Stop relaying sign-in codes; complete Codex listing metadata

[ottomarilves]

Why

We're preparing the Codex plugin (codex/display-dev/) for OpenAI's Plugins Directory, which covers ChatGPT and Codex. OpenAI scans bundled skills automatically, and its plugin guidelines forbid plugins from soliciting "MFA/OTP codes" (Plugin guidelines → Privacy → Restricted data). Our skill told the agent to ask the user for the six-digit email code and pass it to dsp login --code. The directory submission also requires listing fields our Codex manifest lacks.

Spec (display-dev repo): engineering/specs/feat-openai-plugin-directory-listing.md, §10.

What changes

Sign-in (canonical, every mount)

• With a bundled remote MCP server, sign-in uses that connection: the user connects in the host.
• Otherwise the user runs dsp login --client-source display-dev-skill@0.7.4 in their own terminal. The CLI asks for the email and the code there, and SSO organizations open the browser.
• The agent never runs dsp login or ./scripts/login.sh itself, and never asks for, reads, or relays the code. The trust-boundary bullet, the copy section, and the publish.sh post-publish hint now say the same.
• The CLI path only works when the user's terminal shares the agent's machine and OS user. Hosted sandboxes use the host's MCP connection, or publish anonymously and hand over the claim URL.
• tests/run.sh gains a "sign-in codes stay with the user" check. It fails on --code, read and provide, provide the six-digit or login.sh --email in SKILL.md or scripts, and requires both prohibition sentences. Every forbidden pattern matches the 0.7.4 files, so the check would have failed on the old text.
• README sign-in paragraph and CONTRIBUTING audit guidance updated, including a third clean-room prompt with the dsp test double on PATH.

Codex manifest (OpenAI listing fields)

• shortDescription "Publish and share agent docs" (28 chars; the limit is 30 and the old text was 44).
• New capabilities, composerIcon, supportURL, privacyPolicyURL and termsOfServiceURL fields, plus a host-neutral longDescription.
• assets/logo.svg moves to a 64×64 viewBox with the same artwork; OpenAI requires square icons of at least 48×48.

Decision: canonical, not per-mount

We first considered changing only the Codex mount. That breaks this repo's rule (your 2026-07-22 hardening) that the canonical tree is placeholder-free and every mount is byte-identical, so the change applies to every mount. Trade-off: Claude Code and other CLI users now run dsp login themselves instead of pasting the code into chat. That happens once per machine, since the session is long-lived.

Not in this PR

• Release commit (version bump across manifests and attribution literals), per the repo's release convention.
• Removing http_headers from .mcp.json. Only the directory ZIP build strips it, so the GitHub Codex plugin keeps its attribution.
• Listing name ("Display") and the final listing icon. The SVG uses <text> with system-ui, so its rendering depends on OpenAI's rasterizer; swap in the brand badge (paths or PNG) before upload.
• scripts/login.sh is unchanged. SKILL.md no longer references it, but the argv-forwarding tests still use it.

Verification

• bin/sync-mounts.sh --check: 4 mounts match canonical.
• bash tests/run.sh: exit 0, 10 PASS groups.
• node --test tests/archive-candidate.test.mjs: pass.
• jq empty on all manifests and configs. Manifest fields checked against OpenAI's submission limits; support, privacy and terms URLs return 200.
• Not run: the CONTRIBUTING clean-room agent audit. Codex CLI hit its usage limit (resets 2026-10-15), and the Claude CLI can't authenticate in an isolated HOME without copying credentials. Please run the three sign-in prompts before the release commit.

Review

Three independent Fable review rounds. Rounds 1–2 found remaining relay paths (the publish.sh hint, the copy section, a runnable bash fence before the prohibition, the email solicitation) and missing sandbox guidance; all are fixed. The final gate found no blocker, high or medium issues, and its two low wording fixes are applied. The display-dev repo's docs page (website/content/docs/publishing/skill.mdx) and spec are updated separately.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}