oacs 0.3.1a1

Open Agent Context Standard draft reference implementation for governed agent memory and context

OACS v0.1 draft - Open Agent Context Standard

EN

OACS is an open lower-layer draft contract for agent memory and context: MemoryRecord, ContextCapsule, CapabilityGrant, EvidenceRef, auditable memory_calls, and adapter boundaries. The bundled acs CLI is the reference local interface to that contract.

OACS is not a replacement for MCP. MCP describes tool/server interoperability. OACS describes how an agent assembles and governs context before a model or MCP tool is used.

It is also not an agent framework, model backend, vector database, or benchmark harness. Those systems can sit above or beside OACS and call its memory, context, capability, and audit operations.

Standard Draft vs Reference Implementation

• OACS v0.1 draft spec: terminology, lifecycle, capsule format, security model, and JSON contracts in docs/ and schemas/.
• Python reference implementation: local oacs package, acs CLI, FastAPI API, SQLite backend, encryption layer, registries, memory loop, and validation adapters. Storage goes through a thin StorageBackend protocol; SQLite is the bundled reference backend.

The draft can change before v1.0. See docs/COMPATIBILITY.md for breaking change policy.

Core Contracts

The core draft is intentionally small:

• MemoryRecord: lifecycle, depth, scope, encrypted content, and evidence.
• ContextCapsule: portable governed context for one task.
• CapabilityGrant: actor-scoped permission record.
• EvidenceRef and structured evidence items: support for memory and context decisions.
• MemoryOperation, ContextOperation, MemoryLoopRun, and memory_call: auditable operation envelopes.

Benchmarks, LM Studio, MCP execution, repo dogfood, and task packs are reference adapters. They validate or exercise the contract but do not expand it.

Quickstart

This path reaches the first useful OACS result: commit a memory, retrieve it, and build an explainable Context Capsule.

For the public PyPI install path, see docs/QUICKSTART_PYPI.md.

python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev,crypto]"

export OACS_DB=./.oacs/oacs.db
export OACS_PASSPHRASE="<choose-a-local-dev-passphrase>"

acs init --json
acs key init --passphrase "$OACS_PASSPHRASE" --json
acs actor create --type human --name "User" --json

CANDIDATE_ID=$(acs memory propose --type procedure --depth 2 --scope project \
  --text "In project Alpha reports are generated with make report-safe." --json \
  | python3 -c 'import json,sys; print(json.load(sys.stdin)["id"])')

acs memory commit "$CANDIDATE_ID" --json
acs memory query --query "Alpha report" --scope project --json
acs context build --intent answer_project_question --scope project --budget 4000 --json

Expected result: memory query returns the committed procedure and context build returns a ctx_... capsule with that memory included.

Validation Adapters

acs benchmark generate --suite memory_critical --count 20 --json
acs benchmark run --mode baseline_no_memory --json
acs benchmark run --mode oacs_memory_call_loop --json
acs benchmark compare --json

Benchmarks are validation fixtures for the memory/context contract, not the purpose of OACS. oacs_memory_call_loop records deterministic OACS memory_calls such as memory.query and memory.extract_evidence; benchmark scoring stays in the benchmark adapter. Task pack import/download is schema and checksum validated; downloads require explicit --allow-network. oacs_memory_call_loop is the preferred execution path for benchmark and product validation. oacs_memory_loop remains a broad Context Capsule compatibility mode.

Current technical reports:

• examples/benchmarks/memory_calls_gemma_e2b_2026-05-01.md
• examples/benchmarks/full_context_gemma_e2b_2026-05-02.md
• examples/benchmarks/community_memory_gemma_e2b_2026-05-02.md

Killer Demo

The local killer demo proves the public product story without a hosted service, network access, LM Studio, or a running model. It writes one scoped memory, builds and exports a Context Capsule, validates the export envelope, records memory_calls, imports MCP metadata as an adapter boundary, verifies the audit chain, and links the checked-in full-context benchmark comparison.

python3 examples/killer_demo/run_demo.py --out .oacs/killer-demo --force

Raw artifacts are written to the output directory; start with SUMMARY.md and summary.json.

Tool onboarding is documented in docs/TOOL_BINDINGS.md.

Development Dogfood

Optional source-checkout dogfood lives in the removable repo_development_memory skill under examples/skills/. It is not part of the standard surface or the minimal installed-package path:

acs skill scan examples/skills --json
acs skill run repo_development_memory \
  --payload '{"action":"capture","task":"tighten memory_calls","summary":"Removed benchmark-specific shortcuts and kept selector metadata typed.","cwd":"."}' --json
acs skill run repo_development_memory \
  --payload '{"action":"context","task":"continue OACS development","cwd":"."}' --json

The dogfood skill is a source-checkout adapter. Auto mode commits only D1 repo episodes; D2/D3 memory remains explicit review.

LM Studio

Start LM Studio with an OpenAI-compatible server at http://localhost:1234/v1. The model name is configurable:

export OACS_LMSTUDIO_BASE_URL=http://localhost:1234/v1
export OACS_LMSTUDIO_MODEL=gemma-4-e2b
acs benchmark run --mode oacs_memory_call_loop --provider lmstudio --model "$OACS_LMSTUDIO_MODEL" --json

Unit tests do not require LM Studio; integration tests skip when the server is unavailable.

Build Pipeline

GitHub Actions runs lint, typecheck, tests, package build, wheel install, and CLI smoke checks. Public prerelease publishing uses trusted publishing and the checklist in docs/RELEASE.md; see docs/BUILD.md for local build parity.

Security Model

Memory and sensitive capsule payloads are encrypted before they are written to SQLite. The default provider is passphrase-based envelope encryption. PQC is a key-wrapping integration point only; no fake post-quantum claims are made when optional PQ libraries are absent.

Limitations

This is a local POC, not a hosted multi-tenant system. MCP execution is modeled through typed bindings and imported tool metadata, with optional stdio execution guarded by tool.call capabilities. Retrieval is pluggable, but deterministic lexical/structured retrieval is the baseline. Embeddings are optional adapters, not a core requirement. Tiny tasks can show OACS overhead; medium and long memory tasks are the current strength.

RU

OACS - open lower-layer draft contract для агентской памяти и контекста: MemoryRecord, ContextCapsule, CapabilityGrant, EvidenceRef, auditable memory_calls и adapter boundaries. Встроенный CLI acs является reference local interface к этому contract.

OACS не заменяет MCP. MCP описывает совместимость tools/server. OACS описывает, как агент собирает и контролирует контекст до вызова модели или MCP tool.

Также это не agent framework, model backend, vector database или benchmark harness. Такие системы могут находиться выше или рядом с OACS и вызывать его memory, context, capability и audit operations.

Draft стандарта и reference implementation

• OACS v0.1 draft spec: терминология, lifecycle, формат capsule, security model и JSON contracts в docs/ и schemas/.
• Python reference implementation: локальный пакет oacs, CLI acs, FastAPI API, SQLite backend, encryption layer, registries, memory loop и validation adapters. Storage идёт через тонкий StorageBackend protocol; SQLite является bundled reference backend.

До v1.0 draft может меняться. Compatibility policy описана в docs/COMPATIBILITY.md.

Core Contracts

Core draft намеренно небольшой:

• MemoryRecord: lifecycle, depth, scope, encrypted content и evidence.
• ContextCapsule: переносимый управляемый контекст для одной задачи.
• CapabilityGrant: actor-scoped permission record.
• EvidenceRef и structured evidence items: поддержка memory/context decisions.
• MemoryOperation, ContextOperation, MemoryLoopRun и memory_call: auditable operation envelopes.

Benchmarks, LM Studio, MCP execution, repo dogfood и task packs являются reference adapters. Они валидируют или упражняют contract, но не расширяют его.

Quickstart

Этот путь даёт первый полезный результат OACS: записать memory, найти её и построить explainable Context Capsule.

Публичный путь установки через PyPI описан в docs/QUICKSTART_PYPI.md.

python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev,crypto]"

export OACS_DB=./.oacs/oacs.db
export OACS_PASSPHRASE="<choose-a-local-dev-passphrase>"

acs init --json
acs key init --passphrase "$OACS_PASSPHRASE" --json
acs actor create --type human --name "User" --json

CANDIDATE_ID=$(acs memory propose --type procedure --depth 2 --scope project \
  --text "В проекте Alpha отчёты генерируются через make report-safe." --json \
  | python3 -c 'import json,sys; print(json.load(sys.stdin)["id"])')

acs memory commit "$CANDIDATE_ID" --json
acs memory query --query "Alpha отчёты" --scope project --json
acs context build --intent answer_project_question --scope project --budget 4000 --json

Ожидаемый результат: memory query возвращает committed procedure, а context build возвращает capsule ctx_... с этой memory внутри.

Validation adapters

acs benchmark generate --suite memory_critical --count 20 --json
acs benchmark run --mode baseline_no_memory --json
acs benchmark run --mode oacs_memory_call_loop --json
acs benchmark compare --json

Benchmarks - это validation fixtures для memory/context contract, а не цель OACS. oacs_memory_call_loop записывает deterministic OACS memory_calls, например memory.query и memory.extract_evidence; scoring остаётся в benchmark adapter. Import/download task packs валидируется по schema и checksum; downloads требуют явный --allow-network. oacs_memory_call_loop - preferred execution path для benchmark и product validation. oacs_memory_loop остаётся broad Context Capsule compatibility mode.

Текущие technical reports:

• examples/benchmarks/memory_calls_gemma_e2b_2026-05-01.md
• examples/benchmarks/full_context_gemma_e2b_2026-05-02.md
• examples/benchmarks/community_memory_gemma_e2b_2026-05-02.md

Killer Demo

Локальное killer demo доказывает публичный product story без hosted service, network access, LM Studio или запущенной модели. Оно пишет одну scoped memory, строит и экспортирует Context Capsule, валидирует export envelope, записывает memory_calls, импортирует MCP metadata как adapter boundary, проверяет audit chain и ссылается на checked-in full-context benchmark comparison.

python3 examples/killer_demo/run_demo.py --out .oacs/killer-demo --force

Raw artifacts пишутся в output directory; начинать стоит с SUMMARY.md и summary.json.

Tool onboarding описан в docs/TOOL_BINDINGS.md.

Development dogfood

Optional source-checkout dogfood живёт в отключаемом repo_development_memory skill в examples/skills/. Это не часть standard surface и не minimal installed-package path:

acs skill scan examples/skills --json
acs skill run repo_development_memory \
  --payload '{"action":"capture","task":"tighten memory_calls","summary":"Removed benchmark-specific shortcuts and kept selector metadata typed.","cwd":"."}' --json
acs skill run repo_development_memory \
  --payload '{"action":"context","task":"continue OACS development","cwd":"."}' --json

Dogfood skill является source-checkout adapter. Auto mode коммитит только D1 repo episodes; D2/D3 memory остаётся под явным review.

LM Studio

Запустите LM Studio с OpenAI-compatible server на http://localhost:1234/v1. Модель настраивается:

export OACS_LMSTUDIO_BASE_URL=http://localhost:1234/v1
export OACS_LMSTUDIO_MODEL=gemma-4-e2b
acs benchmark run --mode oacs_memory_call_loop --provider lmstudio --model "$OACS_LMSTUDIO_MODEL" --json

Unit tests не требуют LM Studio; integration tests пропускаются, если server недоступен.

Сборочная линия

GitHub Actions запускает lint, typecheck, tests, package build, wheel install и CLI smoke checks. Публичные prerelease публикации используют trusted publishing и checklist в docs/RELEASE.md; локальный build parity описан в docs/BUILD.md.

Security model

Memory и sensitive capsule payloads шифруются до записи в SQLite. Default provider использует passphrase-based envelope encryption. PQC - только integration point для key wrapping; если optional PQ libraries отсутствуют, проект не делает fake post-quantum claims.

Ограничения

Это локальный POC, не hosted multi-tenant service. MCP execution представлен typed bindings и импортированной metadata, а optional stdio execution защищён tool.call capabilities. Retrieval расширяемый, но baseline - deterministic lexical/structured retrieval. Embeddings являются optional adapters, а не core requirement. Tiny tasks могут показывать OACS overhead; medium и long memory tasks сейчас являются сильной стороной.