Context Sentinel

Monitors session context and automatically manages model switching based on a cascading protocol. Use as part of a heartbeat or cron job to maintain session health and optimize token usage.

Security Scan

The skill's stated purpose (automated model switching) is plausible, but the runtime instructions reference missing scripts and files and leave broad, vague system-level actions to the agent — the bundle is internally inconsistent and needs clarification before use.

Purpose & Capability

The name/description (monitor session context and switch models) matches the requested capabilities, but the SKILL.md expects a PowerShell script (scripts/check_context.ps1) and MEMORY.md to exist and drive behavior — neither are bundled. Asking the agent to run 'session_status' and write to a 'handoff' file is plausible for this purpose, but the absence of the script and referenced docs makes the package incomplete and incoherent.

Instruction Scope

Instructions tell the agent to execute a PowerShell script, evaluate its output, run a 'session_status' command, and potentially write to a handoff file. Those are system-level actions (running scripts, writing files, invoking agent commands). The SKILL.md is vague about the handoff file path, the contents of the script, and what 'session_status' does — granting broad discretion and potential to affect session state or system files without explicit boundaries.

Install Mechanism

This is an instruction-only skill with no install spec and no code files, which is low-risk from an installer/extraction perspective. However, because runtime behavior depends on an external script that isn't included, the skill as-published cannot function as-is.

Credentials

The skill requests no environment variables, credentials, or config paths. That is proportionate to its stated goal. Still, the instructions reference reading/writing session state and handoff files (unspecified), so the skill could interact with system files if the external script or agent commands do so.

Persistence & Privilege

The skill does not request always: true and is user-invocable (defaults). Autonomous invocation (disable-model-invocation: false) is the platform default; combined with the vague runtime actions this increases potential impact if the missing script were later supplied or pointed to a different path. On its own this is not a showstopper, but it amplifies the other concerns.

What to consider before installing

Do not install or run this skill as-is. The SKILL.md references a PowerShell script (skills/context-sentinel/scripts/check_context.ps1) and MEMORY.md that are not included — the package is incomplete. Before using: (1) ask the author to provide the exact script and MEMORY.md and review their contents line-by-line to ensure they only read the expected session state and write only to explicit, safe paths; (2) confirm what the 'session_status' command does and where the handoff file lives; (3) run any provided script in a sandboxed environment first (or inspect it locally) rather than allowing automatic/cron invocation; (4) avoid granting persistent/autonomous execution until the implementation is complete and you trust the source. Because the instructions are vague and reference system-level actions, treat this skill as potentially risky until its missing artifacts and exact behaviors are provided and reviewed.

This skill provides a script to automate the "Cascading Model Protocol," ensuring the agent gracefully degrades from high-cost models to high-context models as the session grows.

Protocol Summary

1. Opus 4.6: Use until 80% context, then switch to Opus 4.5.
2. Opus 4.5: Use until 80% context, then switch to Gemini 2.5 Pro.
3. Gemini Pro: Use until 80% context, then trigger a handoff.

This skill operationalizes the logic defined in MEMORY.md.

How to Use

This skill is designed to be run periodically, either via a cron job or as part of the main agent's HEARTBEAT.md checklist.

Workflow

1. Execute the script: Run the check_context.ps1 script to get the current session status and determine the required action.

   powershell -File scripts/check_context.ps1
   

2. Evaluate the output: The script will return one of three possible string commands:

   • SWITCH_TO:<model_id>
   • HANDOFF_NOW
   • STATUS_OK

3. Take Action: Based on the output, execute the appropriate agent command.

   • If SWITCH_TO:<model_id>, run session_status with the new model ID:

     session_status model=<model_id>
     

   • If HANDOFF_NOW, trigger the handoff process by writing to the handoff file. This is typically done by running a specific, pre-defined prompt or script.

   • If STATUS_OK, no action is needed.

Example Usage in HEARTBEAT.md

You can replace the manual checks in your HEARTBEAT.md with a call to this skill's script.

Old HEARTBEAT.md:

## Cascading Model Protocol (Check Every Heartbeat)
1.  **Check Status:** Get current model and context %.
2.  **Opus 4.6:** If model is `Opus 4.6` and context > 80% -> Switch to `Opus 4.5`.
...

New HEARTBEAT.md using this skill:

## Context Sentinel (Check Every Heartbeat)
1. Run `powershell -File skills/context-sentinel/scripts/check_context.ps1`.
2. Evaluate the output and take action (`SWITCH_TO`, `HANDOFF_NOW`, or `STATUS_OK`).

This makes the logic reusable and keeps the HEARTBEAT.md file clean and focused on execution.