agentsclimarketplace

Loop grill

Skill mfittko/dev-loops/skills/loop-grill

Standalone pre-loop Socratic Q&A grill for issues (tracker-first), PR bodies, or local plan files (local-planning). Detects spec gaps, asks clarifying questions (interactive) or self-answers them from codebase context (--auto), then synthesizes Acceptance criteria / Definition of done / Non-goals into the body while keeping the raw Q&A only in an ephemeral tmp artifact.From its SKILL.md

Install
npx -y skills add mfittko/dev-loops --skill loop-grill

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 18 stars18 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

19.9 KB, as published. Nobody here has run it

Loop-grill skill

A standalone, on-demand pre-loop grilling skill. Run it against an issue or a local plan file before the dev loop starts to surface underspecified acceptance criteria, fuzzy scope boundaries, unresolved primary actors, and undocumented hard-to-reverse decisions.

It is entirely separate from the in-loop docs-grill (skills/docs/docs-grill-step.md), which audits code/doc drift while the loop runs. This skill operates on the spec before any implementation begins.

State machine

This grill is the bounded, closed sub-loop modeled by packages/core/src/loop/refinement-grill-state.mjs (detector scripts/loop/detect-refinement-grill-state.mjs), rendered on the State atlas; see docs/refinement-grill-state-graph.md. It obeys GRILL-SUBLOOP-STATE-MACHINE, GRILL-SUBLOOP-ITERATE-TO-CLEAN, GRILL-SUBLOOP-NO-EMBED-SYNTHESIS, GRILL-SUBLOOP-HONEST-HANDOFF, GRILL-SUBLOOP-FULL-REWRITE, GRILL-SUBLOOP-RATIONALE-COMMENT, and GRILL-SUBLOOP-NO-BARE-HASH. Iteration lives in the transition graph; the answer/synthesis is the bounded input consumed at the await_answers state.

Interface

/loop-grill <issue-number>             # tracker-first, interactive
/loop-grill <issue-number> --auto      # tracker-first, auto-answer
/loop-grill <pr-url-or-number>         # PR body, interactive
/loop-grill <pr-url-or-number> --auto  # PR body, auto-answer
/loop-grill <path/to/plan.md>          # local-planning, interactive
/loop-grill <path/to/plan.md> --auto   # local-planning, auto-answer

In a consumer (plugin) install these run as /dev-loops:loop-grill …; the bare /loop-grill form is dev-loops-repo-local (repo-local .claude/commands).

PR bodies are supported the same way as issues: fetch/edit the PR body via the existing PR wrappers (scripts/github/view-pr.mjs / scripts/github/edit-pr.mjs), never raw gh.

Argument validation (fail-closed)

Before doing anything else:

  1. Confirm exactly one positional argument is present (issue number or path). No argument → error, stop.
  2. Confirm the only optional flag is --auto. Any other flag → error, stop.
  3. Tracker-first: verify the issue exists. Non-existent issue → Error: issue #<n> not found., stop.
  4. Local-planning: verify the file exists. Missing file → Error: plan file not found: <path>, stop.
  5. Never mutate any artifact when argument validation fails.

Step 1 — Load the target

  • Tracker-first: fetch the issue body (title + description) via scripts/github/view-issue.mjs (never raw gh).
  • PR body: fetch the PR body via scripts/github/view-pr.mjs (never raw gh).
  • Local-planning: read the plan file from disk.

"Is it refined" is decided by detectIssueRefinementArtifact (packages/core/src/loop/issue-refinement-artifact.mjs) — the single source of truth, same as the enqueue gate; do not add a divergent check. An already-refined artifact (AC/DoD present) is a zero-iteration grill_clean: do not rewrite the body.

Step 1b — Surface external resources

Before detecting gaps, scan the loaded content for external resource references: links, other repo URLs, API endpoints, doc URLs, screenshots, or Playwright navigation descriptors.

  • Interactive mode: if external resources are present, ask the operator to confirm they are accessible or to provide them before the Q&A starts.
  • --auto mode: attempt to fetch each resource using bounded wrapper commands (e.g. gh, API wrapper scripts under scripts/). Do not fetch resources inline with raw curl or token-heavy calls. Flag any inaccessible resource as unresolved in the findings rather than silently skipping it.
  • When no external resources are present, skip this step silently.

Visual resources

Screenshots and Playwright navigation descriptors are visual resources: they let a design/UI gap be answered against the current screen instead of ambiguous prose. Surface them here, before the Q&A, alongside the other external resources.

  • Screenshot (file path or URL): include the referenced image as context when answering design/UI-related gaps. A local path is loaded directly with the read tool (it renders images); a URL is fetched with a bounded wrapper command, never a raw token-heavy call. An unreadable path or an inaccessible URL is flagged unresolved — never describe a screen you could not see.
  • Playwright navigation descriptor (an ordered path to a screen, e.g. "go to /settings, click Edit Profile"):
    • --auto mode: invoke the bounded wrapper dev-loops loop visual-grill-capture --repo-root <p> --app-url <url> --output-dir <p> --descriptor <json|@file> (source-repo fallback: node scripts/loop/visual-grill-capture.mjs …) to drive a headless browser and capture that screen's screenshot as context. The wrapper is a thin adapter over the ui_review drive harness; do not run browser code inline. A @<file> descriptor reference must be a relative, non-traversing path inside the repo root (an absolute or ..-escaping path is rejected unread). Navigation is confined to the running app: only http/https goto targets on the app's own origin are allowed (a file:/data: scheme or a cross-origin override is rejected), steps are limited to DOM-interaction actions that reach a screen (goto/click/fill/select/dispatch — upload and unknown actions are rejected so no local file is read), and the descriptor is capped at a bounded step count. Confinement is also enforced at runtime — if a redirect or a click-navigation leaves the app origin, the capture fails closed rather than screenshotting an off-origin page. Only the final screen is persisted: intermediate step captures (which may hold sensitive state, e.g. a screen after a credential fill) are pruned as the walk advances. When it returns ok: false (the runner is unavailable, login fails, a step cannot be reached, or navigation is rejected), flag the visual gap unresolved — never re-describe the screen in prose. Playwright is an optional peer dependency, so an install that never set it up returns ok: false with the install instructions in stopReason (npm install --save-dev @playwright/test, then npx playwright install webkit); that is a resolvable setup gap, not a defect in the target.
    • Interactive mode: ask the operator to provide the screenshot or confirm the navigation steps before the Q&A starts.
  • Any inaccessible screenshot, or a navigation descriptor that cannot be captured, is flagged unresolved — the same fail-closed degradation rule as every other external resource.

Step 2 — Detect gaps

Scan the loaded content and identify each gap. The minimum required gap detectors are:

Gap kindDetection signal
Missing acceptance criteriaNo ## Acceptance criteria section, or section is empty / stub
Missing scope boundaryNo explicit in-scope / out-of-scope statement or non-goals section
Unresolved primary actorNo named user, system, or role that is the main beneficiary of the feature
Undocumented hard-to-reverse decisionDestructive or irreversible operations described without a rationale or rollback note

Additional gaps discovered through semantic reading of the spec are also recorded.

Count-based acceptance criteria guardrail

<!-- rule: GRILL-COUNT-AC-UNIT-DISPATCH-MODE -->

GRILL-COUNT-AC-UNIT-DISPATCH-MODE: An acceptance criterion that names a count (sentinel count, angle count, dispatch-unit count) MUST specify which unit the count refers to — sentinel vs angle vs dispatch-unit — and MUST be validated against BOTH the per-angle default AND the shipped grouped-dispatch default, not just one. When the repo uses grouped fan-out by default (this repo does — resolveFanoutGroups writes ONE sentinel per group; #1579/#1601), the AC MUST call out the grouped-dispatch interaction explicitly, or it will false-fail every grouped round. Detect a count-based AC that omits the unit or the dispatch-mode interaction as a gap before synthesis, and sharpen the unit + dispatch-mode validation into the synthesized ## Acceptance criteria.

Cautionary case (#1618 AC3): the literal AC "sentinel count short of non-carried angle count" would have regressed the shipped grouped-dispatch default — grouped fan-out writes ONE sentinel per group, so 1 sentinel < N non-carried angles false-fails every grouped round. A fresh-context reviewer caught it as HIGH; the operator chose Option D (thread --expected-dispatch-units). Writing the AC without considering grouped dispatch forced an in-flight scope decision. Count-based ACs that omit the unit + dispatch-mode interaction repeat this failure mode.

For each gap, classify it as either:

  • Bounded choice — the answer is one of a small discrete set (e.g. yes/no, A/B/C).
  • Open-ended — the answer requires free-form elaboration.

Step 3 — Fill gaps

Interactive mode (default)

For each gap, in order:

  • Bounded choice gap: use AskUserQuestion with the question text and the choice options, plus an "Other / free text" option. Block until the user answers.
  • Open-ended gap: present the question as a plain text turn. Block until the user answers.

Record each answer's source as the operator's GitHub handle: resolve it once per run with gh api user --jq .login. This direct self-lookup is a knowingly accepted advisory raw-gh call under the wrapper-for-gh-reads rule owned by the Dev Loop Skill — the operator declared a dedicated wrapper a non-goal for this one read-only self-read; every other gh read still goes through a wrapper. Treat the login as resolved only when the command exits 0 and its trimmed stdout is non-empty, is not the literal null, and matches ^[A-Za-z0-9-]{1,39}$ (a plain GitHub handle — never markdown-significant text); use that handle (e.g. mfittko) as the Source value for every human-answered gap. In every other case fall back to the literal human. --auto mode is unaffected — its evidence-source tokens are not human answers.

AskUserQuestion is a Claude Code–native construct. If you are running outside Claude Code, use --auto mode instead.

Auto mode (--auto)

Answer every grilling question yourself without prompting the user. Source answers from (in priority order):

  1. codebase — inspectable source files, tests, scripts, config in the repository.
  2. docs — markdown files under docs/, skills/docs/, and adjacent contract docs.
  3. context — CONTEXT.md at the repo root, if present. When absent, skip silently — do not crash, no warning required.
  4. inferred — reasoning from the issue/plan text alone, with no external citation.

Record the evidence source for every answer. Flag a question as unresolved when:

  • The only available source is inferred, and
  • No codebase path, doc section, or issue/plan text can be cited as the basis for the answer.

Do not silently guess an inferred answer when no evidence can be cited — flag it unresolved instead.

Step 4 — Write back (synthesize sections; raw Q&A to tmp only)

Synthesize the answers into the body as sharpened ## Acceptance criteria, ## Definition of done, and ## Non-goals sections. Per canonical heading, use replace-section semantics:

  • Find the existing section: the range from that ##-level heading through the next ##-level heading (exclusive) or end of file.
  • Replace that range in place with the synthesized content; if the section is absent, append it.
  • This makes re-runs idempotent — no accumulated noise, no duplicate sections.
  • If parsing a section boundary fails, abort with an error rather than silently truncating.

The synthesized sections carry no rationale scaffolding: do NOT write a ## Grill findings section and do NOT embed the raw Q&A table in the body (GRILL-SUBLOOP-NO-EMBED-SYNTHESIS). (The full rewrite below defines the complete set of content the locked body keeps — context, decided approach, and the canonical sections — so this is a "no embed", not a "sections only".)

If a body migrated from older embed behavior still carries a ## Grill findings section, remove it as part of write-back — strip from that heading through the next ##-level heading (exclusive) or end of file, using the same replace-section boundary logic. This is a removal-only migration, never a re-introduction of the embed.

Write the raw Q&A transcript ONLY to the gitignored, ephemeral, session-scoped artifact tmp/issues/issue-<n>/grill/<timestamp>.md (tmp/ is already gitignored; never committed). For PR-body and plan-file surfaces, use the same tmp path shape scoped by surface (issues: tmp/issues/issue-<n>/grill/; a parallel tmp/... path for PR/plan).

Tracker-first write-back is a full rewrite, not an append (GRILL-SUBLOOP-FULL-REWRITE). After the replace-section synthesis above, scan the ENTIRE remaining description (not just the canonical headings) and resolve it into one locked, unambiguous spec:

  • Any "suggested / option A or B / TBD" phrasing that describes a gap THIS grill run just decided: rewrite it to the decided form only; delete the rejected alternative(s). A leftover undecided option for a gap the grill already resolved is a write-back defect.
  • Any Refinement notes / Grill findings / RFC-style rationale narrative (gap tables, recommendation + rejected alternatives, decision log) anywhere in the description: remove it entirely. That content is never body prose — it moves to the results comment (see below), not the description.
  • Any contradiction between now-stale prose and the locked AC/DoD/approach: resolve in favor of the locked form; delete the stale prose.
  • Any bare #<number> used as a defect/item enumeration rather than a genuine issue/PR reference (GRILL-SUBLOOP-NO-BARE-HASH): rewrite as defect N / item N / backticks. GitHub auto-links a bare #<number> to an unrelated issue/PR in this repo — reserve #<number> only for a real issue/PR cross-reference.

The rewritten description carries ONLY normative locked content: context, the decided approach, ## Acceptance criteria, ## Definition of done, ## Non-goals, and a linked refinement doc reference if present. Write it back with:

dev-loops issue edit --repo <owner/repo> --issue <n> --body-file <tmp-body-path>

(source-repo fallback: node scripts/github/edit-issue.mjs --repo <owner/repo> --issue <n> --body-file <tmp-body-path>)

Post the rationale as a separate results comment (GRILL-SUBLOOP-RATIONALE-COMMENT): the description and the rationale are two distinct artifacts — never merge them. When the operator accepts an escalated RFC-worthy decision from the rationale, it is persisted as an ADR per the Decision record contract. Write the rationale (gaps found and filled, the RFC recommendation and rejected alternatives, and decisions taken) to a second tmp file and post it as its own comment titled 🔬 Grill / refinement results:

node scripts/github/comment-issue.mjs --repo <owner/repo> --issue <n> --body-file <tmp-rationale-path>

Never gh issue comment directly, and never fold this content back into the issue body. The same #<number> hygiene rule (GRILL-SUBLOOP-NO-BARE-HASH) applies to the comment. A zero-iteration grill_clean target (already refined, body left unchanged) has no rationale to post and skips this step; any run that actually filled a gap MUST post the results comment.

PR-body write-back: update the PR body via scripts/github/edit-pr.mjs (never raw gh), same replace-section semantics.

GitHub body size guard: issue/PR bodies are capped at 65,536 characters. Before writing back, check whether the updated body would exceed this limit. If so, warn: Warning: updated body would exceed GitHub's 65,536-character limit — write-back skipped. Sharpen the sections or the body manually. Do not silently truncate.

Local-planning write-back: update the plan file in place using the edit tool.

Output artifact format

Three distinct artifacts for tracker-first (two for PR-body/local-planning, which have no separate results-comment surface in this contract):

  1. Rewritten description (issue/PR/plan body): the fully rewritten, locked spec — context, decided approach, ## Acceptance criteria, ## Definition of done, and ## Non-goals. No raw Q&A, no rationale narrative, no unresolved "suggested … or …" phrasing, no bare non-issue #<number>.

  2. Results comment (tracker-first only, posted separately, titled 🔬 Grill / refinement results): the rationale — gaps found and filled, the RFC recommendation and rejected alternatives, and decisions taken. For an interactive run the preamble reads source: <handle> answers via operator Q&A using the same resolved handle (fallback: source: human answers via operator Q&A). Same #<number> hygiene rule applies.

  3. Raw Q&A transcript (ephemeral tmp/issues/issue-<n>/grill/<timestamp>.md only — never the body, never the comment):

<!-- loop-grill: <timestamp> mode:<interactive|auto> -->

### Resolved gaps

| # | Gap | Question | Answer | Source |
|---|-----|----------|--------|--------|
| 1 | Missing AC | <question text> | <answer text> | codebase \| docs \| context \| inferred \| <operator-handle> (interactive; fallback `human`) |

### Unresolved gaps

| # | Gap | Question | Reason unresolved |
|---|-----|----------|-------------------|
| 1 | Unresolved primary actor | <question text> | No citable evidence found |

### Verdict

grill-clean

Replace grill-clean with N unresolved items when unresolved gaps remain.

Step 5 — Emit verdict

Before emitting the verdict for a tracker-first grill that filled at least one gap, verify the write-back contract and fail closed if any check fails — stop and report the specific violation instead of emitting a verdict:

  1. The rewritten description has no Refinement notes / Grill findings / rationale narrative section.
  2. The rewritten description has no unresolved "suggested … or …" / "option A or B" marker for a gap this run decided.
  3. Neither the rewritten description nor the results comment contains a bare non-issue #<number>.
  4. A 🔬 Grill / refinement results comment was actually posted (skip this check only for the zero-iteration grill_clean path, which has no rationale to post).

After write-back (and, for tracker-first, after the above verification passes), emit the verdict line to stdout:

  • grill-clean when no unresolved gaps remain.
  • N unresolved items (e.g. 3 unresolved items) when gaps remain after all questions are answered.

CONTEXT.md degradation rule

When CONTEXT.md is absent from the repo root, skip the context-source check silently. Do not crash. Do not emit a warning. The grill continues with the remaining sources (codebase, docs, inferred).

Idempotency guarantee

Running /loop-grill twice on the same target must not accumulate duplicate sections: the per-heading replace-section logic replaces the synthesized ## Acceptance criteria / ## Definition of done / ## Non-goals sections in place. An already-refined target is a zero-iteration grill_clean and its body is left unchanged. The raw Q&A transcript is written to a fresh timestamped tmp file per run; the body never carries it.

Non-goals

  • Auto-triggering from issue_intake — this is on-demand only.
  • Replacing or modifying the in-loop docs-grill (skills/docs/docs-grill-step.md, scripts/loop/docs-grill-contract.mjs) — different concern, different firing surface.
  • Full DDD CONTEXT.md management.
  • Scheduling or storing grill runs — stateless and on-demand.
  • Any CI/CD integration.

Keep looking

Skills are one crate of 611,178. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.