Session Watchdog

Monitors session context usage and saves checkpoints when approaching limits to prevent data loss before compaction or long tasks.

Security Scan

The skill's instructions, required resources, and behavior are consistent with a session-monitoring/checkpointing helper; nothing requested is disproportionate, but the runtime instructions are a bit vague about paths and in-scope files so review before enabling.

Purpose & Capability

Name/description (monitor session context and save checkpoints) match the instructions: check session_status contextTokens and write checkpoint files under memory/YYYY-MM-DD.md. No unrelated credentials, binaries, or installs are requested.

Instruction Scope

Instructions explicitly read/write memory/YYYY-MM-DD.md and call session_status to get contextTokens — these are appropriate for a checkpointing watchdog. However the doc is somewhat vague about the exact memory directory root and where 'Files modified and their paths' should be sourced from, which could lead an agent to attempt broader filesystem reads if not constrained. It also assumes a session_status command with a contextTokens field exists.

Install Mechanism

No install spec and no code files — lowest risk. Nothing will be downloaded or installed by this skill.

Credentials

No environment variables, credentials, or external service tokens are requested; the skill's data access is limited to its declared memory file usage.

Persistence & Privilege

Skill writes persistent checkpoint files to memory/YYYY-MM-DD.md which is coherent for a checkpointing tool. 'always' is false and autonomous invocation is allowed by default. Consider whether persistent memory writes are acceptable for your privacy/retention policies.

Assessment

This skill appears to do what it says: monitor context levels and save checkpoints to memory files. Before installing: (1) confirm what 'memory/' directory the agent will use and that the skill is restricted to that path (avoid giving it open filesystem read access); (2) ensure session_status exists and returns contextTokens as expected; (3) review what kinds of information will be saved (decisions, pending tasks, file paths) and avoid storing secrets or credentials in checkpoints; (4) consider access controls or encryption for the memory folder if checkpoint files could contain sensitive data; (5) if you want stricter behavior, ask for the skill to explicitly limit reads to its memory file and not enumerate or read arbitrary system files.

Monitors context levels, warns before compaction, and saves checkpoints to preserve important information.

Context Thresholds

Level	Tokens	Action
Safe	0-140k	Normal operation
Warning	140k-160k	Warn user, save checkpoint
Critical	160k-197k	Warn + stop unless urgent
Full	197k+	Compaction imminent

Check Context

Before each session and periodically during long conversations:

session_status

Check the contextTokens field from the response.

Checkpoint Protocol

When approaching 80% (160k tokens):

1. Save checkpoint to memory file:

   • Read current memory/YYYY-MM-DD.md
   • Add key context: decisions, pending tasks, important details
   • Write back to memory file

2. Alert user: Say: "⚠️ Approaching context limit (~160k tokens). Saving checkpoint to memory before continuing."

3. Ask user:

   • Continue and accept compaction?
   • Summarize and restart fresh?
   • Pause until ready?

What to Save

Essential information that must survive compaction:

• Decisions made in this conversation
• Pending tasks not yet completed
• Important context (project state, configurations, preferences)
• Files modified and their paths
• Unresolved issues requiring follow-up

When to Trigger

• At session start
• After every 30k tokens of conversation
• Before initiating large tasks (file edits, multiple operations)
• When user asks "how much context do we have left?"

Memory File Format

# YYYY-MM-DD

## Session Checkpoint (at X% context)

### Decisions
- Decision 1
- Decision 2

### Pending
- [ ] Task 1
- [ ] Task 2

### Important Context
- Project state: ...
- Last file modified: ...

### Unresolved
- Issue needing follow-up