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(backlog): record today's progress across the telegram, x402scan, and agent index work orders

nirholasde79dcb90f28
diff --git a/CLAUDE.md b/CLAUDE.md
index d80e149d5..9217dd88b 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -305,7 +305,7 @@ The whole build + submit + purge in one command: `npm run deploy:gcp:full`.
- Frontend: vanilla JS modules + Vite (`npm run dev`, port 3000).
- 3D: Three.js with glTF/GLB.
- Backend touchpoints: serverless-style handlers in `api/`, workers in `workers/`.
-- **Production runs on Google Cloud Run, NOT Vercel** (migrated 2026-07-07 after Vercel disabled the deployment). One container ([server/index.mjs](server/index.mjs)) serves the static frontend, the vercel.json route table, and all `api/**` handlers; the crons (112 as of 2026-09-02, see vercel.json) run on Cloud Scheduler. `vercel.json` is a LIVE config file: `server/index.mjs` reads its `routes` array on boot (split at the `{handle:"filesystem"}` marker into pre- and post-filesystem phases), and `scripts/create-gcp-scheduler.mjs` reads its `crons` array to sync Cloud Scheduler jobs. The server itself never reads `crons`. Never delete `vercel.json` as a leftover. GCP builds/deploys must pin the `three-ws-build@` (build) and `three-ws@` (runtime) service accounts; the project's default compute SA was deleted. Deploys: see the "Deploy runbook" section above.
+- **Production runs on Google Cloud Run, NOT Vercel** (migrated 2026-07-07 after Vercel disabled the deployment). One container ([server/index.mjs](server/index.mjs)) serves the static frontend, the vercel.json route table, and all `api/**` handlers; the crons (113 as of 2026-09-02, see vercel.json) run on Cloud Scheduler. `vercel.json` is a LIVE config file: `server/index.mjs` reads its `routes` array on boot (split at the `{handle:"filesystem"}` marker into pre- and post-filesystem phases), and `scripts/create-gcp-scheduler.mjs` reads its `crons` array to sync Cloud Scheduler jobs. The server itself never reads `crons`. Never delete `vercel.json` as a leftover. GCP builds/deploys must pin the `three-ws-build@` (build) and `three-ws@` (runtime) service accounts; the project's default compute SA was deleted. Deploys: see the "Deploy runbook" section above.
- **Env-var trap:** `vercel env pull` returns EMPTY for secret-type vars. Never trust a Vercel env export as complete. Production env lives on the Cloud Run service (`gcloud run services describe/update three-ws-api --region us-central1`).
- Solana/agent SDKs in `sdk/`, `solana-agent-sdk/`, `agent-payments-sdk/`.
- Real APIs in use: Pump.fun feed, Solana RPC, OpenAI/Anthropic via worker proxies. Never mock these.
diff --git a/docs/mobile-launch/coinmarketcap-article.md b/docs/mobile-launch/coinmarketcap-article.md
index ceee12c16..22fcf03c2 100644
--- a/docs/mobile-launch/coinmarketcap-article.md
+++ b/docs/mobile-launch/coinmarketcap-article.md
@@ -172,7 +172,7 @@ Solana Mobile built the phone this app was designed for, the dApp Store it ships
NVIDIA: three.ws is a member of NVIDIA Inception. Every 3D generation lane runs on NVIDIA silicon, the free hosted NIM lane behind the Forge runs TRELLIS, and Nemotron models handle the LLM, vision, embedding, reranking, and safety work around it. Membership is a startup programme, not an endorsement.
-OpenAI: three.ws is an OpenAI Select Partner. The free 3D Studio connector gives ChatGPT nine keyless 3D tools with an AR handoff on every generation. We are an independent member of the partner network.
+OpenAI: three.ws is an OpenAI Select Partner. The free 3D Studio connector gives ChatGPT eleven keyless 3D tools with an AR handoff on every generation. We are an independent member of the partner network.
Google Cloud: production runs on Cloud Run, the GPU fleet runs on Cloud Run, Vertex AI provides the Gemini and image lanes, and three.ws is a member of Google Cloud for Web3 Startups, whose grant funds the compute behind every model you make.
diff --git a/docs/mobile-launch/x-article.md b/docs/mobile-launch/x-article.md
index d24a7bb30..5f711eb83 100644
--- a/docs/mobile-launch/x-article.md
+++ b/docs/mobile-launch/x-article.md
@@ -128,7 +128,7 @@ Generation is free on three.ws because serious people backed us early, and every
**NVIDIA.** three.ws is a member of NVIDIA Inception, and every 3D generation lane runs on NVIDIA silicon: text-to-3D and image-to-3D on Cloud Run L4s and Blackwell, the free hosted NIM lane behind the Forge running TRELLIS, and Nemotron models doing the LLM, vision, embedding, reranking, and safety work around it. Membership is a startup programme, not an endorsement, and we say so every time.
-**OpenAI.** three.ws is an OpenAI Select Partner. The free 3D Studio connector gives ChatGPT nine keyless 3D tools, and every generation carries a place-in-your-room AR link. We are an independent member of the partner network; nothing here is an OpenAI product.
+**OpenAI.** three.ws is an OpenAI Select Partner. The free 3D Studio connector gives ChatGPT eleven keyless 3D tools, and every generation carries a place-in-your-room AR link. We are an independent member of the partner network; nothing here is an OpenAI product.
**Google Cloud.** Production runs on Cloud Run, the GPU fleet runs on Cloud Run, Vertex AI provides the Gemini and image lanes, and a Google for Startups Web3 cloud grant pays for the compute behind every model you make.
diff --git a/docs/nvidia-apps-catalog-listing.md b/docs/nvidia-apps-catalog-listing.md
index fbaad1c7e..999dbcebb 100644
--- a/docs/nvidia-apps-catalog-listing.md
+++ b/docs/nvidia-apps-catalog-listing.md
@@ -61,7 +61,7 @@ The portal has no "publish to catalog" button, so the request goes to the progra
> three.ws is the AI-agent layer for the open web: a browser-native platform where anyone can build, embed, and monetize autonomous agents with real 3D bodies. Type a prompt at three.ws/forge and a textured 3D model appears in seconds; upload a photo or a sketch and it becomes a mesh; one more click auto-rigs it with a full skeleton and ARKit-52 blendshapes, ready to talk, walk, and be embedded anywhere with a single script tag. No plugins, no installs, no account required for the free lane.
>
-> The entire generation stack runs on NVIDIA accelerated computing, in two layers. The self-hosted layer is a fleet of NVIDIA L4 GPU workers on Google Cloud Run running PyTorch with custom-compiled CUDA extensions and NVIDIA Kaolin: Hunyuan3D 2.1 and Microsoft TRELLIS for image-to-3D, TripoSG for sketch-to-3D, TripoSR for fast drafts, Make-It-Animatable for auto-rigging, Motion Diffusion Model for text-to-animation, and a video-to-point-cloud world scanner accelerated with FlashInfer. The Blackwell-generation RTX PRO 6000 path is validated with sm_120 kernel builds. The hosted layer is one NVIDIA API key wide: TRELLIS text-to-3D on NVCF, Nemotron LLMs for agent brains, NV-EmbedQA embeddings and Mistral reranking for agent memory, NemoGuard content safety on every public publish, Riva Magpie multilingual TTS and Riva ASR for voice, and Audio2Face-3D driving real-time facial animation in the browser.
+> The entire generation stack runs on NVIDIA accelerated computing, in two layers. The self-hosted layer is a fleet of NVIDIA L4 GPU workers on Google Cloud Run running PyTorch with custom-compiled CUDA extensions and NVIDIA Kaolin: Hunyuan3D 2.1 and Microsoft TRELLIS for image-to-3D, TripoSG for sketch-to-3D, TripoSR for fast drafts, Make-It-Animatable for auto-rigging, and Motion Diffusion Model for text-to-animation. The Blackwell-generation RTX PRO 6000 path is validated with sm_120 kernel builds. The hosted layer is one NVIDIA API key wide: TRELLIS text-to-3D on NVCF, Nemotron LLMs for agent brains, NV-EmbedQA embeddings and Mistral reranking for agent memory, NemoGuard content safety on every public publish, Riva Magpie multilingual TTS and Riva ASR for voice, and Audio2Face-3D driving real-time facial animation in the browser.
>
> Around the bodies sits a full agent economy: on-chain identity (ERC-8004 on EVM, Metaplex Core on Solana), x402 pay-per-call so agents can charge for and pay for services autonomously, an MCP server exposing free text-to-3D to any AI assistant, and 60+ open npm SDKs. three.ws is an NVIDIA Inception member; the engineering field notes on running image-to-3D on L4 and Blackwell are published on the three.ws blog and the NVIDIA Developer Forums.
@@ -77,9 +77,9 @@ Pick the closest available options in the form; the taxonomy varies.
Every item below is live in production unless marked otherwise. Do not claim TensorRT, Triton, or self-hosted NIM microservices; the GPU workers are PyTorch + CUDA today (TensorRT-LLM is an Inception roadmap item).
-Also do not claim the **video-to-scene world scanner** or the **texture worker**, and do not re-add **FlashInfer** on their behalf. Both have a GPU `cloudbuild.yaml` in the tree but neither is deployed in any region (verified 2026-08-17 with the recount command below), so they are built code, not a running lane. Deploy them and they belong here; until then a curator who checks would find nothing behind the claim.
+Also do not claim the **video-to-scene world scanner** or the **texture worker**, and do not re-add **FlashInfer** on their behalf. Both have a GPU `cloudbuild.yaml` in the tree but neither is deployed in any region (verified 2026-08-17 and re-verified 2026-09-02 with the recount command below), so they are built code, not a running lane. The long description above used to carry the world-scanner claim in defiance of this rule; it was removed on 2026-09-02, because a submission is exactly where an unbacked claim costs the most. Deploy them and they belong here; until then a curator who checks would find nothing behind the claim.
-- **NVIDIA L4 GPUs** (Google Cloud Run GPU) running the entire self-hosted 3D generation fleet: image-to-3D, sketch-to-3D, auto-rigging, text-to-motion. Eight GPU workers, twelve deployments across `us-central1` and `us-east4` (eleven L4, one RTX PRO 6000). Recount before submitting: `gcloud run services list --project aerial-vehicle-466722-p5 --format="csv[no-heading](metadata.name,region,spec.template.spec.nodeSelector)" | grep -i accelerator`
+- **NVIDIA L4 GPUs** (Google Cloud Run GPU) running the entire self-hosted 3D generation fleet: image-to-3D, sketch-to-3D, auto-rigging, text-to-motion. Eight GPU workers, twelve deployments across `us-central1` and `us-east4` (eleven L4, one RTX PRO 6000). Recount before submitting: `gcloud run services list --project aerial-vehicle-466722-p5 --format="csv[no-heading](metadata.name,region,spec.template.spec.nodeSelector)" | grep -i accelerator` Re-counted 2026-09-02: unchanged, and every public link in the [Links for the form](#links-for-the-form) section answered 200 the same day.
- **CUDA** with custom-compiled extensions (nvdiffrast, diffoctreerast, diff-gaussian-rasterization, torchmcubes)
- **NVIDIA Kaolin** (TRELLIS worker)
- **NIM APIs** (integrate.api.nvidia.com): Nemotron Super and Nano LLMs, Nemotron Nano VL vision, NV-EmbedQA-E5-v5 embeddings, Mistral 4B reranking, NemoGuard 8B content safety
diff --git a/docs/open-source-ecosystem.md b/docs/open-source-ecosystem.md
index aaf29c1dc..860550fce 100644
--- a/docs/open-source-ecosystem.md
+++ b/docs/open-source-ecosystem.md
@@ -131,7 +131,7 @@ Beyond the official registry: 18 of these are indexed on PulseMCP and 10 on Glam
x402 is the HTTP 402 pay-per-call protocol. three.ws runs on it, and every piece of the implementation is in the repo.
- The facilitator is self-hosted, in api/_lib/x402/. It has processed 110,416 on-chain USDC settlements and 803,483 verifications. Nothing routes through a third party to settle on Solana.
-- The discovery catalog at three.ws/.well-known/x402.json lists 4,519 priced endpoints, every one on Solana mainnet. It is also indexed by the Coinbase CDP Bazaar, by x402scan (where we have an upstream pull request), and by 402index.
+- The discovery catalog at three.ws/.well-known/x402.json lists 4,519 priced endpoints, every one on Solana mainnet. x402scan indexes it and, since our facilitator registration merged upstream on 2026-08-11, counts our Solana settlements against it: 18,636 transactions and $1,055 of USDC volume as of 2026-09-02. 402index carries the endpoints too. The Coinbase CDP Bazaar does not: it indexes a service only after that service settles a payment through the CDP facilitator on Base, and every settlement here runs on our own Solana rail.
- The Receipt Vault at three.ws/receipts holds 58,907 signed Offer and Receipt artifacts, retrievable forever.
- The client side is published: x402-fetch for code, the two modals for any web page, x402-mcp and mcp-bridge for agents, and the VS Code extension (on the VS Code Marketplace and Open VSX) for people.
- The merchant side is published too: x402-server turns any endpoint into a paid one, and agentcore-payments-mcp gives an agent a budgeted session.
diff --git a/docs/open-source-footprint.md b/docs/open-source-footprint.md
index 1d2a3047d..0d11f902b 100644
--- a/docs/open-source-footprint.md
+++ b/docs/open-source-footprint.md
@@ -57,7 +57,7 @@ Not published anywhere yet (do not claim): PyPI, crates.io, Docker Hub, Smithery
|---|---|---|
| Discovery catalog | https://three.ws/.well-known/x402.json | 4,519 priced endpoints, all on Solana mainnet |
| Self-hosted facilitator | in the repo, `api/_lib/x402/` | 110,416 on-chain USDC settlements, 803,483 verifications |
-| x402scan | https://www.x402scan.com/server/17cbd874-52ac-4920-a020-b22ff2489a07 | server page, plus an upstream [PR](https://github.com/Merit-Systems/x402scan/pull/1032) |
+| x402scan | https://www.x402scan.com/server/17cbd874-52ac-4920-a020-b22ff2489a07 | server page (60 registered resources), plus [our facilitator](https://www.x402scan.com/facilitator/three-ws) since the upstream [PR](https://github.com/Merit-Systems/x402scan/pull/1032) merged 2026-08-11 |
| Coinbase CDP Bazaar | https://api.cdp.coinbase.com/platform/v2/x402/discovery/resources | settlement-indexed |
| 402index.io | registered via `scripts/x402-register-directories.mjs` | 19 endpoints, domain claimed |
| Receipts | https://three.ws/receipts | 58,907 signed Offer and Receipt artifacts |
diff --git a/prompts/finish/backlog-00-INDEX.md b/prompts/finish/backlog-00-INDEX.md
index 67485c0de..e25cfb0de 100644
--- a/prompts/finish/backlog-00-INDEX.md
+++ b/prompts/finish/backlog-00-INDEX.md
@@ -44,9 +44,9 @@ observable health, so 01 to 03 come first.
| 05 | [R2 bucket CORS: verify, then fix at the origin](backlog-05-r2-bucket-cors.md) | one credential | mint an R2 admin token |
| 07 | [BNB testnet: deploy the two finished contracts](backlog-07-bnb-testnet-deploys.md) | one funded EOA | send tBNB to `0x1C4918894dfA5eE11cfF9629B458b5169Cfa3871` (faucet is reCAPTCHA-gated) |
| 08 | [OKX chat bot: move off the codespace](backlog-08-okx-chat-bot-always-on.md) | nothing | one email OTP login |
-| 09 | [Telegram bots: durable hosting for both feeds](backlog-09-telegram-bots-durability.md) | nothing | none |
-| 10 | [x402scan listing: finish the last three steps](backlog-10-x402scan-listing.md) | one PAT or one comment | classic PAT, or the owner comments |
-| 11 | [Agent index: bring the on-chain crawl back from `down`](backlog-11-agent-index-lag.md) | nothing | none |
+| 09 | [Telegram bots: durable hosting for both feeds](backlog-09-telegram-bots-durability.md) | **done**, verified live 2026-09-02 | clear the commit gate on its file update |
+| 10 | [x402scan listing: finish the last three steps](backlog-10-x402scan-listing.md) | the deploy (the facilitator listing itself is live) | approve the deploy, then one wallet signature to re-register the origin |
+| 11 | [Agent index: bring the on-chain crawl back from `down`](backlog-11-agent-index-lag.md) | one production deploy (all code, tests, docs and the migration landed 2026-09-02) | approve the deploy; the commands are in [PROGRESS.md](backlog-PROGRESS.md) |
---
@@ -71,11 +71,19 @@ token. 07 needs one faucet claim and nothing else: a deployer key now exists in
balance 0 tBNB), both dry runs simulate green against the live chain-97 RPC, and the
`--broadcast` path is proven end to end on a local chain-97 node. Fund that address, not
the retired 2026-08-02 one.
-08's worker is built and committed but has never reported a heartbeat, so it is not running
-anywhere. 09's deliverables live in a sibling repository that is not checked out in this
-workspace, so nothing here can verify them. 10's remaining external step resolved on its own
-(the upstream pull request merged 2026-08-11 and the live discovery endpoint matches it);
-its file is retired as soon as the owner clears the commit gate that its content sits behind.
+08's worker is built and committed; on 2026-09-02 it beat for the first time, from this
+codespace, so `/api/healthz` now carries the `okx_chat_bot` subsystem and reports it as the
+stopgap it is (`hostDurable=false`). The Cloud Run host is still undeployed.
+09 is done: the sibling repository is checked out at `/workspaces/pump-fun-sdk` after all, and
+both feeds were verified running on Cloud Run on 2026-09-02 (`Ready=True`, websocket transport,
+25 h uptime each) while this codespace's own rebuild killed their local processes.
+10's remaining external step resolved on its own
+(the upstream pull request merged 2026-08-11, the registry attributes 18,636 settlements and
+$1,055 of volume to our facilitator, and their own crawler replayed against production returns
+4,519 stable items with no duplicates). What was left was on our side: their registration flow
+reads `/openapi.json`, which hand-enumerated 24 of the 75 live paid services, so 52 endpoints
+answered a valid 402 and could not be listed. That is fixed in the tree and ships with the next
+deploy, after which one wallet signature re-registers the origin.
## Shared rules (every work order obeys these)
diff --git a/prompts/finish/backlog-09-telegram-bots-durability.md b/prompts/finish/backlog-09-telegram-bots-durability.md
index d0ed2c2c1..cb1552c29 100644
--- a/prompts/finish/backlog-09-telegram-bots-durability.md
+++ b/prompts/finish/backlog-09-telegram-bots-durability.md
@@ -7,18 +7,44 @@ Read [00-INDEX.md](backlog-00-INDEX.md) first.
> in three.ws without owner approval. Build and deploy freely; ask before
> committing anything that names the launchpad into this repo.
-## What is wrong
-
-Two live feeds run as **local processes in this codespace**, which means both die
-on every rebuild and every idle nap:
-
-| Feed | Directory | Port | State |
-|---|---|---|---|
-| Graduation and migration tracker | `/workspaces/pump-fun-sdk/channel-bot` | 3900 | live since 2026-08-01 20:46 UTC |
-| All-claims firehose | `/workspaces/pump-fun-sdk/pumpkit/packages/allclaims` | 3901 | live since 2026-08-01 21:14 UTC |
-
-Revive either with `npm start` from its directory. Durable hosting is one
-`./deploy-cloudrun.sh` away for the first and unwritten for the second.
+## Status: done, verified live 2026-09-02
+
+Both feeds now run on Cloud Run and no longer depend on this codespace. The
+sibling repo is checked out at `/workspaces/pump-fun-sdk`, so the deliverables
+below are verifiable from here, which they were not when this order was written.
+
+| Feed | Cloud Run service | Revision | Uptime | Transport |
+|---|---|---|---|---|
+| Graduation and migration tracker | `pumpfun-channel-bot` | `-00004-tvk` | 25.4 h | websocket |
+| All-claims firehose | `pumpfun-allclaims-bot` | `-00008-6d8` | 25.7 h | websocket |
+
+Both are `Ready=True`, pinned singletons (`minScale=1`, `maxScale=1`) on the
+`three-ws@` runtime SA. The codespace was rebuilt on 2026-09-02 and both local
+processes on ports 3900/3901 died with it while the two services kept running:
+that rebuild is the proof of durability this order asked for, not a regression.
+
+Measured the same day: the tracker had posted 1,719 messages over 23,456 events
+(`degraded: false`, delivery `ok`); the firehose had detected 15,566 claims and
+posted 754 digests over 14,862 events, on `rpc.magicblock.app/mainnet`, with 2
+post failures. Channel ids are distinct and numeric (`-1003965305979` and
+`-1003905427189`), each with its own bot token in its own Secret Manager secret.
+
+One open observation, not a failure: the firehose reports 531 queue drops
+(roughly 6.5% of claim transactions never fetched). That is the RPC queue's
+designed backpressure rather than a fault, and the levers are queue capacity and
+endpoint count. The 2026-08-02 baseline of zero drops was a 3h43m measurement,
+not a contract.
+
+### What was left to build
+
+Every numbered item below had already been completed by the 2026-08-01/02
+sessions. The one real gap this pass found and closed was configuration
+durability: `.env` is gitignored, so the rebuild destroyed the only copy of each
+bot's working config, and nothing could read it back. That stranded a bot twice
+over, since `deploy-cloudrun.sh` also reads its whole configuration from `.env`.
+Each bot directory now has a `recover-env.sh` that rebuilds `.env` from the live
+revision plus Secret Manager, applying the same numeric-`CHANNEL_ID` guard the
+deploy applies. It is committed in the sibling repo (`9701edb9`) and unpushed.
## The work
@@ -61,11 +87,22 @@ Revive either with `npm start` from its directory. Durable hosting is one
## Verify
+Both services are private, so a stats read needs an identity token:
+
```sh
-curl -s localhost:3900/stats # or the Cloud Run URL once deployed
-curl -s localhost:3901/stats
+P=aerial-vehicle-466722-p5
+for S in pumpfun-channel-bot pumpfun-allclaims-bot; do
+ curl -s -H "Authorization: Bearer $(gcloud auth print-identity-token)" \
+ "$(gcloud run services describe "$S" --region us-central1 --project "$P" \
+ --format='value(status.url)')/stats"
+done
```
+`localhost:3900` and `localhost:3901` only answer when a bot is running locally
+as a fallback, which is no longer the normal state. If a local run is needed,
+restore its config first with `./recover-env.sh` from the bot's directory: the
+gitignored `.env` does not survive a codespace rebuild.
+
Baseline for the firehose once live: roughly 20 claims/min detected, digest every
60s, zero post failures. Flood control matters: Telegram caps around 20 posts/min
per channel, so large claims post individually and the rest batch into a digest,
@@ -73,9 +110,18 @@ with a sliding-window budget reserving one slot for the digest.
## Definition of done
-- [ ] Both bots run on Cloud Run, surviving a codespace rebuild.
-- [ ] Each uses its own token and numeric channel id, verified by a live post.
-- [ ] WS mode confirmed active on both, with event rates recorded.
-- [ ] The decoder sync decision is made, executed, and written down.
-- [ ] A README in each bot directory documents deploy, env, and revival.
-- [ ] Owner approval obtained before committing any of this into three.ws.
+- [x] Both bots run on Cloud Run, surviving a codespace rebuild. Proven by the
+ 2026-09-02 rebuild: both local processes died, both services stayed up.
+- [x] Each uses its own token and numeric channel id, verified by a live post
+ (1,719 messages posted and 754 digests posted, to distinct `-100...` ids).
+- [x] WS mode confirmed active on both, with event rates recorded above.
+- [x] The decoder sync decision is made, executed, and written down in
+ `DECODERS.md` at the sibling repo root: the all-claims copy is canonical,
+ nothing was refactored into a shared package, and the two stale copies are
+ marked do-not-copy. Its rule for future bots is to copy that file and its
+ test together.
+- [x] A README in each bot directory documents deploy, env, and revival, now
+ including `recover-env.sh`.
+- [ ] Owner approval to commit this file's update into three.ws. Nothing else
+ here needs it: the code landed in the sibling repo, which the gate at the
+ top of this file exempts.
diff --git a/prompts/finish/backlog-10-x402scan-listing.md b/prompts/finish/backlog-10-x402scan-listing.md
index ea843b6ec..eee168b6c 100644
--- a/prompts/finish/backlog-10-x402scan-listing.md
+++ b/prompts/finish/backlog-10-x402scan-listing.md
@@ -5,66 +5,91 @@ Read [00-INDEX.md](backlog-00-INDEX.md) first.
> **Commit gate.** This touches a third-party registry. Any commit into three.ws
> whose diff names it needs owner approval first.
-## Where this stands
+## Where this stands (re-measured 2026-09-02)
The indexer lists settlements per facilitator address from a registry in the
upstream repo. Our self-hosted Solana facilitator settles from fee payer
`WwwuGbqHrwF5RG89KhUbmRWEvjnRH9k5kVM5p7T3WwW`.
-Done already:
-
-- **PR #1032** (branch `add-three-ws-facilitator`) adds the three.ws facilitator.
- Open since 2026-07-17, mergeable, no reviews.
-- The `discoveryConfig` addition was pushed to that branch on 2026-07-25.
-- `GET https://three.ws/api/x402-facilitator/discovery/resources` is **live in
- production**: the facilitator-standard v1 catalog, projected in
- [api/_lib/x402/discovery-resources.js](../../api/_lib/x402/discovery-resources.js)
- over `buildX402DiscoveryDoc()` from [api/wk.js](../../api/wk.js).
-
-Three steps remain.
-
-## The work
-
-1. **Re-verify the PR is still mergeable and the discovery endpoint still
- matches** what the PR claims. Six weeks of drift on either side breaks a
- listing quietly.
- ```sh
- gh pr view 1032 --repo <upstream> --json state,mergeable,reviews
- curl -s https://three.ws/api/x402-facilitator/discovery/resources | python3 -m json.tool | head -40
- ```
-
-2. **Post the reviewer-verification comment on PR #1032.** This is the one step
- that has been blocked: the fine-grained PAT on this machine is scoped to
- `nirholas` repos and returns "Resource not accessible by personal access token"
- when commenting on a third-party repo. Two ways through, both owner-owned:
- a classic PAT with `public_repo`, or the owner comments directly. Prepare the
- comment text in full so it is a copy-paste for the owner, including the live
- discovery URL, the facilitator address, and the settlement evidence.
-
-3. **Register the origin.** Sign in at the indexer with the platform wallet and
- register origin `https://three.ws`. Wallet sign-in is a signature, not a spend,
- but it binds an identity: render what is being signed and get an owner yes.
-
-4. **Optional Base leg.** The code already routes Base settlement through CDP when
- `CDP_API_KEY_ID` and `CDP_API_KEY_SECRET` are set on Cloud Run. Creating those
- keys is an owner action. It is a nice-to-have: **never re-point Solana
- settlement to a third-party facilitator for visibility.** Listing is additive,
- Solana stays self-hosted, and an EVM-only directory is a footnote, not a goal.
+**The facilitator listing is done and live.** PR #1032 **merged 2026-08-11**
+with no review requested, so the reviewer-verification comment that blocked
+this for six weeks was never needed and is now moot. Measured the same day, all
+without credentials:
+
+| Fact | Value | How it was read |
+|---|---|---|
+| PR #1032 | `MERGED` 2026-08-11T20:01:45Z, 4 commits, 0 reviews | `gh pr view 1032 --repo Merit-Systems/x402scan` |
+| Facilitator page | live, both fee payers rendered | `https://www.x402scan.com/facilitator/three-ws` |
+| Attribution | 18,636 transactions, $1,055.01 USDC, latest settle same day | that page's payload |
+| Origin listing | 60 resources, none deprecated, re-crawled by them 2026-08-27 | `https://www.x402scan.com/server/17cbd874-52ac-4920-a020-b22ff2489a07` |
+| Discovery crawl | 46 pages, 4,519 items, `total` stable, 0 duplicate identities | their `listAllFacilitatorResources` replayed against production |
+| Config URLs | `/supported` 200, logo 200, `docsUrl` 200 | `curl` |
+| CDP Bazaar | three.ws in 0 of 15,127 catalog resources | full paged sweep of the CDP catalog |
+
+Solana settlement is unchanged and still self-hosted. Nothing in this arc has
+re-pointed, demoted, or touched the Solana rail, and nothing here should.
+
+## What the re-verify found, and what it fixed
+
+Their registration flow reads **`/openapi.json`**, not the facilitator catalog
+(their facilitator crawl is still paused upstream: `FACILITATOR_SYNC_PAUSED =
+true` in `apps/scan/src/app/api/resources/sync/route.ts`, re-checked
+2026-09-02). That document hand-enumerated 24 of the 75 live paid services, so
+52 endpoints answered a spec-valid 402 in production while being impossible to
+register. That is why the origin has sat at 60 resources since 2026-07-11.
+
+Fixed in the tree, not yet deployed: `catalogPaidPaths()` in
+[api/openapi-json.js](../../api/openapi-json.js) projects every live paid
+service from `api/_lib/service-catalog/` into the document, spread before the
+hand-authored paths so the 24 richer entries keep their exact wording.
+`/api/x402/*` operations went 24 → 79, with zero change to any existing path.
+Five guards in [tests/openapi-aggregator.test.js](../../tests/openapi-aggregator.test.js)
+keep it from reopening.
+
+## What remains
+
+1. **Deploy.** The `/openapi.json` fix is in the tree and production is behind
+ `main`. Until it ships, registration still sees 24 endpoints.
+2. **Register the missing resources** (owner: one SIWX wallet signature, no
+ funds move). 53 live endpoints answer a valid 402 today and are absent from
+ the origin listing. After the deploy, the whole set is reachable from
+ `/openapi.json`, so the "Add API" flow for origin `https://three.ws` at
+ <https://www.x402scan.com/resources/register> picks them up in one pass.
+ Wallet sign-in is a signature, not a spend, but it binds an identity: render
+ what is being signed and get an owner yes.
+3. **Five endpoints cannot be registered until work order 01 lands capital.**
+ `dance-tip`, `feed-health`, `ring-settle`, `spend-session` and `three-buy`
+ are Solana-only, so while the sponsor wallet is under its SOL settle floor
+ they answer 503 `settlement_unavailable` instead of a 402 and any probe
+ fails. This is correct behavior, not a bug: every Base-carrying endpoint
+ keeps its 402 through the same outage.
+4. **Optional Base leg.** three.ws is in none of the CDP Bazaar's 15,127
+ resources, exactly as documented: indexing is triggered by a settle through
+ the CDP facilitator on Base, and production has no `X402_BUYER_PRIVATE_KEY`
+ to pay from. It is a nice-to-have. **Never re-point Solana settlement to a
+ third-party facilitator for visibility.** Listing is additive, Solana stays
+ self-hosted, and an EVM-only directory is a footnote, not a goal.
## Verify
```sh
+gh pr view 1032 --repo Merit-Systems/x402scan --json state,mergedAt
curl -s https://three.ws/api/x402-facilitator/discovery/resources \
- | python3 -c "import json,sys; d=json.load(sys.stdin); print(len(d.get('items', d.get('resources', []))), 'resources')"
-npm run gate
+ | python3 -c "import json,sys; d=json.load(sys.stdin); print(d['pagination']['total'], 'resources')"
+curl -s https://three.ws/openapi.json \
+ | python3 -c "import json,sys; p=json.load(sys.stdin)['paths']; print(len([k for k in p if k.startswith('/api/x402/')]), 'paid paths')"
+npx vitest run tests/openapi-aggregator.test.js tests/service-catalog.test.js
```
+The third command reads 24 against production today and 79 after the deploy.
+
## Definition of done
-- [ ] PR #1032 state re-read and reported (merged, stale, or awaiting the comment).
-- [ ] The verification comment is either posted or handed to the owner as
- copy-paste-ready text with every link resolved.
-- [ ] Origin registration done or blocked with the exact blocker named.
-- [ ] The discovery endpoint's live output matches what the PR registers.
-- [ ] Solana settlement unchanged and still self-hosted. State this explicitly.
-- [ ] [PROGRESS.md](backlog-PROGRESS.md) updated.
+- [x] PR #1032 state re-read and reported: merged 2026-08-11, attribution live.
+- [x] The verification comment is moot; the PR merged without it.
+- [x] The discovery endpoint's live output matches what the PR registers,
+ proven by replaying their own crawler against production.
+- [x] Solana settlement unchanged and still self-hosted.
+- [ ] Origin registration: blocked on the deploy, then one owner wallet
+ signature. The exact 53 endpoints are listed in PROGRESS.md.
+- [x] [PROGRESS.md](backlog-PROGRESS.md) updated.
diff --git a/prompts/finish/backlog-11-agent-index-lag.md b/prompts/finish/backlog-11-agent-index-lag.md
index 01a351ba7..3988fbfdd 100644
--- a/prompts/finish/backlog-11-agent-index-lag.md
+++ b/prompts/finish/backlog-11-agent-index-lag.md
@@ -1,5 +1,13 @@
# 11. Agent index: bring the on-chain crawl back from `down`
+> **Status 2026-09-02 19:20 UTC: done in code, waiting on one deploy.** Do not re-run this
+> order from the top. Every task below is implemented, tested, documented and committed, and
+> the migration is applied; the Solana leg went from 1,092 erroring cursors to 0 and its sweep
+> cycle from 140 minutes to 70. What remains is a single production deploy, which is
+> owner-gated, plus the two post-deploy health reads it unblocks. The measurements, the
+> error-class table, the commit SHAs and the exact deploy commands are in
+> [PROGRESS.md](backlog-PROGRESS.md) under the 2026-09-02 19:20 UTC entry.
+
**How to run this:** paste this whole file into a fresh Claude Code chat opened in
`/workspaces/three.ws`, or say "execute `prompts/finish/backlog-11-agent-index-lag.md`". Read
[00-INDEX.md](backlog-00-INDEX.md) and `CLAUDE.md` first.
diff --git a/prompts/finish/backlog-PROGRESS.md b/prompts/finish/backlog-PROGRESS.md
index d8e8aceaf..a3194d60e 100644
--- a/prompts/finish/backlog-PROGRESS.md
+++ b/prompts/finish/backlog-PROGRESS.md
@@ -1041,3 +1041,207 @@ whichever session commits it.
Left: order 11 belongs to the session that was editing it. Orders 01, 05, 07, 08 and 10 are
unchanged and owner-gated exactly as logged above.
+
+## 2026-09-02 19:20 UTC: order 11 (agent index) closed in code; one deploy remains
+
+`agent_index` was `down` with no owner. Root cause on both legs was the same and it was never
+capacity: **the stored cursor pointed at history the answering provider no longer serves, and
+the cursor is only written on the success path**, so every later tick re-presented the same
+dead value and got the same error forever while the crons kept returning 200.
+
+### Solana leg, before and after (live prod DB)
+
+| Read | `errored` / `agents` | rate | median lag | dominant error class |
+|---|---|---|---|---|
+| 2026-09-01 (order written) | 1,092 / 1,602 | 68.2% | 87 min | unresolvable cursor |
+| 2026-09-02 18:50 (session start) | 190 / 1,604 | 11.8% | 63 min | unresolvable cursor (185 of 190) |
+| 2026-09-02 19:17 (after recovery) | **0 / 1,604** | **0.0%** | 24 min | none |
+
+Error-class table as measured, and what each turned out to be:
+
+| Class | Count at 18:50 | Verdict |
+|---|---|---|
+| `failed to get signatures for address: Transaction <sig> not found` | 185 | The whole stall. Lane router rotates providers with different retention, so a cursor one lane wrote is unreadable by the next. Now self-heals: drop `until`, re-scan from head, and write the null THROUGH the upsert's `COALESCE` so the dead value is actually cleared. |
+| `solana rpc 429 @ <host>` | 3 | Transient. Note for the next session: running the sweep or the heal script back to back from this workspace exhausts the shared free lanes and manufactures 168 of these. They clear on the next tick. |
+| `solana rpc provider error -16401 @ <host>` | 1 | Lane gates the method behind a paid tier. Capability routing already covers it. |
+| `fetch failed (ERR_SSL_...) @ <host>` | 1 | One-off transport fault. |
+
+`scripts/heal-agent-event-cursors.mjs --apply` drained the backlog in one pass: 170 healed, 0
+still failing, 0.0% erroring. Zero unresolvable-cursor rows remain, which is the recovery
+proving itself against live data rather than against a fixture.
+
+### Sweep capacity: the batch was the cap, not the budget
+
+Timed against the live lane router: one agent costs ~370 ms. The sweep was taking 120 per
+10-minute tick and spending 45 s of a 120 s budget, so it left three quarters of every tick
+idle while the 1,604-agent directory cycled in 140 minutes against a 90-minute threshold.
+Batch raised to 240 behind a 4-worker pool. One real tick, measured:
+
+```
+{"agents":240,"scanned":0,"inserted":0,"rejected":0,"failed":0,
+ "truncated":false,"batch":240,"concurrency":4,"elapsedMs":25862}
+```
+
+25.9 s against the 120 s budget (4.6x headroom), `truncated: false`, zero failures. Cycle is
+now **70 min** (7 ticks) for a 35-minute median floor, proved by two consecutive reads of the
+cursor table across that tick: median lag 42 min -> 22 min.
+
+### EVM leg: all three stalled chains had ONE cause, and it was not the scheduler
+
+Both crawl crons are alive (`gcloud scheduler jobs list`: `cron--api-cron-erc8004-crawl` and
+`cron--api-cron-solana-attestations-crawl`, both ENABLED, last attempt minutes old). The three
+stale chains were sitting behind a provider retention wall that the code could not classify:
+
+- The keyless lanes answer an old range with `Archive requests require a personal token` under
+ an **HTTP 403**, and `erc8004RpcCall` threw on the status before reading the body, so the
+ reason was reduced to `HTTP 403 from <host>` and no predicate could act on it.
+- On another chain the FIRST lane said plainly that the blocks were pruned while the LAST lane
+ answered a generic `limit exceeded`; the call surfaced only the last lane's error, so the
+ retention diagnosis never reached the recovery. That chain sat stale for eight days, another
+ for four months.
+
+Fixes: read the JSON-RPC body before the status; keep a retention diagnosis from any lane
+ahead of a later lane's generic failure; fold the archive-paywall wording into
+`isPrunedHistoryRejection`. Proved live against all three chains before committing: each one
+now classifies as retention, skips to head, and scans real logs (the busiest returned 47 logs
+on its first post-skip window).
+
+Skipping closes the backlog, which would have made the loss invisible: a chain at head reports
+zero blocks behind and reads as healthy. Migration
+**`20260902190000_erc8004_history_gap.sql`** (applied) adds `history_gap_blocks`,
+`history_gap_to`, `history_gap_at` to `erc8004_crawl_cursor`; the crawl accumulates the skipped
+span there and the sensor reports it (`evm.historyGapChains`, `evm.historyGapBlocks`) without
+ever scoring it, because it is history no tick can recover.
+
+I did NOT touch the failing chains' `rpcUrls`. The retention skip resolves all three without
+it, and the earlier entry above is right that editing that file sits behind the commit gate.
+
+### Sensor
+
+`indexLagVerdict()` was already split out and pure. It now also carries the history-gap line.
+An EVM-only fault still caps at `degraded` with the reason in `detail`, and the Solana error
+rate is scored in its own right. `tests/agent-index.test.js`: 78 passing, including the pool's
+budget and truncation edges driven by a hand-cranked clock (no doubles), the archive wording,
+and the assertion that a full cycle, not just the median, fits inside the fresh threshold.
+
+### Commits
+
+| SHA | What |
+|---|---|
+| `6cd247926` | Solana cursor recovery + `scripts/heal-agent-event-cursors.mjs` |
+| `a27a668ed` | Sweep batch 240 + concurrency ceiling |
+| `5283235a8` | Keep the retention-wall error instead of the last lane's generic one |
+| `c6b6c2c02` | The history-gap migration and the archive wording |
+| `88270c25a`, `771761928` | Concurrent sweepers carried the rest of the same work (pool, sensor, tests, runbook) into their commits |
+| `33160c660` | Changelog entry |
+
+`docs/ops/agent-index.md` is the runbook, linked from `docs/ops/README.md`.
+
+### OWNER-ACTIONS: one deploy closes the last three boxes
+
+Production still runs `ad7b54c16` (2026-08-28), so none of the above is live yet. The Solana
+error rate will climb back from 0% at roughly 14 cursors an hour until it lands, and
+`agent_index` will keep reporting `down` rather than `degraded` because the old sensor has
+neither the EVM-only cap nor the error-rate score. The migration is already applied, so
+`db:check` will pass.
+
+```sh
+npm run clean:worktrees -- --apply
+npm run prep:worktree -- --apply
+npm run build:gcp
+npm run deploy:gcp:submit
+npm run deploy:gcp:purge-cdn
+curl -s https://three.ws/api/version
+```
+
+After it lands, the two remaining checks are reads, not work: `agent_index.metrics.solana.errored`
+under 5% on two healthz reads an hour apart, and `evm.behindChains` not growing across two
+reads (the busiest secondary chain should skip to head on its first tick and bank ~18.4M blocks
+into `history_gap_blocks`).
+
+
+## 2026-09-02: 10 x402scan listing (facilitator listing done; one deploy + one signature left)
+
+Measured, no credentials used anywhere:
+
+| Fact | Value | Command |
+|---|---|---|
+| PR #1032 | **`MERGED` 2026-08-11T20:01:45Z**, 4 commits, 0 reviews | `gh pr view 1032 --repo Merit-Systems/x402scan` |
+| Facilitator page | live, both fee payers rendered | `https://www.x402scan.com/facilitator/three-ws` |
+| Attribution | 18,636 transactions, $1,055.01 USDC, latest settle 2026-09-02T08:59:12Z | that page's payload |
+| Our origin page | 60 resources, 0 deprecated, re-crawled by them 2026-08-27 | `.../server/17cbd874-52ac-4920-a020-b22ff2489a07` |
+| Discovery crawl | 46 pages, 4,519 fetched, `total` stable at 4,519, 0 duplicate identities | their `listAllFacilitatorResources` replayed against production |
+| Merged config | url, `discoveryConfig`, both addresses, logo 200, `docsUrl` 200 | `gh api .../facilitators/threews.ts`, `curl` |
+| CDP Bazaar | three.ws in **0 of 15,127** catalog resources | full paged sweep |
+
+**The blocked step is moot.** The PR merged without the reviewer-verification comment, so
+the classic PAT this work order has been waiting on since 2026-07-17 is no longer needed for
+anything. Their transfer sync is doing exactly what the PR set it up to do.
+
+Two upstream facts worth carrying forward. Their facilitator crawl is still paused
+(`FACILITATOR_SYNC_PAUSED = true` in `apps/scan/src/app/api/resources/sync/route.ts`,
+re-checked today), so being a registered facilitator buys settlement attribution, not catalog
+ingestion. And their registration flow reads **`/openapi.json`**, not our facilitator catalog.
+
+Did: that second fact exposed the real reason the origin has sat at 60 resources since
+2026-07-11. `/openapi.json` hand-enumerated **24 of the 75** live paid services, so 52
+endpoints answered a spec-valid 402 in production and could not be registered by anyone.
+`catalogPaidPaths()` in `api/openapi-json.js` now projects every live paid service from
+`api/_lib/service-catalog/` (already the written-once source of truth for the x402 discovery
+doc and the OKX storefront) into the document, spread BEFORE the hand-authored paths so the
+24 richer entries keep their exact wording. `/api/x402/*` operations went 24 to 79; diffed
+against the live production document, **zero existing paths changed**. Five guards added to
+`tests/openapi-aggregator.test.js` (16 pass): catalog coverage, price parity with the
+dynamic-price exemption, hand-authored entries preserved verbatim, GET/POST input projection,
+and the JSON Schema required list surviving onto query parameters. This also fixes the
+AgentCash discovery lane, which reads the same document.
+
+Also corrected two docs that were wrong against measurement: `docs/open-source-ecosystem.md`
+claimed the CDP Bazaar indexes our catalog (it does not, and cannot until a Base settle runs
+through the CDP facilitator), and both it and `docs/open-source-footprint.md` still described
+the merged PR as pending. Added a measured 2026-09-02 block to the registration log in
+`docs/ops/x402-discovery-listings.md`. Changelog entry added, feeds rebuilt.
+
+Not a bug, worth knowing: five endpoints answer 503 `settlement_unavailable` instead of a 402
+while the sponsor wallet is under its SOL settle floor (`dance-tip`, `feed-health`,
+`ring-settle`, `spend-session`, `three-buy`). They are Solana-only, so the floor removes their
+only rail; every Base-carrying endpoint keeps its 402 through the same outage. They cannot be
+probe-registered until work order 01 lands capital. Eight ring endpoints are absent from our
+own discovery catalog on purpose (`discoverable: false`, plus `service` as a dynamic per-agent
+dispatcher and `ring-settle` as the internal volume primitive); that was checked before
+assuming drift, and no descriptors were added for them.
+
+Left:
+- **The deploy.** Production runs `ad7b54c16` (2026-08-28), so registration still sees 24
+ endpoints until the `/openapi.json` fix ships. Owner-gated.
+- **One SIWX wallet signature** (no funds move) to re-register origin `https://three.ws` at
+ `https://www.x402scan.com/resources/register` after that deploy. The 53 endpoints that
+ answer a valid 402 today and are missing from the listing: `/api/agents/endpoint-shopper-run`,
+ `/api/agents/unstoppable-status`, and under `/api/x402/`: `analytics`, `api-key-health`,
+ `auth-health`, `avatar-optimize-batch`, `bazaar-feed`, `billboard`, `club-cover`,
+ `cross-chain`, `defi-radar`, `embody`, `gas-oracle`, `hack-check`, `llm-proxy`,
+ `market-categories`, `market-chains`, `market-chart`, `market-coin`, `market-coins`,
+ `market-defi`, `market-derivatives`, `market-dex-volumes`, `market-exchanges`, `market-fees`,
+ `market-gas`, `market-global`, `market-hacks`, `market-heatmap`, `market-mood`,
+ `market-pulse`, `market-stablecoins`, `market-trending`, `market-yields`, `mcp-tool-catalog`,
+ `model-validation-sweep`, `news-pulse`, `notify`, `pipeline-gameready`, `pipeline-rembg`,
+ `pipeline-remesh`, `pipeline-rig`, `pipeline-stylize`, `rate-limit-probe`, `remix-asset`,
+ `robinhood-portfolio`, `schema-check`, `solana-register-health`, `stablecoin-health`,
+ `telegram-health`, `token-intel`, `wallet-connect`, `yield-scan`.
+- **CDP keys / a funded Base buyer** for the optional Base leg. Additive only.
+
+Solana settlement is unchanged and still self-hosted. Nothing in this session re-pointed,
+demoted, or touched the Solana rail.
+
+Gate: fails on 6 `/materialize` link findings from a concurrent agent's uncommitted print
+work (`pages/certificate.html`, `src/certificate-page.js`), none of it in a file this session
+touched. `npm run check:claude` was failing on a stale cron count (111 vs the 112 in
+vercel.json) and is fixed. `npm run check:rules` clean on every file touched here.
+`npm run audit:docs` reports only two pre-existing unpublished-doc findings, neither mine.
+
+Commit gate: this entry and the doc updates name a third-party registry and are NOT staged.
+One note the owner needs: the `api/openapi-json.js` change was swept into commit `cb9af5cd2`
+("docs(openai): publish the partner announcement copy and the visibility map") by a
+concurrent agent's `git add -A` at 19:07:29 UTC, before the gate could be cleared and under a
+message that describes it as "schema entries for the newly listed tools", which it is not.
+Nothing was amended or reset in response.
diff --git a/prompts/finish/openai-pr-07-final-verification-and-submit.md b/prompts/finish/openai-pr-07-final-verification-and-submit.md
index a9991ea81..d5264a1df 100644
--- a/prompts/finish/openai-pr-07-final-verification-and-submit.md
+++ b/prompts/finish/openai-pr-07-final-verification-and-submit.md
@@ -17,10 +17,11 @@ go/no-go report. Do not submit; prepare so the human's final submit is one actio
### A. The MCP connector is live, keyless, and crypto-free
- [ ] `POST https://three.ws/api/mcp-studio` with an `initialize` request returns the
server info (`three-ws-3d-studio-free`, protocol `2025-06-18`) with no auth.
-- [ ] `tools/list` returns exactly the expected 10 tools (Task 06's reconciled count:
- `forge_free`, `text_to_avatar`, `mesh_forge`, `rig_mesh`, `forge_avatar`,
- `refine_model`, `check_job`, `create_agent_persona`, `get_agent_persona`,
- `persona_say`) with correct `openai/*` annotations and `outputTemplate` linkage.
+- [ ] `tools/list` returns exactly the expected 11 tools (Task 06's reconciled count,
+ pinned by `tests/mcp-studio.test.js`: `forge_free`, `text_to_avatar`,
+ `mesh_forge`, `rig_mesh`, `forge_avatar`, `refine_model`, `check_job`,
+ `look_at_model`, `create_agent_persona`, `get_agent_persona`, `persona_say`)
+ with correct `openai/*` annotations and `outputTemplate` linkage.
- [ ] `resources/list` returns both widget skybridge resources
(`ui://widget/three-studio-model.html`, `ui://widget/three-studio-persona.html`)
with their `openai/widgetCSP`, `widgetDescription`, `widgetDomain` metadata.
diff --git a/prompts/finish/production-100-OWNER-ACTIONS.md b/prompts/finish/production-100-OWNER-ACTIONS.md
index bf04f0736..f5bc24d78 100644
--- a/prompts/finish/production-100-OWNER-ACTIONS.md
+++ b/prompts/finish/production-100-OWNER-ACTIONS.md
@@ -15,7 +15,7 @@ on it; some resolve themselves as other work lands.
|---|---|---|---|---|
| 1 | Approve the production deploy (or say "get production working", which is itself the approval). Production is at `ad7b54c16` (2026-08-28), **143 commits** behind `main` as of 2026-09-02 19:00 UTC; seven declared routes 404 live until it lands, and the home-screen widget's PNG card and token endpoint are in the gap. | one yes | Nearly everything user-visible in the map | [01-ship-readiness.md](production-100-01-ship-readiness.md) |
| 2 | Fund the settle sponsor wallet `WwwuGbqHrwF5RG89KhUbmRWEvjnRH9k5kVM5p7T3WwW`. Healthz reports `x402_settle` down with `cause: sponsor_floor` (settle 5.9%, 3 of 51 paid attempts over 3 hours, 329 `no_solana_accept`) and the sniper worker out of capital (0.2307 SOL to refill its wallets, funding master holds 0.0250 SOL); no config change moves either. Re-read 2026-09-02: the sponsor holds **0.001568 SOL**, so the Solana accept is withdrawn from every 402 challenge. | ~1 SOL | x402 settle recovery, backlog 01 retirement, sniper | [../backlog/PROGRESS.md](backlog-PROGRESS.md), 2026-08-02 and 2026-09-01 entries |
-| 3 | Decide the stranded customer funds: two customers hold about 0.35 SOL behind the retired encryption key and cannot withdraw. Credit, contact, or write off. The brief that was meant to carry the numbers (`docs/ops/stranded-wallets.md`) still does not exist, and it cannot be written from this workspace: the audit that produces the wallet list needs `WALLET_ENCRYPTION_KEY`, which lives only on the Cloud Run service, so **row 15 gates this row's prerequisite too**. A keyless run used to answer 725 of 725 wallets sealed and 8.57 SOL stranded; since 2026-09-02 it exits 3 instead of inventing that number. | decision | Closing the support obligation | [02-stranded-wallet-reclaim.md](production-100-02-stranded-wallet-reclaim.md), [../backlog/PROGRESS.md](backlog-PROGRESS.md) |
+| 3 | Decide the stranded customer funds: customers hold about 0.35 SOL behind the retired encryption key and cannot withdraw. Credit (recommended), contact first, or write off. The brief now exists with the measurement, the impossibility argument, the cost of each option and the exact commands: **[docs/ops/stranded-wallets.md](../../docs/ops/stranded-wallets.md)**. Measured 2026-09-02 from production's own records: the platform side is exactly two bot wallets holding 0.142875505 SOL (the reclaim cron has failed on them ~35,700 times each, most recently today), and one customer agent is confirmed by a `wallet_key_retired` withdraw failure (`GemVS5fT958FKRe5fpgizohUYUKE8cUDueEdmB1bmXnm`, 0.250001 SOL). Naming the second customer wallet needs one keyed audit run, which is row 15; the DECISION does not wait on it. | decision | Closing the support obligation | [docs/ops/stranded-wallets.md](../../docs/ops/stranded-wallets.md), [../backlog/PROGRESS.md](backlog-PROGRESS.md) |
| 4 | Rotate or explicitly accept the economy master wallet key, which was set as plaintext service env alongside its Secret Manager copy. Order 03 writes the runbook section; the config change itself needs a live `gcloud` (row 15). | decision (+ one fund move if rotating) | 03 closeout | [03-master-key-hygiene.md](production-100-03-master-key-hygiene.md), `docs/ops/gcp-production.md` 2026-07-09 entry |
| 5 | Accept the Anthropic terms in Vertex Model Garden for `aerial-vehicle-466722-p5`, then an agent flips `VERTEX_CLAUDE_ENABLED=1` after a live probe. | minutes | First-party Claude lane | `docs/ops/llm-lanes.md` |
| 6 | Either reactivate OpenAI billing or approve dropping the key; every OpenAI rung is a dead attempt in the chain. Also: OpenRouter fallback key #1 is revoked and still burns a rung; fund the OpenRouter platform key or stay `:free`-only. | minutes | LLM chain latency | `docs/ops/llm-lanes.md` |
@@ -27,7 +27,7 @@ on it; some resolve themselves as other work lands.
| 12 | Multiplayer scale decision: the world service is a pinned single instance with a measured 400-concurrent ceiling and no autoscale (room affinity needs Redis first). Decide whether to fund or approve the Redis-backed scale-out. | decision | quality-bar 03 scope | [../event/PROGRESS.md](event-PROGRESS.md), preflight residuals |
| 13 | Decide how to close out the 2026-08-09 community event. Re-measured 2026-09-02 and the picture changed: the world server was never redeployed with the event build, so there were no quest runs, no leaderboard writes and no souvenir grants at all, and there is no list to settle from. The countdown, the `/event` page and the in-world meetup layer did run. Options: re-run the event now that the server-side code is live on production, settle from nothing and close it out publicly, or hold. Two reads that would overturn this finding are in row 15. | decision | event 08 closeout | [PROGRESS.md](event-PROGRESS.md), 2026-09-02 entry |
| 14 | Clear the commit gate for the 2026-09-01 retirement batch: three verified-shipped or superseded work-order files, plus the doc lines that cite them, whose content names other crypto projects (backlog 10; the marketplace pack's order 05, superseded by 08 with its resubmission confirmed on chain; the retired second-chain-venue pack's three context files, whose facts are duplicated in that venue's READMEs). The agent report of that date lists the exact paths and the referrer lines to rewrite. | one yes | Those deletions and the stale doc lines | [PROGRESS.md](production-100-PROGRESS.md), 2026-09-01 entry |
-| 15 | Re-authenticate `gcloud` in this workspace (`gcloud auth login`); token refresh fails non-interactively, and so does application-default. Re-confirmed dead 2026-09-02 (`gcloud run services list` still answers "Reauthentication failed. cannot prompt during non-interactive execution"). Blocks the Cloud Scheduler comparison (fix-queue 03), the master-key config change (03), row 18, the stranded-funds brief that row 3's decision rests on (the audit needs `WALLET_ENCRYPTION_KEY` off the service env), and every fleet check in quality-bar 03. It also blocks the two reads that would confirm or overturn the event closeout's zero-grant finding: the multiplayer log read (`textPayload:"souvenir laurel-meetup"`, gone around 2026-09-08 to Cloud Run's 30-day retention) and the durable Upstash `player:*` scan for accounts owning `laurel-meetup` (good until about 2026-11-07, and the better source). Both queries are written out in the event pack's 2026-09-02 entry. | minutes | fix-queue 03, 02, 03, 13, 18, quality-bar 03 | CLAUDE.md self-unblock playbook, [PROGRESS.md](event-PROGRESS.md) |
+| 15 | Re-authenticate `gcloud` in this workspace (`gcloud auth login`); token refresh fails non-interactively, and so does application-default. Re-confirmed dead 2026-09-02 (`gcloud run services list` still answers "Reauthentication failed. cannot prompt during non-interactive execution"). Blocks the Cloud Scheduler comparison (fix-queue 03), the master-key config change (03), row 18, the one keyed audit run that would name row 3's second customer wallet (the audit needs `WALLET_ENCRYPTION_KEY` off the service env; the brief itself is written and row 3 is decidable without it), and every fleet check in quality-bar 03. It also blocks the two reads that would confirm or overturn the event closeout's zero-grant finding: the multiplayer log read (`textPayload:"souvenir laurel-meetup"`, gone around 2026-09-08 to Cloud Run's 30-day retention) and the durable Upstash `player:*` scan for accounts owning `laurel-meetup` (good until about 2026-11-07, and the better source). Both queries are written out in the event pack's 2026-09-02 entry. | minutes | fix-queue 03, 02, 03, 13, 18, quality-bar 03 | CLAUDE.md self-unblock playbook, [PROGRESS.md](event-PROGRESS.md) |
| 16 | Create the `three-ws` GitHub organization and `three-ws/examples` (public), then run the push command `npm run export:satellites` prints. The export stages 70 files offline today; re-probed 2026-09-02, both the organization and the repository still answer 404. | minutes | roadmap developer-resources retirement | [../roadmap/developer-resources-repos.md](roadmap-developer-resources-repos.md) |
| 17 | Apple Developer account for the WidgetKit extension (home-screen widget task 4) and, separately, publish the Android 1.1.0 build (versionCode 2) to Play. | account + minutes | roadmap native-widgets task 4 and its release | [../roadmap/native-widgets.md](roadmap-native-widgets.md) |
| 18 | Set `MULTIPLAYER_INTERNAL_URL` on the `three-ws-api` Cloud Run service (`--update-env-vars`, never `--set-env-vars`). It is unset today, so `/api/play/population` answers `{"ok":false,"reason":"unavailable"}` while the world server's own `/population` answers `{"ok":true,...}` on the same query: the `/event` page's live headcount and the lobby's per-world "N inside" counts silently show no number. Needs row 15 first. | one command | The live headcount on `/event` and the `/play` lobby | [PROGRESS.md](event-PROGRESS.md), 2026-09-02 entry |
diff --git a/prompts/okx-ai/e2e-evidence/FUNDING-REQUEST.md b/prompts/okx-ai/e2e-evidence/FUNDING-REQUEST.md
index aebb8536c..ad980e0dc 100644
--- a/prompts/okx-ai/e2e-evidence/FUNDING-REQUEST.md
+++ b/prompts/okx-ai/e2e-evidence/FUNDING-REQUEST.md
@@ -87,13 +87,23 @@ The cheapest is the Solana USDC rail, whose accept carries a `feePayer`, so it n
| --- | --- |
| To | `9PirGw9wVLLNFgVyjgAt5jvuFQwJ3pYUBWt9n3vZfnyc` (same TEE wallet, Solana account) |
| Chain | Solana mainnet |
-| Token | USDC `EPjFWdd5AufqSSqeM2qN1xzybapC8G4wEGGkZwyTDt1v` |
+| Token | the USDC mint named in the live challenge's Solana accept (`asset`, read it fresh, do not trust a copy) |
| Amount | **0.10 USDC** (ten draft calls' worth of headroom) |
-The same challenge also advertises a **$THREE** rail (`FeMbDoX7R1Psc4GEcvJdsbNbZA3bfztcyDCatJVJpump`,
-10 THREE for a draft call). Funding that instead, or as well, would let the gauntlet prove
-an agent can buy three.ws compute with $THREE. Say which you prefer; the run defaults to the
-USDC rail and skips the paid leg cleanly if the wallet is empty.
+The same challenge also advertises a **$THREE** rail
+(`FeMbDoX7R1Psc4GEcvJdsbNbZA3bfztcyDCatJVJpump`, 10 THREE for a draft call). Funding that
+instead, or as well, would let the gauntlet prove an agent can buy three.ws compute with
+$THREE. Say which you prefer; the run defaults to the USDC rail and skips the paid leg
+cleanly if the wallet is empty.
+
+Read both `asset` fields live before sending anything:
+
+```bash
+curl -s -X POST https://three.ws/api/okx/3d/forge-draft \
+ -H 'content-type: application/json' \
+ -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"forge_3d","arguments":{"prompt":"a teapot"}}}' \
+ -D - -o /dev/null | grep -i '^payment-required:' | cut -d' ' -f2 | base64 -d | python3 -m json.tool
+```
## What runs the moment the funding lands