Telemetry

Kognisant records execution statistics locally for every interaction. This data helps you understand model performance, track token usage, monitor system health, and make informed decisions about model selection.

The /telemetry Command

Summary view

code

/telemetry

Shows an aggregate summary of recent execution statistics:

code

Telemetry Summary (last 50 executions):

  Total executions:    50
  Success rate:        94%
  Avg response time:   7.8s
  Total tokens in:     142,000
  Total tokens out:    28,400

  Models used:
    gemma4:latest      38 calls | 95% success | avg 6.2s
    deepseek-chat      12 calls | 92% success | avg 12.1s

  Classifications:
    SIMPLE:   12 (24%)
    CONTEXT:  18 (36%)
    COMPLEX:  17 (34%)
    AUTONOMOUS: 3 (6%)

  Valence trend: Good (+22)

Per-model deep dive

code

/telemetry gemma4:latest

Shows detailed statistics for a specific model:

code

Model: gemma4:latest

  Executions:       38
  Successes:        36
  Failures:         2
  Reliability:      0.93 (Bayesian)
  Avg response:     6.2s
  Token calibration: 1.08

  Failure breakdown:
    Timeout:   1
    Empty:     1
    Error:     0

  Response time distribution:
    <5s:   14 (37%)
    5-15s: 20 (53%)
    15-30s: 3 (8%)
    >30s:   1 (3%)

  Circuit breaker: CLOSED (healthy)
  Last failure: 2025-06-14T09:22:00Z

---

What Is Recorded Per Execution

Every time you send a message and receive a response, a telemetry record is appended:

Field	Description
:---	:---
`timestamp`	When the execution started (ISO-8601 UTC)
`model`	Which model handled the request
`classification`	SIMPLE, CONTEXT, COMPLEX, or AUTONOMOUS
`success`	Whether the execution completed without error
`response_time_ms`	Total wall-clock time in milliseconds
`tokens_in`	Estimated input tokens
`tokens_out`	Output tokens received
`tools_called`	Number of tool calls made
`tool_names`	List of tool names invoked
`valence_before`	Valence score before this execution
`valence_after`	Valence score after this execution
`error_type`	If failed: timeout, empty, api_error, etc.
`thinking_duration_ms`	Time spent in reasoning (if applicable)
`phase_times`	Breakdown of time per phase (bootstrap, plan, execute, reflect)

---

File Format

Location

code

~/.kognisant_core/telemetry.jsonl

The file uses JSON Lines format (one JSON object per line). This makes it easy to append without reading/rewriting the entire file.

Example record

json

{"timestamp":"2025-06-15T14:30:22Z","model":"gemma4:latest","classification":"COMPLEX","success":true,"response_time_ms":18200,"tokens_in":2100,"tokens_out":420,"tools_called":3,"tool_names":["read_project_file","edit_project_file","create_project_file"],"valence_before":22,"valence_after":27,"thinking_duration_ms":12400}

Rotation at 5MB

When telemetry.jsonl reaches 5MB, Kognisant rotates it:

The current file is renamed to telemetry.jsonl.1

A fresh telemetry.jsonl is created

Older rotated files (.2, .3, etc.) are not kept; only one backup is retained

This prevents unbounded growth while preserving recent history.

---

Valence Tracking

Valence is a system-wide mood score ranging from -100 to +100. It reflects how well things have been going:

How valence changes

Event	Delta
:---	:---
Success, fast (<10s)	+5
Success, moderate (10-30s)	+3
Success, slow (>30s)	+1
Timeout	-15
Empty response	-10
Generic error	-10
User cancelled	-5
Background pressure (failed jobs, stale world model)	up to -5

Temporal decay

Valence decays 10% toward zero per calendar day of inactivity. If you do not use Kognisant for 3 days, a valence of +50 would decay to approximately +36.

Interpreting valence

Range	Status	Meaning
:---	:---	:---
+50 to +100	Excellent	Consistent success, fast responses
+20 to +49	Good	Mostly working well
0 to +19	Neutral	Mixed results
-20 to -1	Declining	Some failures accumulating
-50 to -21	Poor	Frequent issues, consider switching models
-100 to -51	Critical	Persistent failures, intervention needed

---

Model Reliability Scores

Each model has a Bayesian reliability score:

code

reliability = (successes + 1) / (successes + failures + 2)

This formula (Laplace smoothing) starts at 0.5 and moves toward the true success rate as more data accumulates. It means:

A new model starts at 0.5 (no opinion)

After 10 successes and 0 failures: 0.92

After 10 successes and 2 failures: 0.79

After 5 successes and 5 failures: 0.5

Reliability is used by:

Circuit breaker logic (auto-switch on low reliability)

Agent model selection (prefer high-reliability models for workers)

WARM reflection advisories (warning when reliability drops below 0.3)

---

Token Usage Trends

Track your token consumption over time by examining the telemetry file:

bash

# Count tokens used today
grep "$(date -u +%Y-%m-%d)" ~/.kognisant_core/telemetry.jsonl | \
  python3 -c "
import sys, json
records = [json.loads(l) for l in sys.stdin]
total_in = sum(r.get('tokens_in', 0) for r in records)
total_out = sum(r.get('tokens_out', 0) for r in records)
print(f'Today: {total_in:,} in / {total_out:,} out ({len(records)} executions)')
"

Or just use the /telemetry command for a formatted summary.

Token calibration

Each model has a per-model token calibration factor that improves estimation accuracy over time:

code

calibration = calibration  0.8 + (actual/estimated)  0.2

The 📋 line uses this calibration to show more accurate token estimates. After 10-20 executions with a model, estimates become quite precise.

---

Background Pressure

The valence system also accounts for background system health:

Signal	Pressure
:---	:---
Average tool reliability below 50%	-2
Failed daemon jobs (per job)	-1
World model stale (>24h without update)	-1
Goal acceptance rate below 30%	-1
Combined maximum	-5

This means even successful executions can have slightly reduced valence gains if background signals indicate system stress.

---

When to Check Telemetry

Situation	What to look at
:---	:---
Responses feel slow	`/telemetry` - check avg response time trend
Getting frequent errors	`/telemetry` - check success rate and failure breakdown
Choosing between models	Compare reliability scores and response times
System auto-switched models	Check circuit breaker state in per-model view
Cost tracking (cloud models)	Look at token totals combined with pricing info
Verifying a fix worked	Check if recent executions show improved success rate

---

Privacy

All telemetry data is:

Stored locally only (~/.kognisant_core/telemetry.jsonl)

Never transmitted anywhere

Does not contain message content (only metadata)

Under your full control (delete the file anytime)

The telemetry file records statistical metadata about executions, not the actual messages or responses. Your conversations remain in session history only.

Runtime Lifecycle Slash Commands