dev-issues — brain-dump → distributed issues

A maintainer skill. It does not touch .ralphy/workspaces/<ws>/projects/ or call media models. It converts an unstructured stream of thoughts into a set of correctly-shaped notes/issues/*.md work items. Pairs with /dev-loop, which executes the issues this skill files.

When it fires

• The user tags /dev-issues and then writes a lot — a wall of ideas, gripes, product direction, "and also we should…". The longer and less structured the input, the more this skill earns its keep.
• ALSO FIRE when a normal dev conversation produces ≥3 distinct actionable threads the user wants preserved before they evaporate.
• DO NOT FIRE for one scoped task (do it, or file a single note via dev-tasks), for content production, or for one-off idea/decision capture into the notes/ inbox (dev-tasks).

Workflow

1. Read the inputs first. Read notes/README.md (the note shape + the Issues folder layout), docs/developing-ralphy.md (English-only rule, the lint suite), and skim the active backlog (the flat top level of notes/issues/) plus notes/ideas/ so you know what already exists. notes/issues/done/ and notes/issues/deprecated/ are the closed archive — skim them only during collision-check (step 3). dev-tasks is the sibling skill for notes/ inbox capture (ideas / decisions).

   Status-by-folder. notes/issues/ is split into the flat top level (the live backlog), done/ (resolved / promoted-and-landed), and deprecated/ (superseded / won't-do). New issues are always filed at the flat top level. Each issue keeps its > **Status:** line in sync with its folder.

2. Decompose the brain-dump into discrete items. One issue = one coherent unit of work (one verb / file / feature / refactor / bug). Split compound thoughts; merge near-duplicates. Aim for issues a single sub-agent could execute end-to-end.

3. Collision-check before filing each. Search all of notes/issues/ (active top level + done/ + deprecated/) and notes/ideas/ for an existing entity that already covers the thought. If one exists, update it (or note the overlap) rather than spawning a duplicate. If the match lives in done/, the thought is likely a regression or follow-up — git mv it back to the active top level and reset its > **Status:** line, or file a fresh issue that cross-links it. Never let the same thing live in two places.

4. Ask clarifying questions only where the answer changes what gets filed — a genuine fork the user must decide (scope, irreversibility, a product direction with no sensible default). Use AskUserQuestion, batch the questions, and offer a recommended option. Do NOT ask about things you can decide from the repo or sensible defaults. If the user wrote in Russian (or any non-English), that is fine in chat — but everything you WRITE to disk is English (translate/paraphrase).

5. Write each issue at the flat top level as notes/issues/NNN-kebab-slug.md, numbered monotonically across all three locations (fd -e md . notes/issues | rg -o '[0-9]{3}' | sort -n | tail -1, increment — never reuse a number freed by a move into done//deprecated/). Follow the notes/README.md shape (Status / Filed / Folder / Context / What / Why it matters / Notes) plus a concrete ## Scope / acceptance section listing what "done" means — file paths, gates to pass, acceptance criteria — so a downstream agent can execute without re-deriving scope. Add severity + category when useful. Cross-link related issues by number.

6. Order + dependencies. When the batch has a natural execution order (foundational schema before consumers, a cross-cutting rename before dependent work, plan-only/docs as low-risk), note it — either in a short lead issue or in each issue's Notes ("sequence after #NNN"). /dev-loop reads this ordering.

7. Commit the batch. git add notes/issues/<new files> and commit with a docs(notes): … message summarizing the set. Run the Cyrillic gate (rg '\p{Cyrillic}' --pcre2 notes/issues/<files>) before committing — empty is the gate. Push only if the user works commit-to-main (this repo does; see the user's memory).

8. Report back a thought→issue map: for each filed issue, one line (number, slug, the thought it captures), plus the recommended execution order and any item you deliberately dropped or merged.

HARD INVARIANTS

• English-only on disk. Translate any non-English brain-dump before writing.
• One thought, one home. Collision-check first; update over duplicate.
• Every issue is executable. If it lacks a concrete ## Scope / acceptance, it is not ready — sharpen it or ask.
• Capture, don't execute. This skill files issues; it does NOT implement them. Hand execution to /dev-loop or a targeted dev session.
• Numbering is monotonic across active + done/ + deprecated/. Never reuse a number freed by a move, and never renumber existing issues.
• File active issues at the flat top level only. done/ and deprecated/ are written to by /dev-loop (or a manual resolve), never by initial filing.