llm(claude-config-validator): post validate-ai findings inline on the diff

[SaintPatrck]

🎟️ Tracking

Pairs with bitwarden/gh-actions#959, which grants the inline comment and reply tools. Without them every finding's details nest in the summary, so this one merges first.

📔 Objective

validate-ai reports in one long comment that buries each finding in a dense paragraph, so reviewers map every path:line back to the diff by hand (the one on #209 ran to 11k characters). We post CRITICAL and IMPORTANT findings inline in the shape code-review uses, a one-line severity title with the details and fix collapsed, and keep the sticky comment to the verdict and one line per finding. Findings carry code-review's ❌ ⚠️ 🎨 labels so both reviews read the same way, and a re-run replies in a resolved thread whose problem persists rather than reposting a finding the pull request already carries.

🤖 AI-assisted review

A multi-agent review (Substantial path) ran on the branch before opening, and every finding it raised was addressed in this change.

[github-actions]

Claude Code validation

Result: Pass

Validated the claude-config-validator plugin changes in this pull request against merge base main (head c0967f9): the plugin manifest and version bump, the reviewing-claude-config skill, both /validate-ai and /validate-ai-local command definitions, their sibling READMEs, and the rewritten reference/validate-ai-scope.md support file. No critical findings, no secrets, and nothing that widens a permission or tool grant — the version bump is in lockstep across all four required files with a dated changelog entry, and the critical/major/minor → CRITICAL/IMPORTANT/SUGGESTED migration leaves no stale references behind. Two non-blocking findings, both on lines this changeset introduced.

Critical

None.

Major

• plugins/claude-config-validator/commands/validate-ai/validate-ai.md:40 — An absent or unreadable PR THREADS FILE: silently disables the only guard against duplicate inline comments. The new bullet says "If it is missing or unreadable, continue without it," while step 5 still posts every CRITICAL and IMPORTANT finding inline. If the workflow ever runs without supplying that line — an older bitwarden/gh-actions revision, or a failed get-pull-request-threads step — each re-push reposts the same findings as new inline comments, with nothing bounding the duplication and no signal to the reader that deduplication did not happen. This also contradicts the report contract added in the same changeset at reference/validate-ai-scope.md:153-157, which requires a check that could not run to be named on the **Not covered:** line rather than silently omitted. Fix: make the absent or unreadable threads file put a reason on the report's **Not covered:** line (for example, "existing pull request threads were unavailable, so findings were not deduplicated") and fall back to the summary's nested form on that run instead of posting inline.

Minor

• plugins/claude-config-validator/CHANGELOG.md:12 — The 2.2.0 entry documents behavior that only works when the consuming workflow supplies two claude-code-action MCP tools (mcp__github_inline_comment__create_inline_comment, mcp__github_replies__add_reply_to_pull_request_comment) and a PR THREADS FILE: prompt line, but records none of that coupling. An operator pinned to an older bitwarden/gh-actions validate-ai revision silently gets the degraded path — no inline comments and no deduplication — with nothing in the changelog to explain why. MINOR remains the correct bump, since the command falls back rather than failing. Fix: add one bullet under ### Changed naming the minimum bitwarden/gh-actions validate-ai revision, or the workflow inputs 2.2.0 expects (inline-comment tool, reply tool, PR THREADS FILE:).

Findings dropped after documentation check

• plugins/claude-config-validator/commands/validate-ai/validate-ai.md:3 allowed-tools (claimed that omitting mcp__github_inline_comment__create_inline_comment and mcp__github_replies__add_reply_to_pull_request_comment from the grant would restrict the command and silently disable the inline-comment path in CI): "allowed-tools | No | Tools Claude can use without asking permission during the turn that invokes this skill. The grant clears when you send your next message. Accepts a space- or comma-separated string, or a YAML list." The field pre-approves tools rather than restricting them, so the reasoning at commands/validate-ai/README.md:138-140 holds and the omission is not a defect.

Checks run

Check	Status
Plugin structure	Skipped — run by the workflow as a dedicated step before this review; its result reaches the pull request via the check status
Marketplace	Skipped — same, a dedicated workflow step
Version bump	Skipped — same, a dedicated workflow step (the AI plugin validation independently confirmed 2.1.0 → 2.2.0 is in lockstep)
Plugin validation (AI)	Passed — manifest, semantic versioning, directory structure, command frontmatter, and credential scan all clean; no hooks or MCP config to validate
Skill review (AI)	Passed — the single changed SKILL.md line is accurate, consistent with the rewritten reference, and introduces no broken references
Configuration & security	Ran — 1 major, 1 minor. Covered the two command definitions, both command READMEs, and the reference/validate-ai-scope.md support file; secret scan and settings.local.json check clean

Scope and coverage notes

• No CLAUDE.md, .claude/, agent, hook, or settings file changed, so those buckets were empty and needed no review. A .claude-pr/ snapshot exists but holds no path this pull request touched.
• The skill support file skills/reviewing-claude-config/reference/validate-ai-scope.md was substantially rewritten. It falls in no change-detection bucket of its own, so it was read directly in the configuration and security pass for instruction content and dangerous guidance rather than being left uncovered.
• Prompt-injection check (CWE-1427): negative. Nothing in the changed files addresses or attempts to direct a reviewing agent. The imperative text in the command and reference files is the /validate-ai runtime's own operating prose, and this changeset extends the existing data-not-instructions boundary to cover pull request thread bodies at reference/validate-ai-scope.md:276-279.
• Findings are limited to lines this changeset introduced or worsened. Pre-existing issues on untouched lines were deliberately not reported.

[github-actions]

🤖 Bitwarden Claude Code Review

Overall Assessment: APPROVE

Documentation-and-prompt change to claude-config-validator 2.2.0: /validate-ai posts CRITICAL and IMPORTANT findings as inline comments, the sticky comment shrinks to a verdict plus one line per finding, and the checks table and critical/major/minor mapping are replaced by the ❌/⚠️/🎨 labels bitwarden-code-review uses. No executable code changed — both command allowed-tools frontmatters are untouched, the new PR THREADS FILE: input is read with the already-granted unscoped Read, no shell-interpolation path is added, and the scope reference keeps the untrusted-data boundary by naming the threads file as data. The version bump is consistent at 2.2.0 across marketplace.json, plugin.json, the README catalog row, and CHANGELOG.md; the plugin has no agents/ directory, so no AGENT.md bump applies. The tool grants this depends on are confirmed present in bitwarden/gh-actions#959 — it registers a github_replies MCP server, allowlists mcp__github_inline_comment__create_inline_comment, adds database_id to get-pull-request-threads, and passes PR THREADS FILE: — so the open questions from the earlier validate-ai comment on this PR resolve once that lands, and every finding degrades to the documented nested form until it does.

Not covered: Skill review did not run — this review path cannot launch plugin-dev:skill-reviewer, so plugins/claude-config-validator/skills/reviewing-claude-config/SKILL.md was not checked for description quality, length, or progressive disclosure. performing-multi-agent-code-review covers them where plugin-dev is installed.

Code Review Details

• ⚠️ : Interactive /validate-ai runs none of the three script checks and the report no longer has a slot to say so, so a sticky comment reading Result: Pass hides three unevaluated deterministic gates
  • plugins/claude-config-validator/commands/validate-ai/validate-ai.md:165-166 (posted inline)
• 🎨 : /validate-ai-local runs the three scripts itself, but the new report shape records only failures (as findings) and skips (on **Not covered:**) — a check that ran and passed has nowhere to appear, while validate-ai-local.md:88 still says to "report them all together"
  • plugins/claude-config-validator/commands/validate-ai-local/README.md:124-127

Verified and not flagged: the write-once rule and the <!-- validation-complete --> terminator survive the report rewrite, so the action's completion check is unaffected; posting inline before the single summary write is ordered deliberately and leaves a killed session with comments plus a failed check rather than a false pass; priority-framework.md is still resolvable from the scope reference; thread matching plus the resolved-thread reply path keep a re-run from reposting; and SKILL.md:177-178 correctly assigns inline-comment posting to the command, which SKILL.md:170-173 forbids the skill from doing.