Harness Source Map

Live site: harness.dtmont.com · Claude Code · Codex/ChatGPT · OpenCode · Cursor · Trace

What Claude Code, Codex/ChatGPT, OpenCode and Cursor put in front of the model: prompts, reminders, tool descriptions, settings, flags and environment variables, read from shipped binaries, apps and pinned public source, each record with its provenance. Plus Trace, which opens your own session log in the browser and shows how those pieces actually arrive. One repo, one site: https://harness.dtmont.com.

• / asks which harness you want.
• /claude-code/ is the Claude Code reference, including What wins.
• /codex/ is the Codex/ChatGPT reference.
• /opencode/ is the OpenCode reference, including request assembly and provider routing.
• /cursor/ is the Cursor reference, including shipped desktop and Agent CLI records and source-text archives.
• /trace/ is Trace, which reads real sessions and optional network captures from all four products.

For agents, start at /llms.txt. Each harness has its own /<product>/llms.txt, organized by the existing documentation sections. It links to /<product>/<slug>.md counterparts of the reference pages, generated from the same resolved content as the HTML. The indexes explain provenance and the distinction between a static source map and an observed session. Source-text archives remain accessible through the evidence pages; the entry point does not concatenate the full corpus. The indexes and Markdown pages regenerate with every site build, including watcher builds. The linked Trace operating guide covers computer use, shortcuts (generated from Trace's actual key map), session loading, local sources and network recording.

The two older sites, ccprompts.dtmont.com and gpt6aeon.dtmont.com, redirect here path for path.

Layout

Path	What it is
claude-code/	Claude Code extraction (extract/), its published records (outputs/) and the gitignored work/ (the extracted binary).
codex/	Codex/ChatGPT extraction (extract/codex, extract/codex-config), records (outputs/) and the gitignored work/.
opencode/	Release-pinned public source extraction (extract/), records (outputs/) and the gitignored source checkout (work/).
site/	The one site. src/shared/ holds the site config (site.mjs: domain, repo, sections), the landing page and the Trace build. src/claude-code/, src/codex/ and src/opencode/ are each section's generator. trace/ is the viewer. redirects/ holds the retired-host Worker. test/ holds the tests.
tools/	trace-local.mjs (open local sessions by id), check-links.mjs and leak-check.mjs.
watch/	The release watcher. It is disabled; see its README before re-enabling.
video/	The teaser and explainer video pipelines, as code only.
docs/	Specs, plans and design notes.
private/	Gitignored and never published: frozen sessions, video media, research.

Commands

npm ci && (cd site && npm ci)
npm run build        # site/dist: landing + /claude-code/ + /codex/ + /opencode/ + /trace/
npm test             # extraction tests + site, Trace and redirect tests + tools
npm run check        # build + tests + link check + leak check: the gate before any deploy or push

Re-extract a product from a new release:

• Claude Code: see claude-code/README.md. Runs python3 claude-code/extract/bun-extract.py <claude.exe>, then node claude-code/extract/<area>.mjs.
• Codex/ChatGPT: see codex/README.md. Runs node codex/extract/codex/refresh.mjs and bash codex/extract/codex-config/run_all.sh.
• OpenCode: see opencode/README.md. Extract from a checked-out upstream release with node opencode/extract/extract.mjs.

Record an OpenCode run explicitly:

npm run trace:opencode -- --open -- --model openrouter/deepseek/deepseek-v3.2 "Your real task"

This keeps the native session JSON and credential-redacted HAR in a private local bundle. Drop both files into Trace to inspect received reasoning and the actual outbound requests. Recording stays opt-in and changes trust only for the launched process. The network view shows the observed destination, requested model and reported serving provider; a model publisher's name does not establish geography or retention. See capture setup and usage for requirements and incomplete recordings.

Deploy (by hand)

cd site && npx wrangler deploy                                    # the one site, harness.dtmont.com
cd site/redirects && npx wrangler deploy -c wrangler.ccprompts.jsonc   # ccprompts -> /claude-code/
cd site/redirects && npx wrangler deploy -c wrangler.gpt6aeon.jsonc    # gpt6aeon  -> /codex/

The redirect Workers reuse the retired sites' Worker names, which already own those custom domains. Run npm run check first.