Parity Gap

Parity Gap

Bring this program (default: cwd) up to (and past) a reference. Map both sides, diff them, confirm scope in rounds of 4 questions before writing anything, then author tasks into docs/todo.md, via the todo skill if installed or directly otherwise.

1. Resolve the Inputs First

• this program defaults to the working directory. Name another explicitly to override.
• reference is one of:
  • a remote repository: clone it into a temp dir then treat as local codebase.
  • a local codebase (path on disk): read its tree, entrypoints, commands, modules.
  • a docs/spec/feature list (file or pasted text): treat every command, endpoint, flag, or behavior it describes as a reference feature.
• If either is ambiguous (which repo? which spec?) or the reference can't be reached once identified (private, no network, no read access), ask anything and no guesses before mapping.

2. Map Each Side (Inline, Single-Pass)

Map what each side exposes to an end user; skip internals and file structure. The user-reachable surface:

• commands / subcommands / flags
• exported APIs, routes, endpoints
• config keys, env vars
• UI/TUI/CLI surfaces and the actions they expose
• notable behaviors, formats, integrations

Use Grep/Read/semantic code search inline. The reference is the spec side, so read it thoroughly; the program is the have side. If the reference is too big to hold inline, offload its mapping to a read-only exploration subagent (e.g. Explore) and diff from the returned inventory.

3. Diff into Two Buckets

• Parity gaps: capabilities the reference has that this program lacks. The core of the output.
• Improvements: things beyond the reference (better defaults, affordances the reference itself lacks, wins spotted during mapping). List these in their own section; never mix them into the parity list.

Drop anything the program already has. Note cascade order where one gap blocks others.

4. Confirm Scope

Before writing a single task, run clarifying questions through AskUserQuestion (max 4 per round, batched rather than trickled), looping rounds until scope is nailed. (On other harnesses, check for a native equivalent first: opencode's question tool, gemini-cli's ask_user tool, or codex's request_user_input [Plan Mode only]. Fall back to a plain numbered message and wait for answers only if none is available.) This gate is the point of the skill: the user reviews the gap list as questions, not as a finished file.

Each round, target the unknowns that matter most:

• which gaps are in scope versus intentionally omitted (the program may skip a reference feature on purpose)
• ambiguous parity calls (match the reference's behavior exactly, or adapt it?)
• which improvements to include vs drop
• depth (every gap, or only a named subset / surface)

Guidance:

• Keep looping rounds while real ambiguity remains and stop as soon as it doesn't. Four is a ceiling, not a quota; never pad a round to fill it.
• Use multiSelect: true when the user is picking which of several gaps/improvements to keep.
• Put a recommended option first, labelled (Recommended).
• Never park an unresolved decision inside a task; resolve it here in a round instead. Next session should be able to start implementing everything that gets decided now.

5. Write a Todo

Once scope is confirmed, write the gap list: if the todo skill is installed, load it; otherwise write straight to docs/todo.md. Either way, map the buckets onto the same two-bucket structure:

• Parity gaps -> punch-list under # <reference> parity punch-list, with one-line provenance (what was compared against what, date). Group by feature area; note cascade order.
• Improvements -> its own checkbox-spec section (e.g. ## Beyond <reference>), placed after the parity tasks.
• If docs/todo.md already has content, append the new punch-list section to it. Never overwrite the file or delete existing tasks (same no-delete rule the todo skill applies to itself).

Task rules (per the todo skill if installed): each task independently executable, says what to change not where, sized to one commit. Out-of-scope tasks aren't written.

Gotchas

• Rerunning against an existing docs/todo.md: append the new punch-list section, don't overwrite the file or delete prior tasks (Section 5).
• Reference identified but unreachable (private, no network, no read access) is a different blocker than the reference not being identified in the first place. Both stop at asking, never guess or substitute a stand-in (Section 1).

Boundaries

• This skill owns analysis + the question gate; the todo skill (or plain docs/todo.md) owns the file format; your multi-task runner runs one task each.
• Plain note-dumps with no reference to catch up to -> go straight to todo (or docs/todo.md directly), skip this skill.