OS-Agent Mesh is a reproducible prototype for the competition topic
“面向多智能体协作的低开销通信、状态传递与共享记忆机制”. It compares a
verbose text collaboration baseline with a structured protocol that passes
compact actions, results, fixed-size non-text state, and reusable shared memory.
What is implemented
Four roles: planner, retriever, tool executor, and summarizer.
Versioned JSON messages with handshake and capability discovery.
Structured ACTION, PLAN, MEMORY_HIT, TOOL_RESULT, and RESULT messages.
Fixed-size float32 semantic sketches stored out-of-band and referenced by ID.
Optional file-backed mmap state store: an executor process writes a binary vector once and downstream agents reopen it by state_ref.
Optional named Unix Domain Socket transport for actual Linux/openEuler protocol frames, with per-frame byte and latency metrics.
SQLite shared memory with metadata, evidence, and cosine retrieval.
Retry and failure-injection boundary for agent execution.
Deterministic offline backend plus optional Ollama/OpenAI-compatible adapters.
Reproducible OS startup diagnostic fixture and read-only systemd-analyze adapter.
Benchmark modes: text_only, structured_only, and structured_state_memory.
The same commands work on Linux after replacing the PowerShell environment
assignment with PYTHONPATH="$PWD/src".
Important evidence boundary
The default demo uses a deterministic fixture so that the protocol and
benchmark can run on Windows. Use --real-systemd on an actual systemd host to
collect read-only startup evidence. Acceptance evidence from the official
openEuler 24.03-LTS-SP3 WSL image is recorded in
evidence/openeuler/20260713T074259Z/; it includes the OS fingerprint, 11
passing tests, a real UDS+mmap demo, a 10-round benchmark, and SHA-256 hashes.
The recorded kernel is WSL2; it is not presented as bare-metal validation.
Layout
src/os_agent_mesh/: runtime implementation.
tests/: standard-library unit and integration tests.
docs/: design, deployment, experiment, and demo material.
OS-Agent Mesh
OS-Agent Mesh is a reproducible prototype for the competition topic “面向多智能体协作的低开销通信、状态传递与共享记忆机制”. It compares a verbose text collaboration baseline with a structured protocol that passes compact actions, results, fixed-size non-text state, and reusable shared memory.
What is implemented
ACTION,PLAN,MEMORY_HIT,TOOL_RESULT, andRESULTmessages.mmapstate store: an executor process writes a binary vector once and downstream agents reopen it bystate_ref.systemd-analyzeadapter.text_only,structured_only, andstructured_state_memory.Quick start
On Linux/openEuler, run the system-level path rather than the Windows fixture:
The same commands work on Linux after replacing the PowerShell environment assignment with
PYTHONPATH="$PWD/src".Important evidence boundary
The default demo uses a deterministic fixture so that the protocol and benchmark can run on Windows. Use
--real-systemdon an actual systemd host to collect read-only startup evidence. Acceptance evidence from the official openEuler 24.03-LTS-SP3 WSL image is recorded inevidence/openeuler/20260713T074259Z/; it includes the OS fingerprint, 11 passing tests, a real UDS+mmap demo, a 10-round benchmark, and SHA-256 hashes. The recorded kernel is WSL2; it is not presented as bare-metal validation.Layout
src/os_agent_mesh/: runtime implementation.tests/: standard-library unit and integration tests.docs/: design, deployment, experiment, and demo material.evidence/openeuler/: timestamped, reviewable openEuler acceptance evidence.results/: generated benchmark artifacts; no hand-written performance values.License
Apache License 2.0. Replace
[TEAM_NAME]before submission and keep all submission materials anonymous as required by the competition.