[kyle-sexton]

No related issue: part of the upstream-pointer sweep over the Claude Code docs.

Summary

Fixed the drifted Desktop enabling path and artifacts availability statements, replaced restated skill-resolution, plugin-dependency, credential-storage and CLI-scope specifics with anchored pointer records, and added changelog fragments per plugin.

Fix

• docs(visual): point computer-use, visualization and prototype at live artifacts and computer-use sections
• docs(visual): point playwright, playgrounds, animation, pixel-art and miro at live sections
• docs(visual): finish pointer shapes left by the first pass
• docs(visual): drop gate categories and outcome claims left beside pointers
• docs(visual): give added pointers and internal relays their own as-of and trigger
• docs(visual): correct the plugin-root substitution comment and anchor internal relays
• docs(prototype): complete the relay records and the recheck trigger
• docs(miro): address review findings on upstream pointers (miro README disabled-server pointer now targets "Disable a server without removing it"; decision-matrix restores the probed /_runtime/mermaid-11.16.1.min.js path; miro fragment wording corrected)

Verification

• The fresh-context verifier ran two fix rounds and did not converge, so the orchestrator fixed the last remainders directly in 4316f61, and those fixes were not re-verified.
• The changed-skills gate has not been run on this branch yet (it runs serially before ready).
• Eval run owed before ready: learn-style, diagnose, setup (computer-use), setup (miro), use, check (playwright), demo-video, present, visualize
• Review findings (quality-gate pr mode) addressed in ad3760f; not re-reviewed.
• Not in this PR: detect-ecosystems.test.sh:6-8 and :142 in both prototype skills still say an allowed-tools rule leaves ${CLAUDE_PLUGIN_ROOT} literal.

Related

The upstream-pointer sweep.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-12T03:18:05.130027Z	df0d3f6	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[claude]

Claude finished @kyle-sexton's task in 2m 16s —— View job

---

Security review complete

• Gather context (PR docs: point restated upstream docs at live sections in the visual plugins #7271, docs-only changes across visual plugins)
• Invoke /review:security-review per custom instructions
• Post findings

Scope reviewed: full diff against origin/main (30 files, .changes/* fragments + docs/SKILL.md/README.md edits across animation, computer-use, miro, pixel-art, playgrounds, playwright, prototype, visualization). The only non-prose edits are comment-only changes in plugins/prototype/skills/{explore-directions,pressure-test}/scripts/detect-ecosystems.sh (no logic touched).

No security issues found. This PR is documentation-only: it replaces restated upstream specifics (skill resolution, plugin dependencies, credential storage, CLI scope flags) with anchored pointers to live docs. Per the skill's instruction-surface-deletion lens, I checked every narrowed/removed standing instruction against the Gate 0 protected classes (secret-handling, agent-authority, external-publication, etc.) and confirmed the actual guardrails survive, only the restated specifics were trimmed:

• secret-handling — miro's "never read or reveal" instruction for the stored token is unchanged (plugins/miro/skills/setup/SKILL.md); only the Keychain/.credentials.json path description became a pointer. Same for the README's token-rotation guidance.
• agent-authority — computer-use:diagnose's "do not call request_access during a check" is untouched (plugins/computer-use/skills/setup/SKILL.md); the playwright storage-and-auth "agents never run these steps" rule for state-save is kept and actually stated more plainly (plugins/playwright/skills/playwright/reference/storage-and-auth.md); the visualize ClaudeDesign "report the refusal and never retry around it" rule for gated writes is kept verbatim (plugins/visualization/skills/visualize/SKILL.md).
• external-publication — the decision-matrix's dropped "a connector-backed page can never be shared to a public link" claim is a factual correction (flagged in the PR's own Fixed notes as inaccurate), not a removed control: it's replaced with an instruction to read the live sharing/approval conditions and tell the person what an unconnected viewer sees before recommending a page they'll share (plugins/visualization/skills/visualize/context/decision-matrix.md) — net effect is an equal-or-stronger advisory, not a weaker one.

No injection, authz, or trust-boundary surfaces are touched, and the shell-script edits are comments only.

[claude]

Claude finished @kyle-sexton's task in 4m 49s —— View job

---

Review complete

• Gather context (PR body, diff, custom instructions)
• Invoke /review:code-review per custom instructions
• Verify anchored pointer citations against live upstream pages
• Post summary comment

Scope

Applied the code-review lane criteria from REVIEW.md (this repo runs a dedicated security lane — pr-review-security-hosted.yml — so security findings are out of scope here; none were seen in this docs-only diff anyway). No paved-path row in the standards index applies to this change.

What I checked

This PR replaces restated upstream specifics (CLI/Desktop enabling paths, artifact availability, skill/plugin-dependency/credential-storage mechanics) with anchored pointer records across 8 plugins, per .claude/rules/skill-bodies-state-current-rules.md. Given the nature of the change, the highest-value review was verifying the new anchors actually resolve and that the "Fixed" claims are genuinely corrections rather than regressions. I fetched the live pages for every substantially-cited target (mcp, skills, artifacts, plugins/manifest-reference, computer-use, settings-reference, permission-modes, plugins/cli-reference, plugins/dependencies, plugins/install, vision guide) and cross-checked headings/anchors and content against the claims:

• Confirmed correct fixes, not just rewording:
  • plugins/miro/README.md — "Server deduplication" and "Plugin MCP tool names" are not real headings on /docs/en/mcp anymore; the new pointers (#plugin-provided-mcp-servers, #scope-hierarchy-and-precedence, #disable-a-server-without-removing-it) all resolve, and the per-project disabled-server storage claim matches the live section verbatim.
  • plugins/computer-use/README.md / diagnose/SKILL.md / setup/SKILL.md — the live "Differences from the Desktop app" table now reads Settings > This computer > System, not "Settings → General" as the old restated text said. Genuine drift; the fix (pointing live instead of restating) is correct.
  • plugins/playwright/skills/demo-video/SKILL.md — old URL .../plugins-reference#reference-a-saved-value doesn't exist; new .../plugins/manifest-reference#reference-a-saved-value does.
  • plugins/visualization/skills/visualize/context/decision-matrix.md — the live Artifacts "Availability" table's version floor is now v2.1.281 (not the previously-restated v2.1.183), and "Share an artifact" confirms a connector-backed page can be shared publicly (viewers without claude.ai sign-in just lose the live sections). The old "can never be shared to a public link" claim was wrong; removing it was correct.
  • plugins/prototype/.../detect-ecosystems.sh (both skills) — confirmed /docs/en/skills#available-string-substitutions does say ${CLAUDE_PLUGIN_ROOT} substitutes in a plugin skill's allowed-tools Bash rules, matching the new "repo convention, not a platform limit" framing. Also confirmed the cited lines 14-24 of plugins/prototype/scripts/allowed-tools-pairing.test.sh say exactly that.
• All other anchors I sampled (skills#control-who-invokes-a-skill, #add-supporting-files, #resolve-skills-that-share-a-name; artifacts#draft-a-design-canvas, #start-from-a-slides-design-or-docs-template, #pull-live-data-with-mcp-connectors, #control-connector-calls-from-artifacts; manifest-reference#user-configuration, #where-values-are-stored, #limit-a-field-to-fixed-options, #defaultenabled; settings-reference#disablebundledskills, #skilloverrides; permission-modes#how-auto-mode-evaluates-actions) resolve to the sections the pointers claim.
• docs/native-surfaces/records.json exists and holds design/ClaudeDesign/design-sync records the probe citations point at.
• The PR body's own disclosure that detect-ecosystems.test.sh:6-8 and :142 in both prototype skills still assert the old (now-contradicted) "${CLAUDE_PLUGIN_ROOT} stays literal" claim is accurate — but those lines aren't touched by this diff, so under the high-signal bar (no findings on untouched lines) that's not reportable here; it's already tracked as out-of-scope by the author.
• One citation I could not verify: the mermaid-runtime probe in plugins/visualization/.../decision-matrix.md now points at "pull request feat(visualization): absorb code-shape sketch forms into visualize (0.5.0) #3729" instead of the prior binary-read basis — I don't have access to fetch this repo's own PR history from this sandbox, so I can't confirm that PR actually records the 11.16.1 binary-string read. Not flagging as a defect, just noting it's unverified by me.

Findings

No correctness or maintainability findings meeting the high-signal bar (something a careful senior reviewer would block or flag). The changelog fragments in .changes/* accurately describe each plugin's diff, and the anchor/pointer format consistently follows the required Pointer/As-of/Recheck-trigger shape.

[github-actions]

claude-security-review has reviewed this pull request through 5090693; a later push is reviewed from there.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: df0d3f6bef

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[github-actions]

claude-review has reviewed this pull request through 5090693; a later push is reviewed from there.

[claude]

Claude finished @kyle-sexton's task in 34s —— View job

---

Security review complete

• Gather context (incremental review scope: only plugins/visualization/skills/visualize/context/decision-matrix.md changed since df0d3f6)
• Read .claude-lane/incremental.diff
• Invoke /review:security-review per custom instructions
• Post findings

Scope reviewed: incremental diff only, per the lane's custom instructions — this lane last reviewed df0d3f6bef69694ab0455c6c9fd6b6f59d5d5732, and the only file changed since then is plugins/visualization/skills/visualize/context/decision-matrix.md. The diff is a single hunk:

-  - **Pointer**: probe: the binary string read recorded in pull request #3729.
+  - **Pointer**: probe: the binary string read recorded in
+    <https://github.com/melodic-software/claude-code-plugins/pull/3729>.

No security issues found. This is a formatting-only change: it turns the bare text reference "pull request #3729" into a markdown link pointing at the same PR (https://github.com/melodic-software/claude-code-plugins/pull/3729). No information content changes, no standing instruction is deleted, narrowed, or softened, and no injection, authz, secret-handling, or trust-boundary surface is touched. Per the skip gate, this hunk has no security-relevant surface.

Note: this does not change the substance of the earlier Codex finding on this line (the pointer still names a PR rather than a durable in-repo artifact for the binary-string probe) — that's a documentation/evidence-durability concern, not a security one, so it's out of scope for this lane.

[claude]

Claude finished @kyle-sexton's task in 1m 46s —— View job

---

Incremental code review — df0d3f6 → 478d148

• Gather context (incremental diff, custom instructions, PR body/comments)
• Invoke /review:code-review criteria against the one changed file
• Post findings

Scope: per the lane's custom instructions, this pass reviewed only the one file changed since the last review at df0d3f6: plugins/visualization/skills/visualize/context/decision-matrix.md (confirmed via git diff df0d3f6..478d148 --stat — single file, +2/-1). All other files were already reviewed at df0d3f6 and are out of scope here.

Finding (posted inline)

plugins/visualization/skills/visualize/context/decision-matrix.md#L31-L32 — the hunk turns the bare text pull request #3729 into a hyperlink to the same PR, but doesn't address the substance of Codex's P1 comment on the prior commit: a link to a whole pull request still isn't "inspectable probe evidence" someone can use to reproduce the binary string read when the Recheck trigger fires. .claude/rules/skill-bodies-state-current-rules.md governs this file (it's a skill reference under plugins/*/skills/**) and expects a pointer at the exact evidence, not an issue/PR number wrapped in markdown link syntax.

I couldn't fetch PR #3729 from this sandbox (gh isn't in this lane's tool allowlist) to check whether a more specific comment/diff anchor exists there — noted as unverified in the inline comment, same caveat the earlier full review on this PR gave.

No other findings in the changed hunk (markdown link syntax and line-wrap formatting are both correct and consistent with the surrounding bullets).

[claude]

Claude finished @kyle-sexton's task in 1m 22s —— View job

---

Incremental code review — 478d148 → 5090693

• Gather context (incremental diff, custom instructions, PR body/comments)
• Invoke /review:code-review criteria against the one changed file
• Post findings

Scope: per the lane's custom instructions, this pass reviewed only the one file changed since the last review at 478d148: plugins/visualization/skills/visualize/context/decision-matrix.md (confirmed via git diff 478d148..HEAD --stat — single file, +3/-2). All other files were already reviewed at 478d148 and are out of scope here. No paved-path row in the standards index applies to this file.

What changed

-  - **Pointer**: probe: the binary string read recorded in
-    <https://github.com/melodic-software/claude-code-plugins/pull/3729>.
+  - **Pointer**: probe: `grep -a '/_runtime/mermaid-'` over the installed Claude Code
+    binary, recorded in
+    <https://github.com/melodic-software/claude-code-plugins/pull/3729#issuecomment-5546567555>.

Assessment

This directly resolves the open P1 from the earlier full review (Codex's comment and my own inline comment on 478d148): the pointer previously named a whole pull request, which isn't inspectable probe evidence. This hunk fixes both gaps I flagged:

• Names the probe: grep -a '/_runtime/mermaid-' over the installed binary — satisfies .claude/rules/skill-bodies-state-current-rules.md's requirement that a probe observation name its probe.
• Anchors to the exact evidence: the link now targets a specific comment (#issuecomment-5546567555) inside PR feat(visualization): absorb code-shape sketch forms into visualize (0.5.0) #3729, not the PR as a whole — this is the "link the specific comment/diff" fix I suggested as an alternative to an in-repo record.

The As-of and Recheck trigger lines are untouched and still correctly bracket the pointer.

Unverified: I don't have gh/web access in this lane's sandbox to fetch issuecomment-5546567555 and confirm it actually contains the grep -a output for /_runtime/mermaid-11.16.1.min.js — same caveat the earlier full review and my prior inline comment gave for PR #3729 generally. Flagging as unverified, not as a defect; the fix is structurally correct for the finding it addresses.

Findings

None. No correctness or maintainability issues in this hunk — it's a precise, scoped fix for the previously-reported finding, with no other formatting or content changes in the diff.

[claude]

Claude finished @kyle-sexton's task in 54s —— View job

---

Security review complete

• Gather context (incremental review scope: only plugins/visualization/skills/visualize/context/decision-matrix.md changed since 478d148)
• Read .claude-lane/incremental.diff
• Invoke /review:security-review per custom instructions
• Post findings

Scope reviewed: incremental diff only, per the lane's custom instructions — this lane last reviewed 478d14839ffddc91e8428423576f3d40f4de7fdd, and the only file changed since then is plugins/visualization/skills/visualize/context/decision-matrix.md. The diff is a single hunk:

-  - **Pointer**: probe: the binary string read recorded in
-    <https://github.com/melodic-software/claude-code-plugins/pull/3729>.
+  - **Pointer**: probe: `grep -a '/_runtime/mermaid-'` over the installed Claude Code
+    binary, recorded in
+    <https://github.com/melodic-software/claude-code-plugins/pull/3729#issuecomment-5546567555>.

No security issues found. This is a documentation-only change: it names the exact probe command (grep -a '/_runtime/mermaid-') and narrows the pointer from the whole PR to a specific comment anchor (#issuecomment-5546567555). No standing instruction is added, removed, or softened, and no injection, authz, secret-handling, or trust-boundary surface is touched. A literal grep -a invocation documented as prose in a markdown reference file is not executable in this context and poses no command-injection risk.

This also resolves the earlier Codex P1 and code-review lane findings on this line (a bare PR reference wasn't "inspectable probe evidence"): the pointer now names the exact command and a specific comment anchor, not just a whole pull request.