deepstrike 0.2.2__tar.gz → 0.2.3__tar.gz

{deepstrike-0.2.2 → deepstrike-0.2.3}/Cargo.lock

@@ -62,33 +62,12 @@ version = "1.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c08606f8c3cbf4ce6ec8e28fb0014a2c086708fe954eaa885384a6165172e7e8"
 
-[[package]]
-name = "base64"
-version = "0.21.7"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9d297deb1925b89f2ccc13d7635fa0714f12c87adce1c75356b39ca9b7178567"
-
 [[package]]
 name = "base64"
 version = "0.22.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "72b3254f16251a8381aa12e40e3c4d2f0199f8c6508fbecb9d91f575e0fbb8c6"
 
-[[package]]
-name = "bit-set"
-version = "0.5.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0700ddab506f33b20a03b13996eccd309a48e5ff77d0d95926aa0210fb4e95f1"
-dependencies = [
- "bit-vec",
-]
-
-[[package]]
-name = "bit-vec"
-version = "0.6.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "349f9b6a179ed607305526ca489b34ad0a41aed5f7980fa90eb03160b69598fb"
-
 [[package]]
 name = "bitflags"
 version = "1.3.2"
@@ -101,17 +80,6 @@ version = "2.11.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c4512299f36f043ab09a583e57bceb5a5aab7a73db1805848e8fef3c9e8c78b3"
 
-[[package]]
-name = "bstr"
-version = "1.12.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "63044e1ae8e69f3b5a92c736ca6269b8d12fa7efe39bf34ddb06d102cf0e2cab"
-dependencies = [
- "memchr",
- "regex-automata",
- "serde",
-]
-
 [[package]]
 name = "bumpalo"
 version = "3.20.2"
@@ -226,10 +194,9 @@ dependencies = [
 
 [[package]]
 name = "deepstrike-core"
-version = "0.2.2"
+version = "0.2.3"
 dependencies = [
  "compact_str",
- "deepstrike-tokenizer",
  "pretty_assertions",
  "serde",
  "serde_json",
@@ -239,11 +206,10 @@ dependencies = [
 
 [[package]]
 name = "deepstrike-node"
-version = "0.2.2"
+version = "0.2.3"
 dependencies = [
  "compact_str",
  "deepstrike-core",
- "deepstrike-tokenizer",
  "napi",
  "napi-build",
  "napi-derive",
@@ -252,11 +218,10 @@ dependencies = [
 
 [[package]]
 name = "deepstrike-py"
-version = "0.2.2"
+version = "0.2.3"
 dependencies = [
  "compact_str",
  "deepstrike-core",
- "deepstrike-tokenizer",
  "pyo3",
  "serde_json",
  "uuid",
@@ -264,7 +229,7 @@ dependencies = [
 
 [[package]]
 name = "deepstrike-sdk"
-version = "0.2.2"
+version = "0.2.3"
 dependencies = [
  "async-stream",
  "async-trait",
@@ -296,21 +261,12 @@ dependencies = [
  "tokio",
 ]
 
-[[package]]
-name = "deepstrike-tokenizer"
-version = "0.2.2"
-dependencies = [
- "pretty_assertions",
- "tiktoken-rs",
-]
-
 [[package]]
 name = "deepstrike-wasm"
-version = "0.2.2"
+version = "0.2.3"
 dependencies = [
  "compact_str",
  "deepstrike-core",
- "deepstrike-tokenizer",
  "js-sys",
  "serde",
  "serde-wasm-bindgen",
@@ -368,17 +324,6 @@ dependencies = [
  "windows-sys 0.61.2",
 ]
 
-[[package]]
-name = "fancy-regex"
-version = "0.13.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "531e46835a22af56d1e3b66f04844bed63158bc094a628bec1d321d9b4c44bf2"
-dependencies = [
- "bit-set",
- "regex-automata",
- "regex-syntax",
-]
-
 [[package]]
 name = "fastrand"
 version = "2.4.1"
@@ -708,7 +653,7 @@ version = "0.1.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "96547c2556ec9d12fb1578c4eaf448b04993e7fb79cbaad930a656880a6bdfa0"
 dependencies = [
- "base64 0.22.1",
+ "base64",
  "bytes",
  "futures-channel",
  "futures-util",
@@ -921,12 +866,6 @@ dependencies = [
  "libc",
 ]
 
-[[package]]
-name = "lazy_static"
-version = "1.5.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bbd2bcb4c963f2ddae06a2efc7e9f3591312473c50c6685e1f298068316e66fe"
-
 [[package]]
 name = "leb128fmt"
 version = "0.1.0"
@@ -1379,7 +1318,7 @@ version = "0.12.28"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "eddd3ca559203180a307f12d114c268abf583f59b03cb906fd0b3ff8646c1147"
 dependencies = [
- "base64 0.22.1",
+ "base64",
  "bytes",
  "encoding_rs",
  "futures-core",
@@ -1430,12 +1369,6 @@ dependencies = [
  "windows-sys 0.52.0",
 ]
 
-[[package]]
-name = "rustc-hash"
-version = "1.1.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "08d43f7aa6b08d49f382cde6a7982047c3426db949b1424bc4b7ec9ae12c6ce2"
-
 [[package]]
 name = "rustix"
 version = "1.1.4"
@@ -1771,22 +1704,6 @@ dependencies = [
  "syn",
 ]
 
-[[package]]
-name = "tiktoken-rs"
-version = "0.6.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "44075987ee2486402f0808505dd65692163d243a337fc54363d49afac41087f6"
-dependencies = [
- "anyhow",
- "base64 0.21.7",
- "bstr",
- "fancy-regex",
- "lazy_static",
- "parking_lot",
- "regex",
- "rustc-hash",
-]
-
 [[package]]
 name = "tinystr"
 version = "0.8.3"

{deepstrike-0.2.2 → deepstrike-0.2.3}/Cargo.toml

@@ -1,16 +1,15 @@
 [workspace]
 resolver = "2"
-members = ["crates/deepstrike-core", "crates/deepstrike-tokenizer", "crates/deepstrike-py"]
+members = ["crates/deepstrike-core", "crates/deepstrike-py"]
 
 [workspace.package]
-version = "0.2.2"
+version = "0.2.3"
 edition = "2024"
 license = "MIT"
 repository = "https://github.com/anthropics/deepstrike"
 
 [workspace.dependencies]
-deepstrike-core = { path = "crates/deepstrike-core", version = "0.2.2" }
-deepstrike-tokenizer = { path = "crates/deepstrike-tokenizer", version = "0.2.2" }
+deepstrike-core = { path = "crates/deepstrike-core", version = "0.2.3" }
 notify = "6"
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"

{deepstrike-0.2.2 → deepstrike-0.2.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: deepstrike
-Version: 0.2.2
+Version: 0.2.3
 Requires-Dist: httpx>=0.27
 Requires-Dist: pyyaml>=6.0
 Requires-Dist: anyio>=4.0
@@ -87,6 +87,29 @@ All providers accept `RetryConfig` for exponential backoff and share a `CircuitB
 
 ---
 
+## Context model (four slots)
+
+The kernel renders context as four LLM API slots — only **history** is compressed.
+
+| Slot | Source | Role |
+|------|--------|------|
+| `system_stable` | system partition | Identity, rules — never changes within a run |
+| `system_knowledge` | knowledge partition | Preloaded memory, skill defs — low frequency |
+| `turns[0]` | `task_state` + signals | Goal, plan, progress, compression log, runtime signals |
+| `turns[1..N]` | history | Conversation transcript |
+
+```python
+runner = RuntimeRunner(RuntimeOptions(
+    initial_memory=["User prefers chartreuse."],  # → Slot 2
+    system_prompt="You are a helpful assistant.",  # → Slot 1
+    # ...
+))
+```
+
+Full reference: [docs/context-partition-compression.md](../docs/context-partition-compression.md)
+
+---
+
 ## Runtime options
 
 ```python
@@ -104,6 +127,11 @@ runner = RuntimeRunner(RuntimeOptions(
     signal_source=gateway,      # SignalGateway or any SignalSource
     dream_store=my_store,       # DreamStore for long-term memory
     agent_id="my-agent",        # required with dream_store for memory meta-tool
+    initial_memory=["..."],     # preloaded blocks → Slot 2 (system_knowledge)
+    sub_agent_harness=SubAgentHarnessConfig(  # optional: HarnessLoop for spawned sub-agents
+        eval_provider=eval_provider,
+        max_attempts=3,
+    ),
 ))
 ```
 
@@ -151,7 +179,7 @@ effort: 1
 
 ## Knowledge
 
-Implement `KnowledgeSource` — the kernel injects a `knowledge` meta-tool.
+Implement `KnowledgeSource` — the kernel injects a `knowledge` meta-tool. **Runtime retrieval results land in history** as tool results. Use `initial_memory` for durable preload into Slot 2.
 
 ```python
 from deepstrike import KnowledgeSource
@@ -177,7 +205,9 @@ runner = RuntimeRunner(RuntimeOptions(
 
 ## Memory
 
-### WorkingMemory (in-session scratch pad)
+### WorkingMemory (SDK-side scratch pad)
+
+`WorkingMemory` is an SDK helper — not the kernel `working` partition (removed). Kernel task state renders into Slot 3 (`turns[0]`).
 
 ```python
 from deepstrike import WorkingMemory
@@ -210,7 +240,8 @@ runner = RuntimeRunner(RuntimeOptions(
     agent_id="my-agent",
 ))
 
-# In-session: LLM calls memory(query) → DreamStore.search()
+# In-session: LLM calls memory(query) → DreamStore.search() → history tool result
+# Preload:    initial_memory → Slot 2 (system_knowledge)
 # Post-session: trigger memory consolidation
 result = await runner.dream("my-agent", now_ms=int(time.time() * 1000))
 ```
@@ -295,6 +326,16 @@ outcome = await EvalLoopHarness(runner, gate=ContainsHello(), max_attempts=3).ru
 
 # 3. HarnessLoop — LLM-as-judge with feedback injection + skill extraction
 loop = HarnessLoop(runner, eval_provider=eval_provider, max_attempts=3, skill_dir="./skills")
+
+# Sub-agents: pass sub_agent_harness on RuntimeOptions to auto-evaluate spawned children
+from deepstrike import SubAgentHarnessConfig
+runner = RuntimeRunner(RuntimeOptions(
+    provider=provider,
+    session_log=InMemorySessionLog(),
+    execution_plane=plane,
+    sub_agent_harness=SubAgentHarnessConfig(eval_provider=eval_provider, max_attempts=3),
+    # ...
+))
 async for event in loop.run_streaming(HarnessRequest(goal="Write a haiku")):
     if event.type == "done":
         print(event.verdict.passed, event.verdict.feedback)

{deepstrike-0.2.2 → deepstrike-0.2.3}/README.md

@@ -74,6 +74,29 @@ All providers accept `RetryConfig` for exponential backoff and share a `CircuitB
 
 ---
 
+## Context model (four slots)
+
+The kernel renders context as four LLM API slots — only **history** is compressed.
+
+| Slot | Source | Role |
+|------|--------|------|
+| `system_stable` | system partition | Identity, rules — never changes within a run |
+| `system_knowledge` | knowledge partition | Preloaded memory, skill defs — low frequency |
+| `turns[0]` | `task_state` + signals | Goal, plan, progress, compression log, runtime signals |
+| `turns[1..N]` | history | Conversation transcript |
+
+```python
+runner = RuntimeRunner(RuntimeOptions(
+    initial_memory=["User prefers chartreuse."],  # → Slot 2
+    system_prompt="You are a helpful assistant.",  # → Slot 1
+    # ...
+))
+```
+
+Full reference: [docs/context-partition-compression.md](../docs/context-partition-compression.md)
+
+---
+
 ## Runtime options
 
 ```python
@@ -91,6 +114,11 @@ runner = RuntimeRunner(RuntimeOptions(
     signal_source=gateway,      # SignalGateway or any SignalSource
     dream_store=my_store,       # DreamStore for long-term memory
     agent_id="my-agent",        # required with dream_store for memory meta-tool
+    initial_memory=["..."],     # preloaded blocks → Slot 2 (system_knowledge)
+    sub_agent_harness=SubAgentHarnessConfig(  # optional: HarnessLoop for spawned sub-agents
+        eval_provider=eval_provider,
+        max_attempts=3,
+    ),
 ))
 ```
 
@@ -138,7 +166,7 @@ effort: 1
 
 ## Knowledge
 
-Implement `KnowledgeSource` — the kernel injects a `knowledge` meta-tool.
+Implement `KnowledgeSource` — the kernel injects a `knowledge` meta-tool. **Runtime retrieval results land in history** as tool results. Use `initial_memory` for durable preload into Slot 2.
 
 ```python
 from deepstrike import KnowledgeSource
@@ -164,7 +192,9 @@ runner = RuntimeRunner(RuntimeOptions(
 
 ## Memory
 
-### WorkingMemory (in-session scratch pad)
+### WorkingMemory (SDK-side scratch pad)
+
+`WorkingMemory` is an SDK helper — not the kernel `working` partition (removed). Kernel task state renders into Slot 3 (`turns[0]`).
 
 ```python
 from deepstrike import WorkingMemory
@@ -197,7 +227,8 @@ runner = RuntimeRunner(RuntimeOptions(
     agent_id="my-agent",
 ))
 
-# In-session: LLM calls memory(query) → DreamStore.search()
+# In-session: LLM calls memory(query) → DreamStore.search() → history tool result
+# Preload:    initial_memory → Slot 2 (system_knowledge)
 # Post-session: trigger memory consolidation
 result = await runner.dream("my-agent", now_ms=int(time.time() * 1000))
 ```
@@ -282,6 +313,16 @@ outcome = await EvalLoopHarness(runner, gate=ContainsHello(), max_attempts=3).ru
 
 # 3. HarnessLoop — LLM-as-judge with feedback injection + skill extraction
 loop = HarnessLoop(runner, eval_provider=eval_provider, max_attempts=3, skill_dir="./skills")
+
+# Sub-agents: pass sub_agent_harness on RuntimeOptions to auto-evaluate spawned children
+from deepstrike import SubAgentHarnessConfig
+runner = RuntimeRunner(RuntimeOptions(
+    provider=provider,
+    session_log=InMemorySessionLog(),
+    execution_plane=plane,
+    sub_agent_harness=SubAgentHarnessConfig(eval_provider=eval_provider, max_attempts=3),
+    # ...
+))
 async for event in loop.run_streaming(HarnessRequest(goal="Write a haiku")):
     if event.type == "done":
         print(event.verdict.passed, event.verdict.feedback)

{deepstrike-0.2.2 → deepstrike-0.2.3}/crates/deepstrike-core/Cargo.toml

@@ -11,7 +11,6 @@ serde_json = { workspace = true }
 thiserror = { workspace = true }
 compact_str = { workspace = true }
 uuid = { workspace = true }
-deepstrike-tokenizer = { workspace = true }
 
 [dev-dependencies]
 pretty_assertions = "1"

{deepstrike-0.2.2 → deepstrike-0.2.3}/crates/deepstrike-core/src/context/compression.rs

@@ -22,6 +22,7 @@ pub trait Compressor: Send + Sync {
         partitions: &mut ContextPartitions,
         target_tokens: u32,
         max_tokens: u32,
+        preserve_k: usize,
         summarizer: &dyn Summarizer,
         engine: &ContextTokenEngine,
     ) -> CompressResult;
@@ -38,6 +39,7 @@ impl Compressor for SnipCompactor {
         partitions: &mut ContextPartitions,
         _target_tokens: u32,
         max_tokens: u32,
+        _preserve_k: usize,
         _summarizer: &dyn Summarizer,
         engine: &ContextTokenEngine,
     ) -> CompressResult {
@@ -93,6 +95,14 @@ impl Compressor for SnipCompactor {
         }
 
         partition.token_count = partition.token_count.saturating_sub(saved);
+
+        if saved > 0 {
+            partitions.task_state.log_compression(
+                "snip_compact",
+                format!("{saved} tokens truncated from oversized messages"),
+            );
+        }
+
         CompressResult {
             tokens_saved: saved,
             summary: None,
@@ -195,6 +205,7 @@ impl Compressor for MicroCompactor {
         partitions: &mut ContextPartitions,
         _target_tokens: u32,
         _max_tokens: u32,
+        _preserve_k: usize,
         _summarizer: &dyn Summarizer,
         engine: &ContextTokenEngine,
     ) -> CompressResult {
@@ -269,6 +280,14 @@ impl Compressor for MicroCompactor {
         }
 
         partition.token_count = partition.token_count.saturating_sub(saved);
+
+        if saved > 0 {
+            partitions.task_state.log_compression(
+                "micro_compact",
+                format!("{saved} tokens excerpted from tool results"),
+            );
+        }
+
         CompressResult {
             tokens_saved: saved,
             summary: None,
@@ -286,6 +305,7 @@ impl Compressor for CollapseCompactor {
         partitions: &mut ContextPartitions,
         target_tokens: u32,
         _max_tokens: u32,
+        preserve_k: usize,
         summarizer: &dyn Summarizer,
         engine: &ContextTokenEngine,
     ) -> CompressResult {
@@ -293,8 +313,8 @@ impl Compressor for CollapseCompactor {
         let mut saved = 0u32;
         let mut n = 0usize;
 
-        // Preserve last K = 2 turns (4 messages) to maintain conversational context
-        let limit = partition.messages.len().saturating_sub(4);
+        let keep = preserve_k * 2; // turns → messages (user + assistant per turn)
+        let limit = partition.messages.len().saturating_sub(keep);
         for i in 0..limit {
             if partition.token_count.saturating_sub(saved) <= target_tokens {
                 break;
@@ -315,27 +335,20 @@ impl Compressor for CollapseCompactor {
         let archived: Vec<Message> = partition.messages.drain(..n).collect();
         let summary_text =
             summarizer.summarize(&archived, PressureAction::ContextCollapse, target_tokens);
-        let summary_tokens = engine.count(&summary_text);
-
-        let mut summary_msg = Message::assistant(summary_text.clone());
-        summary_msg.token_count = Some(summary_tokens);
 
-        partition.messages.insert(0, summary_msg);
+        partition.token_count = partition.token_count.saturating_sub(saved);
 
-        partition.token_count = partition
-            .token_count
-            .saturating_sub(saved)
-            .saturating_add(summary_tokens);
+        partitions.task_state.log_compression("context_collapse", summary_text.clone());
 
         CompressResult {
-            tokens_saved: saved.saturating_sub(summary_tokens),
+            tokens_saved: saved,
             summary: Some(summary_text),
             archived,
         }
     }
 }
 
-/// rho > auto_threshold: collapse history entirely except last K turns, updating scratchpad.
+/// rho > auto_threshold: collapse history entirely except last K turns, updating compression log.
 pub struct AutoCompactor;
 
 impl Compressor for AutoCompactor {
@@ -344,6 +357,7 @@ impl Compressor for AutoCompactor {
         partitions: &mut ContextPartitions,
         _target_tokens: u32,
         _max_tokens: u32,
+        preserve_k: usize,
         summarizer: &dyn Summarizer,
         engine: &ContextTokenEngine,
     ) -> CompressResult {
@@ -357,9 +371,8 @@ impl Compressor for AutoCompactor {
         }
 
         let original_tokens = partition.token_count;
-
-        // Preserve last K = 2 turns (4 messages)
-        let limit = partition.messages.len().saturating_sub(4);
+        let keep = preserve_k * 2;
+        let limit = partition.messages.len().saturating_sub(keep);
         let (archived, kept): (Vec<Message>, Vec<Message>) = if limit > 0 {
             let archived_msgs = partition.messages.drain(..limit).collect();
             let kept_msgs = partition.messages.drain(..).collect();
@@ -381,12 +394,6 @@ impl Compressor for AutoCompactor {
 
         let summary_text =
             summarizer.summarize(&archived, PressureAction::AutoCompact, _max_tokens);
-        let summary_tokens = engine.count(&summary_text);
-
-        partitions.task_state.scratchpad = summary_text.clone();
-
-        let working_msg = Message::system(summary_text.clone());
-        partitions.working.push(working_msg, summary_tokens);
 
         let kept_tokens: u32 = partition
             .messages
@@ -395,6 +402,8 @@ impl Compressor for AutoCompactor {
             .sum();
         partition.token_count = kept_tokens;
 
+        partitions.task_state.log_compression("auto_compact", summary_text.clone());
+
         CompressResult {
             tokens_saved: original_tokens.saturating_sub(kept_tokens),
             summary: Some(summary_text),
@@ -406,11 +415,13 @@ impl Compressor for AutoCompactor {
 /// Compression pipeline — operates on history partition but can reference full partitions.
 pub struct CompressionPipeline {
     stages: Vec<(PressureAction, Box<dyn Compressor>)>,
+    preserve_recent_turns: usize,
 }
 
 impl CompressionPipeline {
     pub fn new(config: &ContextConfig) -> Self {
         Self {
+            preserve_recent_turns: config.preserve_recent_turns,
             stages: vec![
                 (
                     PressureAction::SnipCompact,
@@ -444,12 +455,17 @@ impl CompressionPipeline {
 
         for (stage_action, compressor) in &self.stages {
             if *stage_action <= action {
-                // 如果当前总 token 已经满足预算目标，则无需继续进行更重级别的压缩
                 if partitions.total_tokens(engine) <= target_tokens {
                     break;
                 }
-                let res =
-                    compressor.compress(partitions, target_tokens, max_tokens, &summarizer, engine);
+                let res = compressor.compress(
+                    partitions,
+                    target_tokens,
+                    max_tokens,
+                    self.preserve_recent_turns,
+                    &summarizer,
+                    engine,
+                );
                 total_saved += res.tokens_saved;
                 if let Some(s) = res.summary {
                     if !s.is_empty() {
@@ -507,7 +523,7 @@ mod tests {
         };
         let mut ctx = ContextPartitions::new(&cfg);
         ctx.history.push(Message::user("a".repeat(800)), 200);
-        let result = compactor.compress(&mut ctx, 0, MAX, &summarizer(), &engine());
+        let result = compactor.compress(&mut ctx, 0, MAX, 2, &summarizer(), &engine());
         assert!(result.tokens_saved > 0);
         if let Content::Text(ref t) = ctx.history.messages[0].content {
             assert!(t.contains("… [… 100 tokens omitted …] …"), "got: {t}");
@@ -525,7 +541,7 @@ mod tests {
         };
         let mut ctx = ContextPartitions::new(&cfg);
         ctx.history.push(Message::user("short"), 5);
-        let result = compactor.compress(&mut ctx, 0, MAX, &summarizer(), &engine());
+        let result = compactor.compress(&mut ctx, 0, MAX, 2, &summarizer(), &engine());
         assert_eq!(result.tokens_saved, 0);
     }
 
@@ -550,7 +566,7 @@ mod tests {
         ctx.history.messages.push(msg);
         ctx.history.token_count = 300;
 
-        let result = compactor.compress(&mut ctx, 0, MAX, &summarizer(), &engine());
+        let result = compactor.compress(&mut ctx, 0, MAX, 2, &summarizer(), &engine());
         assert!(result.tokens_saved > 0);
         let text = ctx.history.messages[0].content.as_text().unwrap();
         assert!(
@@ -566,9 +582,10 @@ mod tests {
         for _ in 0..8 {
             ctx.history.push(Message::user("msg"), 50);
         }
-        let result = compactor.compress(&mut ctx, 250, MAX, &summarizer(), &engine());
+        let result = compactor.compress(&mut ctx, 250, MAX, 2, &summarizer(), &engine());
         assert!(result.tokens_saved > 0);
         assert!(ctx.history.messages.len() < 8);
+        assert!(ctx.task_state.compression_log.iter().any(|e| e.action == "context_collapse"));
     }
 
     #[test]
@@ -618,7 +635,7 @@ mod tests {
         ctx.history.messages.push(msg);
         ctx.history.token_count = 300;
 
-        let result = compactor.compress(&mut ctx, 0, MAX, &summarizer(), &engine());
+        let result = compactor.compress(&mut ctx, 0, MAX, 2, &summarizer(), &engine());
         // Since call_id "keep_me" is in preserved_refs, it should not be replaced!
         assert_eq!(result.tokens_saved, 0);
         let text_opt = ctx.history.messages[0].content.as_text();
@@ -635,12 +652,12 @@ mod tests {
         for i in 0..10 {
             ctx.history.push(Message::user(format!("msg {i}")), 10);
         }
-        let result = compactor.compress(&mut ctx, 0, MAX, &summarizer(), &engine());
+        let result = compactor.compress(&mut ctx, 0, MAX, 2, &summarizer(), &engine());
         assert!(result.tokens_saved > 0);
         assert_eq!(ctx.history.messages.len(), 4); // kept last 2 turns = 4 messages
         assert!(result.summary.is_some());
-        assert!(!ctx.task_state.scratchpad.is_empty());
-        assert_eq!(ctx.working.messages.len(), 1); // summary pushed to working partition
+        // Summary now routes through compression_log → systemVolatile
+        assert!(ctx.task_state.compression_log.iter().any(|e| e.action == "auto_compact"));
     }
 
     #[test]