deepstrike 0.2.82__tar.gz → 0.2.83__tar.gz

{deepstrike-0.2.82 → deepstrike-0.2.83}/Cargo.lock

@@ -293,7 +293,7 @@ dependencies = [
 
 [[package]]
 name = "deepstrike-core"
-version = "0.2.82"
+version = "0.2.83"
 dependencies = [
  "compact_str",
  "deepstrike-tokenizer",
@@ -307,7 +307,7 @@ dependencies = [
 
 [[package]]
 name = "deepstrike-lab"
-version = "0.2.82"
+version = "0.2.83"
 dependencies = [
  "serde",
  "serde_json",
@@ -315,7 +315,7 @@ dependencies = [
 
 [[package]]
 name = "deepstrike-node"
-version = "0.2.82"
+version = "0.2.83"
 dependencies = [
  "compact_str",
  "deepstrike-core",
@@ -327,7 +327,7 @@ dependencies = [
 
 [[package]]
 name = "deepstrike-py"
-version = "0.2.82"
+version = "0.2.83"
 dependencies = [
  "compact_str",
  "deepstrike-core",
@@ -338,7 +338,7 @@ dependencies = [
 
 [[package]]
 name = "deepstrike-sdk"
-version = "0.2.82"
+version = "0.2.83"
 dependencies = [
  "async-stream",
  "async-trait",
@@ -376,7 +376,7 @@ dependencies = [
 
 [[package]]
 name = "deepstrike-tokenizer"
-version = "0.2.82"
+version = "0.2.83"
 dependencies = [
  "pretty_assertions",
  "tiktoken-rs",
@@ -384,7 +384,7 @@ dependencies = [
 
 [[package]]
 name = "deepstrike-wasm"
-version = "0.2.82"
+version = "0.2.83"
 dependencies = [
  "compact_str",
  "deepstrike-core",

{deepstrike-0.2.82 → deepstrike-0.2.83}/Cargo.toml

@@ -3,13 +3,13 @@ resolver = "2"
 members = ["crates/deepstrike-core", "crates/deepstrike-tokenizer", "crates/deepstrike-py"]
 
 [workspace.package]
-version = "0.2.82"
+version = "0.2.83"
 edition = "2024"
 repository = "https://github.com/kongusen/deepstrike"
 
 [workspace.dependencies]
-deepstrike-core = { path = "crates/deepstrike-core", version = "0.2.82" }
-deepstrike-tokenizer = { path = "crates/deepstrike-tokenizer", version = "0.2.82" }
+deepstrike-core = { path = "crates/deepstrike-core", version = "0.2.83" }
+deepstrike-tokenizer = { path = "crates/deepstrike-tokenizer", version = "0.2.83" }
 notify = "6"
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"

{deepstrike-0.2.82 → deepstrike-0.2.83}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: deepstrike
-Version: 0.2.82
+Version: 0.2.83
 Requires-Dist: httpx>=0.27
 Requires-Dist: pyyaml>=6.0
 Requires-Dist: anyio>=4.0
@@ -250,7 +250,7 @@ The kernel renders context as four LLM API slots — only **history** is compres
 |------|--------|------|
 | `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[0]` | `task_state` + signals | Goal, plan, progress, the model's notes, compression log, runtime signals |
 | `turns[1..N]` | history | Conversation transcript |
 
 ```python
@@ -268,6 +268,47 @@ Full reference: [docs/concepts/context-slots-compression.md](../docs/concepts/co
 
 ---
 
+## Notes and model-proposed memory (0.2.82)
+
+The model decides what is worth keeping; the kernel keeps it.
+
+- **Notes.** Every request carries a `note` tool. The model writes, rewrites and retires notes — a
+  running count, what is done, a rule it learned. Notes render in the state turn (`notes:`), and no
+  compaction tier or renewal touches them. Over the budget (default 600 tokens) the kernel refuses
+  the write instead of deleting for the model. Each note is a context-graph item with
+  `derived_from` edges to the tool results it came from.
+- **Memory the model proposes.** With a candidate sink, the model calls `propose_memory` for what
+  should outlive the task; the proposal goes through the attested candidate ledger, and later
+  sessions recall it as an index line.
+
+```python
+RuntimeOptions(..., context_policy={"notes_budget_tokens": 600})  # default; 0 turns notes off
+```
+
+```python
+RuntimeOptions(
+    ...,
+    memory_candidate_sink=sink,
+    memory_runtime_policy={"retrieval": "host", "model_writes": "candidate_sink"},
+)
+```
+
+Measured on DeepSeek `deepseek-flash`, paired against the same tasks without the mechanism
+(cost is price-weighted, per correct answer):
+
+| Scenario | Without | With | Cost per correct answer |
+|---|---|---|---|
+| long-horizon (30 pairs) | 15/30 | **20/30** with notes | **−43%** |
+| ledger (30 pairs) | 25/30 | **30/30** with notes | −3% |
+| multi-session (48 pairs) | 9/48, no memory | **33/48** with notes + `propose_memory` | **−68%** |
+
+The kernel never initiates a model call to manage context: kernel-side summaries, cache-account
+compaction, note diagnostics and a post-session consolidation run were measured and deleted. Why,
+and how a mechanism earns its place: [DEC-13: Model-Led Context](https://github.com/kongusen/deepstrike/blob/main/docs/en/decisions/013-model-led-context.md). Details:
+[Context Engineering — notes](https://github.com/kongusen/deepstrike/blob/main/docs/en/guides/context-engineering.md).
+
+---
+
 ## Runtime options
 
 ```python

{deepstrike-0.2.82 → deepstrike-0.2.83}/README.md

@@ -235,7 +235,7 @@ The kernel renders context as four LLM API slots — only **history** is compres
 |------|--------|------|
 | `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[0]` | `task_state` + signals | Goal, plan, progress, the model's notes, compression log, runtime signals |
 | `turns[1..N]` | history | Conversation transcript |
 
 ```python
@@ -253,6 +253,47 @@ Full reference: [docs/concepts/context-slots-compression.md](../docs/concepts/co
 
 ---
 
+## Notes and model-proposed memory (0.2.82)
+
+The model decides what is worth keeping; the kernel keeps it.
+
+- **Notes.** Every request carries a `note` tool. The model writes, rewrites and retires notes — a
+  running count, what is done, a rule it learned. Notes render in the state turn (`notes:`), and no
+  compaction tier or renewal touches them. Over the budget (default 600 tokens) the kernel refuses
+  the write instead of deleting for the model. Each note is a context-graph item with
+  `derived_from` edges to the tool results it came from.
+- **Memory the model proposes.** With a candidate sink, the model calls `propose_memory` for what
+  should outlive the task; the proposal goes through the attested candidate ledger, and later
+  sessions recall it as an index line.
+
+```python
+RuntimeOptions(..., context_policy={"notes_budget_tokens": 600})  # default; 0 turns notes off
+```
+
+```python
+RuntimeOptions(
+    ...,
+    memory_candidate_sink=sink,
+    memory_runtime_policy={"retrieval": "host", "model_writes": "candidate_sink"},
+)
+```
+
+Measured on DeepSeek `deepseek-flash`, paired against the same tasks without the mechanism
+(cost is price-weighted, per correct answer):
+
+| Scenario | Without | With | Cost per correct answer |
+|---|---|---|---|
+| long-horizon (30 pairs) | 15/30 | **20/30** with notes | **−43%** |
+| ledger (30 pairs) | 25/30 | **30/30** with notes | −3% |
+| multi-session (48 pairs) | 9/48, no memory | **33/48** with notes + `propose_memory` | **−68%** |
+
+The kernel never initiates a model call to manage context: kernel-side summaries, cache-account
+compaction, note diagnostics and a post-session consolidation run were measured and deleted. Why,
+and how a mechanism earns its place: [DEC-13: Model-Led Context](https://github.com/kongusen/deepstrike/blob/main/docs/en/decisions/013-model-led-context.md). Details:
+[Context Engineering — notes](https://github.com/kongusen/deepstrike/blob/main/docs/en/guides/context-engineering.md).
+
+---
+
 ## Runtime options
 
 ```python

{deepstrike-0.2.82 → deepstrike-0.2.83}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "maturin"
 
 [project]
 name = "deepstrike"
-version = "0.2.82"
+version = "0.2.83"
 requires-python = ">=3.10"
 readme = "README.md"
 license = {file = "LICENSE"}