(bug) Pi agents run with no read, bash, edit or write tools outside the playground

[mmabrouk]

When I build an agent in the playground it can run shell commands and write files. When the same
saved agent runs anywhere else, for example from a schedule, it answers that it has no tools:

I couldn't run this because no shell or filesystem tool is available in the current session.
No workspace, .env, script, or outputs were accessed.

I expect a saved agent to have the same tools it had while I was building it.

Steps to reproduce

1. Create an agent with the default configuration. The harness is pi_core.
2. In the playground, ask it to run a shell command or write a file. It works.
3. Save the agent, then run it from a schedule (an automation), or invoke its committed
   revision any other way that is not the playground.
4. The agent replies that it has no shell or filesystem tool available and does nothing.

Where it reproduces

Reproduced on the Pi harness. Claude agents are not affected: Claude brings its own Bash, Read
and Write, so this is specific to Pi.

Prompt used in the failing run

Run the daily GitHub measurement snapshot now. Work in the durable agent workspace at
$AGENTA_AGENT_MOUNT_DIR. Do exactly this: (1) cd into the workspace; (2) source the token with
set -a && . ./.env && set +a; (3) run python3 06-measurement/github-snapshot/snapshot.py ...

The run reached the model, produced reasoning about being unable to execute, and ended without
touching the workspace.

Notes

Related to #5562, which reported the same failure from the automation side.

[linear-code]

AGE-4029