fix(agent): bound researcher web tool loop (#4313)

src/openhuman/agent_registry/agents/loader.rs

@@ -1077,17 +1077,18 @@ mod tests {
     }
 
     #[test]
-    fn researcher_has_curl_for_artifact_downloads() {
+    fn researcher_is_bounded_to_search_and_fetch() {
         let def = find("researcher");
+        assert_eq!(
+            def.max_iterations, 10,
+            "researcher keeps enough turns to recover from bad search results without broadening its tool surface"
+        );
         match &def.tools {
             ToolScope::Named(tools) => {
-                assert!(
-                    tools.iter().any(|t| t == "curl"),
-                    "researcher needs curl for artifact downloads"
-                );
-                assert!(
-                    tools.iter().any(|t| t == "http_request"),
-                    "researcher still needs http_request"
+                assert_eq!(
+                    tools,
+                    &vec!["web_search_tool".to_string(), "web_fetch".to_string()],
+                    "researcher must stay limited to search+fetch so simple lookups do not fan out into deep research loops"
                 );
             }
             ToolScope::Wildcard => panic!("researcher must have Named tool scope"),

src/openhuman/agent_registry/agents/researcher/agent.toml

@@ -3,7 +3,7 @@ display_name = "Researcher"
 delegate_name = "research"
 when_to_use = "Web & docs crawler — reads real documentation, compresses to dense markdown. Use for any task that requires looking up external knowledge."
 temperature = 0.4
-max_iterations = 8
+max_iterations = 10
 iteration_policy = "extended"
 max_result_chars = 8000
 sandbox_mode = "none"
@@ -20,31 +20,10 @@ hint = "agentic"
 
 [tools]
 named = [
-    "http_request",
-    "curl",
-    "web_search",
+    # Keep the default researcher narrow: search to locate sources, fetch to read
+    # selected pages. Deep Parallel, market, filesystem, and general HTTP/curl
+    # tools belong to more specific agents so simple research tasks don't expand
+    # into open-ended search loops.
     "web_search_tool",
-    "file_read",
-    # Coding-harness read-only primitives from #1208. `web_fetch` is the
-    # simple URL-GET sibling of `http_request` (capped body, same
-    # allowed_domains gate); grep/glob/list let the researcher navigate
-    # cached docs in the workspace without falling through to shell.
     "web_fetch",
-    "grep",
-    "glob",
-    "list",
-    # Parallel — full surface (search/extract are also delegated by web_search_tool;
-    # chat/research/enrich/dataset unlock grounded answers and deep multi-step research).
-    "parallel_search",
-    "parallel_extract",
-    "parallel_chat",
-    "parallel_research",
-    "parallel_enrich",
-    "parallel_dataset",
-    # Stock / FX / crypto / commodity market data via the financial-apis backend.
-    "stock_quote",
-    "stock_exchange_rate",
-    "stock_options",
-    "stock_crypto_series",
-    "stock_commodity",
 ]

src/openhuman/agent_registry/agents/researcher/prompt.md

@@ -6,7 +6,6 @@ You are the **Researcher** agent. You find accurate, up-to-date information.
 
 - Web search for current information
 - HTTP requests to fetch documentation
-- Read local files for project context
 - Memory recall for prior research
 
 ## Rules
@@ -16,3 +15,18 @@ You are the **Researcher** agent. You find accurate, up-to-date information.
 - **Compress output** — Distill long documents into dense, factual markdown summaries.
 - **Cite sources** — Include URLs or file paths for information you reference.
 - **Stay focused** — Answer the specific question asked, not everything tangentially related.
+
+## Research Loop Contract
+
+- Use `web_search_tool` to find likely sources, then `web_fetch` to read only the pages needed to answer.
+- For simple factual requests, one focused search plus one or two fetched sources is enough unless results are empty or contradictory.
+- Do not keep broadening, re-searching, or chasing tangents once you have source-backed evidence for the requested answer.
+- Prefer fetching authoritative or primary sources over reading many secondary summaries.
+- If search or fetch fails, return what happened under `Failed tool calls`; do not silently keep trying unrelated queries.
+
+## Output Contract
+
+- Always return an output to the orchestrator, even when the answer is incomplete.
+- If you could answer, put the answer first, then list the URLs you used.
+- If you could not answer, say exactly what is missing and what you tried.
+- Never finish with only tool calls or internal notes; the orchestrator needs a compact synthesis it can pass on or evaluate.