Issue triage

How reports reach this repository and how they are sorted: issue forms, labels, what counts as a bug.

A skill deviation is its own report form and label, separate from a bug

Type: decision Status: active Evidence: confirmed Source: maintainer decision, 2026-09-09, while labeling the three findings of the 0.16.0 release series (#354–#356) and rewriting the issue templates (#358, #359) Revisit when: a skill-behavior report turns out to be deterministic often enough that the skill-wording auto-label is wrong more than it is right — then the form should stop labeling and leave it to triage

The issue chooser has three forms: Skill behavior report (label skill-wording) for "the skill was loaded and the agent still did something SKILL.md says it should not"; Bug report (label bug) for the linter, the GitHub Action, the docs and the evals tooling; Feature request (label enhancement). A second label, evals, marks a finding's origin — a full-suite run documented in docs/evals.md — independent of its nature. An open design question raised by a run, such as #356 on what Evidence grades when an entry has a confirmed and an unknown half, is question + evals until decided.

Reason: a rule that is stated in the skill and not followed reliably by the model is not a bug in the software sense — the files do what they say, the deviation is probabilistic (the 0.16.0 series shows 1–2 of 3 runs, never 3 of 3), and the fix is wording, not code. Calling it bug tells an outside reader that something is deterministically broken and then hands them a prose discussion. The two kinds of report also need different evidence: a skill deviation is reproduced from skill version, install route, agent tool, model, settings and a transcript excerpt; a linter or Action bug from a version, an OS, a Python version, a command and its output. One form would either make everything optional or ask a linter user for a model and a transcript.

Rejected alternative: one Bug report form with a component dropdown (skill / linter / Action / docs) and bug on everything. Rejected for the two reasons above: the label would be misleading for the largest class of reports, and the field set cannot serve both without going optional throughout. A form without any auto-label was also considered; skill-wording is applied on the assumption that most such reports are that, and a deterministic one is relabeled bug at triage — the cost of a wrong auto-label is one edit, the cost of no label is every report starting unsorted.

Consequence: bug is reserved for things that fail the same way every time. Eval-series findings get evals plus whichever nature label fits, never bug on the strength of a flip alone. GitHub adds its own "Report a security vulnerability" link to the chooser from SECURITY.md, so config.yml carries no security link of its own (#359).