kramme-pr-verify-description

name	kramme:pr:verify-description
description	Compare an existing PR's title and body against the actual branch diff and report drift — false claims, missing major changes, stale scope, missing risk callouts. Use after pushing changes to a branch with an open PR, or before requesting review. Read-only by default; add --fix to delegate to kramme:pr:generate-description for an updated description. Complements kramme:pr:code-review (which checks description accuracy as one signal among many code-quality checks) by being a fast, focused, single-purpose check that runs in seconds.
argument-hint	[--fix] [--base <ref>] [--strict]
disable-model-invocation	false
user-invocable	true

PR Description Verifier

Parse Arguments

Parse $ARGUMENTS for flags:

• --fix: After reporting drift, ask the user once whether to delegate to kramme:pr:generate-description --auto to regenerate and update the PR body. Default is report-only.
• --base <ref>: Use <ref> as the base branch for diff computation instead of auto-detecting.
• --strict: Tighten the accuracy bar. Flag every bullet in the description that cannot be tied to a diff hunk. Default mode is loose — only contradictions, material omissions, and missing risk callouts surface.

Set FIX_MODE=true / BASE_BRANCH_OVERRIDE=<ref> / STRICT_MODE=true from the flags and remove each flag (and its value) from remaining arguments.

When to Use This Skill

Use this skill when:

• You just pushed new commits to a branch that already has an open PR and want a quick sanity check that the description still matches the code.
• You're about to request review (or move the PR out of draft) and want to confirm the body is honest about scope, risks, and what's in the diff.
• A reviewer left a comment about the description being unclear or wrong and you want a structured drift report before rewriting.

When NOT to use this skill:

• The PR doesn't exist yet — use kramme:pr:generate-description to draft one instead.
• You want a full code review — use kramme:pr:code-review (it includes description accuracy as one of many checks).
• You just want to rewrite the description from scratch — use kramme:pr:generate-description --auto directly.

Scope and Rubric

This skill is single-purpose: it answers "does the PR body honestly describe the current diff?" It does not review code quality, suggest implementation changes, or rewrite the description on its own.

Classify each potential drift point against the rubric below. Severity drives the verdict; mode controls which severities surface in the report.

Drift type	Severity	Visible in loose mode
Body contradicts diff (flag default flipped, "no migration" claim refuted, "no breaking changes" claim with a rename in the diff)	Critical	Yes
Undisclosed migration, breaking API rename, missing required migration note, or new required env var	Critical	Yes
Undisclosed SemVer-major implication, or body claims no breaking changes when a versioned artifact surface or durable public contract breaks	Critical	Yes
Undisclosed new endpoint, dependency, feature flag, or removed code path	Important	Yes
Versioned artifact or durable public contract change lacks release impact, SemVer, changelog, or migration story needed for release-note accuracy	Important	Yes
Potential concerns: None despite a migration, flag flip, partial coverage, or known follow-up	Important	Yes
Stale claim — body describes work that was removed or refactored away	Important	Yes
Things I didn't touch excludes adjacent work the diff did touch	Important	Yes
Title Conventional Commit type wrong for the dominant change (fix: for a feature addition)	Important	Yes
Title type defensible but not ideal for a mixed change (refactor: vs feat:)	Suggestion	Strict only
Per-area technical notes missing when 3+ areas changed	Suggestion	Strict only
Test Plan missing manual scenarios when reviewer-visible UI changed	Suggestion	Strict only
Body claim cannot be traced to a specific diff hunk	Suggestion	Strict only
Wording, tone, polish, bullet ordering, formatting	Not flagged	Never
Description shorter than the diff (concise but accurate)	Not flagged	Never
Missing Linear ID link	Side note in report header	Not a finding

Severity definitions — apply to new drift types not enumerated above:

• Critical — a reviewer relying on the body would approve a PR that breaks production, ships a flag in the wrong default state, or merges undisclosed scope.
• Important — misleads about scope, test coverage, or risk in a way that affects review depth or release-note accuracy but won't immediately break production.
• Suggestion — minor inaccuracy a careful reviewer would catch from the diff itself.

Workflow

Phase 1: Branch and PR Detection

1. ALWAYS confirm the gh CLI is installed and authenticated before any other step:

   if ! command -v gh > /dev/null; then
     echo "MISSING REQUIREMENT: gh CLI not installed. Install from https://cli.github.com." >&2
     exit 1
   fi
   if ! gh auth status > /dev/null 2>&1; then
     echo "MISSING REQUIREMENT: gh CLI not authenticated. Run \`gh auth login\` first." >&2
     exit 1
   fi
   

2. Confirm the current branch:

   git branch --show-current
   

3. Resolve the base branch with the shared plugin script. It uses the same 3-tier strategy: explicit --base override, PR target branch, then origin/HEAD/origin/main/origin/master. It runs in strict mode, so fetch failures stop the workflow with the script's stderr message instead of being silently swallowed:

   RESOLVE_ARGS=(--strict)
   [ -n "${BASE_BRANCH_OVERRIDE:-}" ] && RESOLVE_ARGS+=(--base "$BASE_BRANCH_OVERRIDE")
   
   RESOLVED=$("${CLAUDE_PLUGIN_ROOT}/scripts/resolve-base.sh" "${RESOLVE_ARGS[@]}") || {
     echo "Error: Could not resolve base branch; see the message above. Re-run with --base <ref>." >&2
     exit 1
   }
   eval "$RESOLVED"
   

   The script exports BASE_REF, BASE_BRANCH, and MERGE_BASE for the diff commands in Phase 2.

4. ALWAYS confirm a PR exists for the current branch and is in an open state:

   gh pr view --json number,url,title,body,baseRefName,headRefName,state
   

   If no PR exists, stop with:

   MISSING REQUIREMENT: no PR found for the current branch.
   Run `/kramme:pr:create` or `/kramme:pr:generate-description` first.
   

   If state is not OPEN (i.e. MERGED or CLOSED), warn but continue — verifying a merged PR's drift is occasionally useful (e.g. when preparing a follow-up). Prepend a PR state: <STATE> (verification on non-open PR) line to the report header so the user notices.

   Capture PR_NUMBER, PR_URL, PR_TITLE, PR_BODY, PR_STATE for downstream phases.

Phase 2: Diff and Commit Gathering

1. Get the branch diff against the base:

   git diff origin/$BASE_BRANCH...HEAD
   git diff origin/$BASE_BRANCH...HEAD --stat
   git log origin/$BASE_BRANCH..HEAD --format="%h %s%n%b%n"
   

2. Include local uncommitted work in the diff scope (the PR body should match what will be on the branch after the next push):

   git status --porcelain
   git diff HEAD       # staged + unstaged
   

   If local changes exist, ALWAYS note this in the report header (Local uncommitted changes included in scope: <N> files). If they're substantial and the user is verifying "after pushing", warn them that the comparison includes work not yet on the remote.

3. Categorize changed files for the rubric:

   • Migrations / schema changes
   • New or removed endpoints / routes
   • New or removed dependencies (package.json, requirements.txt, go.mod, etc.)
   • Feature-flag definitions or default flips
   • Public API surface changes (exported symbols renamed, removed, or new)
   • Versioned artifact surfaces (package/CLI/API/SDK/schema/integration contracts) and their SemVer implications
   • Changelog, release-note, tag/version, or migration-guide files
   • Env var additions
   • Test additions / removals
   • Documentation changes

Phase 3: Drift Analysis

Walk the description body section-by-section and the diff in parallel. Classify each potential drift point against the rubric in ## Scope and Rubric above. For each finding, record:

• Location — section of the PR body (e.g. ## Summary, ### Potential concerns, Title) or PR description if global.
• Type — the drift type from the rubric (e.g. Contradiction, Material omission, Stale claim, Missing risk callout, Scope misrepresentation, Title drift).
• Severity — from the rubric column.
• Evidence — what the body says vs. what the diff shows (cite file paths and a hunk summary, not full diffs).
• Recommended fix — concrete text edit or "regenerate via kramme:pr:generate-description --auto".

For diffs that change a versioned artifact surface or durable public contract, verify that the body either states the release impact or explicitly explains why there is no versioned consumer contract. Flag undisclosed SemVer implications, missing migration/upgrade notes for breaking changes, missing curated changelog/release-note expectations, and release-impact claims that are not supported by the diff or gathered PR context.

In loose mode, suppress any finding whose rubric row says "Strict only" or "Never". In --strict mode, surface those too.

Phase 4: Report

Present an inline report (do not write a separate file). Use this structure:

# PR Description Verification — #<PR_NUMBER>

**PR:** <PR_URL>
**Title:** <current title>
**Base:** <BASE_BRANCH> · **Head:** <current branch>
**Mode:** loose | strict
**PR state:** <STATE> (only show this line when state is not OPEN)
**Local uncommitted changes included in scope:** <N> files (only show when N > 0)

## Verdict

**VERDICT_TAG:** ACCURATE | MINOR_DRIFT | MATERIAL_DRIFT | INACCURATE

<prose verdict: "Accurate" | "Minor drift" | "Material drift" | "Inaccurate — do not merge as-is">

## Findings

### Critical

- **[Type]** *(Location)* — <one-sentence summary>
  - Body says: "<quoted or paraphrased>"
  - Diff shows: <evidence with file paths>
  - Fix: <concrete recommendation>

### Important

(same structure)

### Suggestions

(only emitted in --strict mode)

## Not flagged (loose-mode skips)

- <one line per skipped potential issue with a brief reason, e.g. "Wording in summary is dry but accurate">
- <rank by closest-to-flagging; if list grows past ~8 entries, truncate the tail with "+<N> more low-signal skips">

## Next steps

- <if no findings: "Description matches the diff. Safe to request review.">
- <if findings: numbered list of fix options, ending with "Run `/kramme:pr:verify-description --fix` to regenerate via kramme:pr:generate-description --auto">

ALWAYS include both the VERDICT_TAG: line (for tooling) and the prose verdict line. Map findings to verdict like this:

Highest finding severity	VERDICT_TAG	Prose verdict
No findings	ACCURATE	Accurate
Only Suggestions	MINOR_DRIFT	Minor drift
Important (no Critical)	MATERIAL_DRIFT	Material drift
Any Critical	INACCURATE	Inaccurate — do not merge as-is

Phase 5: Optional Fix Delegation

Skip this phase if FIX_MODE is not set.

If FIX_MODE=true and at least one Important or Critical finding was reported, ask the user once:

Found <N> finding(s). Regenerate the PR description by running
`/kramme:pr:generate-description --auto`? [y/N]

• On y: invoke the kramme:pr:generate-description skill with --auto (it will detect the existing PR and update it directly). Then re-run Phase 1-4 of this skill to verify the regenerated body actually resolves the findings — if any Critical or Important finding persists, report it and stop.
• On n or no response: stop. Print the report and exit.
• If FIX_MODE=true but the verdict was Accurate, do not prompt — just confirm "Nothing to fix."
• If kramme:pr:generate-description is not available in this environment, report the drift findings only, note MISSING REQUIREMENT: kramme:pr:generate-description not installed; --fix unavailable, and stop. Do not attempt to rewrite the PR body manually.

NEVER invoke kramme:pr:generate-description automatically without the y/N confirmation, even in FIX_MODE. The user opted into fixing, not into a silent rewrite.

Output Markers

Use these uppercase markers in the report (and in conversation output around it) so the user can audit decisions:

• UNVERIFIED — a description claim you could not confirm or refute from the diff alone. UNVERIFIED: body says "matches behavior on staging" — cannot validate from diff.
• CONFUSION — diff evidence that contradicts the body or the commit log. CONFUSION: body says "flag defaults OFF" but flag definition sets default=true.
• NOTICED BUT NOT TOUCHING — drift you spotted but deliberately did not flag in loose mode (move to a Not flagged line in the report). NOTICED BUT NOT TOUCHING: title says "fix" but change is small enough that "fix" is defensible.
• MISSING REQUIREMENT — context the user must supply before verification can complete. MISSING REQUIREMENT: no open PR for current branch.

Common Rationalizations

Watch for these — each one means a finding is about to be wrongly suppressed:

• "The diff is small, the description doesn't need to cover everything." → Size doesn't excuse contradictions. A two-line diff that flips a default still needs that line in the body.
• "The reviewer can see the migration in the file tree." → The body's Potential concerns block is the contract; visible-in-diff does not equal disclosed.
• "The title says fix but it's basically a fix." → Conventional Commit type drives changelogs and release notes. If the dominant change is a feature, the title is wrong regardless of how the author thinks of it.
• "SemVer can wait until release day." → SemVer is the consumer promise. A PR that changes a public contract must disclose the version and migration implications while reviewers can still evaluate them.
• "Things I didn't touch: None is fine — the author probably considered it." → Only fine when nothing adjacent was changed. If the diff touches adjacent files, the block needs an entry.
• "The author will rewrite the description before merge anyway." → Maybe — but the point of this skill is to remove that step or to make it explicit now, not to assume future cleanup.

Red Flags — STOP

Pause and re-examine the diff if any of these are true while drafting the report:

• You're about to return Accurate but the diff contains a migration, feature-flag default, or breaking change.
• You're about to return Accurate but the diff changes a versioned artifact surface or durable public contract such as a public API, package, CLI, integration contract, or schema, and the body has no release impact, SemVer, changelog, or migration story.
• You're tempted to soften a Critical finding to Important because "the author probably knows".
• The report has more Not flagged entries than Findings — that usually means the bar is set too low; reconsider whether several of the skipped items are actually material.
• You're about to invoke kramme:pr:generate-description --auto without explicit user confirmation.
• The diff is empty or near-empty (only whitespace, comments, or formatting) — there's nothing to verify against; return Accurate and stop.

Verification

Before presenting the report, self-check:

• gh CLI was confirmed installed and authenticated; missing-tool exit was clean (MISSING REQUIREMENT), not an opaque stderr dump.
• PR existence was confirmed; if none, the skill exited with a MISSING REQUIREMENT message rather than improvising.
• Non-open PR state was surfaced in the report header rather than ignored.
• Base branch was detected and origin/$BASE_BRANCH was fetched successfully.
• Diff scope explicitly includes local uncommitted changes if any exist, and the report header says so.
• Every finding has a Location, Type, Severity, Evidence (with file paths), and Recommended fix.
• Severity assignments follow the rubric in ## Scope and Rubric — Critical is reserved for merge-blocking misinformation.
• Loose-mode reports contain only Important and Critical findings; Suggestions appear only under --strict.
• Both VERDICT_TAG: and the prose verdict are present and match the highest finding severity.
• --fix flow asked for y/N confirmation before delegating to kramme:pr:generate-description, and stopped cleanly if the sibling skill is unavailable.
• Report was emitted inline; no file was written.
• No code review findings (style, types, security, performance) appear in the report — that's kramme:pr:code-review's job.

Notes

• This skill is read-only by default. The only side effect is the optional --fix delegation, which itself requires explicit user confirmation.
• The diff is the source of truth; the description is the suspect. When the body and the diff disagree, the recommended fix is always to update the body, not the code.
• Suggestions are intentionally suppressed in loose mode to keep the signal-to-noise ratio high. Reviewers rerunning this skill should not be drowning in nits.
• This skill complements kramme:pr:code-review, which checks description accuracy as one signal among many. Use this skill when you only need the description check and want it fast; use pr:code-review when you want a full quality pass.