feat(harness): reduce agent & sub-agent output verbosity (#4099) (#4114)

src/openhuman/agent/prompts/builder.rs

@@ -11,7 +11,14 @@ use anyhow::Result;
 pub const GLOBAL_STYLE_SUFFIX: &str = "## Output style\n\n\
     - Do **not** use em-dashes (`—`). Replace them with commas, colons, \
     parentheses, or two short sentences. This applies to every output \
-    you produce: chat replies, summaries, tool args, and file contents.\n";
+    you produce: chat replies, summaries, tool args, and file contents.\n\
+    - **Be concise.** Lead with the answer, then only the detail the task \
+    needs. No preamble, no recap of the request, no filler. Make the \
+    response as long as the task requires and no longer: a one-line answer \
+    for a simple ask, fuller treatment only when the task genuinely needs it.\n\
+    - **Do not repeat yourself.** Don't restate facts, context, or results \
+    already shown earlier in this conversation; reference them instead of \
+    pasting them again.\n";
 
 #[derive(Default)]
 pub struct SystemPromptBuilder {

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

@@ -4,6 +4,11 @@ delegate_name = "archive_session"
 when_to_use = "Background librarian — extracts lessons from a completed session, updates MEMORY.md, and indexes to FTS5. Runs cheap and slow."
 temperature = 0.4
 max_iterations = 3
+# Bound the archiving summary that flows up to the orchestrator verbatim
+# (delegate_archivist). 8000 chars (~2000 tokens) matches the normal
+# sub-agent cap and stops a verbose memory-write confirmation from
+# bloating the orchestrator's context (#4099).
+max_result_chars = 8000
 sandbox_mode = "none"
 background = true
 omit_identity = true

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

@@ -4,6 +4,11 @@ delegate_name = "review_code"
 when_to_use = "Adversarial reviewer — reviews diffs and code against project rules, flags vulnerabilities, regressions, and missing tests. Read-only."
 temperature = 0.4
 max_iterations = 5
+# Bound the review verdict that flows up to the orchestrator verbatim
+# (delegate_critic). 8000 chars (~2000 tokens) matches the normal
+# sub-agent cap (planner/researcher) and keeps long diffs from leaking
+# an unbounded critique into the orchestrator's context (#4099).
+max_result_chars = 8000
 sandbox_mode = "read_only"
 omit_identity = true
 omit_memory_context = true

src/openhuman/agent_registry/agents/loader.rs

@@ -1017,6 +1017,25 @@ mod tests {
         );
     }
 
+    #[test]
+    fn chatty_sub_agents_have_bounded_output() {
+        // critic + archivist results flow up to the orchestrator verbatim
+        // (delegate_critic / delegate_archivist). Without a cap their output
+        // is unbounded and bloats the orchestrator's context (#4099). Both
+        // must carry the normal sub-agent cap so a long diff review or a
+        // verbose memory-write confirmation can't leak unbounded text.
+        assert_eq!(
+            find("critic").max_result_chars,
+            Some(8000),
+            "critic output must be bounded so reviews don't leak unbounded text up"
+        );
+        assert_eq!(
+            find("archivist").max_result_chars,
+            Some(8000),
+            "archivist output must be bounded so memory summaries stay concise"
+        );
+    }
+
     #[test]
     fn researcher_has_curl_for_artifact_downloads() {
         let def = find("researcher");

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

@@ -47,7 +47,7 @@ Follow this sequence for every user message:
    - If complex multi-step decomposition is required, use `delegate_plan`.
    - If code review is requested, use `delegate_critic`.
    - If memory archiving or distillation is required, use `delegate_archivist`.
-5. **After delegation**, summarise results clearly and concisely.
+5. **After delegation, distill — never forward verbatim.** A sub-agent's reply is raw material, not your answer. Extract only the parts that answer the user's question and present them in as few words as carry the meaning. Drop the sub-agent's working notes, restated context, and any detail the user already has. If the useful answer is two sentences, send two sentences, even when the sub-agent returned eight paragraphs. Never paste a sub-agent's full response back to the user.
 
 Default bias: **do not spawn a sub-agent when a direct response or direct tool call is sufficient** — but live external-service, scheduling, desktop-control, presentation, product-docs, code-repo, market, and crypto requests belong to their specialists.