OpenSage boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

This run analyses OpenSage as an enclosing runtime, not only its agent class or memory API. Its intended use is to characterize the shipped model-driven execution, coordination and retained-state mechanisms, including the drivers that create and later consume that state. The boundary is whole-system for the repository's own implementation, prompts, configurations, agent library, evaluation drivers and training integration surfaces. It does not include implementation bodies of external dependencies.

The evidence is frozen at 54d64702e18d195c8f0fed711c6db5dc1d62ab29, with an applicability cutoff of 2026-10-14. The acquired checkout's HEAD equals that pin. The canonical identity remains https://github.com/opensage-agent/OpenSage; references to the older opensage-adk name in source documentation do not establish another source. The evidence tier is code-grounded: source inspection establishes implemented paths and declared policy, not successful execution, model compliance or measured benefit.

Entry points and state loop

The installed opensage command resolves to opensage.cli.opensage_cli:main (SRC-1, pyproject.toml). The CLI loads an agent factory, registers its tree, spawns a root instance, starts the manager and sends user input. It restores saved session state on resume and saves a session snapshot on shutdown when cleanup is disabled (SRC-1, src/opensage/cli/opensage_cli.py). These shipped caller responsibilities belong to OpenSage rather than an excluded application.

OpenSage depends on Google ADK for the underlying model-agent execution machinery, but supplies session services, resource managers, orchestration and patches around it. The session constructs an agent manager; the manager constructs per-instance ADK runners with OpenSage services and plugins (SRC-1, src/opensage/session/opensage_session.py, src/opensage/orchestration/manager.py). This makes the target an enclosing runtime assembled around a dependency, not merely a returning computation or runtime client.

return Runner( app_name=self._app_name, agent=agent, session_service=self._opensage_session.session_service, artifact_service=self._opensage_session.artifact_service, memory_service=self._opensage_session.memory_service, credential_service=self._opensage_session.credential_service, plugins=plugins, ) --- src/opensage/orchestration/manager.py

The memory boundary includes creation of sandbox memory directories, runtime path injection, trajectory persistence, agent-written knowledge files and their index, compaction and pinned content, host-side instance state and resume. The import path applies patches; the run wrapper injects runtime memory information and configurable policy, initializes layouts and persists trajectories (SRC-1, src/opensage/__init__.py, src/opensage/patches/agent_run_async.py). Knowledge-store initialization and runtime path disclosure are separate from the prompt's agent-maintenance policy (SRC-1, src/opensage/memory/file_based/long_term/knowledge_store.py, src/opensage/memory/file_based/short_term/prompt.py, src/opensage/templates/auto_insert_prompts/default.md). Policy is evidence of intended maintenance, not proof that agents performed it.

1. Skim index.md. Did you learn anything that future-you (or another agent) would want and can't find here?
2. If yes, write a new .md and append a line to index.md. Be concise — one short markdown file beats a long one.
3. If you found an existing entry to be wrong or stale, update or remove it (and its index line) instead of leaving the contradiction. --- src/opensage/templates/auto_insert_prompts/default.md

State consumers include the CLI's resume/review paths, web routes that read sessions and disk instance trees, and evaluation drivers that retain event traces. The ordinary web /run route submits input through the manager. Evaluation registers and spawns agents through that manager and subsequently collects their session events (SRC-1, src/opensage/cli/opensage_web_app.py, src/opensage/evaluation/base.py). These are included producers and consumers, not external hosts.

await opensage_session.agent_manager.spawn( agent_name=agent.name, session_id=task.session_id, parent_session_id=None, ) --- src/opensage/evaluation/base.py

Not every shipped driver uses the same enclosing loop. The local-support CLI creates an ADK runner directly, saves a trace and may issue repeated continuation calls; the vulnerability-memory benchmark also contains direct ADK runner paths. They remain included, rather than being attributed automatically to the manager loop (SRC-1, agent_library/local_support/local_cli.py, benchmarks/secodeplt/vul_detection_memory.py). The RL client resolves shipped agents and benchmark interfaces; training launchers may delegate to an external trainer (SRC-1, src/opensage/evaluation/rl_adapters/client.py, rl/areal/train.sh). Later mechanism findings must distinguish these paths.

Tracked top-level coverage

“Included” defines functional coverage, not a claim that every blob has already been read. The tree was inventoried at the frozen commit. Initial inspection traced the principal executable and retained-state loop above, plus alternate callers. Detail not yet inspected is explicitly identified below and remains available under SRC-1; it is not an access gap.

Dependencies and evidence limits

External dependencies include ADK, LiteLLM/model providers, sandbox/container services, Neo4j for supported graph tooling, MCP services and evaluation/training ecosystems. The package declares ADK and LiteLLM dependencies along with optional Harbor, AgentBeats, Nitrobox and OpenSandbox integrations (SRC-1, pyproject.toml). This source set supports OpenSage-side wiring and contracts, not those services' implementation behavior. External model outputs, datasets, deployed sandboxes, live sites and trainer repositories are not captures in this boundary.

There is no supplied execution trace or causal experiment. The target, its tests and examples were not executed. Thus this boundary supports analysis of implemented and intended operation, not claims about model adherence to knowledge maintenance, successful cross-restart recovery in a deployed environment, benchmark performance or improvement caused by memory. Available but unread files in the coverage table are inspection work for subsequent source-grounded jobs, not unavailable evidence. No material OpenSage-owned execution or retained-state responsibility has been deliberately excluded.

Source register