Optimization
Keep work focused and easier to maintain.
Reusing content already read
When an agent asks for content it has already read, ClevAgent checks whether that content is still reusable.
For clear, substantial repeats, including reading back a file it just wrote, it can interrupt and redirect the agent to existing context. Smaller overlaps get guidance; uncertain cases stay quiet.
Keeping growing code manageable
Large code files can make new work harder to follow.
ClevAgent guides new logic into smaller units where practical. It favors a small local edit when splitting would add risk or unnecessary work.
Following up on missing checks
A finished coding task should include the checks relevant to its changes.
At task completion, ClevAgent asks the agent to run the checks that are still missing, using recorded results for the latest changes. If a check ran but its result could not be read, ClevAgent asks the agent to show it plainly instead of interrupting you. Checks already running are left alone.
Breaking repeated failure loops
Retrying the same failed step without progress can leave a task stuck.
Repeated failures trigger guidance to investigate the cause or change approach. Code failures and environment problems get different advice.
Reading the relevant part of a large file
A focused question may need one section rather than an entire large file.
After a large full-file read, ClevAgent encourages reuse and focused reads for the next step. Explicit requests to review the whole file are respected.
Security
Control access to secrets, files and agent settings.
Enable the protections you need. Depending on the rule, ClevAgent warns, pauses, or blocks supported actions.
Protecting secrets
API keys, access tokens and private credential files need care when an agent reads or sends them.
With Security Mode on, ClevAgent blocks recognized secrets from being sent to the model and pauses identified private credential reads for approval.
Resisting hidden instructions
Webpages and tool results can carry instructions that try to redirect your agent.
ClevAgent warns you and tells the agent to treat that content as untrusted and keep going.
Protecting recent work
A cleanup or reset can erase changes that have not been saved in Git.
ClevAgent checks the files at risk and pauses supported destructive cleanup. You can review the affected files before deciding.
Keeping control of agent settings
Settings can change what an agent can access or do automatically.
Recognized writes to protected Claude and Codex configuration files pause for your approval. The affected paths are shown without exposing file contents.
Controlling startup and background changes
Some changes can keep scripts running or preserve access to your computer.
ClevAgent pauses writes to protected startup files and other settings that can preserve access. You decide whether to allow the change.
Stopping destructive commands
Some commands can erase files or make large changes that are hard to reverse.
Recognized risky commands are blocked for review. You can allow a specific action when approval is available, or refuse it.
Holding requests that cannot be checked
A request too large for ClevAgent to check should not slip through unscreened.
If a request or reply is too large for ClevAgent to check, it is held and your agent pauses. You decide whether to let it through unchecked or stop it.
Cleva
These rules apply to agent actions through Cleva.
If payment details or deletion targets are unclear, the action stays blocked for you to review.
Reviewing payments before they proceed
A final payment or subscription action should stay under your control.
When Cleva detects a final payment or subscription action, ClevAgent holds it for your approval. Approval covers that specific action once.
Reviewing deletions before they happen
Deleting data or closing an account should be a deliberate decision.
When Cleva detects a deletion or account-removal action, ClevAgent holds it before proceeding. You can approve that specific action once or refuse it.