fix: auto-delegation to the crew + crew-chat visibility #110

Merged

rawan merged 3 commits from int_delegate_crew into integration

2026-06-11 12:53:29 +00:00

rawan commented

2026-06-11 12:31:01 +00:00

Member

Summary

Fixes the "no auto delegation to the crew" investigation in #99. Two root causes were found and fixed:

The orchestrator never knew the installed crew existed. The base prompt told it to consult agent.list (a server JSON-RPC method, not an LLM tool) and no context provider injected the roster, so the model assumed "no crew installed" and did everything itself.
Even when it did delegate, delegate_task spawns an in-process sub-agent that never writes to the durable agent_messages mailbox, so delegated work never appeared in the per-member crew-chat threads.

Changes

Inject an [installed crew] roster into the chat-path system prompt (new CrewRosterProvider) listing each registered member by name and role, so the model delegates to them by name via delegate_task { agent: "<name>" }.
Inject the same roster into the autonomy executor prompt so multi-step autonomy runs also delegate to the crew.
Rewrite the misleading delegation guidance in agent_minimal_base.md (drop the dead agent.list reference; key off the injected roster).
Bridge delegate_task / delegate_parallel into the crew mailbox: when a delegation resolves to a registered crew member (by explicit name, or by mapping a specialist profile to the member of that role), mirror the task and its result into that member's crew-chat thread.

Tests

Unit tests for the roster body gating/wording.
Integration test proving a mirrored task + reply land in the member's thread via the same query the crew UI uses.
Full engine lib suite green.

Related to #99

#99

## Summary Fixes the "no auto delegation to the crew" investigation in #99. Two root causes were found and fixed: 1. The orchestrator never knew the installed crew existed. The base prompt told it to consult `agent.list` (a server JSON-RPC method, not an LLM tool) and no context provider injected the roster, so the model assumed "no crew installed" and did everything itself. 2. Even when it did delegate, `delegate_task` spawns an in-process sub-agent that never writes to the durable `agent_messages` mailbox, so delegated work never appeared in the per-member crew-chat threads. ## Changes - Inject an `[installed crew]` roster into the chat-path system prompt (new `CrewRosterProvider`) listing each registered member by name and role, so the model delegates to them by name via `delegate_task { agent: "<name>" }`. - Inject the same roster into the autonomy executor prompt so multi-step autonomy runs also delegate to the crew. - Rewrite the misleading delegation guidance in `agent_minimal_base.md` (drop the dead `agent.list` reference; key off the injected roster). - Bridge `delegate_task` / `delegate_parallel` into the crew mailbox: when a delegation resolves to a registered crew member (by explicit name, or by mapping a specialist profile to the member of that role), mirror the task and its result into that member's crew-chat thread. ## Tests - Unit tests for the roster body gating/wording. - Integration test proving a mirrored task + reply land in the member's thread via the same query the crew UI uses. - Full engine lib suite green. Related to #99 https://forge.ourworld.tf/lhumina_code/hero_shrimp/issues/99

rawan added 2 commits

2026-06-11 12:31:02 +00:00