diff --git a/src/harness/middleware/mod.rs b/src/harness/middleware/mod.rs
index 5e8b323..1c12eee 100644
--- a/src/harness/middleware/mod.rs
+++ b/src/harness/middleware/mod.rs
@@ -40,6 +40,7 @@ use crate::error::{Result, TinyAgentsError};
 use crate::harness::cache::{CacheLayoutEvent, PromptCacheLayout};
 use crate::harness::context::RunContext;
 use crate::harness::events::AgentEvent;
+use crate::harness::message::Message;
 use crate::harness::model::{ModelDelta, ModelRequest, ModelResponse};
 use crate::harness::summarization::{
     ConcatSummarizer, SummarizationPolicy, Summarizer, SummaryRecord, TrimStrategy,
@@ -720,6 +721,98 @@ impl<State: Send + Sync, Ctx: Send + Sync> Middleware<State, Ctx> for ContextCom
     }
 }
 
+// ── MicrocompactMiddleware ────────────────────────────────────────────────────
+
+impl MicrocompactMiddleware {
+    /// Creates a micro-compaction middleware that keeps the newest `keep_recent`
+    /// tool-result bodies verbatim and blanks older ones with `placeholder`.
+    /// Event emission is off by default; enable it with
+    /// [`MicrocompactMiddleware::with_events`].
+    pub fn new(keep_recent: usize, placeholder: impl Into<String>) -> Self {
+        Self {
+            label: "microcompact",
+            keep_recent,
+            placeholder: placeholder.into(),
+            emit_events: false,
+        }
+    }
+
+    /// Enable or disable emitting an
+    /// [`AgentEvent::Compressed`][crate::harness::events::AgentEvent::Compressed]
+    /// event whenever at least one tool body is cleared. Off by default so the
+    /// middleware can be a silent transcript rewrite.
+    pub fn with_events(mut self, emit_events: bool) -> Self {
+        self.emit_events = emit_events;
+        self
+    }
+
+    /// The number of most-recent tool-result bodies kept verbatim.
+    pub fn keep_recent(&self) -> usize {
+        self.keep_recent
+    }
+
+    /// The placeholder text swapped in for cleared tool-result bodies.
+    pub fn placeholder(&self) -> &str {
+        &self.placeholder
+    }
+}
+
+#[async_trait]
+impl<State: Send + Sync, Ctx: Send + Sync> Middleware<State, Ctx> for MicrocompactMiddleware {
+    fn name(&self) -> &str {
+        self.label
+    }
+
+    async fn before_model(
+        &self,
+        ctx: &mut RunContext<Ctx>,
+        _state: &State,
+        request: &mut ModelRequest,
+    ) -> Result<()> {
+        // Indices of every tool-result message, oldest → newest.
+        let tool_idxs: Vec<usize> = request
+            .messages
+            .iter()
+            .enumerate()
+            .filter(|(_, m)| matches!(m, Message::Tool(_)))
+            .map(|(i, _)| i)
+            .collect();
+        if tool_idxs.len() <= self.keep_recent {
+            return Ok(());
+        }
+
+        let from_tokens = if self.emit_events {
+            total_message_tokens(&request.messages)
+        } else {
+            0
+        };
+
+        let cut = tool_idxs.len() - self.keep_recent;
+        let mut cleared = 0usize;
+        for &i in &tool_idxs[..cut] {
+            // Skip messages already reduced to the placeholder; otherwise swap the
+            // body for it (idempotent, preserves the tool_call_id).
+            if request.messages[i].text() == self.placeholder {
+                continue;
+            }
+            if let Message::Tool(t) = &request.messages[i] {
+                let id = t.tool_call_id.clone();
+                request.messages[i] = Message::tool(id, self.placeholder.clone());
+                cleared += 1;
+            }
+        }
+
+        if self.emit_events && cleared > 0 {
+            let to_tokens = total_message_tokens(&request.messages);
+            ctx.emit(AgentEvent::Compressed {
+                from_tokens,
+                to_tokens,
+            });
+        }
+        Ok(())
+    }
+}
+
 // ── PromptCacheGuardMiddleware ────────────────────────────────────────────────
 
 impl PromptCacheGuardMiddleware {
diff --git a/src/harness/middleware/test.rs b/src/harness/middleware/test.rs
index 8b9b10b..e58fbe8 100644
--- a/src/harness/middleware/test.rs
+++ b/src/harness/middleware/test.rs
@@ -348,6 +348,157 @@ async fn context_compression_none_window_falls_back_to_trigger_tokens() {
     assert_eq!(mw.records().len(), 1);
 }
 
+// ── MicrocompactMiddleware ────────────────────────────────────────────────────
+//
+// These cases are the byte-for-byte parity contract ported from the OpenHuman
+// in-house `MicrocompactMiddleware` this type replaces: older tool-result bodies
+// are blanked to the caller's placeholder, the newest `keep_recent` are kept
+// verbatim, non-tool messages are never touched, and the pass is idempotent.
+
+const CLEARED: &str = "[Old tool result content cleared]";
+
+#[tokio::test]
+async fn microcompact_clears_older_tool_bodies_and_keeps_recent() {
+    let mw = MicrocompactMiddleware::new(1, CLEARED);
+    let mut request = ModelRequest {
+        messages: vec![
+            Message::system("sys"),
+            Message::user("hello"),
+            Message::tool("t1", "FIRST_BODY"),
+            Message::assistant("thinking"),
+            Message::tool("t2", "SECOND_BODY"),
+            Message::tool("t3", "THIRD_BODY"),
+        ],
+        ..Default::default()
+    };
+
+    mw.before_model(&mut ctx(), &(), &mut request)
+        .await
+        .unwrap();
+
+    // 3 tool messages, keep_recent=1 → the two oldest cleared, newest kept.
+    assert_eq!(request.messages[2].text(), CLEARED);
+    assert_eq!(request.messages[4].text(), CLEARED);
+    assert_eq!(request.messages[5].text(), "THIRD_BODY");
+    // Non-tool messages are never touched.
+    assert_eq!(request.messages[0].text(), "sys");
+    assert_eq!(request.messages[1].text(), "hello");
+    assert_eq!(request.messages[3].text(), "thinking");
+    // Cleared tool messages keep their tool_call_id.
+    match &request.messages[2] {
+        Message::Tool(t) => assert_eq!(t.tool_call_id, "t1"),
+        other => panic!("expected tool message, got {other:?}"),
+    }
+}
+
+#[tokio::test]
+async fn microcompact_is_a_noop_when_within_keep_recent() {
+    let mw = MicrocompactMiddleware::new(5, CLEARED);
+    let mut request = ModelRequest {
+        messages: vec![Message::tool("t1", "A"), Message::tool("t2", "B")],
+        ..Default::default()
+    };
+    mw.before_model(&mut ctx(), &(), &mut request)
+        .await
+        .unwrap();
+    assert_eq!(request.messages[0].text(), "A");
+    assert_eq!(request.messages[1].text(), "B");
+}
+
+#[tokio::test]
+async fn microcompact_is_idempotent() {
+    let mw = MicrocompactMiddleware::new(1, CLEARED);
+    let mut request = ModelRequest {
+        messages: vec![Message::tool("t1", "FIRST"), Message::tool("t2", "SECOND")],
+        ..Default::default()
+    };
+    mw.before_model(&mut ctx(), &(), &mut request)
+        .await
+        .unwrap();
+    assert_eq!(request.messages[0].text(), CLEARED);
+    // Second pass leaves the already-cleared body as the placeholder.
+    mw.before_model(&mut ctx(), &(), &mut request)
+        .await
+        .unwrap();
+    assert_eq!(request.messages[0].text(), CLEARED);
+    assert_eq!(request.messages[1].text(), "SECOND");
+}
+
+#[tokio::test]
+async fn microcompact_emits_no_event_by_default() {
+    // Default construction is silent: bodies are still cleared, but no
+    // Compressed event is emitted (parity with the OpenHuman in-house version).
+    let mw = Arc::new(MicrocompactMiddleware::new(1, CLEARED));
+    let mut stack: MiddlewareStack<()> = MiddlewareStack::new();
+    stack.push(mw.clone());
+
+    let recorder = Arc::new(RecordingListener::new());
+    let mut c = ctx();
+    c.events.subscribe(recorder.clone());
+
+    let mut request = ModelRequest {
+        messages: vec![Message::tool("t1", "FIRST"), Message::tool("t2", "SECOND")],
+        ..Default::default()
+    };
+    stack
+        .run_before_model(&mut c, &(), &mut request)
+        .await
+        .unwrap();
+
+    assert_eq!(request.messages[0].text(), CLEARED);
+    let events: Vec<AgentEvent> = recorder.events().into_iter().map(|r| r.event).collect();
+    assert!(
+        !events
+            .iter()
+            .any(|e| matches!(e, AgentEvent::Compressed { .. })),
+        "no Compressed event should be emitted when events are off"
+    );
+}
+
+#[tokio::test]
+async fn microcompact_emits_compressed_event_when_enabled() {
+    let mw = Arc::new(MicrocompactMiddleware::new(1, CLEARED).with_events(true));
+    let mut stack: MiddlewareStack<()> = MiddlewareStack::new();
+    stack.push(mw.clone());
+
+    let recorder = Arc::new(RecordingListener::new());
+    let mut c = ctx();
+    c.events.subscribe(recorder.clone());
+
+    let mut request = ModelRequest {
+        messages: vec![
+            Message::tool("t1", "x".repeat(400)),
+            Message::tool("t2", "y".repeat(400)),
+        ],
+        ..Default::default()
+    };
+    stack
+        .run_before_model(&mut c, &(), &mut request)
+        .await
+        .unwrap();
+
+    let compressed: Vec<(u64, u64)> = recorder
+        .events()
+        .into_iter()
+        .filter_map(|r| match r.event {
+            AgentEvent::Compressed {
+                from_tokens,
+                to_tokens,
+            } => Some((from_tokens, to_tokens)),
+            _ => None,
+        })
+        .collect();
+    assert_eq!(
+        compressed.len(),
+        1,
+        "one Compressed event when a body cleared"
+    );
+    assert!(
+        compressed[0].0 > compressed[0].1,
+        "tokens dropped after clear"
+    );
+}
+
 #[tokio::test]
 async fn usage_accounting_accumulates_across_calls() {
     let mw = Arc::new(UsageAccountingMiddleware::new());
diff --git a/src/harness/middleware/types.rs b/src/harness/middleware/types.rs
index 40ca649..f73778c 100644
--- a/src/harness/middleware/types.rs
+++ b/src/harness/middleware/types.rs
@@ -491,6 +491,44 @@ pub struct ContextCompressionMiddleware {
     pub(crate) records: Mutex<Vec<SummaryRecord>>,
 }
 
+// ── MicrocompactMiddleware ────────────────────────────────────────────────────
+
+/// Middleware that clears the bodies of older **tool-result** messages while
+/// keeping the `keep_recent` most recent ones verbatim.
+///
+/// In `before_model` it walks `request.messages`, finds the tool-result
+/// messages, and — once there are more than `keep_recent` of them — replaces the
+/// content of every tool result *except* the newest `keep_recent` with a fixed
+/// [`placeholder`](Self::placeholder) string, preserving each message's
+/// `tool_call_id`. Non-tool messages (system/user/assistant) are never touched
+/// and no chat turn is dropped, so this bounds the cost of a long, tool-heavy
+/// thread without the semantic loss of summarization.
+///
+/// This is the "micro-compaction" companion to
+/// [`ContextCompressionMiddleware`]: the latter summarizes *older chat history*
+/// when the transcript nears the context window, whereas this one only ever
+/// blanks *stale tool payloads* that the model no longer needs verbatim. The two
+/// compose cleanly.
+///
+/// The operation is **idempotent**: a body already equal to the placeholder is
+/// left as-is, so repeated `before_model` passes converge. When there are at
+/// most `keep_recent` tool results the middleware is a complete no-op.
+///
+/// The placeholder text is caller-supplied (via
+/// [`MicrocompactMiddleware::new`]) so host applications can keep their own
+/// model-facing wording stable. Event emission is **opt-in** (default off, see
+/// [`MicrocompactMiddleware::with_events`]): when enabled and at least one body
+/// is cleared, an
+/// [`AgentEvent::Compressed`][crate::harness::events::AgentEvent::Compressed]
+/// event carrying the before/after token estimate is emitted; when disabled the
+/// middleware mutates the request silently.
+pub struct MicrocompactMiddleware {
+    pub(crate) label: &'static str,
+    pub(crate) keep_recent: usize,
+    pub(crate) placeholder: String,
+    pub(crate) emit_events: bool,
+}
+
 // ── PromptCacheGuardMiddleware ────────────────────────────────────────────────
 
 /// Middleware that watches the prompt cache layout for accidental prefix