stop the active/paused turn. The typed draft is not lost or automatically submitted. |
| Agent explains how to compile but produces no result | Inspect the workflow's tools, Work plan and Activity. Look for an actual compile operation. |
| A tool is unavailable in a saved workflow | Review updated starter definitions, save and start a new session. |
| A request asks for Windows or Android | Supply the target; the controller does not infer it from unrelated old artifacts. |
| Approval is stale or a source write conflicts | Inspect the changed source/plan and request fresh review; do not reuse the old approval. |
| The model exceeds context or cannot progress | Shorten the request, narrow evidence/tools and inspect effective context before increasing limits. |
| A result belongs to another project | Open the original workspace named by the result. Do not silently relabel it as the current app. |
What success looks like
The requested result is backed by a scoped operation and inspectable evidence, any write affected only the intended application, and remaining limitations are visible. Use Build for release checks and Test Runner for repeatable app-surface verification.