awecompress 0.2.0__tar.gz

awecompress-0.2.0/LICENSE

@@ -0,0 +1,207 @@
+Mozilla Public License Version 2.0
+
+1. Definitions
+--------------
+
+1.1. "Contributor"
+     means each individual or legal entity that creates, contributes to
+     the creation of, or owns Covered Software.
+
+1.2. "Contributor Version"
+     means the combination of the Contributions of others (if any) used
+     by a Contributor and that particular Contributor's Contribution.
+
+1.3. "Contribution"
+     means Covered Software of a particular Contributor.
+
+1.4. "Covered Software"
+     means Source Code Form to which the initial Contributor has attached
+     the notice in Exhibit A, the Executable Form of such Source Code
+     Form, and Modifications of such Source Code Form, in each case
+     including portions thereof.
+
+1.5. "Incompatible With Secondary Licenses"
+     means that the work at issue cannot be redistributed under the terms
+     of this License without violating, conflicting with, or otherwise
+     being incompatible with the terms of the Secondary License.
+
+1.6. "Larger Work"
+     means a work that combines Covered Software with other material, in
+     a separate file or files, that is not Covered Software.
+
+1.7. "License"
+     means this document.
+
+1.8. "Licensable"
+     means having the right to grant, to the maximum extent possible,
+     whether at the time of the initial grant or subsequently, any and
+     all of the rights conveyed by this License.
+
+1.9. "Modifications"
+     means any of the following:
+
+     (a) any addition to or deletion from the substance or structure of
+         the Covered Software; or
+
+     (b) any new Source Code Form to which the initial Contributor has
+         attached the notice in Exhibit A and which constitutes Covered
+         Software.
+
+1.10. "Patent Claims" of a Contributor
+      means any patent claim, including without limitation, method,
+      process, and apparatus claims, in any patent Licensable by such
+      Contributor that would be infringed, but for the grant of the
+      License, by the making, using, selling, offering for sale, having
+      made, import, or transfer of either its Contributions or its
+      Contributor Version.
+
+1.11. "Secondary License"
+      means either the GNU General Public License, Version 2.0, or any
+      later version of such license published by the Free Software
+      Foundation; or the GNU Lesser General Public License, Version 2.1,
+      or any later version of such license published by the Free Software
+      Foundation; or the GNU Affero General Public License, Version 3.0,
+      or any later version of such license published by the Free Software
+      Foundation.
+
+1.12. "Source Code Form"
+     means the form of software preferred for making modifications.
+
+1.13. "You" (or "Your")
+     means an individual or a legal entity exercising rights under this
+     License.
+
+2. License Grants and Conditions
+--------------------------------
+
+2.1. Grants
+
+     Each Contributor hereby grants You a world-wide, royalty-free,
+     non-exclusive license, subject to third-party intellectual property
+     claims, under intellectual property rights (other than patent or
+     trademark) Licensable by such Contributor to use, reproduce, make
+     available, modify, display, perform, distribute, and otherwise
+     exploit its Contributions, either on an unmodified basis, with
+     Modifications, or as part of a Larger Work; and under Patent Claims
+     of such Contributor to make, use, sell, offer for sale, have made,
+     import, and otherwise transfer either its Contributions or its
+     Contributor Version.
+
+2.2. Effective Date
+
+     The licenses granted in Section 2.1 with respect to any Contribution
+     become effective for each Contribution on the date the Contributor
+     first distributes such Contribution.
+
+2.3. Limitations on Grant Scope
+
+     Nothing in this License shall be construed as excluding or limiting
+     any implied license or other defense to infringement that may
+     otherwise be available to You under applicable patent law.
+
+2.4. Subsequent Licenses
+
+     No Contributor makes additional grants as a result of Your choice to
+     distribute the Covered Software under a subsequent version of this
+     License or under the terms of a Secondary License (if permitted
+     under the terms of this License).
+
+2.5. Representation
+
+     Each Contributor represents that the Contributor believes its
+     Contributions are its original creation(s) or it has sufficient
+     rights to grant the rights to its Contributions conveyed by this
+     License.
+
+3. Obligations
+--------------
+
+3.1. Distribution of Source Code Form
+
+     All distribution of Covered Software in Source Code Form, including
+     any Modifications that You create or to which You contribute, must
+     be under the terms of this License. You must inform recipients that
+     the Source Code Form of the Covered Software is subject to the terms
+     of this License, and how they can obtain a copy of this License.
+
+3.2. Distribution of Executable Form
+
+     If You distribute Covered Software in Executable Form then:
+
+     (a) such Covered Software must also be made available in Source Code
+         Form; and
+
+     (b) You may distribute such Executable Form under the terms of this
+         License; and
+
+     (c) You must include a copy of this License with every copy of the
+         Covered Software in Executable Form You distribute; and
+
+     (d) You may not prevent recipients from copying, modifying, or
+         further distributing the Covered Software in Source Code Form.
+
+3.3. Distribution of a Larger Work
+
+     You may create and distribute a Larger Work under terms of Your
+     choice, provided that You also comply with the requirements of this
+     License for the Covered Software.
+
+3.4. Notices
+
+     You may not remove or alter the substance of any license notices
+     (including copyright notices, patent notices, disclaimers of
+     warranty, or limitations of liability) contained within the Source
+     Code Form of the Covered Software.
+
+3.5. Application of Additional Terms
+
+     You may not offer or impose any additional or different terms on any
+     portion of the Covered Software that would alter or restrict the
+     applicable version of this License or the recipients' rights
+     hereunder.
+
+4. Inability to Comply Due to Statute or Regulation
+---------------------------------------------------
+
+If it is impossible for You to comply with any of the terms of this
+License with respect to some or all of the Covered Software due to
+statute, judicial order, or regulation then You must: (a) comply with
+the terms of this License to the maximum extent possible; and (b)
+describe the limitations and the code they affect.
+
+5. Termination
+--------------
+
+5.1. The rights granted under this License will terminate automatically
+     if You fail to comply with any of its terms. However, if You become
+     compliant, then the rights granted under this License are
+     reinstated.
+
+5.2. If You initiate litigation against any entity by asserting a patent
+     infringement claim (excluding declaratory judgment actions,
+     counter-claims, and cross-claims) alleging that a Contributor
+     Version directly or indirectly infringes any patent, then the rights
+     granted to You by any and all Contributors for the Covered Software
+     under Sections 2.1 and 2.2 shall terminate.
+
+6. Disclaimer of Warranty
+-------------------------
+
+Covered Software is provided under this License on an "as is" basis,
+without warranty of any kind, either express or implied.
+
+7. Limitation of Liability
+---------------------------
+
+To the extent permitted by applicable law, in no event shall any
+Contributor be liable to You for any indirect, special, incidental, or
+consequential damages.
+
+8. Miscellaneous
+----------------
+
+8.1. This License represents the complete agreement concerning subject
+     matter hereof.
+
+8.2. If any provision of this License is held to be unenforceable, it
+     shall be reformed only to the minimum extent necessary.

awecompress-0.2.0/PKG-INFO

@@ -0,0 +1,206 @@
+Metadata-Version: 2.4
+Name: awecompress
+Version: 0.2.0
+Summary: Context compression proxy: freeze old turns into one summary before they reach your provider.
+Author: Peng
+License: MPL-2.0
+Keywords: ai,llm,proxy,claude,claude-code,context,compression
+Classifier: Development Status :: 3 - Alpha
+Classifier: Environment :: Console
+Classifier: Intended Audience :: Developers
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3 :: Only
+Classifier: Topic :: Utilities
+Requires-Python: >=3.9
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: click>=8.1
+Requires-Dist: aiohttp>=3.9
+Provides-Extra: dev
+Requires-Dist: pytest>=8; extra == "dev"
+Requires-Dist: pytest-asyncio>=0.23; extra == "dev"
+Dynamic: license-file
+
+<div align="center">
+  <h1>awecompress: Context Compression Proxy</h1>
+  <p><strong>Freeze old turns into one summary before they reach your provider.</strong></p>
+  <p>Local context compression for coding agents — three wire protocols, standalone proxy or in-process inside awerouter. When a session's history crosses a token threshold, the oldest whole turns are replaced by a single frozen LLM summary — cached, so every later request reuses the same bytes and your provider's prompt cache stays warm.</p>
+  <p>
+    <strong>English</strong> ·
+    <a href="./README_cn.md">简体中文</a>
+  </p>
+  <p>
+    <img src="https://img.shields.io/badge/version-0.2.0-7C3AED?style=flat-square" alt="Version">
+    <img src="https://img.shields.io/badge/python-%E2%89%A53.9-0EA5E9?style=flat-square" alt="Python">
+    <img src="https://img.shields.io/badge/license-MPL--2.0-22C55E?style=flat-square" alt="License">
+  </p>
+  <p>
+    <img src="https://img.shields.io/badge/status-alpha-c96a3d?style=flat-square" alt="Status">
+    <img src="https://img.shields.io/badge/install-pip-22C55E?style=flat-square" alt="pip">
+    <img src="https://img.shields.io/badge/platform-terminal-334155?style=flat-square" alt="Platform">
+    <img src="https://img.shields.io/github/stars/wehuman01/awecompress?style=flat-square" alt="Stars">
+  </p>
+</div>
+
+> Compress long coding-agent context: old turns become one frozen summary, requests shrink, sessions run for days without a `/clear`. Standalone proxy, or one flag inside awerouter.
+
+## How it works
+
+Claude Code resubmits the whole conversation every turn. Hours in, most of that is dead weight — old file reads, finished exploration, failed attempts.
+
+awecompress sits between the agent and whatever speaks its protocol upstream — Anthropic Messages, OpenAI Chat Completions, or OpenAI Responses:
+
+Responses requests that use the valid string form of `input` are forwarded transparently; compression requires the list form so turn boundaries remain explicit.
+
+```
+Claude Code → awecompress (:8808) → awerouter → providers
+```
+
+For each request it estimates the context size. Above a threshold, it picks a cut point on a turn boundary (a message a human actually sent — so a tool call is never separated from its result), summarizes everything before it with one LLM call, replaces those messages with a single summary message, and freezes the result in a local SQLite store. Protected content — todo lists, plans, task/skill outcomes, files you name by pattern — renders into the summarizer uncapped and must survive the summary verbatim.
+
+Three properties matter:
+
+- **Frozen, not recomputed.** The summary is stored once. Every later request reuses the same bytes, so the provider prompt cache sees a stable prefix. Growing the covered span rewrites the summary once — a one-time cache miss.
+- **Fail-open.** Any failure in the compression path forwards the original body untouched. A compression problem never breaks the session.
+- **No auth, no routing.** Auth headers pass through; routing and failover stay in [awerouter](https://github.com/wehuman01/awerouter) (or whatever your upstream is). The summary calls themselves are ordinary requests through that upstream — awerouter's flash routing applies to them like anything else.
+
+Set `X-Awecompress: off` on a request to bypass compression entirely.
+
+## Install
+
+```bash
+pip install awecompress
+```
+
+Or from source:
+
+```bash
+git clone https://github.com/wehuman01/awecompress
+cd awecompress && pip install -e .
+```
+
+## Quick Start
+
+Stack with awerouter (the intended setup):
+
+```bash
+awerouter serve run          # your routing daemon, as usual
+awecompress serve            # the compression proxy, foreground
+
+# point Claude Code at awecompress instead of awerouter
+export ANTHROPIC_BASE_URL=http://127.0.0.1:8808
+claude
+
+# openai-chat / openai-responses clients work the same way
+export OPENAI_BASE_URL=http://127.0.0.1:8808/v1
+```
+
+**Or skip the proxy entirely** — with awerouter installed, flip the profile flag and the compression runs inside the router (see below).
+
+Standalone against any Anthropic-protocol endpoint:
+
+```bash
+awecompress serve --upstream https://api.anthropic.com
+```
+
+Watch it work — one line per compressed request, and stats on demand:
+
+```
+[awecompress] 3f9a2c1b: init — summarized messages 0..61 (est 41200 tok) into 1100
+              via claude-sonnet in 2.8s; body est 48900 -> 8800 tokens
+[awecompress] 3f9a2c1b: applied frozen summary (messages 0..61) — est 48900 -> 8800 tokens
+```
+
+```bash
+awecompress status
+```
+
+## With awerouter (in-process, no proxy)
+
+`awerouter` accepts an `awecompress` profile flag, exactly like `rtk`/`odcp`. The compression core runs inside the router's pipeline — ahead of odcp pruning and rtk compression — so clients keep pointing at the router port and the flag hot-reloads with `routing.json`:
+
+```json
+"cc-router-1": {
+  "protocol": "anthropic",
+  "destinations": { "flash": "stepfun,step-3.7-flash", "pro": "glm,glm-5.3" },
+  "odcp": true,
+  "awecompress": true
+}
+```
+
+An object tunes it — `summaryModel` picks who serves the summary calls: `"flash"` (default, the flash destination — routed directly, never re-priced to pro by the long-context rule), `"pro"`, or any model a provider declares in `providers.json` (validated at serve start). The other keys mirror the standalone config:
+
+```json
+"awecompress": {
+  "summaryModel": "flash",
+  "thresholdTokens": 60000,
+  "keepRecentTurns": 4,
+  "protectedTools": ["task", "skill", "todowrite", "todoread", "updateplan"],
+  "protectedFilePatterns": ["**/*.schema.json"]
+}
+```
+
+Requires the package on the router's side: `pip install awerouter[compress]` (the flag dies at serve start with that hint when it is missing). Savings land in the usage log next to rtk/odcp (`awecompress_saved`, shown by `awerouter usage`), `X-Awerouter-Token-Saver: off` disables all lossy layers at once, and the frozen store is shared with the standalone proxy (`awecompress status` / `clear` manage it either way).
+
+## Config
+
+`~/.config/awecompress/config.json` (or `$AWECOMPRESS_CONFIG_DIR`), written with defaults on first run:
+
+```json
+{
+  "port": 8808,
+  "upstream": "http://127.0.0.1:20128",
+  "thresholdTokens": 60000,
+  "keepRecentTurns": 4,
+  "minSpanTokens": 8000,
+  "summaryModel": "",
+  "summaryMaxTokens": 2048
+}
+```
+
+| Key | Default | Meaning |
+| --- | --- | --- |
+| `port` | `8808` | Listen port. |
+| `upstream` | `http://127.0.0.1:20128` | Where requests go — awerouter by default. |
+| `thresholdTokens` | `60000` | Estimated context above which compression triggers. |
+| `keepRecentTurns` | `4` | Human turns always kept verbatim. |
+| `minSpanTokens` | `8000` | Don't summarize spans smaller than this — not worth a call. |
+| `summaryModel` | `""` | Model for summary calls. Empty = the request's own model, routed by your upstream (usually flash). |
+| `summaryMaxTokens` | `2048` | Max output tokens for a summary. |
+| `summaryTimeoutSeconds` | `60` | Give up on a summary call after this; the request forwards uncompressed. |
+| `protectedTools` | see below | Tools whose calls/results render into the summarizer uncapped and must survive the summary verbatim. |
+| `protectedFilePatterns` | `[]` | Glob patterns; a call whose `file_path`/`path` argument matches renders uncapped too. |
+| `transcriptResultCap` | `4000` | Per-tool-result cap (chars) when flattening history for the summarizer. |
+| `dbPath` | config dir | SQLite store for frozen summaries. |
+
+## Commands
+
+```bash
+awecompress serve                  # run the proxy in the foreground
+awecompress serve --port 8809 --upstream http://127.0.0.1:20128
+awecompress status                 # running state + compression stats
+awecompress config path            # where the config lives
+awecompress config show            # print it
+awecompress clear --yes            # drop all frozen summaries
+```
+
+## Notes and limits
+
+- **Three protocols** — Anthropic Messages, OpenAI Chat Completions, OpenAI Responses. Requests to other paths are relayed untouched.
+- **Compression is lossy by design.** The summarizer prompt demands exhaustive technical detail and verbatim short user messages, but a summary is still a summary. `keepRecentTurns` keeps the working set verbatim; raise it if you want more raw history.
+- **A session rewound to a checkpoint** (changed history under a stored summary) is detected by hash and recompressed from scratch.
+- **`/v1/messages/count_tokens`** applies existing summaries but never triggers a new summary call.
+- Inspired by [DCP](https://github.com/Opencode-DCP/opencode-dynamic-context-pruning)'s Compress strategy (AGPL) and the closed-source Sleev — both harness-integrated. awecompress is an independent, proxy-native implementation; no DCP code is used.
+
+## Development
+
+```bash
+pip install -e ".[dev]"
+pytest
+```
+
+See [docs/CONTRIBUTING.md](docs/CONTRIBUTING.md) for architecture and the design contract.
+
+## License
+
+MPL-2.0. Compression behavior inspired by DCP's public Compress documentation; implementation written from scratch.

awecompress-0.2.0/README.md

@@ -0,0 +1,183 @@
+<div align="center">
+  <h1>awecompress: Context Compression Proxy</h1>
+  <p><strong>Freeze old turns into one summary before they reach your provider.</strong></p>
+  <p>Local context compression for coding agents — three wire protocols, standalone proxy or in-process inside awerouter. When a session's history crosses a token threshold, the oldest whole turns are replaced by a single frozen LLM summary — cached, so every later request reuses the same bytes and your provider's prompt cache stays warm.</p>
+  <p>
+    <strong>English</strong> ·
+    <a href="./README_cn.md">简体中文</a>
+  </p>
+  <p>
+    <img src="https://img.shields.io/badge/version-0.2.0-7C3AED?style=flat-square" alt="Version">
+    <img src="https://img.shields.io/badge/python-%E2%89%A53.9-0EA5E9?style=flat-square" alt="Python">
+    <img src="https://img.shields.io/badge/license-MPL--2.0-22C55E?style=flat-square" alt="License">
+  </p>
+  <p>
+    <img src="https://img.shields.io/badge/status-alpha-c96a3d?style=flat-square" alt="Status">
+    <img src="https://img.shields.io/badge/install-pip-22C55E?style=flat-square" alt="pip">
+    <img src="https://img.shields.io/badge/platform-terminal-334155?style=flat-square" alt="Platform">
+    <img src="https://img.shields.io/github/stars/wehuman01/awecompress?style=flat-square" alt="Stars">
+  </p>
+</div>
+
+> Compress long coding-agent context: old turns become one frozen summary, requests shrink, sessions run for days without a `/clear`. Standalone proxy, or one flag inside awerouter.
+
+## How it works
+
+Claude Code resubmits the whole conversation every turn. Hours in, most of that is dead weight — old file reads, finished exploration, failed attempts.
+
+awecompress sits between the agent and whatever speaks its protocol upstream — Anthropic Messages, OpenAI Chat Completions, or OpenAI Responses:
+
+Responses requests that use the valid string form of `input` are forwarded transparently; compression requires the list form so turn boundaries remain explicit.
+
+```
+Claude Code → awecompress (:8808) → awerouter → providers
+```
+
+For each request it estimates the context size. Above a threshold, it picks a cut point on a turn boundary (a message a human actually sent — so a tool call is never separated from its result), summarizes everything before it with one LLM call, replaces those messages with a single summary message, and freezes the result in a local SQLite store. Protected content — todo lists, plans, task/skill outcomes, files you name by pattern — renders into the summarizer uncapped and must survive the summary verbatim.
+
+Three properties matter:
+
+- **Frozen, not recomputed.** The summary is stored once. Every later request reuses the same bytes, so the provider prompt cache sees a stable prefix. Growing the covered span rewrites the summary once — a one-time cache miss.
+- **Fail-open.** Any failure in the compression path forwards the original body untouched. A compression problem never breaks the session.
+- **No auth, no routing.** Auth headers pass through; routing and failover stay in [awerouter](https://github.com/wehuman01/awerouter) (or whatever your upstream is). The summary calls themselves are ordinary requests through that upstream — awerouter's flash routing applies to them like anything else.
+
+Set `X-Awecompress: off` on a request to bypass compression entirely.
+
+## Install
+
+```bash
+pip install awecompress
+```
+
+Or from source:
+
+```bash
+git clone https://github.com/wehuman01/awecompress
+cd awecompress && pip install -e .
+```
+
+## Quick Start
+
+Stack with awerouter (the intended setup):
+
+```bash
+awerouter serve run          # your routing daemon, as usual
+awecompress serve            # the compression proxy, foreground
+
+# point Claude Code at awecompress instead of awerouter
+export ANTHROPIC_BASE_URL=http://127.0.0.1:8808
+claude
+
+# openai-chat / openai-responses clients work the same way
+export OPENAI_BASE_URL=http://127.0.0.1:8808/v1
+```
+
+**Or skip the proxy entirely** — with awerouter installed, flip the profile flag and the compression runs inside the router (see below).
+
+Standalone against any Anthropic-protocol endpoint:
+
+```bash
+awecompress serve --upstream https://api.anthropic.com
+```
+
+Watch it work — one line per compressed request, and stats on demand:
+
+```
+[awecompress] 3f9a2c1b: init — summarized messages 0..61 (est 41200 tok) into 1100
+              via claude-sonnet in 2.8s; body est 48900 -> 8800 tokens
+[awecompress] 3f9a2c1b: applied frozen summary (messages 0..61) — est 48900 -> 8800 tokens
+```
+
+```bash
+awecompress status
+```
+
+## With awerouter (in-process, no proxy)
+
+`awerouter` accepts an `awecompress` profile flag, exactly like `rtk`/`odcp`. The compression core runs inside the router's pipeline — ahead of odcp pruning and rtk compression — so clients keep pointing at the router port and the flag hot-reloads with `routing.json`:
+
+```json
+"cc-router-1": {
+  "protocol": "anthropic",
+  "destinations": { "flash": "stepfun,step-3.7-flash", "pro": "glm,glm-5.3" },
+  "odcp": true,
+  "awecompress": true
+}
+```
+
+An object tunes it — `summaryModel` picks who serves the summary calls: `"flash"` (default, the flash destination — routed directly, never re-priced to pro by the long-context rule), `"pro"`, or any model a provider declares in `providers.json` (validated at serve start). The other keys mirror the standalone config:
+
+```json
+"awecompress": {
+  "summaryModel": "flash",
+  "thresholdTokens": 60000,
+  "keepRecentTurns": 4,
+  "protectedTools": ["task", "skill", "todowrite", "todoread", "updateplan"],
+  "protectedFilePatterns": ["**/*.schema.json"]
+}
+```
+
+Requires the package on the router's side: `pip install awerouter[compress]` (the flag dies at serve start with that hint when it is missing). Savings land in the usage log next to rtk/odcp (`awecompress_saved`, shown by `awerouter usage`), `X-Awerouter-Token-Saver: off` disables all lossy layers at once, and the frozen store is shared with the standalone proxy (`awecompress status` / `clear` manage it either way).
+
+## Config
+
+`~/.config/awecompress/config.json` (or `$AWECOMPRESS_CONFIG_DIR`), written with defaults on first run:
+
+```json
+{
+  "port": 8808,
+  "upstream": "http://127.0.0.1:20128",
+  "thresholdTokens": 60000,
+  "keepRecentTurns": 4,
+  "minSpanTokens": 8000,
+  "summaryModel": "",
+  "summaryMaxTokens": 2048
+}
+```
+
+| Key | Default | Meaning |
+| --- | --- | --- |
+| `port` | `8808` | Listen port. |
+| `upstream` | `http://127.0.0.1:20128` | Where requests go — awerouter by default. |
+| `thresholdTokens` | `60000` | Estimated context above which compression triggers. |
+| `keepRecentTurns` | `4` | Human turns always kept verbatim. |
+| `minSpanTokens` | `8000` | Don't summarize spans smaller than this — not worth a call. |
+| `summaryModel` | `""` | Model for summary calls. Empty = the request's own model, routed by your upstream (usually flash). |
+| `summaryMaxTokens` | `2048` | Max output tokens for a summary. |
+| `summaryTimeoutSeconds` | `60` | Give up on a summary call after this; the request forwards uncompressed. |
+| `protectedTools` | see below | Tools whose calls/results render into the summarizer uncapped and must survive the summary verbatim. |
+| `protectedFilePatterns` | `[]` | Glob patterns; a call whose `file_path`/`path` argument matches renders uncapped too. |
+| `transcriptResultCap` | `4000` | Per-tool-result cap (chars) when flattening history for the summarizer. |
+| `dbPath` | config dir | SQLite store for frozen summaries. |
+
+## Commands
+
+```bash
+awecompress serve                  # run the proxy in the foreground
+awecompress serve --port 8809 --upstream http://127.0.0.1:20128
+awecompress status                 # running state + compression stats
+awecompress config path            # where the config lives
+awecompress config show            # print it
+awecompress clear --yes            # drop all frozen summaries
+```
+
+## Notes and limits
+
+- **Three protocols** — Anthropic Messages, OpenAI Chat Completions, OpenAI Responses. Requests to other paths are relayed untouched.
+- **Compression is lossy by design.** The summarizer prompt demands exhaustive technical detail and verbatim short user messages, but a summary is still a summary. `keepRecentTurns` keeps the working set verbatim; raise it if you want more raw history.
+- **A session rewound to a checkpoint** (changed history under a stored summary) is detected by hash and recompressed from scratch.
+- **`/v1/messages/count_tokens`** applies existing summaries but never triggers a new summary call.
+- Inspired by [DCP](https://github.com/Opencode-DCP/opencode-dynamic-context-pruning)'s Compress strategy (AGPL) and the closed-source Sleev — both harness-integrated. awecompress is an independent, proxy-native implementation; no DCP code is used.
+
+## Development
+
+```bash
+pip install -e ".[dev]"
+pytest
+```
+
+See [docs/CONTRIBUTING.md](docs/CONTRIBUTING.md) for architecture and the design contract.
+
+## License
+
+MPL-2.0. Compression behavior inspired by DCP's public Compress documentation; implementation written from scratch.

awecompress-0.2.0/pyproject.toml

@@ -0,0 +1,41 @@
+[build-system]
+requires = ["setuptools>=68"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "awecompress"
+version = "0.2.0"
+description = "Context compression proxy: freeze old turns into one summary before they reach your provider."
+readme = "README.md"
+requires-python = ">=3.9"
+license = { text = "MPL-2.0" }
+authors = [
+  { name = "Peng" }
+]
+keywords = ["ai", "llm", "proxy", "claude", "claude-code", "context", "compression"]
+classifiers = [
+  "Development Status :: 3 - Alpha",
+  "Environment :: Console",
+  "Intended Audience :: Developers",
+  "Programming Language :: Python :: 3",
+  "Programming Language :: Python :: 3 :: Only",
+  "Topic :: Utilities"
+]
+dependencies = [
+    "click>=8.1",
+    "aiohttp>=3.9",
+]
+
+[project.optional-dependencies]
+dev = ["pytest>=8", "pytest-asyncio>=0.23"]
+
+[project.scripts]
+awecompress = "awecompress.cli:main"
+
+[tool.setuptools.packages.find]
+where = ["src"]
+
+[tool.pytest.ini_options]
+pythonpath = ["src"]
+testpaths = ["tests"]
+asyncio_mode = "auto"

awecompress-0.2.0/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+

awecompress-0.2.0/src/awecompress/__init__.py

@@ -0,0 +1,14 @@
+"""awecompress: local context compression for coding agents.
+
+Sits between the harness (Claude Code, OpenCode, ...) and its upstream, or
+runs inside awerouter beside odcp/rtk. When a session's history crosses a
+token threshold, the oldest whole turns are replaced by one frozen LLM
+summary — cached, so every later request reuses the same bytes.
+"""
+
+from importlib.metadata import PackageNotFoundError, version
+
+try:
+    __version__ = version("awecompress")
+except PackageNotFoundError:  # running from a source checkout
+    __version__ = "0.2.0"