/diagnose

A bug gets fixed twice when the first fix targets where the pain shows instead of why it happened. This skill is a loop that ends at a root cause carrying an evidence trail, or ends honestly without one — it never ends at a plausible-sounding patch. Two commitments hold for the whole run: nothing gets fixed while the loop is running, and every probe leaves a visible mark in the evidence log.

The loop is stack-agnostic. If the repo pins its stack under docs/agentic-skills/stack.md, read it first and use its recorded commands; otherwise derive commands from the repo's own manifests. No step below assumes a language, framework, or test runner.

The evidence log

The trail is the deliverable. Open a log in the session output at the first hypothesis and append to it after every probe — earlier rows are never rewritten, superseded, or tidied away. The log is append-only so that rejected ideas stay visible; a hypothesis that keeps returning in new wording is the classic sign the loop is circling, and only a complete log exposes it.

Shape:

| # | Hypothesis                                    | Prediction (what would prove it wrong) | Evidence gathered        | Verdict    |
| - | --------------------------------------------- | -------------------------------------- | ------------------------ | ---------- |
| 1 | the sorter's comparison is inverted           | sorting two already-parsed values …    | direct sort of the pair … | eliminated |
| 2 | …                                              | …                                       | …                         | narrowed   |