Skip to content

The fastest, cheapest, most efficient open-source agent harness. Run more than 500 agents on a $10 VPS.

feat(agent): tell the orchestrator when to write facts, blockers and approach

sanil-237eb0fc812b5b
diff --git a/crates/openhuman-core/src/agent_registry/agents/orchestrator/prompt.md b/crates/openhuman-core/src/agent_registry/agents/orchestrator/prompt.md
index 6406fcffdb..fdcfac2312 100644
--- a/crates/openhuman-core/src/agent_registry/agents/orchestrator/prompt.md
+++ b/crates/openhuman-core/src/agent_registry/agents/orchestrator/prompt.md
@@ -23,6 +23,7 @@ First match wins:
- Make a tool call in the message that announces it; keep going until done; batch independent calls.
- 3+ steps: `todo`, then execute. List the request's stated constraints, filters and thresholds as `todo` items too. One step at a time, in order: tick each with `todo {"complete": N}` as soon as it's done, before starting the next. Ask only if the ambiguity changes the tool.
+- Keep notes in any `todo` call as you learn them: `facts` for what a result proved, `blockers` for what does not work or must be avoided here (with why) the moment you find it, `approach` for what you are trying now. They stay in `[run state]` on every call; read its `Blockers` before retrying anything.
- Explicit yes only before moving funds or stopping, uninstalling or updating OpenHuman.
- Tools named by a tool result or `tool_search` are callable by name; other unlisted names always fail, so don't retry them.
- Never invent names, ids, paths, URLs, quotes or numbers; copy figures exactly. Worker summaries are claims: check them against their evidence. Truncated output is incomplete.
diff --git a/crates/openhuman-core/src/agent_registry/agents/orchestrator/prompt_tests.rs b/crates/openhuman-core/src/agent_registry/agents/orchestrator/prompt_tests.rs
index 2e729f166b..49c238f134 100644
--- a/crates/openhuman-core/src/agent_registry/agents/orchestrator/prompt_tests.rs
+++ b/crates/openhuman-core/src/agent_registry/agents/orchestrator/prompt_tests.rs
@@ -507,6 +507,11 @@ fn build_never_mandates_plan_review_and_allows_a_lead_in() {
assert!(!body.contains("before doing any of the work"));
assert!(body.contains("3+ steps: `todo`, then execute."));
assert!(body.contains("tick each with `todo {\"complete\": N}` as soon as it's done"));
+ assert!(
+ body.contains("`blockers` for what does not work or must be avoided here (with why)")
+ && body.contains("read its `Blockers` before retrying anything"),
+ "the orchestrator is told when to write each kind of note"
+ );
assert!(body.contains("Make a tool call in the message that announces it"));
}