Helping us reproduce your issue

Last updated

The fastest path from "customer reported it" to "fix shipped" is a clean reproducer we can run on our side without entering your workspace. This article is the checklist we send when we ask for one.

The minimal reproducer checklist

  1. The exact input. The full natural-language declaration text you submitted. Copy it from the Architect's Ledger (Stage 1 input is the canonical record).
  2. The fleet capability declaration. Your tier mix and model mix from the same Ledger. The compiler's behavior depends on both the input and the declared fleet, so we need both.
  3. The vendor and model used. Anthropic vs OpenAI, and which model. Different vendors and models produce different intermediate outputs.
  4. What you expected. A one-line description of what you wanted to happen.
  5. What actually happened. The Stage where the divergence first appears, plus any error message attached to that Stage.
  6. The run ID. Even if you provide everything above, the run ID is what lets us cross-reference traces and audit-log entries on our side.

If sharing the input is sensitive

Customer declarations sometimes contain intellectual property you cannot share. If that is the case, let us know — we have two options:

  • You can produce a sanitised version that exhibits the bug (often achievable by replacing proprietary names with placeholders).
  • With your written permission, we can use operator-impersonation into your workspace to reproduce on real data. This always leaves an entry in your audit log (Settings → Audit log) with the support ticket ID.

If the bug is not reproducible

Some compilation bugs are flakes — vendor-side latency, transient network issues. If a single rerun of the same input succeeds, the bug is harder to chase but we still want to know: send us the original run ID and the rerun's run ID so we can compare traces.

Reproducer artifacts you do not need to provide

  • We do not need your fleet's code or your codebase. The compilation problem lives entirely in our Stages 1–7; your fleet executes downstream of that.
  • We do not need your full BYOK key. The vendor / model / last-4 is enough for our triage.
  • We do not need a video. A screenshot of the Architect's Ledger is more useful than a screen recording.

What we do with the reproducer

We re-run the reproducer in our own test workspace, capture the divergence, and attach it to the Linear issue we file from your bug report. We do not run reproducers inside your workspace — that would create state in your account that is not from you.


Paired with the SR-REPRO-001 customer-support flow.

← All docs