Context Health

Context health is a 0–100 score that measures how focused the current agent session is. When sessions drift — touching too many modules, too many files, or running too long — agents make more mistakes and context injection becomes less useful. Mindr quantifies this drift so agents can decide when to checkpoint or start fresh.

Score components

The score starts at 100 and applies penalties for each signal:

Signal	Max penalty	Trigger
Modules touched	24	More than 2 distinct modules
Files touched	20	More than 8 distinct files
Off-task ratio	25	Files touched outside the active task
Elapsed time	15	More than 90 minutes
Topic spread	16	Caller-provided dispersion metric

Penalty formulas:

• Modules: min(24, (moduleCount − 2) × 8)
• Files: min(20, (fileCount − 8) × 2)
• Off-task: min(25, offTaskRatio × 25)
• Time: min(15, Math.floor((minutesElapsed − 90) / 30) × 5)
• Topic spread: min(16, topicSpread × 16) (topicSpread is a 0–1 float supplied by the caller)

Recommendations

Score	Recommendation	Meaning
≥ 70	ok	Session is focused; continue
40–69	consider_checkpoint	Session is drifting; checkpoint before continuing
< 40	recommend_fresh_session	Session has drifted significantly; start fresh

Checking health

Via the MCP tool:

{
  "name": "mindr:context_health",
  "arguments": {
    "filesTouched": ["src/auth/login.ts", "src/auth/session.ts", "src/api/users.ts"],
    "modulesTouched": ["auth", "api"],
    "activeTaskFiles": ["src/auth/login.ts", "src/auth/session.ts"],
    "startedAt": "2026-05-07T10:00:00Z",
    "topicSpread": 0.2
  }
}

Response:

{
  "score": 78,
  "recommendation": "ok",
  "breakdown": {
    "modules": 0,
    "files": 0,
    "offTask": 7,
    "time": 0,
    "topicSpread": 3
  }
}

Via CLI:

mindragent session health <session-id>

Via SDK:

const health = await mindr.getContextHealth('session-abc');
console.log(health.score, health.recommendation);

Writing a checkpoint

Before switching context or starting a long tangent, write a checkpoint:

mindragent session checkpoint <session-id>

await mindr.checkpointSession('session-abc');

This stores a session_checkpoint memory that can be retrieved in a later session to resume where you left off:

mindragent memory list --type session_checkpoint

Topic spread

Topic spread is a 0–1 float the caller supplies to represent how dispersed the session's work is across unrelated topics. It is agent-supplied and not computed by Mindr. Agents that track their own task list can compute this as:

topicSpread = (distinct topics - 1) / maxTopics

If you do not track topics, omit topicSpread and Mindr defaults to 0 (no penalty).