A12 dmtool bug report

Filing a bug-and-findings report against dmtool

You turn a "something broke in dmtool" moment into a crisp, reproducible, correctly-scoped bug-and-findings report, and save it to a folder the user can hand on. The prose template is the easy part; the judgment below — separating a real defect from your own mistake, shrinking the repro, not trusting the tool's own pass/fail, and capturing all of it — is the point.

Don't trust the tool's own pass/fail — classify each error

A non-zero exit is usually a legitimate rejection, not a bug. Sort every error you hit into three:

What you see	What it is	Report it?
outcome:"rejected"/"refused" + a real MVK_*/RK_* code + a clear summary/fix	the kernel evaluated your input and it's genuinely wrong	No — unless the message itself misled you
outcome:"error", a diagnostic with code RK_INTERNAL_ERROR, exit 70, or a raw Java stack trace / Exception on stderr	an internal tool error — the tool broke	Yes, always
Unknown option / usage text, or a malformed spec you then fixed	your own call/spec mistake	No — fix the call; list it honestly so the reader can tell it apart

The tool tells you which it is: an internal failure is a structured error envelope (exit 70) with the real cause in its summary and the trace on stderr — it does not masquerade as a valid:false rejection. Run dmtool diagnostics RK_INTERNAL_ERROR for what it means.

The discipline

1. Classify every error, honestly (above). One session surfaces SEVERAL findings + friction — capture all of them, not just the headline or the last one hit.
2. Shrink a real defect to a minimal repro — the smallest dmtool model new + fewest fields that still fails; confirm it reproduces. Don't report against the user's big model.
3. Isolate the trigger — bisect (change one operator / flag / construct at a time). An affected-vs-unaffected table is the most valuable evidence in the report.
4. Scope it — which verbs reach the defect (the verb set depends on the defect class, not a fixed list).
5. Provenance — record dmtool --version (and the plugin/package version if known).
6. Trim the evidence — the distinctive codes/frames, never a whole multi-hundred-line trace.
7. Findings & improvements — every place the tool's own signals misled or slowed the work, each with a concrete suggested fix aimed at one target: the CLI, its material (operator catalog / schemas / --help / diagnostics), or the skill. Reflect on the whole tool-call trail, not just the calls around an error.
8. Replay log — the user's prompts that drove model changes, verbatim and in order, so the session is replayable.
9. Self-audit — did you stay inside dmtool's surface? Two checks, state both: (a) writes — was any model file changed outside dmtool (an editor / sed / echo / cp)? A direct edit can masquerade as a defect and taints the repro. (b) reads — did you read a model's raw .dm.json (via cat / jq / python / an editor) instead of model describe / field read / rule read? You shouldn't have needed to — and if you did, say so and file it as a [material] finding: the structured read that should have shown what you needed is incomplete (exactly the gap a raw read papers over).

Where to save it

Ask the user where to save the report (e.g. "Which folder should I write the bug report to?"). With no answer — or when you're running non-interactively — default to a dmtool-bug-reports/ folder in the current working directory.

Create that folder and write:

• report.md — the report (use the template below);
• any supporting artifacts you trimmed (full traces, the minimal repro model JSON) as sibling files, referenced from the report.

Then tell the user the folder path and that they can upload it, attach it to an issue, or send it however they like — this skill does not transmit anything itself.

Report template

# Bug report: <one-line title>
## Summary           — what breaks, in one paragraph
## Environment       — dmtool --version; OS; kernel/distribution if shown
## Severity / impact — who it blocks, and any workaround
## Steps to reproduce — the minimal, from-scratch commands
## Expected vs. actual
## Affected vs. unaffected — the bisect table (the key evidence)
## Exception (trimmed) — only the distinctive frames/codes
## Findings & improvements — [CLI] / [material] / [skill], each with a suggested fix
## Replay log         — the driving prompts, verbatim, in order
## Self-audit         — out-of-dmtool model writes (taint check) AND raw-JSON reads (→ a [material] finding)