Full workflow — debug first, reproduce before fixing, test thoroughly, then PR.
Phase 1 Phase 2 Phase 3 Phase 4 Phase 5 Phase 6
Debug → Reproduce → Fix → Verify → Regression → PR
Understand Confirm bug Write the Same eval Happy paths All result
root cause on Cekura code fix must PASS + edge cases URLs in PR
BEFORE fix + commit now pass too
The 6 Phases
| Phase | File | What happens |
|---|---|---|
| 1 | phase1-debug.md | Fetch prod call + logs, identify root cause, confirm with user |
| 2 | phase2-reproduce.md | Build evaluator, attach metrics, run — eval must fail before any fix |
| 3 | phase3-fix.md | Write the code fix, commit locally |
| 4 | phase4-verify.md | Re-run same evaluator — eval must pass now |
| 5 | phase5-regression.md | Test all affected happy paths and edge cases |
| 6 | phase6-pr.md | Raise PR with all Cekura result URLs |
Strictness Rules — Read Before Starting
These rules are non-negotiable. Do not proceed past a gate without satisfying it.
Rule 0 — Use the same connection medium as the production call. No exceptions.
Every reproduction, verification, and regression test MUST be a full end-to-end simulation on Cekura using the same transport the agent is configured for. Retrieve the agent record (GET /test_framework/v1/ai-agents/{id}/) to confirm its transport. Most likely telephony, but follow what the agent is actually configured to use.
❌ Text mode is never a valid substitute. ❌ Do not switch transports between phases.
The bug lives in the real call path; only a simulation over the same medium can confirm it.
Rule 1 — Phases are sequential. No skipping.
Each phase has a gate. A gate is not passed by assumption — it is passed by evidence. The sequence exists because:
- You cannot write a good fix without understanding the root cause (Phase 1 gate)
- You cannot trust a fix without first proving the bug exists in a controlled way (Phase 2 gate)
- You cannot call regression tests meaningful without a passing fix verification (Phase 4 gate)
Rule 2 — Phase 2 is the hardest gate. Treat it as such.
Reproducing the bug is the most critical step. Do not move to Phase 3 until the eval definitively fails on Cekura with metric scores showing the failure. If there is any doubt about whether the bug is truly reproduced, stop and ask the user. Do not guess.
Rule 3 — When in doubt, ask.
If you are unsure which metrics to use, whether the root cause is correct, whether the edge conditions are right, or whether a result is ambiguous — stop and ask the user. A wrong assumption here wastes the entire workflow.
Rule 4 — Never push code until Phase 5 is complete.
The commit happens in Phase 3. The push happens only after all regression tests pass in Phase 5.