feat: session-metrics実装

crates/pod/Cargo.toml

@@ -28,6 +28,7 @@ libc = { workspace = true }
 schemars = { workspace = true }
 memory = { workspace = true }
 uuid = { workspace = true, features = ["v7"] }
+session-metrics = { workspace = true }
 
 [dev-dependencies]
 dotenv = "0.15.0"

crates/pod/src/compact/metrics_tracker.rs

@@ -0,0 +1,50 @@
+//! Sync buffer for `session_metrics::Metric` values queued from inside
+//! Worker callbacks (which run synchronously and cannot themselves
+//! perform `async` store writes).
+//!
+//! Pod drains this buffer in `persist_turn` and writes each metric via
+//! `session_metrics::record_metric`, alongside the regular `LlmUsage`
+//! entries.
+
+use std::sync::Mutex;
+
+use session_metrics::Metric;
+
+pub(crate) struct MetricsTracker {
+    pending: Mutex<Vec<Metric>>,
+}
+
+impl MetricsTracker {
+    pub(crate) fn new() -> Self {
+        Self {
+            pending: Mutex::new(Vec::new()),
+        }
+    }
+
+    /// Queue a metric for the next `persist_turn` flush.
+    pub(crate) fn push(&self, metric: Metric) {
+        self.pending.lock().unwrap().push(metric);
+    }
+
+    /// Drain all queued metrics. Called by Pod after a run completes.
+    pub(crate) fn drain(&self) -> Vec<Metric> {
+        std::mem::take(&mut *self.pending.lock().unwrap())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn push_then_drain_returns_in_order_and_clears() {
+        let t = MetricsTracker::new();
+        t.push(Metric::now("a"));
+        t.push(Metric::now("b"));
+        let drained = t.drain();
+        assert_eq!(drained.len(), 2);
+        assert_eq!(drained[0].name, "a");
+        assert_eq!(drained[1].name, "b");
+        assert!(t.drain().is_empty());
+    }
+}

crates/pod/src/compact/mod.rs

@@ -1,3 +1,4 @@
+pub(crate) mod metrics_tracker;
 pub(crate) mod prune;
 pub(crate) mod state;
 pub(crate) mod token_counter;

crates/pod/src/compact/prune.rs

@@ -5,10 +5,16 @@
 //! 直後）。Worker は usage 履歴を知らないので、`min_savings` 判定に使う savings
 //! の見積もりはコールバックで外部から注入する。このモジュールはそのコールバック
 //! を組み立てて Worker に差し込むための `impl Pod` を提供する。
+//!
+//! 同じ経路で `PruneObserver` も install し、評価のたびに `prune.fire` /
+//! `prune.skip` metric を `MetricsTracker` に積む。`Fired` 時は uuid を
+//! `UsageTracker` にも stash しておき、後続の `LlmUsage` と組で
+//! `prune.post_request` を吐けるようにする。
 
 use llm_worker::Item;
 use llm_worker::llm_client::client::LlmClient;
-use llm_worker::prune::{PruneConfig, SavingsEstimator};
+use llm_worker::prune::{PruneConfig, PruneDecision, PruneObserver, SavingsEstimator};
+use session_metrics::Metric;
 use session_store::Store;
 
 use crate::Pod;
@@ -24,6 +30,12 @@ impl<C: LlmClient, St: Store> Pod<C, St> {
     /// Measurement-less estimates (before the first LLM call, or immediately
     /// after a compact) return `0` from the estimator, which naturally
     /// prevents the prune projection from firing until usage data exists.
+    ///
+    /// Also installs a [`PruneObserver`] that pushes `prune.fire` /
+    /// `prune.skip` metrics into the shared [`MetricsTracker`]. On `Fired`
+    /// the observer additionally stashes a fresh correlation_id in
+    /// [`UsageTracker`] so the next `LlmUsage` can be paired with a
+    /// `prune.post_request` metric carrying the same id.
     pub fn attach_prune(&mut self, config: PruneConfig) {
         let usage = self.usage_history_handle();
         let estimator: SavingsEstimator = Box::new(move |history: &[Item], indices| {
@@ -34,9 +46,43 @@ impl<C: LlmClient, St: Store> Pod<C, St> {
                 _ => est.tokens,
             }
         });
+
+        let metrics = self.metrics_tracker_handle();
+        let usage_tracker = self.usage_tracker_handle();
+        let observer: PruneObserver = Box::new(move |eval| {
+            match &eval.decision {
+                PruneDecision::Fired { .. } => {
+                    let correlation_id = uuid::Uuid::now_v7().to_string();
+                    let mut metric = Metric::now("prune.fire")
+                        .with_value(eval.estimated_savings as f64)
+                        .with_correlation_id(&correlation_id)
+                        .with_dimension("candidate_count", eval.candidate_count.to_string());
+                    if let Some(border) = eval.border_turn {
+                        metric = metric.with_dimension("border_turn", border.to_string());
+                    }
+                    metrics.push(metric);
+                    usage_tracker.note_correlation_id(correlation_id);
+                }
+                PruneDecision::SkippedNoCandidates => {
+                    metrics.push(
+                        Metric::now("prune.skip").with_dimension("reason", "no_candidates"),
+                    );
+                }
+                PruneDecision::SkippedBelowMinSavings => {
+                    metrics.push(
+                        Metric::now("prune.skip")
+                            .with_dimension("reason", "below_min_savings")
+                            .with_dimension("candidate_count", eval.candidate_count.to_string())
+                            .with_value(eval.estimated_savings as f64),
+                    );
+                }
+            }
+        });
+
         let worker = self.worker_mut();
         worker.set_prune_config(Some(config));
         worker.set_savings_estimator(Some(estimator));
+        worker.set_prune_observer(Some(observer));
     }
 
     /// If the manifest has a `[compaction]` section, build a `PruneConfig`

crates/pod/src/compact/usage_tracker.rs

@@ -19,19 +19,35 @@ use std::sync::Mutex;
 use llm_worker::UsageRecord;
 use llm_worker::timeline::event::UsageEvent;
 
+/// One drained measurement: the underlying `UsageRecord` plus an optional
+/// `correlation_id` stamped by the prune projection (or any other future
+/// upstream observer) so that downstream metrics emitted alongside this
+/// record can be joined to it after the fact.
+#[derive(Debug, Clone)]
+pub(crate) struct RecordedUsage {
+    pub(crate) record: UsageRecord,
+    pub(crate) correlation_id: Option<String>,
+}
+
 /// Shared between the pre-request hook, the `on_usage` callback, and Pod.
 pub(crate) struct UsageTracker {
     /// `history.len()` captured at the most recent `pre_llm_request`.
     /// Cleared when paired with an incoming `on_usage` event.
     pending_history_len: Mutex<Option<usize>>,
+    /// Optional `correlation_id` set by an upstream observer (currently
+    /// the prune projection on `Fired`). Paired into the next
+    /// `RecordedUsage` and cleared. Skips that don't fire leave this
+    /// `None`, so the resulting record carries no correlation.
+    pending_correlation_id: Mutex<Option<String>>,
     /// Records accumulated during the current run; drained by Pod.
-    pending_records: Mutex<Vec<UsageRecord>>,
+    pending_records: Mutex<Vec<RecordedUsage>>,
 }
 
 impl UsageTracker {
     pub(crate) fn new() -> Self {
         Self {
             pending_history_len: Mutex::new(None),
+            pending_correlation_id: Mutex::new(None),
             pending_records: Mutex::new(Vec::new()),
         }
     }
@@ -41,16 +57,29 @@ impl UsageTracker {
         *self.pending_history_len.lock().unwrap() = Some(history_len);
     }
 
+    /// Stash a `correlation_id` to be paired into the next `RecordedUsage`.
+    /// Currently invoked by the prune observer on `Fired` so that the
+    /// `prune.fire` metric and the `prune.post_request` metric (emitted
+    /// alongside the resulting `LlmUsage`) carry the same join key.
+    ///
+    /// Overwrites any previous unconsumed value — by construction the
+    /// observer fires at most once per outgoing LLM request, immediately
+    /// before the pre-request hook captures `history_len`.
+    pub(crate) fn note_correlation_id(&self, id: String) {
+        *self.pending_correlation_id.lock().unwrap() = Some(id);
+    }
+
     /// Called from the `on_usage` callback with the aggregated final
     /// UsageEvent. If a `history_len` was previously stashed via
-    /// `note_request`, builds a `UsageRecord` and pushes it onto the buffer.
-    /// If not (e.g. test code that fires Usage outside a request), drops
-    /// the event.
+    /// `note_request`, builds a `RecordedUsage` and pushes it onto the
+    /// buffer. If not (e.g. test code that fires Usage outside a request),
+    /// drops the event.
     pub(crate) fn record_usage(&self, event: &UsageEvent) {
         let history_len = match self.pending_history_len.lock().unwrap().take() {
             Some(n) => n,
             None => return,
         };
+        let correlation_id = self.pending_correlation_id.lock().unwrap().take();
         // UsageEvent.input_tokens は scheme 層で「占有量（プロンプト全長）」に
         // 正規化済みである前提（Anthropic は cache_read + cache_creation を
         // 加算して emit する）。
@@ -58,18 +87,21 @@ impl UsageTracker {
         let cache_read = event.cache_read_input_tokens.unwrap_or(0);
         let cache_write = event.cache_creation_input_tokens.unwrap_or(0);
         let output = event.output_tokens.unwrap_or(0);
-        self.pending_records.lock().unwrap().push(UsageRecord {
-            history_len,
-            input_total_tokens: input_total,
-            cache_read_tokens: cache_read,
-            cache_write_tokens: cache_write,
-            output_tokens: output,
+        self.pending_records.lock().unwrap().push(RecordedUsage {
+            record: UsageRecord {
+                history_len,
+                input_total_tokens: input_total,
+                cache_read_tokens: cache_read,
+                cache_write_tokens: cache_write,
+                output_tokens: output,
+            },
+            correlation_id,
         });
     }
 
     /// Drain accumulated records. Called by Pod after a run completes,
     /// before persisting the turn.
-    pub(crate) fn drain(&self) -> Vec<UsageRecord> {
+    pub(crate) fn drain(&self) -> Vec<RecordedUsage> {
         std::mem::take(&mut *self.pending_records.lock().unwrap())
     }
 }
@@ -96,11 +128,12 @@ mod tests {
 
         let records = tracker.drain();
         assert_eq!(records.len(), 1);
-        assert_eq!(records[0].history_len, 5);
-        assert_eq!(records[0].input_total_tokens, 1000);
-        assert_eq!(records[0].cache_read_tokens, 800);
-        assert_eq!(records[0].cache_write_tokens, 100);
-        assert_eq!(records[0].output_tokens, 42);
+        assert_eq!(records[0].record.history_len, 5);
+        assert_eq!(records[0].record.input_total_tokens, 1000);
+        assert_eq!(records[0].record.cache_read_tokens, 800);
+        assert_eq!(records[0].record.cache_write_tokens, 100);
+        assert_eq!(records[0].record.output_tokens, 42);
+        assert!(records[0].correlation_id.is_none());
     }
 
     #[test]
@@ -129,8 +162,25 @@ mod tests {
 
         let records = tracker.drain();
         assert_eq!(records.len(), 2);
-        assert_eq!(records[0].history_len, 5);
-        assert_eq!(records[1].history_len, 10);
-        assert_eq!(records[1].cache_read_tokens, 50);
+        assert_eq!(records[0].record.history_len, 5);
+        assert_eq!(records[1].record.history_len, 10);
+        assert_eq!(records[1].record.cache_read_tokens, 50);
+    }
+
+    #[test]
+    fn correlation_id_pairs_with_next_record_only() {
+        let tracker = UsageTracker::new();
+        // Stash an ID, then run a request → the ID should land on this record.
+        tracker.note_correlation_id("abc".into());
+        tracker.note_request(5);
+        tracker.record_usage(&make_event(100, 0, 0, 20));
+        // Next request without a fresh stash → no correlation_id.
+        tracker.note_request(10);
+        tracker.record_usage(&make_event(200, 50, 0, 30));
+
+        let records = tracker.drain();
+        assert_eq!(records.len(), 2);
+        assert_eq!(records[0].correlation_id.as_deref(), Some("abc"));
+        assert!(records[1].correlation_id.is_none());
     }
 }

crates/pod/src/pod.rs

@@ -77,6 +77,11 @@ pub struct Pod<C: LlmClient, St: Store> {
     /// Captures `(history_len, UsageEvent)` pairs during a run; drained
     /// in `persist_turn` and persisted as `LogEntry::LlmUsage` entries.
     usage_tracker: Arc<UsageTracker>,
+    /// Sync-side buffer for `Metric` values queued from inside Worker
+    /// callbacks (currently the prune observer). Drained in `persist_turn`
+    /// and written via `session_metrics::record_metric` alongside
+    /// `LogEntry::LlmUsage`. Always present after construction.
+    metrics_tracker: Arc<crate::compact::metrics_tracker::MetricsTracker>,
     /// Cumulative Usage measurement timeline, one entry per LLM call.
     /// Restored from session log on `restore`, appended on each persist.
     /// Read by token-accounting APIs (`Pod::total_tokens`, etc.).
@@ -203,6 +208,7 @@ impl<C: LlmClient, St: Store> Pod<C, St> {
             interceptor_installed: false,
             compact_state: None,
             usage_tracker: Arc::new(UsageTracker::new()),
+            metrics_tracker: Arc::new(crate::compact::metrics_tracker::MetricsTracker::new()),
             usage_history: Arc::new(Mutex::new(Vec::<UsageRecord>::new())),
             tracker: None,
             system_prompt_template: None,
@@ -391,6 +397,26 @@ impl<C: LlmClient, St: Store> Pod<C, St> {
         self.usage_history.clone()
     }
 
+    /// Handle to the per-LLM-request `UsageTracker`.
+    ///
+    /// Sibling modules (e.g. the prune observer) clone this `Arc` to stash
+    /// per-request side state (e.g. a `correlation_id`) that pairs with
+    /// the next `LlmUsage`.
+    pub(crate) fn usage_tracker_handle(&self) -> Arc<UsageTracker> {
+        self.usage_tracker.clone()
+    }
+
+    /// Handle to the synchronous `MetricsTracker` buffer.
+    ///
+    /// Worker callbacks (e.g. the prune observer) clone this `Arc` and
+    /// `.push(metric)` into it; Pod drains it in `persist_turn` and
+    /// writes each metric via `session_metrics::record_metric`.
+    pub(crate) fn metrics_tracker_handle(
+        &self,
+    ) -> Arc<crate::compact::metrics_tracker::MetricsTracker> {
+        self.metrics_tracker.clone()
+    }
+
     /// Attach the session-scoped file-operation tracker from the builtin
     /// `tools` crate. Called by the Controller immediately after it
     /// registers the builtin tools on the Worker. Overwrites any
@@ -1068,13 +1094,36 @@ impl<C: LlmClient, St: Store> Pod<C, St> {
         )
         .await?;
 
+        // Flush any sync-buffered metrics from this run first
+        // (currently `prune.fire` / `prune.skip` from the prune observer).
+        // Ordered before LlmUsage so that a `prune.fire` and the
+        // `prune.post_request` derived from the matching usage record
+        // appear in the log close together.
+        let pending_metrics = self.metrics_tracker.drain();
+        for metric in pending_metrics {
+            session_metrics::record_metric(
+                &self.store,
+                self.session_id,
+                &mut self.head_hash,
+                &metric,
+            )
+            .await?;
+        }
+
         // Persist any LLM Usage measurements collected during this run.
         // One LogEntry::LlmUsage per LLM call (the tool loop may have run
         // many calls within a single Pod::run). Each is also appended to
         // the in-memory `usage_history` so token-accounting APIs see it
-        // before the next run.
+        // before the next run. Records carrying a `correlation_id` (set
+        // by an upstream observer such as the prune projection) also get
+        // a paired `prune.post_request` metric so cache_read/write can be
+        // joined back to the originating event.
         let usage_records = self.usage_tracker.drain();
-        for record in usage_records {
+        for recorded in usage_records {
+            let crate::compact::usage_tracker::RecordedUsage {
+                record,
+                correlation_id,
+            } = recorded;
             session_store::save_usage(
                 &self.store,
                 self.session_id,
@@ -1086,6 +1135,20 @@ impl<C: LlmClient, St: Store> Pod<C, St> {
                 record.output_tokens,
             )
             .await?;
+            if let Some(id) = correlation_id {
+                let metric = session_metrics::Metric::now("prune.post_request")
+                    .with_correlation_id(&id)
+                    .with_value(record.cache_read_tokens as f64)
+                    .with_dimension("cache_write_tokens", record.cache_write_tokens.to_string())
+                    .with_dimension("history_len", record.history_len.to_string());
+                session_metrics::record_metric(
+                    &self.store,
+                    self.session_id,
+                    &mut self.head_hash,
+                    &metric,
+                )
+                .await?;
+            }
             self.usage_history
                 .lock()
                 .expect("usage_history poisoned")
@@ -1895,6 +1958,7 @@ impl<St: Store> Pod<Box<dyn LlmClient>, St> {
             interceptor_installed: false,
             compact_state: None,
             usage_tracker: Arc::new(UsageTracker::new()),
+            metrics_tracker: Arc::new(crate::compact::metrics_tracker::MetricsTracker::new()),
             usage_history: Arc::new(Mutex::new(Vec::new())),
             tracker: None,
             system_prompt_template: common.system_prompt_template,
@@ -1956,6 +2020,7 @@ impl<St: Store> Pod<Box<dyn LlmClient>, St> {
             interceptor_installed: false,
             compact_state: None,
             usage_tracker: Arc::new(UsageTracker::new()),
+            metrics_tracker: Arc::new(crate::compact::metrics_tracker::MetricsTracker::new()),
             usage_history: Arc::new(Mutex::new(Vec::new())),
             tracker: None,
             system_prompt_template: common.system_prompt_template,
@@ -2067,6 +2132,7 @@ impl<St: Store> Pod<Box<dyn LlmClient>, St> {
             interceptor_installed: false,
             compact_state: None,
             usage_tracker: Arc::new(UsageTracker::new()),
+            metrics_tracker: Arc::new(crate::compact::metrics_tracker::MetricsTracker::new()),
             usage_history: Arc::new(Mutex::new(state.usage_history)),
             tracker: None,
             // Restore replays the saved system_prompt verbatim — no

crates/pod/tests/session_metrics_test.rs

@@ -0,0 +1,367 @@
+//! End-to-end coverage for the prune-projection metrics path.
+//!
+//! Drives a Pod with a scripted mock LLM client and a custom tool that
+//! returns a long `ToolOutput.content`, then inspects the persisted
+//! session log to verify:
+//!
+//! - `prune.skip { reason: "no_candidates" }` lands when the protected-turn
+//!   window covers the entire history.
+//! - `prune.fire` lands once enough turns + usage measurements exist for
+//!   the projection to actually apply.
+//! - The fire metric and the immediately-following `prune.post_request`
+//!   metric share the same `correlation_id`, so cache_read / cache_write
+//!   from the LlmUsage that triggered the projection can be joined back
+//!   to the originating event.
+//! - `prune.skip { reason: "below_min_savings" }` lands when candidates
+//!   exist but their estimated savings are below the configured floor.
+
+use std::pin::Pin;
+use std::sync::Arc;
+use std::sync::atomic::{AtomicUsize, Ordering};
+
+use async_trait::async_trait;
+use futures::Stream;
+use llm_worker::Worker;
+use llm_worker::llm_client::event::{
+    Event as LlmEvent, ResponseStatus, StatusEvent, UsageEvent,
+};
+use llm_worker::llm_client::{ClientError, LlmClient, Request};
+use llm_worker::tool::{Tool, ToolDefinition, ToolError, ToolMeta, ToolOutput};
+use session_metrics::{DOMAIN, Metric, metrics_from_extensions};
+use session_store::FsStore;
+
+use pod::{Pod, PodManifest};
+
+#[derive(Clone)]
+struct MockClient {
+    responses: Arc<Vec<Vec<LlmEvent>>>,
+    call_count: Arc<AtomicUsize>,
+}
+
+impl MockClient {
+    fn new(responses: Vec<Vec<LlmEvent>>) -> Self {
+        Self {
+            responses: Arc::new(responses),
+            call_count: Arc::new(AtomicUsize::new(0)),
+        }
+    }
+}
+
+#[async_trait]
+impl LlmClient for MockClient {
+    fn clone_boxed(&self) -> Box<dyn LlmClient> {
+        Box::new(self.clone())
+    }
+
+    async fn stream(
+        &self,
+        _request: Request,
+    ) -> Result<Pin<Box<dyn Stream<Item = Result<LlmEvent, ClientError>> + Send>>, ClientError>
+    {
+        let count = self.call_count.fetch_add(1, Ordering::SeqCst);
+        if count >= self.responses.len() {
+            return Err(ClientError::Config("mock client exhausted".into()));
+        }
+        let events = self.responses[count].clone();
+        let stream = futures::stream::iter(events.into_iter().map(Ok));
+        Ok(Box::pin(stream))
+    }
+}
+
+/// Tool that returns a fixed `ToolOutput { summary, content: Some(big) }`.
+/// `content` is long enough for prune savings to comfortably clear small
+/// `min_savings` thresholds.
+struct BigContentTool {
+    summary: &'static str,
+    content: String,
+}
+
+#[async_trait]
+impl Tool for BigContentTool {
+    async fn execute(&self, _input: &str) -> Result<ToolOutput, ToolError> {
+        Ok(ToolOutput {
+            summary: self.summary.into(),
+            content: Some(self.content.clone()),
+        })
+    }
+}
+
+fn big_content_tool_definition(name: &'static str) -> ToolDefinition {
+    Arc::new(move || {
+        let summary = "tool result summary";
+        let content = "x".repeat(2048);
+        (
+            ToolMeta::new(name)
+                .description("test tool that returns a long content")
+                .input_schema(serde_json::json!({"type": "object"})),
+            Arc::new(BigContentTool { summary, content }) as Arc<dyn Tool>,
+        )
+    })
+}
+
+fn usage_event(input_total: u64, cache_read: u64, cache_write: u64, output: u64) -> LlmEvent {
+    LlmEvent::Usage(UsageEvent {
+        input_tokens: Some(input_total),
+        output_tokens: Some(output),
+        total_tokens: Some(input_total + output),
+        cache_read_input_tokens: Some(cache_read),
+        cache_creation_input_tokens: Some(cache_write),
+    })
+}
+
+/// Tool-call response from the assistant: emits a `tool_use` block then a
+/// usage event so usage_history gains a measurement on this turn.
+fn tool_use_response(call_id: &str, tool_name: &str) -> Vec<LlmEvent> {
+    vec![
+        LlmEvent::tool_use_start(0, call_id, tool_name),
+        LlmEvent::tool_input_delta(0, "{}"),
+        LlmEvent::tool_use_stop(0),
+        usage_event(500, 0, 0, 10),
+        LlmEvent::Status(StatusEvent {
+            status: ResponseStatus::Completed,
+        }),
+    ]
+}
+
+/// Plain text response with explicit cache_read/cache_write so that
+/// `prune.post_request` can carry meaningful values when this is the
+/// LLM call that follows a `prune.fire` event.
+fn text_response_with_cache(text: &str, cache_read: u64, cache_write: u64) -> Vec<LlmEvent> {
+    vec![
+        LlmEvent::text_block_start(0),
+        LlmEvent::text_delta(0, text),
+        LlmEvent::text_block_stop(0, None),
+        usage_event(800, cache_read, cache_write, 5),
+        LlmEvent::Status(StatusEvent {
+            status: ResponseStatus::Completed,
+        }),
+    ]
+}
+
+fn manifest_toml(prune_protected_turns: usize, prune_min_savings: u64) -> String {
+    format!(
+        r#"
+[pod]
+name = "test-pod"
+pwd = "./"
+
+[model]
+scheme = "anthropic"
+model_id = "test-model"
+
+[worker]
+max_tokens = 100
+
+[compaction]
+prune_protected_turns = {prune_protected_turns}
+prune_min_savings = {prune_min_savings}
+
+[[scope.allow]]
+target = "./"
+permission = "write"
+"#
+    )
+}
+
+async fn make_pod(
+    manifest_toml: String,
+    client: MockClient,
+    tool_name: &'static str,
+) -> (Pod<MockClient, FsStore>, tempfile::TempDir, tempfile::TempDir) {
+    let manifest = PodManifest::from_toml(&manifest_toml).unwrap();
+    let store_tmp = tempfile::tempdir().unwrap();
+    let store = FsStore::new(store_tmp.path()).await.unwrap();
+    let pwd_tmp = tempfile::tempdir().unwrap();
+    let pwd = pwd_tmp.path().to_path_buf();
+    let scope = pod::Scope::writable(&pwd).unwrap();
+
+    let mut worker = Worker::new(client);
+    worker.register_tool(big_content_tool_definition(tool_name));
+
+    let pod = Pod::new(manifest, worker, store, pwd, scope).await.unwrap();
+    (pod, store_tmp, pwd_tmp)
+}
+
+/// Drive Pod through enough runs to exercise both skip-no_candidates and
+/// fire branches, then read the session log back and assert the metric
+/// stream.
+#[tokio::test]
+async fn prune_metrics_emit_skip_then_fire_with_post_request_join() {
+    // Run 1 (request 0): tool_use → triggers tool execution → request 1
+    //   on the second iteration to produce the assistant reply.
+    // Run 2 (request 2): plain assistant text. Prune evaluation here
+    //   sees user1's tool_result outside the 1-protected-turn window and
+    //   should fire.
+    let client = MockClient::new(vec![
+        tool_use_response("call-1", "big_tool"),
+        text_response_with_cache("ok", 0, 200),
+        text_response_with_cache("done", 1234, 50),
+    ]);
+    let (mut pod, _store_tmp, _pwd_tmp) =
+        make_pod(manifest_toml(1, 1), client, "big_tool").await;
+    let session_id = pod.session_id();
+    // Cloning the store handle to read the session log back after the
+    // runs complete — the Pod retains its own copy.
+    let store = pod.store().clone();
+
+    pod.run_text("first").await.unwrap();
+    pod.run_text("second").await.unwrap();
+
+    let state = session_store::restore(&store, session_id).await.unwrap();
+    let metrics = metrics_from_extensions(&state.extensions);
+
+    // Run 1 has 2 LLM iterations (tool loop), each evaluates prune with
+    // only one user-message turn → 2x skip{no_candidates}.
+    // Run 2 has 1 LLM iteration with enough turns → 1x fire +
+    // 1x post_request paired by correlation_id.
+    let names: Vec<&str> = metrics.iter().map(|m| m.name.as_str()).collect();
+    assert!(
+        names.contains(&"prune.skip"),
+        "expected prune.skip in {names:?}"
+    );
+    assert!(
+        names.contains(&"prune.fire"),
+        "expected prune.fire in {names:?}"
+    );
+    assert!(
+        names.contains(&"prune.post_request"),
+        "expected prune.post_request in {names:?}"
+    );
+
+    // All skips in run 1 must record reason=no_candidates.
+    for m in metrics.iter().filter(|m| m.name == "prune.skip") {
+        assert_eq!(
+            m.dimensions.get("reason").map(String::as_str),
+            Some("no_candidates"),
+            "skip metric should be no_candidates here, got {m:?}"
+        );
+        assert!(m.correlation_id.is_none());
+    }
+
+    // The fire metric carries dimensions and correlation_id.
+    let fire = metrics
+        .iter()
+        .find(|m| m.name == "prune.fire")
+        .expect("prune.fire missing");
+    assert!(
+        fire.dimensions.contains_key("candidate_count"),
+        "fire missing candidate_count: {fire:?}"
+    );
+    assert!(
+        fire.dimensions.contains_key("border_turn"),
+        "fire missing border_turn: {fire:?}"
+    );
+    assert!(
+        fire.value.is_some(),
+        "fire missing estimated_savings value"
+    );
+    let fire_id = fire
+        .correlation_id
+        .as_ref()
+        .expect("fire metric missing correlation_id");
+
+    // Exactly one post_request metric should exist with the same id, and
+    // its value/dimension should reflect the cache numbers from the
+    // text_response_with_cache call (cache_read=1234, cache_write=50).
+    let post = metrics
+        .iter()
+        .find(|m| m.name == "prune.post_request")
+        .expect("prune.post_request missing");
+    assert_eq!(post.correlation_id.as_ref(), Some(fire_id));
+    assert_eq!(post.value, Some(1234.0));
+    assert_eq!(
+        post.dimensions.get("cache_write_tokens").map(String::as_str),
+        Some("50")
+    );
+    assert!(post.dimensions.contains_key("history_len"));
+}
+
+/// `min_savings` set high enough that candidates exist but the estimated
+/// savings always fall short → the second run should record
+/// `prune.skip { reason: "below_min_savings" }`.
+#[tokio::test]
+async fn prune_metrics_record_below_min_savings_skip() {
+    let client = MockClient::new(vec![
+        tool_use_response("call-1", "big_tool"),
+        text_response_with_cache("ok", 0, 100),
+        text_response_with_cache("done", 0, 0),
+    ]);
+    let (mut pod, _store_tmp, _pwd_tmp) =
+        make_pod(manifest_toml(1, u64::MAX), client, "big_tool").await;
+    let session_id = pod.session_id();
+    let store = pod.store().clone();
+
+    pod.run_text("first").await.unwrap();
+    pod.run_text("second").await.unwrap();
+
+    let state = session_store::restore(&store, session_id).await.unwrap();
+    let metrics = metrics_from_extensions(&state.extensions);
+    let below = metrics
+        .iter()
+        .find(|m| {
+            m.name == "prune.skip"
+                && m.dimensions.get("reason").map(String::as_str) == Some("below_min_savings")
+        })
+        .expect("expected prune.skip with reason=below_min_savings");
+    assert!(
+        below.dimensions.contains_key("candidate_count"),
+        "below_min_savings skip should report candidate_count: {below:?}"
+    );
+    assert!(
+        below.value.is_some(),
+        "below_min_savings skip should report estimated savings as value: {below:?}"
+    );
+    // No prune.fire for this scenario.
+    assert!(metrics.iter().all(|m| m.name != "prune.fire"));
+    // No prune.post_request either (no fire to join with).
+    assert!(metrics.iter().all(|m| m.name != "prune.post_request"));
+}
+
+/// Sessions that have no metrics in the log restore cleanly: the
+/// `RestoredState.extensions` simply contains no `metrics` domain
+/// payloads, and `metrics_from_extensions` returns an empty Vec.
+/// Backward-compatibility check for old logs predating this feature.
+#[tokio::test]
+async fn old_sessions_without_metrics_replay_cleanly() {
+    // Manifest without any `[compaction]` section → prune (and therefore
+    // the prune observer) is never installed, so no metrics get written.
+    let manifest_toml = r#"
+[pod]
+name = "test-pod"
+pwd = "./"
+
+[model]
+scheme = "anthropic"
+model_id = "test-model"
+
+[worker]
+max_tokens = 100
+
+[[scope.allow]]
+target = "./"
+permission = "write"
+"#;
+    let client = MockClient::new(vec![text_response_with_cache("hi", 0, 0)]);
+    let manifest = PodManifest::from_toml(manifest_toml).unwrap();
+    let store_tmp = tempfile::tempdir().unwrap();
+    let store = FsStore::new(store_tmp.path()).await.unwrap();
+    let pwd_tmp = tempfile::tempdir().unwrap();
+    let pwd = pwd_tmp.path().to_path_buf();
+    let scope = pod::Scope::writable(&pwd).unwrap();
+    let worker = Worker::new(client);
+    let mut pod = Pod::new(manifest, worker, store.clone(), pwd, scope)
+        .await
+        .unwrap();
+    let session_id = pod.session_id();
+    pod.run_text("hello").await.unwrap();
+
+    let state = session_store::restore(&store, session_id).await.unwrap();
+    let metrics = metrics_from_extensions(&state.extensions);
+    assert!(metrics.is_empty(), "no metrics should be recorded: {metrics:?}");
+    // And no extension entries at all in the metrics domain.
+    assert!(state.extensions.iter().all(|(d, _)| d != DOMAIN));
+
+    // Smoke check that fold helper is robust on a sentinel Metric value:
+    let m = Metric::now("smoke");
+    assert_eq!(m.name, "smoke");
+}