When something does not work

Most problems in pcx say what went wrong in plain words. This page lists the ones people meet most often, what they mean, and what to do.

Memory and search

My agent writes nothing to memory. Check that it is connected: Connections shows when each key was last used. A key that is read only can search but not write. A key scoped to some workspaces cannot write to others.

Search does not find a record I know is there. A fast search can miss a record written in other words; try the accurate mode, or the exact name of a function or a file, which search by words always finds. An ended record ranks lower: look for what superseded it. Why: Search.

The agent forgot my rules after a long session. Put them in the cards, not in the conversation: cards come back after every compaction.

Modules

A page of a module is missing from the left. A module’s page shows only when the module is on: turn it on in All modules.

A module asks for a setting before it starts. Settings marked with a star are required, such as the workspace a module keeps its documents in, or the board where its approvals go. Workflows and Custom modules need a board for approvals.

I cannot see a secret I saved. That is by design: secrets are write-only. Replace it with a new one.

Agents

A turn ends because a key was refused. The model provider did not take the key. Check it in All modules, Models, and enter it again. A turn that failed this way is not retried.

A turn stops with “its tools kept failing”. The agent called a tool that failed the same way three times in a row, or twelve times in all. The message shows the last failure: fix that, then ask again.

An agent did not do what I asked, and a task waits. It recorded a decision with a risk, which waits for a person in Approvals. Approve it, or answer on the task.

Workflows and jobs

A workflow does not start. A new workflow, or one that asks for a new module, starts only after a person approves its grants: look in Approvals.

A scheduled run says Skipped. It was due while the previous run was still going; a job runs once at a time.

A run waits for ever. It is at an approval step, and its question is on the board. A run that reaches its time limit ends, and its question is taken off.

Custom modules

The install is refused. The message names the file and the reason: a test that failed, a file that is not allowed, a native package in another version than the one every module shares (it says which to lock).

“No module named ‘google’” in a Python module. The module’s lock file has no protobuf. Every Python module brings its own: add it to requirements.in and lock again.

The module cannot reach a website. A custom module has no network. Call the API module through a connection instead.

API connections and mail

A call is refused before it is sent. Its host is not a connection’s and not in the allowed hosts, its address is private, or the day’s limit is reached. The log shows each refusal.

A mail is refused. Its recipient is not among the addresses the settings allow, or the day’s limit is reached. Send a test message shows whether the provider takes mail at all.

A daemon

A connected daemon shows as offline. The organization reaches it only while it runs and is connected. Start it again; the space says when it was last seen. See Connecting a daemon.

Something else? Open an issue on GitHub, with what you did and the message you got.