Skip to content

feat(crosscheck): add the Spec Kit studies and a landing page with a study index - #147

Merged
hnipps merged 4 commits into
mainfrom
claude/crosscheck-study-publication-5yy5i2
Nov 11, 2026
Merged

hnipps merged 4 commits into
mainfrom
claude/crosscheck-study-publication-5yy5i2

Conversation

@hnipps

@hnipps hnipps commented Nov 11, 2026

Copy link
Copy Markdown
Member

Tier: 2
Intent: intent/2026-11-10-spec-kit-studies.md
Spec: intent/2026-11-10-spec-kit-studies.md

Why

On 2026-10-23 Harry filed 28 defect reports against GitHub's Spec Kit (github/spec-kit #5071 to #5098), found with CGV and Crosscheck. The site held one study at its root, so a second study had nowhere to go.

What changed

  • Landing page at /. A study index, three paragraphs on the tool, and the install commands.
  • EVerest study moved to /studies/everest-ocpp/, with its glossary and record. /everest.html, /glossary.html and old fragment links on the root (/#witness, /#q-proof&s=tested) forward there, fragment included. The study page carried its head metadata twice; the second copy is removed.
  • /studies/spec-kit/ lists the 28 reports by area. For each it gives what goes wrong, which line of Spec Kit's own documentation, docstring, test or code it contradicts, and its upstream state on 2026-11-10: 19 with a merged fix, 8 with a fix open, 1 with neither. It uses the EVerest page's questions, strengths and entries.
  • How the defects were found, from the transcript of the 2026-10-22 session:
    • CGV ran first. A triage agent found no real bug among its 39 errors.
    • Eight /reason subagents each searched one slice of the source. A finding counted only when a reproduction showed it.
    • 38 findings were reproduced. 31 became the 28 reports, 6 were left out because an open issue or pull request covered them (listed on the page), and 1 was reported privately to GitHub's security team. The page says nothing about that report beyond its existence.
    • Harry re-ran every reproduction before filing.
  • /studies/spec-kit-expressions/ looks at #5071, #5072 and #5073 on the workflow expression evaluator: the recorded outputs, what Jinja2 gives, and what changed upstream. An agent's reading of main's source is marked Not yet reached.
  • Generated from one record. site/studies/spec-kit/spec-kit-study.json feeds both pages through scripts/site-spec-kit.mjs. check-site.mjs fails if a count the prose repeats drifts from it.
  • Preview cards. Three new cards. render-site-images.mjs gains --cards-only and a PLAYWRIGHT=1 path, because Chrome's command-line screenshot did not finish in a container. The two EVerest cards are unchanged.

Scope

  • Nothing on the Spec Kit pages was rerun for this PR. This session's auto-mode classifier refused to run Spec Kit's code, so the pages report each report's recorded output and say so.
  • Upstream maintainers' triage and merged fixes are reported as their decisions, not as Crosscheck's evidence.
  • Deployment is not part of this PR. See site/LAUNCH.md, updated for the new paths.

Blast radius

Only the static site under site/, its Node build, check and image scripts and their tests, PRODUCT.md, the site journal and the new intent. No plugin, MCP server or CGV code is touched, and no protected surface.

Verification

  • node scripts/build-site.mjs --check, node scripts/check-site.mjs and node --test scripts/*.test.mjs pass (146 tests). Five new site tests cover the Spec Kit record checks, the hand-written counts, rebuild freshness, old-link forwarding and links inside study directories.
  • node scripts/check-public-site.mjs http://127.0.0.1:8765/ passes against a local server.
  • Headless Chromium at 1440px and 390px: no console errors, no horizontal scroll, the question filter works on the Spec Kit page, and old links land on the EVerest study.

Generated by Claude Code

…study index

The root becomes a landing page that lists the studies. The EVerest
study moves to /studies/everest-ocpp/, and /everest.html, /glossary.html
and old root fragment links forward to it.

/studies/spec-kit/ presents the 28 defects reported against GitHub's
Spec Kit (#5071 to #5098): what each contradicts in Spec Kit's own
documentation and its upstream state on 2026-11-10. 19 have a merged
fix, 8 have a fix open and 1 has neither.
/studies/spec-kit-expressions/ looks at the three workflow-expression
reports. Both pages render from spec-kit-study.json. check-site fails
if a count the prose repeats drifts from that record.

The reproductions were not rerun for this change, and the pages say so.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NFQZoTNGDVvV9v3Ycp6Ruc
The method now comes from the transcript of the 2026-10-22 session.
CGV ran first, and an agent triaged its 39 errors as holding no real
bug. Eight /reason subagents then searched one slice each, and a
finding counted only when a reproduction showed it. 31 findings became
the 28 reports. The six findings left unfiled because an open issue or
pull request covered them are now listed.

The pages now attribute the human re-run to the reports, because the
transcript does not show it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NFQZoTNGDVvV9v3Ycp6Ruc
Harry confirmed re-running every reproduction before filing, so the
Spec Kit pages now state it, not attribute it to the reports. The
count now reads 38 reproduced findings: 31 in the 28 reports, 6
already covered upstream, and 1 emailed to GitHub's security team with
no reply. Nothing about that report's subject is published.
check-site fails if those parts stop adding up to 38.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NFQZoTNGDVvV9v3Ycp6Ruc
The Spec Kit page, intent and journal no longer mention whether GitHub's
security team replied.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NFQZoTNGDVvV9v3Ycp6Ruc
@hnipps
hnipps merged commit 49689dc into main Nov 11, 2026
5 checks passed
@hnipps
hnipps deleted the claude/crosscheck-study-publication-5yy5i2 branch November 11, 2026 02:08
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants