Measure AI leverage without the spyware
How Stack Signal reads your local AI coding logs into one private dashboard: 45 anti-pattern rules across five harnesses, zero telemetry in the core.
Every team that has adopted AI coding assistants ends up asking the same question: is this actually working? The honest answer is usually a shrug, because the evidence is scattered. Claude Code keeps its logs in one place, Copilot in another, Codex in a third. The recurring prompt problems that slow a developer down never surface, and the one tool that would tell you often wants to ship your internal coding activity to a server you do not control. So the question goes unanswered.
We built Stack Signal to answer it without that trade. It is a read-only VS Code extension that reads the AI sessions you already generate, on your machine, and turns them into a private dashboard. Core analysis runs locally and sends nothing.
Read what you already have
Your AI sessions are already logged. Stack Signal reads the existing session logs from five harnesses — Claude Code, Copilot, Codex, OpenCode and Xcode — in read-only mode, and normalises them into a single local timeline. You do not instrument anything, change your workflow, or route your prompts through a third party. The data was already sitting on disk; the extension simply reads it.
A tool that asks you to send your coding sessions elsewhere is asking you to trust its retention policy more than your own. Reading logs you already hold removes the question entirely.
That read-only, local posture is the design decision everything else hangs from. Team roll-up, the LimitLens spend view and optional LLM explanations do make network calls — but only when you explicitly enable them. The core, the part that scores your sessions, never leaves the machine.
Forty-five rules, not a vibe
Measuring AI leverage is easy to fake with a single number. Stack Signal instead checks each session against 45 anti-pattern rules, grouped into categories a developer can act on rather than a headline score.
| Category | What it catches |
|---|---|
| Prompt quality | Prompt smells — vague asks, missing constraints |
| Session hygiene | Sessions that sprawl or lose their thread |
| Code review | Weak or absent review of AI-generated output |
| Tool use | Tools left under-used when they would help |
| Context management | Context bloat that degrades the model’s answers |
Alongside the rules, the dashboard shows an activity timeline, how much code came from AI versus your own hands, context-health audits, and the LimitLens view for daily token burndown. A Skill Finder clusters the prompts you keep repeating, so a recurring ask becomes a reusable skill rather than a habit nobody noticed.
The point is not to grade the developer. It is to make the pattern visible, so a team can decide what to change.
Honest about the edges
Stack Signal is tuned to run on Claude Code, so any engineer with a Claude Code login can use it without a separate subscription. We are equally clear about the boundaries. The current log readers cover the five named harnesses and we do not claim others. The networked features — team roll-up, LimitLens and LLM explanations — sit outside the zero-network-call core by design, and product demonstrations use synthetic sample data rather than a customer’s private logs. It is offered today as a no-fee design-partner pilot, available as a VS Code extension.
Stating those limits is the same discipline as the read-only posture. A usage-analytics tool that overclaims its own reach is not one you would trust with your session data.
If you want a private, evidence-based view of how your team actually works with AI — without sending that evidence anywhere — see the Stack Signal product page. To apply for the design-partner pilot, reach us via /contact, or book a call at cal.com/agentix-tech.
