Skip to content

setup-v0.2.4

@paperstreetadmin paperstreetadmin tagged this 21 Apr 00:47
0.2.3 live-test: file_read steering worked (agent picked
agentsid_file_read automatically) but shell steering didn't — "run
sudo echo hi" went to the built-in shell tool instead of
agentsid_shell_run. The model's native-shell bias beat our "prefer"
language.

Strengthen:
  - "MANDATORY TOOL POLICY" framing at the top
  - "MUST use" instead of "prefer"
  - Explicit callout that even trivial shell commands (ls, echo, pwd,
    cat) must route through agentsid_shell_run
  - "hard requirement of this environment, not a preference"

Test now asserts MANDATORY/MUST/hard-requirement phrasing is present
so a future softening regresses immediately.

Still in the "additional instructions" channel (developer_instructions
appends to Codex's built-in guidance). If this still isn't enough, the
escalation path is model_instructions_file which REPLACES Codex's
built-in instructions. Try the softer approach first.
Assets 2
Loading