This Skill is the operating instruction for the Agent; the running code is obtained through a separate Python installation package. The service organizes project progress from multi-turn conversations and, when it finds something worth advancing, makes a suggestion in the Web workbench; after the user accepts, the corresponding Connector hands it back to the original Session to continue execution.
Read instructions by task
- When asking about principles, status, or existing records, only perform the relevant read-only checks; do not install, change configuration, or start services just because the Skill is loaded.
- When the user requests installation, startup, upgrade, or uninstallation, first read General process, then read one system instruction according to the actual runtime environment.
- When configuring conversation access, only read the Connector instructions corresponding to the current Harness; do not load materials for other platforms.
| Condition | Read file |
|---|---|
| Install, run, and maintain the service | install/overview.md |
| macOS | install/macos.md |
| Windows PowerShell or WSL | install/windows.md |
| Hermes access | connectors/hermes.md |
| OpenClaw access | connectors/openclaw.md |
install/ is responsible for the operating system, dependencies, and service lifecycle;connectors/ is responsible for Harness
configuration, conversation collection, and approved continuation. It supports Hermes and OpenClaw; install each according to its corresponding instructions, and do not mix commands.
The target runtime package version is 0.1.4, downloaded and verified via GitHub Release; for release status checks, see the general process.
Hermes uses a Web-only continuation bridge, while OpenClaw uses a Gateway plugin. The current Core's semantic organization and judgment still depend on
the Hermes CLI of the configured model; even if only OpenClaw is monitored, this background worker must be prepared.
Installation package acceptance is not the same as real access acceptance.
The Windows instructions are an installation reference and do not mean full access has been accepted.
Naming and commands
- The directories and file names of this Skill use English; except for the standard entry
SKILL.md, use lowercase and hyphens. - Example commands, parameters, variable names, and comments in command blocks uniformly use English; Chinese explanations are placed outside the command blocks.
- Keep the program's original command names and do not translate parameters. Even if the user's actual path contains Chinese, it should be correctly referenced by its original name; do not change the user's files to conform to example naming.
Runtime boundaries
- Installing the Skill and installing the runtime package are two different things;
pipxdoes not install the Skill into the Harness, nor does it automatically start the service. - Distinguish between the runtime package, installation files, the target window being online, and a real return-flow check.
doctor --session-idreads the actual evidence of the specified window and does not proactively send messages;unverifiedis not equal to passing, and an old package lacking the check capability cannot be regarded as successful access. - Currently, the Web is used to display suggestions, accept or ignore, project progress, and daily reports; the service starts according to
serve --web-only, and the duplicate suggestion interface in the terminal is not enabled. - The Connector automatically reports the complete QA and the new round of input signals; Core is responsible for attribution, status updates, and reminder judgment.
- A suggestion is not authorization. After the user accepts, only continue the original Session through the Connector; the Agent reading the Skill should not separately submit the same action, causing duplicate execution.
- The execution result is returned by the Connector with
source_suggestion_id; this round updates the status and does not continuously generate the next suggestion. - Do not directly edit on behalf of the background service
project.md、item.md、events.mdorruntime.jsonl。 - Do not read, print, or copy API Keys,
.envor other credentials.
View records
Data is placed by default under the current user's Home in .sn-proactive-agent/, not in the installation package or source code directory.
It can be overridden with --data-root or SN_PROACTIVE_AGENT_DATA_ROOT . After the first complete conversation with a clear long-term goal,
the service creates the Project / Item according to ownership; users do not need to manually create projects; one-off questions can be recorded only in the run log.
Upgraded users who already have .proactive-memory/ and for whom the new directory does not exist will automatically continue to use the old directory, without migration or merging.
The old PROACTIVE_MEMORY_* environment variable can still be read; a new SN_PROACTIVE_AGENT_* setting with the same name takes precedence.
The following paths are relative to the data root directory:
projects/<project-id>/project.md: project overview and Item index.projects/<project-id>/items/<item-id>/item.md: latest status of the item.projects/<project-id>/items/<item-id>/events.md: related Q&A and status changes, used for traceability.runtime.jsonl: conversation receipt, ownership, reminder or silence reason, user decision, execution result, and daily report records, used for recovery, deduplication, and audit.
The Web reads the JSON display results of these records and does not directly modify the source files. Suggested areas should remain brief; detailed judgment bases are saved in the run records and can be checked as needed.