llmbuffer 0.1.0__tar.gz

llmbuffer-0.1.0/.claude/settings.local.json

@@ -0,0 +1,11 @@
+{
+  "permissions": {
+    "allow": [
+      "Bash(python3 -m pytest tests/ -q)",
+      "Bash(python3 -m venv .venv)",
+      "Bash(.venv/bin/pip -q install pytest)",
+      "Bash(.venv/bin/pip -q install -e .)",
+      "Bash(.venv/bin/python -m pytest tests/ -q)"
+    ]
+  }
+}

llmbuffer-0.1.0/.gitignore

@@ -0,0 +1,218 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[codz]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#   Usually these files are written by a python script from a template
+#   before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py.cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+# Pipfile.lock
+
+# UV
+#   Similar to Pipfile.lock, it is generally recommended to include uv.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+# uv.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+# poetry.lock
+# poetry.toml
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#   pdm recommends including project-wide configuration in pdm.toml, but excluding .pdm-python.
+#   https://pdm-project.org/en/latest/usage/project/#working-with-version-control
+# pdm.lock
+# pdm.toml
+.pdm-python
+.pdm-build/
+
+# pixi
+#   Similar to Pipfile.lock, it is generally recommended to include pixi.lock in version control.
+# pixi.lock
+#   Pixi creates a virtual environment in the .pixi directory, just like venv module creates one
+#   in the .venv directory. It is recommended not to include this directory in version control.
+.pixi
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# Redis
+*.rdb
+*.aof
+*.pid
+
+# RabbitMQ
+mnesia/
+rabbitmq/
+rabbitmq-data/
+
+# ActiveMQ
+activemq-data/
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.envrc
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#   JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#   be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#   and can be added to the global gitignore or merged into this file.  For a more nuclear
+#   option (not recommended) you can uncomment the following to ignore the entire idea folder.
+# .idea/
+
+# Abstra
+#   Abstra is an AI-powered process automation framework.
+#   Ignore directories containing user credentials, local state, and settings.
+#   Learn more at https://abstra.io/docs
+.abstra/
+
+# Visual Studio Code
+#   Visual Studio Code specific template is maintained in a separate VisualStudioCode.gitignore 
+#   that can be found at https://github.com/github/gitignore/blob/main/Global/VisualStudioCode.gitignore
+#   and can be added to the global gitignore or merged into this file. However, if you prefer, 
+#   you could uncomment the following to ignore the entire vscode folder
+# .vscode/
+# Temporary file for partial code execution
+tempCodeRunnerFile.py
+
+# Ruff stuff:
+.ruff_cache/
+
+# PyPI configuration file
+.pypirc
+
+# Marimo
+marimo/_static/
+marimo/_lsp/
+__marimo__/
+
+# Streamlit
+.streamlit/secrets.toml

llmbuffer-0.1.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Scott Purdy
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

llmbuffer-0.1.0/PKG-INFO

@@ -0,0 +1,244 @@
+Metadata-Version: 2.4
+Name: llmbuffer
+Version: 0.1.0
+Summary: Cache-optimized LLM conversation history management with static/dynamic system prompts, transition modes, and compaction hooks.
+Project-URL: Homepage, https://github.com/scottpurdy/llmbuffer
+Project-URL: Repository, https://github.com/scottpurdy/llmbuffer
+Project-URL: Issues, https://github.com/scottpurdy/llmbuffer/issues
+Author: llmbuffer contributors
+License: MIT License
+        
+        Copyright (c) 2026 Scott Purdy
+        
+        Permission is hereby granted, free of charge, to any person obtaining a copy
+        of this software and associated documentation files (the "Software"), to deal
+        in the Software without restriction, including without limitation the rights
+        to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+        copies of the Software, and to permit persons to whom the Software is
+        furnished to do so, subject to the following conditions:
+        
+        The above copyright notice and this permission notice shall be included in all
+        copies or substantial portions of the Software.
+        
+        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+        IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+        FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+        AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+        LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+        OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+        SOFTWARE.
+License-File: LICENSE
+Keywords: anthropic,conversation,llm,openai,prompt-caching
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.9
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Software Development :: Libraries
+Requires-Python: >=3.9
+Provides-Extra: anthropic
+Requires-Dist: anthropic>=0.30; extra == 'anthropic'
+Provides-Extra: openai
+Requires-Dist: openai>=1.30; extra == 'openai'
+Provides-Extra: transformers
+Requires-Dist: transformers>=4.30; extra == 'transformers'
+Description-Content-Type: text/markdown
+
+# llmbuffer
+
+**Cache-optimized LLM conversation history management.**
+
+Most LLM applications naively concatenate their system prompt, conversation history, and any dynamic context into a single message list — and rebuild it from scratch every turn. This works, but it leaves significant money and latency on the table by constantly invalidating the provider's prompt cache.
+
+`llmbuffer` assembles your messages in the order that maximises cache reuse, manages the boundary between stable and changing content, and handles compaction when history grows too long — all without you having to think about it.
+
+```
+[Static System Prompt] → [Long-Lived History] → [Dynamic Context] → [Recent Messages]
+       cached ✓                  cached ✓             not cached          not cached
+```
+
+The static system prompt and committed conversation history form a **byte-stable prefix** that is never mutated or re-ordered across turns. The frequently-changing parts — RAG results, timestamps, in-flight tool calls — live at the end where they can't invalidate the prefix.
+
+## Install
+
+```bash
+pip install llmbuffer
+```
+
+Optional extras for live benchmarking:
+
+```bash
+pip install "llmbuffer[anthropic]"    # Anthropic prompt caching
+pip install "llmbuffer[openai]"       # OpenAI prefix caching
+```
+
+`llmbuffer` has **zero required dependencies** — just Python 3.9+.
+
+## Quickstart
+
+### Stateful (in-process)
+
+```python
+from llmbuffer import PromptManager, PromptConfig, AnthropicAdapter
+
+manager = PromptManager(PromptConfig(
+    static_system_prompt="You are a senior software engineering assistant...",
+    transition_mode="agent_cycle",   # auto-commit turns to the stable prefix
+    adapter=AnthropicAdapter(),      # inject cache_control markers
+    max_tokens=8_000,                # compact long-lived history beyond this
+))
+
+# Each turn:
+manager.append({"role": "user", "content": user_message})
+messages = manager.build_messages(dynamic_system_prompt=rag_context)
+reply = anthropic_client.messages.create(messages=messages, ...)
+manager.append({"role": "assistant", "content": reply})
+```
+
+### Stateless (web app / serverless)
+
+Pure functions over a JSON-serializable state dict — persist it anywhere between requests:
+
+```python
+from llmbuffer import functional, new_state, dumps, loads, PromptConfig
+
+config = PromptConfig(
+    static_system_prompt="You are a senior software engineering assistant...",
+    transition_mode="manual",
+)
+
+# Load state from DB / session
+state = loads(row.conversation_json) if row else new_state()
+
+# Build messages, call LLM, store updated state
+state = functional.append_message(state, {"role": "user", "content": text}, config)
+messages = functional.build_messages(state, config, dynamic_system_prompt=rag_context)
+# ... call your LLM ...
+state = functional.append_message(state, reply, config)
+row.conversation_json = dumps(state)
+```
+
+## How it works
+
+### Message ordering
+
+`build_messages()` always emits messages in this exact order:
+
+| Position | Content | Cache behaviour |
+|----------|---------|----------------|
+| 1 | **Static system prompt** | Cached — never changes |
+| 2 | **Long-lived history** | Cached — stable, grows slowly |
+| 3 | **Dynamic context** | Not cached — RAG results, timestamps, etc. |
+| 4 | **Short-term history** | Not cached — current turn, tool calls |
+
+### Transition modes
+
+Control when messages graduate from short-term into the stable long-lived history:
+
+| Mode | Behaviour |
+|------|-----------|
+| `none` | Every message goes straight into long-lived history |
+| `manual` | Messages stay short-term until you call `transition()` |
+| `agent_cycle` | Commits automatically when a non-tool-call assistant message ends the turn |
+
+### Compaction
+
+When the long-lived history exceeds `max_tokens`, a compaction hook reduces it to `max_tokens // 2` (configurable). The default hook truncates oldest-first; supply your own to summarise instead:
+
+```python
+def summarise(messages, target_tokens, adapter):
+    summary = call_llm_to_summarise(messages)
+    return [{"role": "system", "content": summary}]
+
+config = PromptConfig(max_tokens=8_000, compaction_hook=summarise)
+```
+
+### Provider adapters
+
+| Adapter | Cache markers | Token counting |
+|---------|--------------|----------------|
+| `OpenAIAdapter` (default) | None needed — automatic prefix caching | ~4 chars/token |
+| `AnthropicAdapter` | `cache_control: {type: ephemeral}` injected at prefix boundaries | ~4 chars/token |
+| `TransformersAdapter(tok)` | None | Exact via HF tokenizer |
+
+Subclass `ProviderAdapter` to add a new provider — override `count_tokens()` and/or `apply_cache_markers()`.
+
+## Benchmark
+
+The benchmark suite runs a multi-turn conversation through both `llmbuffer` and a **naive** approach, and reports cache hits from the provider's own usage metadata.
+
+The naive approach puts the static and dynamic system prompts together at the start of every message list and drops the oldest messages when the context limit is hit — this is the default pattern in most LLM applications today.
+
+### Results (simulated, 15 turns, Anthropic pricing)
+
+> The simulated provider models provider prefix caching exactly: a turn is a cache hit when its message list shares a prefix with a previously-seen turn. Run `--provider anthropic` or `--provider openai` for live numbers.
+
+| Turn | Dynamic changed | llmbuffer cached | naive cached |
+|------|:---------------:|:----------------:|:------------:|
+| 1    | yes             | ✗ 0              | ✗ 0          |
+| 2    | —               | ✓ 1,213          | ✓ 1,340      |
+| 3    | —               | ✓ 1,245          | ✓ 1,368      |
+| 4    | **yes**         | ✓ 1,274          | **✗ 0**      |
+| 5    | —               | ✓ 1,297          | ✓ 1,416      |
+| 6    | —               | ✓ 1,325          | ✓ 1,443      |
+| 7    | **yes**         | ✓ 1,351          | **✗ 0**      |
+| 8    | —               | ✓ 1,379          | ✓ 1,497      |
+| 9    | —               | ✓ 1,403          | ✓ 1,525      |
+| 10   | **yes**         | ✓ 1,430          | **✗ 0**      |
+| 11   | —               | ✓ 1,458          | ✓ 1,568      |
+| 12   | —               | ✓ 1,479          | ✓ 1,597      |
+| 13   | **yes**         | ✓ 1,507          | **✗ 0**      |
+| 14   | —               | ✓ 1,535          | ✓ 1,651      |
+| 15   | —               | ✓ 1,561          | ✓ 1,677      |
+
+| Metric | llmbuffer | naive |
+|--------|----------:|------:|
+| Cache hit ratio | **85.3%** | 66.1% |
+| Total cached tokens | **19,457** | 15,082 |
+| Est. cost (Anthropic, with caching) | **$0.016** | $0.028 |
+| Est. savings vs no caching | **76.7%** | 59.5% |
+
+Every time the dynamic context rotates (turns 4, 7, 10, 13) the naive approach suffers a **full cache miss** — the changed system prompt invalidates the entire prefix. `llmbuffer` keeps the static system and long-lived history stable, so only the new suffix is uncached regardless of what the dynamic context does.
+
+### Run it yourself
+
+```bash
+# No API key needed:
+uv run python -m llmbuffer.benchmark --provider simulated --compare --turns 15
+
+# Live providers (needs API key):
+uv run python -m llmbuffer.benchmark --provider anthropic --compare --turns 15
+uv run python -m llmbuffer.benchmark --provider openai --compare --turns 15
+uv run python -m llmbuffer.benchmark --provider gemini --compare --turns 15
+
+# Ollama (local, needs server log access):
+uv run python -m llmbuffer.benchmark --provider ollama \
+    --ollama-log ~/.ollama/logs/server.log --compare
+
+# JSON output:
+uv run python -m llmbuffer.benchmark --provider anthropic --compare --format json
+```
+
+## Development
+
+```bash
+# Clone and set up:
+git clone https://github.com/scottpurdy/llmbuffer
+cd llmbuffer
+uv sync
+
+# Run tests:
+uv run pytest
+
+# Run benchmark (simulated, no API key needed):
+uv run python -m llmbuffer.benchmark --provider simulated --compare
+```
+
+The test suite includes explicit **cache-stability tests** asserting that the static system prompt and long-lived history are byte-identical across turns — verifying the cache prefix is never accidentally mutated.
+
+## License
+
+MIT

llmbuffer-0.1.0/README.md

@@ -0,0 +1,195 @@
+# llmbuffer
+
+**Cache-optimized LLM conversation history management.**
+
+Most LLM applications naively concatenate their system prompt, conversation history, and any dynamic context into a single message list — and rebuild it from scratch every turn. This works, but it leaves significant money and latency on the table by constantly invalidating the provider's prompt cache.
+
+`llmbuffer` assembles your messages in the order that maximises cache reuse, manages the boundary between stable and changing content, and handles compaction when history grows too long — all without you having to think about it.
+
+```
+[Static System Prompt] → [Long-Lived History] → [Dynamic Context] → [Recent Messages]
+       cached ✓                  cached ✓             not cached          not cached
+```
+
+The static system prompt and committed conversation history form a **byte-stable prefix** that is never mutated or re-ordered across turns. The frequently-changing parts — RAG results, timestamps, in-flight tool calls — live at the end where they can't invalidate the prefix.
+
+## Install
+
+```bash
+pip install llmbuffer
+```
+
+Optional extras for live benchmarking:
+
+```bash
+pip install "llmbuffer[anthropic]"    # Anthropic prompt caching
+pip install "llmbuffer[openai]"       # OpenAI prefix caching
+```
+
+`llmbuffer` has **zero required dependencies** — just Python 3.9+.
+
+## Quickstart
+
+### Stateful (in-process)
+
+```python
+from llmbuffer import PromptManager, PromptConfig, AnthropicAdapter
+
+manager = PromptManager(PromptConfig(
+    static_system_prompt="You are a senior software engineering assistant...",
+    transition_mode="agent_cycle",   # auto-commit turns to the stable prefix
+    adapter=AnthropicAdapter(),      # inject cache_control markers
+    max_tokens=8_000,                # compact long-lived history beyond this
+))
+
+# Each turn:
+manager.append({"role": "user", "content": user_message})
+messages = manager.build_messages(dynamic_system_prompt=rag_context)
+reply = anthropic_client.messages.create(messages=messages, ...)
+manager.append({"role": "assistant", "content": reply})
+```
+
+### Stateless (web app / serverless)
+
+Pure functions over a JSON-serializable state dict — persist it anywhere between requests:
+
+```python
+from llmbuffer import functional, new_state, dumps, loads, PromptConfig
+
+config = PromptConfig(
+    static_system_prompt="You are a senior software engineering assistant...",
+    transition_mode="manual",
+)
+
+# Load state from DB / session
+state = loads(row.conversation_json) if row else new_state()
+
+# Build messages, call LLM, store updated state
+state = functional.append_message(state, {"role": "user", "content": text}, config)
+messages = functional.build_messages(state, config, dynamic_system_prompt=rag_context)
+# ... call your LLM ...
+state = functional.append_message(state, reply, config)
+row.conversation_json = dumps(state)
+```
+
+## How it works
+
+### Message ordering
+
+`build_messages()` always emits messages in this exact order:
+
+| Position | Content | Cache behaviour |
+|----------|---------|----------------|
+| 1 | **Static system prompt** | Cached — never changes |
+| 2 | **Long-lived history** | Cached — stable, grows slowly |
+| 3 | **Dynamic context** | Not cached — RAG results, timestamps, etc. |
+| 4 | **Short-term history** | Not cached — current turn, tool calls |
+
+### Transition modes
+
+Control when messages graduate from short-term into the stable long-lived history:
+
+| Mode | Behaviour |
+|------|-----------|
+| `none` | Every message goes straight into long-lived history |
+| `manual` | Messages stay short-term until you call `transition()` |
+| `agent_cycle` | Commits automatically when a non-tool-call assistant message ends the turn |
+
+### Compaction
+
+When the long-lived history exceeds `max_tokens`, a compaction hook reduces it to `max_tokens // 2` (configurable). The default hook truncates oldest-first; supply your own to summarise instead:
+
+```python
+def summarise(messages, target_tokens, adapter):
+    summary = call_llm_to_summarise(messages)
+    return [{"role": "system", "content": summary}]
+
+config = PromptConfig(max_tokens=8_000, compaction_hook=summarise)
+```
+
+### Provider adapters
+
+| Adapter | Cache markers | Token counting |
+|---------|--------------|----------------|
+| `OpenAIAdapter` (default) | None needed — automatic prefix caching | ~4 chars/token |
+| `AnthropicAdapter` | `cache_control: {type: ephemeral}` injected at prefix boundaries | ~4 chars/token |
+| `TransformersAdapter(tok)` | None | Exact via HF tokenizer |
+
+Subclass `ProviderAdapter` to add a new provider — override `count_tokens()` and/or `apply_cache_markers()`.
+
+## Benchmark
+
+The benchmark suite runs a multi-turn conversation through both `llmbuffer` and a **naive** approach, and reports cache hits from the provider's own usage metadata.
+
+The naive approach puts the static and dynamic system prompts together at the start of every message list and drops the oldest messages when the context limit is hit — this is the default pattern in most LLM applications today.
+
+### Results (simulated, 15 turns, Anthropic pricing)
+
+> The simulated provider models provider prefix caching exactly: a turn is a cache hit when its message list shares a prefix with a previously-seen turn. Run `--provider anthropic` or `--provider openai` for live numbers.
+
+| Turn | Dynamic changed | llmbuffer cached | naive cached |
+|------|:---------------:|:----------------:|:------------:|
+| 1    | yes             | ✗ 0              | ✗ 0          |
+| 2    | —               | ✓ 1,213          | ✓ 1,340      |
+| 3    | —               | ✓ 1,245          | ✓ 1,368      |
+| 4    | **yes**         | ✓ 1,274          | **✗ 0**      |
+| 5    | —               | ✓ 1,297          | ✓ 1,416      |
+| 6    | —               | ✓ 1,325          | ✓ 1,443      |
+| 7    | **yes**         | ✓ 1,351          | **✗ 0**      |
+| 8    | —               | ✓ 1,379          | ✓ 1,497      |
+| 9    | —               | ✓ 1,403          | ✓ 1,525      |
+| 10   | **yes**         | ✓ 1,430          | **✗ 0**      |
+| 11   | —               | ✓ 1,458          | ✓ 1,568      |
+| 12   | —               | ✓ 1,479          | ✓ 1,597      |
+| 13   | **yes**         | ✓ 1,507          | **✗ 0**      |
+| 14   | —               | ✓ 1,535          | ✓ 1,651      |
+| 15   | —               | ✓ 1,561          | ✓ 1,677      |
+
+| Metric | llmbuffer | naive |
+|--------|----------:|------:|
+| Cache hit ratio | **85.3%** | 66.1% |
+| Total cached tokens | **19,457** | 15,082 |
+| Est. cost (Anthropic, with caching) | **$0.016** | $0.028 |
+| Est. savings vs no caching | **76.7%** | 59.5% |
+
+Every time the dynamic context rotates (turns 4, 7, 10, 13) the naive approach suffers a **full cache miss** — the changed system prompt invalidates the entire prefix. `llmbuffer` keeps the static system and long-lived history stable, so only the new suffix is uncached regardless of what the dynamic context does.
+
+### Run it yourself
+
+```bash
+# No API key needed:
+uv run python -m llmbuffer.benchmark --provider simulated --compare --turns 15
+
+# Live providers (needs API key):
+uv run python -m llmbuffer.benchmark --provider anthropic --compare --turns 15
+uv run python -m llmbuffer.benchmark --provider openai --compare --turns 15
+uv run python -m llmbuffer.benchmark --provider gemini --compare --turns 15
+
+# Ollama (local, needs server log access):
+uv run python -m llmbuffer.benchmark --provider ollama \
+    --ollama-log ~/.ollama/logs/server.log --compare
+
+# JSON output:
+uv run python -m llmbuffer.benchmark --provider anthropic --compare --format json
+```
+
+## Development
+
+```bash
+# Clone and set up:
+git clone https://github.com/scottpurdy/llmbuffer
+cd llmbuffer
+uv sync
+
+# Run tests:
+uv run pytest
+
+# Run benchmark (simulated, no API key needed):
+uv run python -m llmbuffer.benchmark --provider simulated --compare
+```
+
+The test suite includes explicit **cache-stability tests** asserting that the static system prompt and long-lived history are byte-identical across turns — verifying the cache prefix is never accidentally mutated.
+
+## License
+
+MIT