Codex plugin manifest is Claude-Code-shaped — MCP won't register (PUBLISH BLOCKER)

[seanchiuai]

Codex plugin manifest is Claude-Code-shaped — MCP won't register (PUBLISH BLOCKER)

Found by code-review-lead during #10 validation (validated against the real OpenAI Codex plugin spec via context7 + full-page fetch, not the self-attestation). Verdict: issues/10#issuecomment-4987419442.

The blocker

.codex-plugin/plugin.json is shaped like a Claude Code plugin, not a Codex plugin. It declares an inline "mcp": {servers:…} block, but Codex requires "mcpServers": "./.mcp.json" (a path pointer) and no .mcp.json exists. Net effect: the MCP server never registers → the plugin is inert (tools + Response Widget never load).

Required fix (Codex format)

• Add .mcp.json — server minds, streamable-http, url https://getminds.ai/mcp, bearer token via the .mcp.json mechanism.
• plugin.json: set "mcpServers": "./.mcp.json", DROP the inline "mcp" block.
• Move the top-level "config" secret block into .mcp.json (not a Codex top-level field).
• commands/ + "commands": Codex has no commands concept → those Claude-Code files are dead. Remove, OR keep only if the Claude-Desktop target needs them (see design question).
• Add a Codex interface object (composerIcon / logo / privacyPolicyURL); iconUrl is not a Codex field.
• Trim description to ≤200 chars (currently 206).
• Empirical proof: real codex plugin install smoke test → MCP registers + tools list. Needs Codex installed = Sean-gated capture.

Cleared clean (no change): marketplace source:url+ref:main valid for a root-located plugin; homepage/repo resolve; all 4 SKILL.md valid + no hardcoded keys; repo-wide grep clean.

⚠️ OPEN DESIGN QUESTION (dual-client) — gated on Sean

The plugin must work in BOTH the ChatGPT/Codex desktop app AND Claude Desktop. The current inline-mcp shape IS the Claude-Code shape. Does dropping it break the Claude path? Determine whether one manifest can dual-target, or whether Codex vs Claude need distinct packaging. Resolve before restructuring.

⚠️ DECONFLICT (before ANY edit) — gated on Sean

This is a prime candidate for WHY the ChatGPT/Codex app rejected the plugin. The chatgpt-submission team (separate worktree) is investigating that rejection and may already be editing plugin.json. Confirm ownership before two teams touch the same manifest.

Constraints

Hard gates: no minds-ai-co/webapp edits; no publish; no release tag. Fix lands as a DRAFT PR (stays draft; only Sean un-drafts). Owner TBD pending deconfliction (default: frontend-lead as plugin owner; code-review-lead re-verifies — keep author≠reviewer).

[seanchiuai]

#12 read-only finding — does restructuring plugin.json regress the Claude path?

Bottom line: NO regression. The "dual-client tension" collapses — Claude Desktop never read this repo's plugin.json to begin with.

Claude Desktop installs Minds via a mechanism entirely independent of this repo

A remote streamable-HTTP MCP server like https://getminds.ai/mcp is added to the consumer Claude Desktop app through:

• Custom Connectors — Settings → Connectors → Add custom connector → enter the remote URL. (src: https://modelcontextprotocol.io/docs/develop/connect-remote-servers)
• .mcpb Desktop Extensions — double-click a .mcpb bundle (ZIP w/ its own manifest.json); replaced the older .dxt. (src: https://www.anthropic.com/engineering/desktop-extensions ; https://support.claude.com/en/articles/12922929-building-desktop-extensions-with-mcpb)

Note: adding a remote server directly to claude_desktop_config.json via a url field is NOT the supported path (some versions wipe the mcpServers section). Not relevant to us — we don't ship that file either.

It does NOT read .codex-plugin/plugin.json

.codex-plugin/ is an OpenAI Codex format. Claude Code (the CLI) uses a differently-named .claude-plugin/plugin.json; Claude Desktop (consumer) uses Connectors + .mcpb. No Anthropic product reads .codex-plugin/plugin.json. (src: https://code.claude.com/docs/en/plugins-reference — documents .claude-plugin/, no .codex-plugin/)

Primary evidence (repo file-independence — strongest, doc-independent)

This repo ships exactly two manifest artifacts: .codex-plugin/plugin.json + marketplace.json. Confirmed via git ls-files + untracked scan: no .claude-plugin/, no .mcpb, no manifest.json, no .mcp.json, no claude_desktop_config.json anywhere. So whatever mechanism any Claude client uses, it sources nothing from this repo. The docs above are corroboration; this fact alone settles it.

Repo intent confirms the split: CONTEXT.md frames Claude Desktop and the ChatGPT/Codex app as two hosts that both call the same hosted MCP endpoint; one-pager.md says the Claude side is a separate "Claude Code skill, live in the skill marketplace" — a distinct artifact, not derived from this plugin.json.

Corollary — the inline mcp block serves no one today

The inline "mcp":{servers:…} block in this Codex plugin.json is read by nobody as-is: Codex wants mcpServers→.mcp.json (that's blocker #12), and Claude Desktop uses Connectors/.mcpb. So it is NOT holding up a Claude path. Restructuring it for Codex → zero Claude regression.

One nuance to surface (not a blocker, not a regression)

The Codex fix ADDS a repo-root .mcp.json. The Claude Code CLI / Code tab reads a project-root .mcp.json when run inside a project dir. (src: https://code.claude.com/docs/en/mcp) So if someone git-clones this repo and runs Claude Code inside it, the new .mcp.json could be picked up. That's a new potential pickup, not a regression of an existing path (no Claude path runs through this repo today), and the Claude distribution is a separate skill-marketplace artifact anyway. Flagging so nobody's surprised.

Remaining empirical gaps (carried, per orchestrator)

• (a) Codex unknown-field behavior (reject vs ignore) → the coexistence smoke test — needs Codex installed = Sean-gated. Still the one open empirical question for the Codex side.
• (b) Claude side here is confirmed from current cited docs, not a live install.

Method: context7 (modelcontextprotocol.io) + claude-code-guide agent (WebFetch of Anthropic/MCP docs) + primary repo evidence. Docs-only; no files edited, no branch, no .mcp.json created.

[seanchiuai]

ORCHESTRATOR STATE CHECKPOINT (2026-07-16, pre-compaction rehydration anchor)

Resolved scope (Sean, dsp-20260716-002): this workspace = desktop-app experience + plugin publish-readiness for the native Response Widget; NO webapp edits; #3 closed; #4/#7 HELD.

Mission state across issues:

• Render-verification: current Response Widget renders in both desktop clients #11 render-verify — DONE. Widget is PRODUCTION-ready on getminds.ai/mcp (bridge-adaptive, vendor-inlined; live-verified). Every remaining render gate is host/connector CONFIG or upstream, NOT our code. Matrix in Render-verification: current Response Widget renders in both desktop clients #11 comment. Per host: ChatGPT/Codex desktop = renders once on Codex ≥26.616.51431 + apps flags (Sean-gated screenshot); Claude desktop = blocked on upstream ext-apps #671 + 2 minds-SERVER-side gaps (webapp/#2427, not us); chatgpt.com web = connector currently OFF (Sean re-enable). 3 webapp/backend report-items flagged for the #2427 team.
• Manifest + marketplace.json validation for publish-readiness #10 manifest validation — DONE. Verdict NOT PUBLISHABLE → this blocker (Codex plugin manifest is Claude-Code-shaped — MCP won't register (PUBLISH BLOCKER) #12).
• Submission packet → submit-ready against compliance-checklist.md #9 submission packet — IN FLIGHT (frontend-lead, dsp-058). Scoped to Codex plugin manifest is Claude-Code-shaped — MCP won't register (PUBLISH BLOCKER) #12-independent parts (honest per-host matrix framing, PDFs, stale-checklist fixes); DEFERRING any 'submit-ready' claim + manifest-shape text until Codex plugin manifest is Claude-Code-shaped — MCP won't register (PUBLISH BLOCKER) #12 lands.
• Codex plugin manifest is Claude-Code-shaped — MCP won't register (PUBLISH BLOCKER) #12 (this) — DESIGN RESOLVED, awaiting Sean. Fix = clean Codex-conformant restructure (add .mcp.json + mcpServers pointer + interface object; drop inline mcp + non-schema config/commands; trim description ≤200). code-review-lead PROVED zero Claude-regression risk (repo ships only .codex-plugin/; Claude installs via Connectors/.mcpb — separate path; inline mcp block serves nobody).

OPEN — awaiting Sean (2 decisions):

1. Codex plugin manifest is Claude-Code-shaped — MCP won't register (PUBLISH BLOCKER) #12 fix ownership/deconfliction — is the manifest fix OURS (route to frontend-lead, code-review-lead re-verifies) or is the parallel chatgpt-submission team already editing plugin.json? (Collision risk — NOT dispatched pending this.)
2. Route the 3 webapp render-items to the #2427/backend team? (orchestrator offered.)

Remaining empirical gap (Sean-gated): codex plugin install smoke test = proof the #12 fix registers the MCP; needs Codex installed on Sean's machine.

Leads: frontend-lead → #9 in progress. code-review-lead → idle, standing by to re-verify #9 PR + #12 fix.

[seanchiuai]

CHECKPOINT UPDATE (2026-07-16) — decisions resolved, #12 fix dispatched.

Sean's decisions (via watchtower dsp-072):

1. Codex plugin manifest is Claude-Code-shaped — MCP won't register (PUBLISH BLOCKER) #12 is OURS — manifest fix is our codex-plugin packaging lane (not chatgpt-submission's). → DISPATCHED to frontend-lead (dsp-076): clean Codex-conformant restructure per the Required-fix checklist above; draft PR into widgets; code-review-lead re-verifies; codex-plugin-install smoke test is Sean-gated proof.
2. Standing rule: any render-item needing a webapp change → route to webapp side (#2427 / webapp MCP owner) as a flagged hand-off; no reaching into webapp.

Open items for the (post-compaction) orchestrator to pick up:

• ROUTE to #2427/webapp: the 3 render items from Render-verification: current Response Widget renders in both desktop clients #11 that need webapp changes — (a) server omits claudemcpcontent.com domain-hash _meta; (b) widget doesn't emit ui/notifications/initialized (both = Claude-desktop mount blockers); (c) Iconify runtime icon-fetch may be CSP-blocked (cosmetic). Flagged hand-off, don't fix here.
• QUEUE to frontend-lead (non-blocking, behind submit gate): 2 small Submission packet → submit-ready against compliance-checklist.md #9 fixes on PR docs(#9): submission packet — honest per-host status, re-attested checklist, PDFs, CONTRIBUTING #13 — (i) security-privacy.md self-contradiction (claims 'no third-party origin in sandbox' but discloses runtime icon-API fetch) → reword; (ii) checklist '✅ name matches repo slug' is false → reword line (don't rename).
• PR docs(#9): submission packet — honest per-host status, re-attested checklist, PDFs, CONTRIBUTING #13 (Submission packet → submit-ready against compliance-checklist.md #9): reviewed-clean as DRAFT; stays draft (only Sean un-drafts); NOT submit-ready (behind Codex plugin manifest is Claude-Code-shaped — MCP won't register (PUBLISH BLOCKER) #12). one-pager.pdf provisional (regen after Codex plugin manifest is Claude-Code-shaped — MCP won't register (PUBLISH BLOCKER) #12).
• Framing item resolved: Render-verification: current Response Widget renders in both desktop clients #11 native-widget framing supersedes Submission packet → submit-ready against compliance-checklist.md #9's stale Path-B/Claude-only criterion (Path B Path B: ship one skill-driven Claude artifact scorecard widget (plugin-only) #3 closed); no BD overclaim.

Leads: frontend-lead → building #12 fix. code-review-lead → idle, standing by to re-verify #12 + do the 2 #9 cosmetic fixes' review.

[seanchiuai]

🚦 Publish-readiness gap checklist — go/no-go for Sean

Compiled by code-review-lead, 2026-07-16. This comment is the rehydration anchor for "what stands between us and a marketplace-publishable Minds Codex plugin." Every item is tagged OWNER · GATE · STATE.

STATE legend: ✅ DONE · 🔴 blocked-on-Sean · 🟠 blocked-on-other-team/upstream · 🟡 in-flight (being worked now)

---

⏱️ TWO GATES — Sean makes two different go/no-go calls, at two different times

The common mistake is bundling these. Keep them separate:

GATE A — marketplace-publishable (the plugin installs, registers, functions)

The only gate that touches what Codex + the marketplace actually read (plugin.json + .mcp.json + skills/ + marketplace.json). In order:

1. 🔴 Sean runs the codex plugin install smoke test — the REAL go/no-go. PR fix(#12): restructure manifest to Codex format so the MCP server registers #14 restructured the manifest to Codex format (reviewed-clean), but "the MCP registers" is proven only to the docs' limit — the remote-HTTP .mcp.json shape is inferred, and the plugin was inert until now so nothing was ever exercised live. PASS = an authenticated tool call returns a real Minds result (@minds ask_mind … / query_panel …), not -32001. (Registration-only / tools/list is a false pass — it's public on getminds.ai/mcp.) If 4b -32001 while 4a works → MINDS_API_KEY isn't injecting → escalate the ON_INSTALL env-bind question. Until this passes, publish is NO-GO.
2. 🔴 Merge + un-draft PR fix(#12): restructure manifest to Codex format so the MCP server registers #14 (manifest) → promote widgets → main. fix(#12): restructure manifest to Codex format so the MCP server registers #14 is reviewed-clean but DRAFT (only Sean un-drafts). Because the marketplace pins ref: main, the published artifact lives on main — so the up-merge into main is a code-review-lead-gated step: I sign off the promotion.
3. 🔴 Cut v0.1.0 tag + flip marketplace.json ref main → v0.1.0. Pin a stable release, not a moving branch. After 1–2.

GATE B — submission/pitch-ready (can actually send to OpenAI BD)

None of these affect whether the plugin installs — they're pitch collateral (same category as screenshots/demo). Gate B can trail Gate A:

• 🟡→🔴 PR docs(#9): submission packet — honest per-host status, re-attested checklist, PDFs, CONTRIBUTING #13 (packet) + its 2 must-fixes → merge + un-draft (§2)
• 🔴 3 screenshots + 60s demo — render-gated (§5)
• 🟠→🟡 8-tool reconcile so skills reference only advertised tools (§4)
• 🔴 Sean end-to-end review + send pitch email (§5)

---

1 · MANIFEST (plugin core function) — #12 / PR #14 @ deb70d0

Item	Owner · Gate · State
Restructure to Codex format (mcpServers→.mcp.json, drop inline mcp/config/commands, add interface, description 206→192)	frontend-lead / code-review-lead verified · ✅ DONE (reviewed-clean, issues/#14#issuecomment-4987858550)
codex plugin install smoke test — authenticated tool call passes	Sean · needs Codex installed + key · 🔴 blocked-on-Sean ← critical-path #1
Merge PR #14 + un-draft	Sean · after smoke test · 🔴 blocked-on-Sean

2 · SUBMISSION PACKET — #9 / PR #13 @ e48f031

Item	Owner · Gate · State
Re-attested checklist, PDFs + generator, CONTRIBUTING, honest per-host matrix	frontend-lead / code-review-lead verified · ✅ DONE (reviewed-clean, pull/#13#issuecomment-4987626354)
Must-fix before submit: (a) widget-framing reconciliation — assumed #11's per-host matrix supersedes #9's older "Claude-artifact/Path-B, no Codex equivalent" blurb (per the adopted #11 acceptance bar), BUT this was flagged an explicit Sean call in the #13 verdict: if Path-B framing is still current, the matrix is a material overclaim → Sean to confirm whether the one-pager keeps any Path-B framing; (b) reword security-privacy.md self-contradiction ("no external loads" vs disclosed Iconify runtime fetch)	frontend-lead → code-review-lead re-verify; (a) needs Sean · 🟡 in-flight / 🔴 (a) Sean-confirm (dsp-140)
New (from #14): commands were permanently deleted → the /minds-panel /minds-create slash-command pitch in one-pager.md/pitch-email.md/demo-script.md is now wrong-forever (not "pending #12") → reframe to skills	frontend-lead · 🟡 in-flight
Cosmetic: reword the residual false ✅ name matches repo slug (Codex name is a free identifier); regenerate provisional one-pager.pdf after #12 + Sean fills XXXX traction	frontend-lead · carry-forward · 🟡 in-flight
Merge PR #13 + un-draft	Sean · after the two must-fixes re-verify · 🔴 blocked-on-Sean

3 · RENDER — rich Response Widget per host — #11

Not a hard publish-blocker for the plugin (skills work as text tools without widget render), but it gates the rich-widget value prop + the screenshots/demo below.

Host / item	Owner · Gate · State
Widget bundle production-ready — bridge-adaptive + vendor-inlined, live in prod, tier-2 5/5	webapp (done) · ✅ DONE (our-side blocker resolved & shipped)
chatgpt.com web — re-enable the Minds MCP connector in Sean's session → live-confirm + screenshot (no regression found; likely-yes)	Sean · config-tier · 🔴 blocked-on-Sean
Codex Desktop — update to ≥ 26.616.51431 + config.toml [features] apps=true + enable_mcp_apps=true; upstream openai/codex #21019 OPEN; CSP-gap caveat	Sean (capture) + OpenAI (upstream GA) · 🔴 blocked-on-Sean / 🟠 upstream
Claude Desktop — 2 minds-side gaps (domain-hash _meta.ui.domain, emit ui/notifications/initialized) + upstream ext-apps #671 OPEN	webapp #2427 (2 gaps) + Anthropic (#671) · 🟠 blocked-on-other-team
Iconify runtime icon-fetch (cosmetic degrade in sandboxed iframe)	webapp #2427 · 🟠 blocked-on-other-team

4 · MCP TOOL SURFACE — chatgpt-submission lane

Item	Owner · Gate · State
Curate advertised MCP to the lean 8-tool core (server currently exposes ~36)	chatgpt-submission team · 🟠 blocked-on-other-team
Reconcile: the 4 skills reference 10 distinct MCP tools (query_panel, get_panel_results, get_panel_status, list_panels, chat_with_mind, list_groups, create_group_from_brief, list_minds, get_mind_status, create_mind). If curation drops any to non-advertised/unavailable, those skills call missing tools. Verify whether curation removes server-side availability or only the advertised list. (Per Sean, curation does not affect the Response Widget — but it can affect skills.)	chatgpt-submission (final list) → frontend-lead (skill fix) · 🟠 → 🟡

5 · SEAN-GATED FINALIZERS (pre-send)

Item	Owner · State
3 reference screenshots (install flow · panel query · results synthesis) — render-gated on §3	Sean · 🔴
60-second demo recording (URL into pitch-email.md) — render-gated	Sean · 🔴
v0.1.0 git tag + marketplace.json ref flip	Sean · 🔴 ← critical-path #3
Sean end-to-end review of plugin contents	Sean · 🔴
Send pitch email (PDFs attached + demo URL) to partnerships@openai.com	Sean · 🔴

✅ Already cleared (no action — recorded so nobody re-opens them)

• marketplace.json source:url + ref:main valid for a root-located plugin; homepage/repo resolve; all 4 SKILL.md valid + no hardcoded keys; repo-wide grep clean (localhost/app.getminds.ai/non-HTTPS/keys) — Manifest + marketplace.json validation for publish-readiness #10.
• No "dual-client" regression: restructuring plugin.json for Codex doesn't touch any Claude path (repo ships no Claude artifact; Claude installs via Connectors/.mcpb) — Codex plugin manifest is Claude-Code-shaped — MCP won't register (PUBLISH BLOCKER) #12.
• (Cosmetic, non-blocking) marketplace.json still carries an undocumented interface.iconUrl (favicon.ico) vs plugin.json's new icon.png — align eventually; not a gate.

---

Bottom line

NO-GO to publish today. Gate A (publishable) hinges on one uncertain gate — Sean's authenticated smoke test; until it passes we do not know the plugin functions at all — plus two mechanical steps: merge/un-draft #14 + promote widgets→main (my sign-off), then cut v0.1.0. Gate B (pitch-ready) trails it and is entirely Sean-gated or cross-team: the #9 packet fixes, screenshots/demo, the 8-tool reconcile, and the email. I'll re-verify frontend-lead's #9 fixes (§2) and sign off the widgets→main promotion the moment each is ready.