Skip to content

Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.

docs(store-submissions): correct the registry batch, record the free-lane escape hatch, retire the order

nirholas090bb353cf14
diff --git a/prompts/finish/store-submissions-01-submission-closeout.md b/prompts/finish/store-submissions-01-submission-closeout.md
deleted file mode 100644
index 8b0e8925c..000000000
--- a/prompts/finish/store-submissions-01-submission-closeout.md
+++ /dev/null
@@ -1,141 +0,0 @@
-# SS-01: Store submission close-out
-
-**How to run this:** paste this whole file into a fresh Claude Code chat opened in
-`/workspaces/three.ws`, or say "execute `prompts/finish/store-submissions-01-submission-closeout.md`".
-It is complete on its own. Also read [`store-submissions-00-README.md`](store-submissions-00-README.md) (the strategy and the store
-policy differences) and `CLAUDE.md`.
-
-Every numbered work order in this pack shipped and was retired. What remains is close-out: a few
-real code gaps, a stale tracker, and a list of steps only a human can click. This work order
-owns all three.
-
-## Binding operating clause
-
-1. Finish 100% of what does not need a human account action. Never end with a question about
- scope, design or priority.
-2. The genuinely human steps (portal submissions, `mcp-publisher login`, directory account
- claims, funding a wallet) get batched into ONE closing list with the exact command or URL for
- each. Do not stall on them and do not do them yourself.
-3. CLAUDE.md hard rules: no mocks, no TODO comments, no em-dash or en-dash characters. Stage
- explicit paths only. The OpenAI surface must stay coin-clean: zero token, wallet or x402
- strings anywhere a reviewer can reach.
-
-## Step 0: refresh the truth (the tracker is the deliverable, not a note)
-
-`_generated/TRACKER.md` is the live state of every submission target and it was last verified
-2026-07-15. Re-verify every row today, by running things, not by reading them:
-
-```bash
-# every hosted remote answers tools/list the way a registry crawler asks
-for u in mcp mcp-3d mcp-studio mcp-agent mcp-bazaar ibm-mcp pump-fun-mcp; do
- echo "== $u"; curl -s -X POST "https://three.ws/api/$u" -H 'accept: application/json' \
- -H 'content-type: application/json' \
- -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | head -c 200; echo; done
-
-# the free lane a reviewer will actually exercise, end to end
-curl -s -X POST https://three.ws/api/mcp-studio -H 'content-type: application/json' \
- -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"forge_free","arguments":{"prompt":"a small ceramic teapot"}}}'
-
-curl -sI "https://three.ws/legal/privacy" | head -3
-curl -sI "https://three.ws/support" | head -3
-npm run audit:mcp && npm run audit:mcp-golden && npm run audit:mcp-catalog && npm run audit:mcp-safety
-ls scripts/ | grep -i mcp # the live generator set; see the note below
-git diff --stat prompts/store-submissions/_generated/
-```
-
-**Generator note, verified 2026-08-01:** the two scripts the tracker's older entries name
-(`build-mcp-listing-source.mjs`, `build-mcp-directory-docs.mjs`) no longer exist in the tree and
-nothing else writes `_generated/mcp-listing-source.json` any more, so that file and
-`_generated/mcp-registry-republish.sh` are now hand-maintained artifacts. The live tooling is
-`npm run build:mcp-catalog` (writes `public/mcp-catalog.json`, gated by `audit:mcp-catalog`),
-`npm run audit:mcp` (manifests), `npm run audit:mcp-golden` (contract snapshot) and
-`npm run audit:mcp-safety` (annotations derived from each handler's AST). Reconcile the
-`_generated/` artifacts against the real `server*.json` manifests plus `public/mcp-catalog.json`
-by hand, and say in the tracker that they are hand-maintained, so the next agent does not go
-looking for a generator that was retired.
-
-Rewrite `_generated/TRACKER.md` from what you measured. Every row gets a status, today's date,
-and the command that produced it. Any claim you cannot reproduce is downgraded, not carried
-forward.
-
-## Task 1: close the real code gaps
-
-1. **`@x402/mcp` auto-pay rejects `mcp://` tool URLs.** This is the interop bug that keeps the
- agent-commerce demo from settling through the official client. Reproduce it, fix it on our
- side if the defect is ours, and if it is upstream, open a minimal reproduction and file it
- per the ecosystem rule in CLAUDE.md (contribute upstream rather than working around).
-2. **Claude submission package** (`_generated/claude-submission.md`) is the one prerequisite the
- tracker still lists as not started. Regenerate it against the current live surface: privacy
- URL, allowed links, server metadata, the exact form-fill content, and the compliance
- acknowledgments. Every string in it must match what production serves today.
-3. **Devnet mint E2E** for the tokenized-3D lane: run `node scripts/tokenize-3d-devnet-e2e.mjs`.
- If the faucet refuses (documented daily per-IP limit), say so with the exact error and leave
- the pre-funded-payer command (`E2E_PAYER_SECRET=...`) in the human list.
-4. **Reviewer path**: confirm every paid tool still returns a clean `PaymentRequired` unpaid and
- that the review-mode entitlement in `mcp-server/src/payments.js` still works. A reviewer
- hitting a hang is how this pack lost a month once already.
-
-## Task 2: keep the surfaces consistent
-
-- `_generated/mcp-listing-source.json` is the canonical listing metadata and is hand-maintained
- (see the generator note above). Reconcile it against the real `server*.json` manifests, the
- live `tools/list` counts you captured in Step 0, and `public/mcp-catalog.json`. Commit only if
- something actually changed.
-- `_generated/mcp-registry-republish.sh` holds the staged `mcp-publisher` commands. Check each
- entry against the live registry (`https://registry.modelcontextprotocol.io/?q=io.github.nirholas`)
- so the batch is accurate, and leave it unrun.
-- Compliance grep across manifests, directory docs and the canonical source: zero references to
- any crypto project other than `$THREE`, and zero coin surface anywhere in the OpenAI lane.
-- `STRUCTURE.md` and `docs/mcp.md` server counts must match the canonical source. Fix drift in
- the same change.
-
-## Definition of done
-
-- [ ] `_generated/TRACKER.md` rewritten from measurements taken today, every row carrying its
- command and result.
-- [ ] The interop bug is fixed or filed upstream with a minimal reproduction linked in the
- tracker.
-- [ ] The Claude submission package regenerated and matching production string for string.
-- [ ] Devnet mint either completed with a signature or blocked with the exact error and the
- unblock command.
-- [ ] All four MCP audits green; generator output committed only if it changed.
-- [ ] Compliance grep clean; server counts consistent across `STRUCTURE.md`, `docs/mcp.md` and
- the canonical source.
-- [ ] `npm run build:pages` green; `npm run audit:docs` clean;
- `npm run check:rules -- --paths <files you touched>` clean.
-- [ ] One closing list of human steps, each with its exact command or URL.
-
-## Never blocked (pre-answered)
-
-| Blocker | Do this |
-|---|---|
-| A portal needs an account or a role | That is a human step. Put it in the closing list with the URL and what to click. Never stall the session on it. |
-| `mcp-publisher` needs a login | Same. The commands are already staged; leave them unrun and say so. |
-| A hosted endpoint hangs | Check `api/_lib/http.js` `readBody` handling first; a body already drained by the server's parsers is the historical cause of exactly this symptom. |
-| The devnet faucet refuses | Documented daily per-IP limit. Record it and move on with a pre-funded payer note. |
-| A tool's annotations look wrong | `npm run audit:mcp-safety` derives them from the handler AST. Trust it over the name and fix the handler or the annotation, not the audit. |
-| An `npx` package test fails inside the workspace | In-workspace `npx` resolves to hoisted `node_modules` and breaks ESM resolution. Re-run outside the repo before reporting a package defect. |
-
-## Report format
-
-The rewritten tracker's diff, what you fixed, what you could not reproduce, and the single
-closing list of human steps. No recap of this file.
-
-## Retire this prompt when it is done (required)
-
-1. Verify every Definition of done line against actual command output in front
- of you. Never claim a line you did not verify.
-2. Record the outcome in this campaign's PROGRESS or INDEX file if it has one.
-3. Commit with explicit paths and a subject that describes the diff (house
- style: type(scope): what changed and why a reader cares), and delete this
- prompt file in that same commit:
-
- git rm prompts/finish/store-submissions-01-submission-closeout.md
-
- A finished order left on disk reads as open work to the next agent, so the
- shrinking directory is the campaign's progress ledger.
-
-If a line genuinely cannot pass inside this session (an external party must
-respond, or an owner-gated action is the final step), finish everything else,
-leave this file in place, and state exactly which line remains and who owns it.
-Never delete this file on a partial.
diff --git a/prompts/store-submissions/_generated/TRACKER.md b/prompts/store-submissions/_generated/TRACKER.md
index f3c2d9fd3..646c45678 100644
--- a/prompts/store-submissions/_generated/TRACKER.md
+++ b/prompts/store-submissions/_generated/TRACKER.md
@@ -1,19 +1,58 @@
# three.ws Store Submission Tracker
-**Last verified: 2026-09-03.** Every row below was re-measured today by running the command in its
-own "Verified by" cell against production, npm, or the live MCP registry. Nothing was carried
-forward on the strength of an earlier note. Claims that could not be reproduced were downgraded,
-and two of them turned out to be wrong in our favour (see "Corrections" at the end).
+**Rewritten 2026-09-03 from measurements taken that day.** Every row below carries the command
+that produced it and what that command actually returned. Nothing was carried forward on trust:
+where a prior claim could not be reproduced today it is downgraded here, not repeated.
+
+Production at the time of measurement: `GET https://three.ws/api/version` returned commit
+`19906ce52` on Cloud Run revision `three-ws-api-00411-m6h`.
Legend: **not-started** | **blocked** | **ready-to-submit** | **submitted** | **live**
---
-## How this pack is generated (read before touching `_generated/`)
+## The one finding that gates almost everything
+
+**Vertex AI is denied project-wide on a billing hold, and it is degrading the generation lanes a
+reviewer will exercise.** Measured 2026-09-03:
+
+```
+gcloud logging read 'resource.type="cloud_run_revision"
+ resource.labels.service_name="three-ws-api" (textPayload:"forge" OR textPayload:"vertex")'
+ --freshness=2h
+```
+
+returns, repeatedly and within the last hour:
+
+- `vertex imagen failed, falling back: Lightning dunning decision is deny for project: projects/93741856042`
+- `[forge-reference-image] vertex lane failed, falling through: Vertex reference image returned 403`
+- `[llm] chain exhausted: groq=http 429 | openrouter=http 404 | openrouter#2=http 401 | openrouter#3..5=http 404 | nvidia=unreachable (timeout) | ovh=http 429 | vertex-gemini=http 403 | pollinations=http 402 | openai=http 429`
+- `[forge] free NVIDIA NIM lane unavailable: NVIDIA returned 504`
+
+Three consequences, all reproduced live today rather than inferred:
+
+1. **The OpenAI B3 blocker is not clearable in code.** The scorer fix IS deployed (`db9fd6191`,
+ an ancestor of the live commit) and it works: today's free-lane teapot scored
+ `flatness 0.153, thinAxis y, planar: true, reasons:["planar"]`, and
+ `shouldEscalateToVisionQA` returned `true` for it. The mesh still shipped, because vision QA
+ fails open: rung 1 (Vertex Gemini) is 403 on the billing hold and rung 2 (the free NVIDIA NIM
+ VLM) returned 504, so `runQualityGate` answers `qa_available:false` and the router correctly
+ ships what it has. The degenerate slab reaches the caller with `score 0.976 / flag ok`.
+2. **The free lane is slow, not broken.** `forge_free` completed in **147s** with a real
+ 3,674,712-byte GLB. That is the NIM 504 plus fallback latency, not the 2026-07-08 hang.
+3. **The OpenRouter rungs are separately misconfigured**, independent of billing: five model IDs
+ answer 404 and one answers 401. Restoring those would give the LLM chain a working rung even
+ while Vertex is held. Filed here as a finding; it is a generation-lane fix, not a submission
+ task.
+
+Owner action, not a code change: clear the GCP billing hold on project `93741856042`.
+
+---
+
+## Which generator owns which artifact
-An earlier revision of the closing work order stated that the listing generators had been retired
-and that `mcp-listing-source.json` and `mcp-registry-republish.sh` were hand-maintained. **That was
-wrong, and acting on it would have hand-edited generated files.** The real state, verified today:
+The closing work order asserted that these generators had been retired and that the artifacts were
+hand-maintained. They are not, and editing them by hand would be overwritten on the next run:
| Artifact | Generated by | Gate |
|---|---|---|
@@ -21,216 +60,209 @@ wrong, and acting on it would have hand-edited generated files.** The real state
| `_generated/mcp-directories/*.md` | `node prompts/store-submissions/_generated/build-mcp-directory-docs.mjs` | none, run it after any manifest change |
| `_generated/mcp-registry-republish.sh` | `node prompts/store-submissions/_generated/build-registry-republish.mjs` | none, reads the live registry |
| `public/mcp-catalog.json` | `npm run build:mcp-catalog` | `npm run audit:mcp-catalog` |
+| `tests/fixtures/mcp-golden-tools.json` | `node scripts/audit-mcp-golden.mjs --update` | `npm run audit:mcp-golden` |
-A stale duplicate of the listing generator was sitting in `_generated/`. It produced 49 servers and
-six missing overlay entries where the authoritative `scripts/` copy produces a clean 50, and it is
-what made the work order believe the generator set had been retired. It was deleted on 2026-09-03.
-There are **five** MCP gates, not four: `audit:mcp`, `audit:mcp-golden`, `audit:mcp-catalog`,
-`audit:mcp-safety`, and `audit:mcp-listing`.
+There are **six** MCP gates, not four: `audit:mcp`, `audit:mcp-golden`, `audit:mcp-catalog`,
+`audit:mcp-safety`, `audit:mcp-listing` and `audit:mcp-reviewer`.
---
## Submission status
-| Target | Status | Blocking items | Verified by (2026-09-03) |
+| Target | Status | Verified how, 2026-09-03 | Blocking items |
|---|---|---|---|
-| **Claude Connectors Directory** | ready-to-submit | Package regenerated today against production, string for string (`_generated/claude-submission.md`). The long-standing free-lane blocker is **cleared**: see "Free lane" below. Two human steps remain, both account actions: the Team/Enterprise org + Owner role for the in-app portal (or use the public form), and the reviewer credential decision in section 6 of the answer sheet. | `curl` sweep of all 7 remotes, 27/27 paid-tool challenge sweep, live `forge_free` + `check_job` round trip |
-| **Claude plugin marketplace** | ready-to-submit | Nothing to build. `[HUMAN: submit via https://platform.claude.com/plugins/submit]` | unchanged since 2026-07-08; no code touched since |
-| **Claude Agent Skills pack** | ready-to-submit | Nothing to build. The live E2E that was blocked on the old hang now passes. | live `forge_free` returned a 2,881,160-byte GLB, `glTF` magic, viewer 200 |
-| **OpenAI App Directory** | held (B3 output quality) | Held on ONE product blocker, B3 in `openai-submission.md` section 0: the free lane intermittently returns a degenerate slab that `scoreGlbQuality` rated `ok`. The scorer fix is landed in `api/_lib/glb-quality.js` and needs a deploy plus one clean re-verification generation. Deploys are owner-gated. | not re-measured today; B3 is a deploy-gated item, unchanged |
-| **OpenAI GPT Store** | live | Published: `https://chatgpt.com/g/g-6a563a3b49a88191abf346245491a444-three-ws-3d-studio`. `[HUMAN: delete the draft duplicate GPT `g-6a5672fbf3f48191b559e482c7fcbf51` in My GPTs]`, which is publicly reachable and renders the same title. | unchanged since 2026-07-14 |
-| **Official MCP Registry** | ready-to-submit | **Far better than previously recorded.** 47 of 50 servers are current in the registry; all 7 hosted remotes are published and current. Only 3 new servers remain: `herald-mcp`, `knock-mcp`, `home-mcp`. `[HUMAN: mcp-publisher login]`, then run the staged batch. | `node scripts/publish-mcp-servers.mjs --dry-run`, and the regenerated `mcp-registry-republish.sh` |
-| **Smithery** | ready-to-submit | `_generated/mcp-directories/smithery.md` regenerated today from 50 servers (was built from 42). | `build-mcp-directory-docs.mjs` |
-| **Glama** | blocked | Doc regenerated; `glama.json` committed. `[HUMAN: sign in to Glama with the `nirholas` GitHub account to claim the listing]` | `build-mcp-directory-docs.mjs` |
-| **mcp.so** | ready-to-submit | Doc regenerated today from 50 servers. | `build-mcp-directory-docs.mjs` |
-| **PulseMCP** | ready-to-submit | Doc regenerated today from 50 servers. | `build-mcp-directory-docs.mjs` |
-| **LobeHub** | ready-to-submit | Doc regenerated today from 50 servers. | `build-mcp-directory-docs.mjs` |
+| **Claude Connectors Directory** | ready-to-submit (pending the human items) | `POST /api/mcp tools/list` unauthenticated returned **200 with 53 tools in 0.10s**; `initialize` likewise. `GET /.well-known/oauth-protected-resource` 200 with 13 scopes. `/legal/privacy`, `/legal/tos`, `/support`, `/docs/mcp` all 200, and privacy still carries the "MCP Connectors, AI Processing & Payments" section. | The 2026-06-28 blocker that discovery was auth-gated is **resolved and no longer true**: the portal can sync tools with no credential. Remaining: `[HUMAN: Team/Enterprise org + Owner role, or use the public form]` and `[HUMAN: generate a real MCP_REVIEW_SECRET and decide which vendor credentials a reviewer gets]`. Note the 147s free-lane latency above before demoing live. |
+| **Claude plugin marketplace** | ready-to-submit | `.claude-plugin/marketplace.json` carries 4 plugins (`three-ws-core`, `three-ws-developer`, `three-ws-pump-fun`, `three-ws-3d`). The end-to-end CLI install evidence from 2026-07-08 stands in `_generated/claude-plugin-marketplace-e2e.md` and was not re-driven today. | `[HUMAN: submit via https://platform.claude.com/plugins/submit per marketplace/plugins/three-ws-3d/SUBMISSION.md]` |
+| **Claude Agent Skills pack** | ready-to-submit | `node scripts/build-skills-pack.mjs --check` → `skills-pack manifest up to date (42 skills)`; `.agents/skills/` holds 44 entries. | None in code. The live free-lane E2E it wanted is now green but slow (147s, see above). |
+| **OpenAI App Directory** | blocked (B3, infrastructure) | Free studio surface re-verified live: `initialize` 200, `tools/list` **11 tools**, `resources/list` **2 widget resources** (`ui://widget/three-studio-model.html`, `ui://widget/three-studio-persona.html`), `/.well-known/3d-studio-openapi.yaml` and `/.well-known/ai-plugin.json` both 200, and a real keyless `forge_free` returned a 3.67 MB GLB whose `/viewer?src=` link returned 200. | **B3 is now understood and is NOT a code gap.** The scorer flags the slab correctly; vision QA cannot run because Vertex is 403 on the billing hold and the NIM fallback 504s. Clearing the hold, then one clean re-verification generation, unblocks the submission. `[HUMAN: clear the GCP billing hold]` then `[HUMAN: final submit in the partner portal]`. |
+| **OpenAI GPT Store** | live, with a duplicate to remove | `https://chatgpt.com/g/g-6a563a3b49a88191abf346245491a444-three-ws-3d-studio` returns 200 logged-out. The old draft `g-6a5672fbf3f48191b559e482c7fcbf51` **301s to a 200 page titled "ChatGPT - three.ws 3D Studio"**, so the 2026-09-02 escalation stands: two public GPTs share one name. | `[HUMAN: delete the draft duplicate in My GPTs]` (no API for it) |
+| **Official MCP Registry** | ready-to-submit | `node prompts/store-submissions/_generated/build-registry-republish.mjs` against the live registry, after the pagination fix below: **3 servers need a publish (0 stale, 3 new: `herald-mcp`, `knock-mcp`, `home-mcp`), 47 are already current**, including all 7 hosted remotes. Cross-checked with `node scripts/publish-mcp-servers.mjs --dry-run`, which agrees. The batch is staged in `_generated/mcp-registry-republish.sh` and was NOT run. | `[HUMAN: mcp-publisher login github, then review and run the staged batch]` |
+| **Smithery** | ready-to-submit | `_generated/mcp-directories/smithery.md` regenerated today from the corrected canonical source; it now covers **50 servers (7 remote + 43 stdio)**. | `[HUMAN: create/claim the Smithery account]` |
+| **Glama** | ready-to-submit | `_generated/mcp-directories/glama.md` regenerated today; `glama.json` and `public/.well-known/glama.json` are committed. Glama auto-ingests `io.github.nirholas/*` once the registry batch runs. | `[HUMAN: sign in to Glama as `nirholas` to claim the listing]` |
+| **mcp.so** | ready-to-submit | `_generated/mcp-directories/mcp-so.md` regenerated today. | `[HUMAN: create the account]` |
+| **PulseMCP** | ready-to-submit | `_generated/mcp-directories/pulsemcp.md` regenerated today. | `[HUMAN: create the account]` |
+| **LobeHub** | ready-to-submit | `_generated/mcp-directories/lobehub.md` regenerated today; `public/lobehub/plugin.json` current. | `[HUMAN: create the account]` |
---
-## Live surface, measured 2026-09-03
-
-**All 7 hosted remotes answer `tools/list` unauthenticated, in under a second**, which is the shape a
-registry crawler sends:
+## Server and tool inventory, measured today
-| Endpoint | Tools |
-|---|---|
-| `/api/mcp` | 53 |
-| `/api/mcp-3d` | 35 |
-| `/api/mcp-studio` | 11 |
-| `/api/mcp-agent` | 6 |
-| `/api/mcp-bazaar` | 5 |
-| `/api/ibm-mcp` | 6 |
-| `/api/pump-fun-mcp` | 21 |
-
-```
-for u in mcp mcp-3d mcp-studio mcp-agent mcp-bazaar ibm-mcp pump-fun-mcp; do
- curl -s -X POST "https://three.ws/api/$u" -H 'accept: application/json' \
- -H 'content-type: application/json' -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'; done
-```
-
-**Pages:** `https://three.ws/legal/privacy` 200, `https://three.ws/legal/tos` 200,
-`https://three.ws/support` 200, `https://three.ws/docs/mcp` 200,
-`https://three.ws/three-ws-mcp-icon.svg` 200,
-`https://three.ws/.well-known/oauth-protected-resource` 200.
+| Surface | Command | Result |
+|---|---|---|
+| Hosted remotes | `POST https://three.ws/api/<path> tools/list` (plain `Accept: application/json`, the shape a registry crawler sends) | All 7 answered **HTTP 200 in under 0.12s**: `/api/mcp` 53, `/api/mcp-3d` 35, `/api/mcp-studio` 11, `/api/mcp-agent` 6, `/api/mcp-bazaar` 5, `/api/ibm-mcp` 6, `/api/pump-fun-mcp` 21. Zero hangs. |
+| Manifests | `npm run audit:mcp` | **50 MCP manifests consistent** (was 49; `robinhood/hood-mcp` was outside the audit's discovery and is now inside it). |
+| Tool contracts | `npm run audit:mcp-golden` | **307** hosted tool contracts match the golden fixture. |
+| Public catalog | `npm run audit:mcp-catalog` | catalog matches source: **307** tools across 40 servers. |
+| Safety annotations | `npm run audit:mcp-safety` | **307** annotations verified against handler behaviour across **248** files, **6** reviewed exemptions. |
+| Canonical listing source | `npm run audit:mcp-listing` | **50 servers (7 remote, 43 stdio)**, up to date. |
+| Reviewer path | `npm run audit:mcp-reviewer` | **20 of 20 paid stdio tools returned a clean `PaymentRequired`**; 4 free tools ran; the review entitlement ran the real handler with no charge, and a mismatched secret did not lift the paywall. |
+| Reviewer path, hosted remotes | unpaid `tools/call` against every paid tool on `/api/mcp`, `/api/mcp-3d` and `/api/ibm-mcp` | **27 of 27 returned a clean x402 `PaymentRequired`** (`x402Version: 2`, four `accepts` entries), no hang and no credential material in any body. This is the remote surface; the row above is the stdio package. |
+| Docs | `npm run audit:docs` | clean (1511 markdown files). |
-**Paid-tool reviewer path: 27 of 27 clean.** Every paid tool on the hosted remotes was called unpaid
-and returned a structured x402 `PaymentRequired` (`x402Version: 2`, four `accepts` entries), with no
-hang and no credential material in any response body.
+---
-**Review-mode entitlement: verified live over a real stdio MCP handshake**, all four cases, against
-`mcp-server/src/index.js` calling the paid `ens_sns_resolve`:
+## What changed in the tree today
+
+- **The `mcp://` interop bug is closed, and it was not what the note said.** The official
+ `@x402/mcp` auto-pay client handles `mcp://tool/<name>` resource URLs fine: driven against the
+ published stdio server it returned `paymentMade: true` and the server answered
+ `transaction_simulation_failed`, the correct result for a deliberately unfunded ephemeral buyer.
+ What WAS broken on our side is that the bazaar discovery extension declared
+ `transport: "stdio"`, a value the spec does not allow (`VALID_MCP_TRANSPORTS` is
+ `streamable-http | sse`), so `validateDiscoveryExtensionSpec` rejected every row and an indexer
+ would drop it. The field is optional and a stdio server has no valid value for it, so it is now
+ omitted in `mcp-server/src/payments.js`, `packages/ibm-x402-mcp/src/payments.js` and
+ `examples/paid-mcp-server/src/payments.js`. Re-driven live: `bazaar spec {"valid":true}`,
+ `bazaar schema {"valid":true}`. The client half of that claim is now pinned by a second
+ re-runnable check, `node scripts/x402-mcp-url-interop-check.mjs`, which stands an in-memory MCP
+ server up with upstream's own `createPaymentWrapper` and drives it with a real signer: it exits 0
+ today, and a URL rejection would make it throw before any signing with `paymentMade` false.
+- **One genuine upstream defect remains, with a runnable reproduction.**
+ `node scripts/x402-bazaar-mcp-url-repro.mjs` shows that `@x402/mcp`'s own
+ `createToolResourceUrl` mints `mcp://tool/<name>`, and `@x402/extensions`' bazaar
+ `extractDiscoveryInfo` canonicalizes it to `"null/<name>"` because `mcp:` is not a WHATWG
+ special scheme and `URL.origin` is the string `"null"`. Two packages from one monorepo disagree
+ about a URL neither treats as invalid; every MCP row an indexer keys on it collides on `null`.
+ The script exits 1 while the bug is live and 0 once it is fixed. Filing it is a human step (it
+ posts to an external tracker).
+- **The canonical listing source has a generator again.** `scripts/build-mcp-listing-source.mjs`
+ reads every factual field from the manifests and keeps only curated copy in its OVERLAY;
+ `npm run audit:mcp-listing` (wired into `npm run gate`) fails when the committed JSON drifts. It
+ found the file six servers short and thirty-five versions behind. The untracked leftover copy at
+ `_generated/build-mcp-listing-source.mjs` was deleted, so one generator owns the file.
+- **`robinhood/hood-mcp` is now audited and publishable.** Adding it to the manifest audit
+ surfaced two real defects it had been hiding behind the audit's discovery gap: a missing
+ `mcpName` in its `package.json` and its absence from `scripts/publish-mcp-servers.mjs`. Both
+ fixed; it also gained the `title` every sibling manifest carries. Note its local manifest is
+ `0.1.0` while the registry already serves `0.1.1`, so this repo's copy lags the published
+ package by a patch.
+- **The staged registry batch was overstating the work by 16 publishes, and the cause was
+ pagination.** `build-registry-republish.mjs` read a single `limit=100` page of
+ `registry.modelcontextprotocol.io`. That page returns 100 VERSION rows covering only ~44 distinct
+ server names and hands back a `nextCursor` of `provenance-mcp:0.1.0`, so every server
+ alphabetically at or after that cursor was silently classified NEW. That is why the batch listed
+ all 7 hosted remotes as needing a publish when the registry already served them. The loader now
+ follows the cursor to exhaustion (160 version rows across 72 names), and the batch went from 20
+ commands to a correct **3**. The embedded verification snippet had the same one-page flaw and now
+ points at `node scripts/publish-mcp-servers.mjs --dry-run`, which queries per server.
+- **The free lane is slow but never hangs, and it has a documented escape hatch.** Six back-to-back
+ `forge_free` calls returned **HTTP 200 six times out of six**: one inline at 143s with a real
+ 2,881,160-byte GLB (`glTF` magic, `/viewer?src=` 200), and five that crossed the server's cutoff
+ and correctly returned a structured `pending` result carrying a `jobId`. `STUDIO_FORGE_TIMEOUT_MS`
+ is unset in production, so the 180,000 ms default in `api/_mcp-studio/forge-client.js` applies.
+ Verified end to end: a 205s call's pending handle passed to the **`check_job`** tool returned the
+ finished GLB. Section 6 of the submission package now warns that a client with a 60s default
+ timeout (the `@x402/mcp` default) should raise it or expect the pending path. Backend choice
+ explains the spread: over 24h `forge_creations` shows `hunyuan3d` averaging 94s to done and
+ `trellis_selfhost` 235s.
+- **`grade_sim_readiness` was failing `audit:mcp-safety`**, declaring `readOnlyHint: true` over a
+ `db-write`. The write is `putGrade()`, a read-through cache fill wrapped in try/catch that returns
+ null on failure, so the caller receives the same grade either way. Recorded as a reviewed
+ category-1 exemption in `scripts/audit-mcp-safety.mjs` rather than weakening the annotation or the
+ gate.
+- **`packages/home-mcp/server.json` carried a 131-character description** against the registry's
+ 100-character maximum, which fails `audit:mcp` and would be rejected at publish. Shortened to 93.
+ Its five tools also joined the golden fixture and the public catalog, which is why the tool count
+ moved from 306 to 307.
+- **The submission package's allowed-link list was wrong in both directions.**
+ `https://three-ws-public.r2.dev` is referenced by no handler and returns 500, so it was dropped;
+ generated assets are served from `https://three.ws/cdn/...`. `https://solscan.io` is the explorer
+ the agent registry and persona wallet tools actually emit (58 references) and was **missing**,
+ which would have produced confirm prompts for a reviewer. Added, with `https://explorer.solana.com`.
+- **Two shipped tool titles carried an em-dash**, which the house rules ban in UI copy: `persona_say`
+ and `persona_identity`. Corrected at source in `api/_mcp3d/tools/studio.js`,
+ `api/_mcp-studio/persona-tools.js` and `api/_mcp3d/tools/persona-identity.js`. Production serves
+ the old punctuation until the next deploy, and the submission package says so explicitly.
+- **`npm run audit:mcp-reviewer` is new** (`scripts/audit-mcp-reviewer-path.mjs`). The connector
+ reviewer evidence used to be hand-run curl transcripts that went stale the moment a tool landed;
+ it is one command now. It immediately found that the "16/16 paid tools" figure was four tools
+ out of date.
-| Case | Result |
-|---|---|
-| server secret set, client presents nothing | `PaymentRequired` |
-| server secret set, client presents a wrong secret | `PaymentRequired` |
-| server secret set, client presents the matching secret | real result (live ENS resolution) |
-| server secret unset, client presents a secret | `PaymentRequired` |
+---
-Review mode is off by default and cannot be enabled from the client side alone, exactly as
-`mcp-server/src/payments.js` documents.
+## Prerequisite prompt status
-**All five MCP gates green** after the fixes recorded below:
-`npm run audit:mcp && npm run audit:mcp-golden && npm run audit:mcp-catalog && npm run audit:mcp-safety && npm run audit:mcp-listing`
+| Prompt | Status | Evidence |
+|---|---|---|
+| 01 Tool annotation and title audit | complete, superseded | `_generated/tool-inventory.md` + `remote-tools-list.json` are older than the live surface; §4 of `claude-submission.md` supersedes them and was pulled live 2026-09-03. |
+| 02 Claude reviewer access guide | complete, refreshed today | `_generated/claude-reviewer-guide.md` and `claude-tool-call-evidence.md` updated to the measured 25-tool surface (5 free + 20 paid), with the audit command that reproduces it. |
+| 03 Claude submission package | complete | `_generated/claude-submission.md` regenerated 2026-09-03 against production; independently re-measured here and every number matches (53 and 35 tools, the OAuth metadata, and the 200s). |
+| 04 OpenAI free 3D endpoint | complete | 11 tools live, 2 widget resources, keyless generation proven end to end today. |
+| 05 OpenAI Apps SDK component | complete | GLB viewer component and its four states; screenshots in `_generated/openai-screenshots/`. |
+| 06 OpenAI submission package | package complete, held on B3 | `_generated/openai-submission.md` §0. B3's cause is now known (see the top of this file) and is infrastructure, not code. |
+| 07 Embodied live agent avatar | code-complete | Persona tools live on the studio surface; real Postgres and R2 persistence. Screenshots in `_generated/embodiment/` were captured locally and not re-driven today. |
+| 08 Live agent commerce | loop real, settlement never demonstrated | The auto-pay path is proven to work end to end (see the interop note above); what has never run is a settlement with real funds. Needs `X402_BUYER_SOLANA_SECRET_BASE58` funded and `MCP_AGENT_TALK_TOKEN` set in production. |
+| 09 Conversational remixable 3D | complete; royalty settlement never run | `refine_model` and the lineage core are live. `transferSolanaUSDC` is hardcoded to Solana mainnet with real USDC, so there is no devnet rehearsal for the royalty leg; it is an owner-funding gap, not a code gap. |
+| 10 Claude plugin marketplace | complete | 4 installable plugins. |
+| 11 Agent Skills pack | complete | 42 skills in the pack manifest, `--check` clean. |
+| 12 OpenAI GPT Store Actions | live | Canonical GPT public; Actions lane verified 2026-07-14. Duplicate draft still public (see the table). |
+| 13 MCP registries and directories | complete, republish staged | 50 manifests consistent; 20 publishes staged and unrun. |
+| 16 Tokenized 3D NFT | code-complete, live mint blocked | `node scripts/tokenize-3d-devnet-e2e.mjs` today: 8 airdrop attempts, all `429 Too Many Requests: "You've either reached your airdrop limit today or the airdrop faucet has run dry"`. Unblock with a pre-funded payer: `E2E_PAYER_SECRET=<base58 devnet secret> node scripts/tokenize-3d-devnet-e2e.mjs`. |
+| 17 Embodied on-chain identity | complete; live USDC demo funding-gated | Tools wired into `api/_mcp3d/catalog.js`; a live settle needs a funded wallet. |
+| 18 Token-gated 3D embeds | complete | `create_gated_embed` plus the server-verified balance gate. |
+| 19 Verifiable AI-3D provenance | complete; live anchor funding-gated | `verify_provenance` free, `anchor_provenance` paid; a real anchor needs a funded issuer key. |
+| 20 Spatial MCP standard | complete | Validator, spec, reference renderer, docs. |
+| 21 AR-ready exports | complete | `export_ar` plus device-aware launch; `/api/ar` links are emitted by the live free lane (seen in today's `forge_free` response). |
---
-## Free lane: the "hang" is resolved, and here is what it actually is
-
-This is the item that held the Claude listings longest, so it is written out in full.
+## Test-suite state at close-out (2026-09-03)
-Sampled live on 2026-09-03, the same prompt on `POST /api/mcp-studio tools/call forge_free`:
+`npx vitest run`: **28,388 passed, 17 failed across 7 files.** None of the 17 are from this pass, and
+the three tests that directly cover the tools changed here pass 46 of 46
+(`tests/mcp-studio.test.js`, `tests/mcp3d-persona-identity.test.js`, `tests/mcp-tool-result.test.js`).
-| Attempt | Wall clock | Outcome |
+| Failing file | Count | Cause |
|---|---|---|
-| 1 | over 180s | client gave up at its own 180s ceiling, no server error |
-| 2 | 5.4s | 200, real 2,881,160-byte GLB, `glTF` magic, viewer 200 |
-| 3 | 143.3s | 200, real GLB inline |
-| 4 | 205.2s | 200, structured `pending` result carrying a `jobId` |
-| 5 | 202.4s | 200, `pending` + `jobId` |
-| 6 | 220.9s | 200, `pending` + `jobId` |
-| 7 | 201.8s | 200, `pending` + `jobId` |
-| 8 | 191.8s | 200, `pending` + `jobId` |
-
-A back-to-back run of six generations returned **HTTP 200 six times out of six**, with zero hangs
-and zero errors. One completed inline at 143s; the other five crossed the 180s cutoff and correctly
-handed back a pollable handle instead of blocking, which is the behaviour a reviewer under load will
-actually see.
-
-**It is not a hang and not the old `readBody` bug.** `tools/list` answers in under a second on every
-remote, and every generation call returned 200. The variance is real GPU work: `forge_creations`
-shows the lane failing over across backends with different queue depths. Over the last 24 hours,
-`hunyuan3d` averaged 94s to done and `trellis_selfhost` averaged 235s.
-
-**The server never blocks past 180 seconds.** `STUDIO_FORGE_TIMEOUT_MS` is unset in production, so
-the 180,000 ms default in `api/_mcp-studio/forge-client.js` applies; past it, `pendingResult()`
-returns a `jobId` and the caller collects with the **`check_job`** tool. Verified end to end today:
-attempt 4's pending handle was passed to `check_job`, which returned the finished GLB.
-
-**What a reviewer needs to know**, and what section 6 of the submission package now says: if the MCP
-client's request timeout is near 60 seconds (the `@x402/mcp` default), raise it, or expect the
-pending-plus-`check_job` path rather than one blocking call.
+| `tests/audit-guards.test.js` | 2 | Guard-registry drift: `audit:mcp-listing`, `check:macos-icon` and `check:apple-widget` are gate steps or guard-shaped scripts missing from `data/guards.json`. The three commits that introduced them (`1ed0e5c66`, `9527c2ff0`, `58a64ad88`) all landed AFTER the commit carrying this pass's changes, so they cannot be caused by it. |
+| `tests/avatar-studio-save.test.js` | 4 | `hydrateAppearance` defaults. Avatar studio, untouched here. |
+| `tests/server-404-routes.test.js` | 3 | `.toMatch()` receiving an object. Routing, untouched here. |
+| `tests/home-members-ui.test.js` | 3 | Smart-home household panel, untouched here. |
+| `tests/home-privacy.test.js` | 1 | Smart-home retention bounds, untouched here. |
+| `tests/alpha-drip.test.js` | 3 | Suggested-ladder JSON parsing, untouched here. |
+| `tests/no-nul-bytes.test.js` | 1 | A raw NUL byte in `tests/no-duplicate-stylesheet-links.test.js`, another agent's file. |
----
+`tests/api/x402-facilitator-log-durable.test.js` failed on an earlier run of this suite and passes
+now: commit `aa121030a` ("build the safe-column fragments on first use, not at import") fixed the
+import-time `sql` call in `api/_lib/home/confirm.js` mid-session.
-## Code gaps closed in this pass
-
-1. **`@x402/mcp` "rejects `mcp://` tool URLs": DOES NOT REPRODUCE.** This was recorded as an upstream
- interop defect blocking the agent-commerce listing. At `@x402/mcp` / `@x402/core` / `@x402/svm`
- **2.14.0** the full official auto-pay loop runs end to end with an `mcp://tool/<name>` resource
- URL: the client accepts the envelope, selects the requirements, signs, and submits. It is now
- pinned by a re-runnable check, `node scripts/x402-mcp-url-interop-check.mjs`, which builds an
- in-memory MCP server with `createPaymentWrapper` and drives it with a real signer. It exits 0
- today. The only failure in the loop is `transaction_simulation_failed` from the deliberately
- unfunded buyer keypair, which is the expected pass condition and not an interop defect.
- **Nothing was filed upstream, because there is nothing to file.** The remaining blocker for the
- agent-commerce lane is purely a funded wallet, item 4 in the human list.
-2. **`packages/home-mcp/server.json` description was 131 chars** against a registry maximum of 100,
- which fails `audit:mcp` and would be rejected at publish. Shortened to 93.
-3. **The golden contract fixture and `public/mcp-catalog.json` were stale** against the five new
- `home-mcp` tools. Both regenerated.
-4. **`grade_sim_readiness` failed `audit:mcp-safety`**, declaring `readOnlyHint: true` over a
- `db-write`. The write is `putGrade()`, a read-through cache fill wrapped in try/catch that returns
- null on failure, so the caller gets the same grade either way. Added as a reviewed category-1
- exemption in `scripts/audit-mcp-safety.mjs` rather than weakening the annotation or the gate.
-5. **`build-registry-republish.mjs` read only the first page of the registry.** One `limit=100` page
- returns 100 version rows covering ~44 distinct names and a `nextCursor`, so every server
- alphabetically at or after the cursor was silently marked NEW. The staged batch therefore held
- **16 redundant publishes** for servers that were already current, including all 7 hosted remotes.
- Fixed to follow the cursor; the batch went from 20 commands to a correct 3.
-6. **Two shipped tool titles carried an em-dash**, which the house rules ban in UI copy:
- `persona_say` and `persona_identity`. Corrected at source in `api/_mcp3d/tools/studio.js`,
- `api/_mcp-studio/persona-tools.js` and `api/_mcp3d/tools/persona-identity.js`. Production serves
- the old punctuation until the next deploy; the submission package notes this explicitly.
+These are left for the agents that own them, per the operating rule that another agent's red does
+not stop your green. They do not touch the MCP surface this pack submits.
---
-## Corrections to earlier entries in this tracker
-
-- **"Discovery is auth-gated in production."** No longer true. Unauthenticated `tools/list` returns
- 200 and the full catalog on every remote. This removes the warning that the directory portal could
- not sync tools before OAuth.
-- **"41/42 servers present, 30 stale + 1 missing."** Wrong, and an artifact of the pagination bug in
- item 5. The real state is 50 servers, 47 current, 3 new.
-- **"42 servers (7 remote + 35 stdio)."** The canonical source now holds **50** (7 remote + 43
- stdio). `docs/mcp.md` already says 50 and agrees.
-- **`/api/mcp` tool count.** The submission package said 35; live is **53**.
-- **`/api/mcp-3d` now carries two fund-moving tools**, `persona_tip` and `persona_send`, both
- `destructiveHint: true`, capped at $1 per call and $5 per session with confirmation above $0.25.
- This is a material compliance change and is now disclosed in section 0.3 and acknowledgment 3 of
- the submission package. It strengthens the existing recommendation to submit `/api/mcp` first.
-- **Allowed links were wrong in both directions.** `https://three-ws-public.r2.dev` is referenced by
- no handler and returns 500, so it was dropped; generated assets are served from
- `https://three.ws/cdn/...`. `https://solscan.io` is the explorer the agent registry and persona
- wallet tools actually emit (58 references) and was **missing** from the list, which would have
- produced confirm prompts for a reviewer. Added, along with `https://explorer.solana.com`.
+## `[HUMAN: ...]` items, the whole list
+
+1. **Clear the GCP billing hold** on project `93741856042`. It is denying Vertex project-wide, which
+ is what leaves the OpenAI B3 blocker unclearable and what is degrading the free generation lane.
+2. **Claude org and role.** Team or Enterprise plan with Owner or Directory-management role, or use
+ the public form instead of the in-app portal.
+3. **Claude reviewer credentials.** Generate a real `MCP_REVIEW_SECRET` value and decide which
+ vendor credentials (ideally a scoped, capped `REPLICATE_API_TOKEN`) a reviewer gets for the
+ funded path, then paste both into the submission's private reviewer notes. The mechanism is
+ built and re-verified today by `npm run audit:mcp-reviewer`.
+4. **File the upstream x402 issue** at `https://github.com/x402-foundation/x402/issues/new` with
+ the output of `node scripts/x402-bazaar-mcp-url-repro.mjs`. Posting to an external tracker is
+ an owner-gated action.
+5. **Run the registry batch:** `mcp-publisher login github`, then review and run
+ `prompts/store-submissions/_generated/mcp-registry-republish.sh` (**3** publishes: `herald-mcp`, `knock-mcp`, `home-mcp`).
+6. **Directory accounts:** Smithery, Glama (sign in as `nirholas` to claim), mcp.so, PulseMCP,
+ LobeHub.
+7. **Delete the duplicate GPT** `g-6a5672fbf3f48191b559e482c7fcbf51` in ChatGPT's My GPTs; it is
+ publicly reachable under the same name as the canonical listing.
+8. **Fund a devnet wallet** and run `E2E_PAYER_SECRET=<base58 secret> node scripts/tokenize-3d-devnet-e2e.mjs`.
+ The public faucet has rate-limited this host's IP again.
+9. **Fund a Solana wallet and set `MCP_AGENT_TALK_TOKEN`** in production to demonstrate a real,
+ non-bypassed x402 settlement for the agent-hire loop, and a payout wallet for the remix royalty leg.
+10. **Submit, once the above are done:** the Claude Connectors listing, the Claude plugin
+ (`https://platform.claude.com/plugins/submit`), and the OpenAI App Directory listing.
---
-## Known drift, flagged not fixed
-
-**`hood-mcp` is behind its own published artifact.** `robinhood/hood-mcp/server.json` and its
-`package.json` both say `0.1.0`, but npm and the official registry both hold `0.1.1`. Publishing
-from this tree would attempt a downgrade. Someone published 0.1.1 without the bump landing in the
-repo, or the bump was reverted. Reconciling it needs a human who knows which content is really in
-0.1.1, so it is deliberately left alone rather than guessed at.
-
----
+## Asset kit
-## `[HUMAN: ...]` items: the complete list of what only a person can do
-
-Each carries the exact command or URL. Nothing here is blocked on engineering.
-
-1. **Claude org and role.** A Team or Enterprise Claude org with the Owner role is required for the
- in-app portal. Without it, submit through the public MCP directory form instead; the field
- content in `_generated/claude-submission.md` is identical either way.
-2. **Claude reviewer credentials.** Section 6 of `_generated/claude-submission.md` needs one of two
- things pasted into the private reviewer-notes box: a reviewer OAuth account flagged for
- operator-funded paid tools, or a small pre-funded Solana USDC test wallet. The mechanism is built
- and verified; only the choice and the value remain.
-3. **stdio reviewer secret.** Generate a value for `MCP_REVIEW_SECRET` and decide which scoped
- vendor credentials to hand a reviewer. Only needed if `@three-ws/mcp-server` is submitted as a
- local connector. The entitlement itself is verified working, four cases, above.
-4. **Fund a Solana wallet** (`X402_BUYER_SOLANA_SECRET_BASE58`) and set `MCP_AGENT_TALK_TOKEN` in
- production, to demonstrate a real non-bypassed x402 settlement for the agent-hire loop. This is
- now the *only* remaining blocker on that lane.
-5. **Devnet SOL for the tokenized-3D mint.** Blocked again today, third day running, by the
- documented per-IP faucet limit. Exact error:
- `429 Too Many Requests: "You've either reached your airdrop limit today or the airdrop faucet has
- run dry."` after 8 airdrop attempts. Unblock with a pre-funded payer:
- `E2E_PAYER_SECRET=<base58 devnet secret> node scripts/tokenize-3d-devnet-e2e.mjs`
-6. **`mcp-publisher login`**, then run the staged batch. It is regenerated, correct, and unrun:
- `bash prompts/store-submissions/_generated/mcp-registry-republish.sh`
- It publishes exactly 3 new servers: `herald-mcp`, `knock-mcp`, `home-mcp`.
-7. **Directory account claims.** Glama needs a sign-in with the `nirholas` GitHub account to claim
- the listing: `https://glama.ai`. Smithery, mcp.so, PulseMCP and LobeHub each need an account if
- one does not exist; the per-directory copy is ready in `_generated/mcp-directories/`.
-8. **Delete the duplicate GPT.** `g-6a5672fbf3f48191b559e482c7fcbf51` is publicly reachable and
- renders the same title as the canonical listing. Remove it in ChatGPT under My GPTs; there is no
- API for this.
-9. **Claude plugin marketplace submission.** `https://platform.claude.com/plugins/submit`, per
- `marketplace/plugins/three-ws-3d/SUBMISSION.md`.
-10. **OpenAI App Directory final submit**, in the partner portal, once B3 is cleared by a deploy.
-11. **Deploy** to pick up the two em-dash title corrections and the OpenAI B3 scorer fix. Deploys are
- owner-gated: `npm run deploy:gcp:full`.
+Every asset from prompt 14 is present in `_generated/assets/`: `icon.svg`, `icon-512x512.png`,
+`icon-256x256.png`, `icon-128x128.png`, seven screenshots (`viewer`, `create`, `studio`,
+`validation`, `discover`, `landing`, `embodiment-hero`), and `og-image.png`. What is outstanding is
+not a missing file but a re-capture: `openai-submission.md` §4 holds the listing screenshots until a
+generation clean enough to photograph exists, which is the same B3 blocker as above.