Production Telemetry
Your AI learns from what actually breaks in production — automatically.
When your AI writes code and that code breaks in production, ekkOS detects the failure, links it to the session that wrote the code, and asks you to confirm before creating a permanent anti-pattern. Next time your AI writes similar code, it avoids the mistake. No setup required — it works through the conversation itself.
The Problem
Every AI coding tool today operates in a bubble. Your AI writes code, you deploy it, something breaks at 2am. You get a Sentry alert, dig through logs, find the root cause, go back to your IDE, explain the whole context to your AI, and fix it. The AI learns nothing from the production failure. Next time it writes similar code, it makes the same mistake.
AI writes code → deploy → breaks at 2am → you investigate → explain to AI → fix → AI forgets → repeat
How It Works
Three things happen silently in the background. None of them require setup.
1. Commit Tracking
Every time your AI commits code during a session, ekkOS records which session produced which commit. This is the attribution chain.
2. Failure Detection
When you paste a stack trace, mention a CI failure, or share an error in conversation, ekkOS detects it and links it to the commit that caused it.
3. User Verification
ekkOS asks you to confirm: "Was this failure caused by your code change, or something else?" Only confirmed failures become anti-patterns.
AI commits code → ekkOS tracks commit → CI fails → ekkOS detects → asks you → you confirm → anti-pattern forged → AI never makes that mistake again
Two Ways It Works
Production telemetry works through both the Pulse proxy and MCP tools directly.
Via Pulse (Automatic)
If you route through the ekkOS proxy, everything is automatic. Commits are tracked passively, failures are detected from conversation text, and GitHub App events flow in without any setup.
Via MCP Tools (Universal)
If you use ekkOS via MCP tools only (Cursor, Windsurf, ChatGPT, etc.), three tools give you the same telemetry loop. Your AI calls them when relevant.
MCP Tools
Three tools for MCP-only users. Add these rules to your CLAUDE.md or system prompt.
ekkOS_TrackCommit
Call after any git commit. Links the commit SHA to the current session.
ekkOS_TrackCommit({
commit_sha: "abc1234",
branch: "main",
message: "fix: auth middleware null check"
})ekkOS_ReportFailure
Call when the user reports a production or CI failure. Paste the full error.
ekkOS_ReportFailure({
error_content: "TypeError: Cannot read property...",
commit_sha: "abc1234",
source: "github_actions"
})ekkOS_CheckCandidates
Call at the start of a session. Returns pending failures that need user verification.
ekkOS_CheckCandidates() // Returns: pending failures to ask the user about
CLAUDE.md Instructions
Add these rules so your AI calls the tools automatically:
## Production Telemetry - After any git commit, call ekkOS_TrackCommit with the SHA - When the user reports a production/CI failure, call ekkOS_ReportFailure - At the start of each session, call ekkOS_CheckCandidates
Safety Guarantees
Production telemetry never auto-forges patterns. Every safeguard is designed to prevent false lessons from entering your memory.
User Verification
Nothing is forged without your confirmation
AI Causality Check
Gemini analyzes if the commit plausibly caused the error
Flaky Test Detection
Same error across 3+ users = infrastructure, not code
Noise Filtering
Rate limiting, dedup, and signature normalization
Staleness Decay
Patterns lose priority if not reinforced in 90 days
Collective Gate
3 independent users must confirm before a pattern enters collective intelligence
GitHub Integration
If you've connected GitHub through the ekkOS dashboard, CI failure detection works automatically via the GitHub App. No additional setup needed.
When a GitHub Actions workflow fails, ekkOS receives the event, finds the commit SHA, looks up which session wrote that code, runs a causality check, and stores it as a candidate for your next session.