You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Make MultiAgentV2 spawned-agent context prompt configurable #17323
Make the MultiAgentV2 spawned-agent context prompt configurable, or at least only inject it when history/context is actually forked.
Today codex-rs/core/src/tools/handlers/multi_agents_v2/spawn.rs unconditionally appends SPAWN_AGENT_DEVELOPER_INSTRUCTIONS to every spawned child agent:
<spawned_agent_context>
... Treat the next user message as your assigned task, and use the forked history only as background context.
</spawned_agent_context>
The problem is that this prompt is added even when the child is spawned with no forked history (for example fork_turns = none, or when no forked context is otherwise provided).
That creates two issues:
the prompt describes forked history even when there is none
there is currently no config knob to disable or override this block for custom multi-agent workflows
It looks like later work improved surrounding spawn behavior (parent model/config inheritance, role-level fork defaults, prompt cache keys, etc.), but I could not find a way to:
disable SPAWN_AGENT_DEVELOPER_INSTRUCTIONS
replace it with a custom block
gate it on whether forked context is actually present
A few reasonable options:
only inject this prompt when fork_turns actually results in forked history/context
add a config option to disable it
add a config option to override the text
split it into two variants: one for forked children, one for no-history children
Specifically, the block is appended unconditionally after spawn config is built, independent of fork_turns / fork mode.
This matters for users building custom multi-agent workflows because the injected prompt becomes part of the child agent's behavior contract even when it is semantically wrong for the spawn mode.
Activity