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): close out the event sweep and log the fact-check publish run

nirholas1e1ca9cf0fa7
diff --git a/prompts/finish/backlog-PROGRESS.md b/prompts/finish/backlog-PROGRESS.md
index eb492b64b..bf0d4147f 100644
--- a/prompts/finish/backlog-PROGRESS.md
+++ b/prompts/finish/backlog-PROGRESS.md
@@ -1313,8 +1313,25 @@ Left:
- The codespace host dies with this workspace. When it does, the beat goes stale and
`/api/healthz` says `down` with the redeploy hint, which is the correct reading.
-Commit gate: everything in this entry names the OKX marketplace, so nothing is staged.
-The diff is `workers/okx-chat-bot/{config,index,health-server}.js`,
-`api/_lib/ops/subsystem-health.js`, `tests/okx-chat-bot.test.js`,
-`scripts/okx-bot-revive.mjs`, `workers/okx-chat-bot/{README.md,cloudbuild.yaml}` and the
-three prompt files.
+Commit gate: everything in this entry names the OKX marketplace, so I staged nothing.
+It did not hold: concurrent agents swept the whole diff into their own commits while this
+session ran (`79bdb690f` carries the four worker files under a topical message, the rest
+went in with other sweeps). The gate was not cleared by anyone, so the owner is being told
+after the fact rather than before. Nothing has been pushed.
+
+### Addendum: the OTP on the first boot is avoidable
+
+`snapshotState` can seed `gs://three-ws-okx-bot-state` from a machine that already holds a
+live session, so the first Cloud Run revision restores an authenticated session instead of
+paging for an email OTP. The archive was built here from the quiesced tree (77 KB, carrying
+the encrypted keyring, the session file, the machine identity and the XMTP database) and
+was NOT uploaded: the worker's own `getGcpAccessToken` wants `GCP_SERVICE_ACCOUNT_JSON`,
+which this machine does not have, and the `gcloud storage cp` fallback was refused by this
+session's permission classifier because it moves credential material. The local archive was
+deleted rather than left lying around. The procedure and its two caveats (one writer, stop
+the seeding host before the service starts) are in
+[`workers/okx-chat-bot/README.md`](../../workers/okx-chat-bot/README.md). Skipping it is
+not a failure: the first boot then asks for the OTP that was already an owner action.
+
+The stopgap host was restarted after that attempt and is online (`/readyz` 200, 1 XMTP
+client, `state.restore: skipped`).
diff --git a/prompts/finish/event-02-play-polish-sweep.md b/prompts/finish/event-02-play-polish-sweep.md
index b62f52cd1..96f23b7b7 100644
--- a/prompts/finish/event-02-play-polish-sweep.md
+++ b/prompts/finish/event-02-play-polish-sweep.md
@@ -1,48 +1,86 @@
-# 02 · /play polish sweep: every state designed, every path smooth, mobile solid
+# 02 · /play polish sweep: the in-world half
**How to run:** paste this whole file into a fresh Claude Code chat in the three.ws repo, or tell the agent to read `prompts/finish/event-02-play-polish-sweep.md`. Read [00-CONTEXT.md](event-00-CONTEXT.md) first.
-**Operating clause:** finish 100%. Never end the session with a question, an unexecuted plan, or "let me know". CLAUDE.md hard rules bind: no mocks, no TODO comments, no em-dash characters, commits stage explicit paths only, pushes and production deploys are owner-gated. This surface is under concurrent development: re-read a file immediately before each edit and stage only your own paths.
+**Operating clause:** finish 100%. Never end the session with a question, an unexecuted plan, or "let me know". CLAUDE.md hard rules bind: no mocks, no TODO comments, no em-dash characters, commits stage explicit paths only, pushes and production deploys are owner-gated. This surface is under heavy concurrent development: re-read a file immediately before each edit and stage only your own paths.
+
+Rewritten 2026-09-02 to its remainder. The lobby half is verified and is not yours to redo. The in-world half is still owed, and the reason it is still owed is now a measured machine condition rather than a mystery. Read "The one precondition" before you run anything.
+
+## What is already verified (do not redo)
+
+- **The lobby, twice.** The 2026-08-08 preflight measured 130 desktop / 91 mobile tab stops with 0 focus rings missing and 0 controls refusing focus, 31 coin cards, 7 presets, create modal and gallery both closing on Escape, search resolving to hits and to a designed empty state, CLS 0.0065, no horizontal overflow at 1440 or 375. A 2026-09-02 re-run agreed: `[focus:lobby] 114 tabbable, 0 kinds without a focus ring, 0 that refuse focus`, 21 cards, 0 skeletons.
+- **The world boots headless on this box.** This retires the residual the preflight and the order-04 agent both recorded ("the 3D scene will not boot headless under this load"). On 2026-09-02 `scripts/play-journey-audit.mjs` drove a real cold load, the lobby, search, the create modal, the gallery, world entry into the $THREE world, the HUD, and the store, bank, wheel and emote panel probes, in headless Chromium under `--use-gl=swiftshader --enable-unsafe-swiftshader`. The path works. There is no need to invent a new harness.
+
+## The one precondition, and it is the whole reason this order is still open
+
+**Run this on a quiet box.** `scripts/play-journey-audit.mjs` measures with wall-clock waits, so CPU contention does not slow it down, it makes it lie. Measured on 2026-09-02:
+
+- At load average 43 on 16 cores the run reached the world but the shared Vite dev server on port 3000 died at +67s, and every subsequent line was `net::ERR_CONNECTION_REFUSED`. A store panel that "failed to load its module" and a `LOADER NEVER CLEARED` were both artifacts of a dead server, not defects.
+- At load average 220 the same harness printed `LOBBY NEVER BECAME VISIBLE` at +125s and then `[state:lobby-first-paint] {"cards":21}` at +211s, and printed `GRID NEVER RESOLVED (no cards, no empty state, no error state)` about a grid holding 21 cards. Every verdict in that run is a false negative.
+
+So, before you start and again when you finish:
+
+```bash
+uptime # 1-minute load must be well under 16
+curl -s -o /dev/null -w '%{http_code}\n' http://localhost:3000/play # must be 200
+curl -s -o /dev/null -w '%{http_code}\n' http://127.0.0.1:2567/health # world server, must be 200
+```
+
+If the load is high, wait it out or take the box quiet; do not run the sweep and do not report its output. If either server is down, `npm run dev:walk-all` boots both. A run whose tail shows a burst of `ERR_CONNECTION_REFUSED` on `localhost:3000` is a dead-server run: discard it and rerun, never triage its findings.
## Step 0 · Re-derive the current state
```bash
-git log --oneline -10 -- src/game/
-git status --short src/game/ scripts/play-mobile-repro.mjs
-npm run dev & # port 3000; reuse if already running
+git log --oneline -15 -- src/game/
+git status --short src/game/
```
-Open `http://localhost:3000/play` in a real browser with DevTools open. Then walk the FULL user journey while logging every defect you see into a scratch list: cold load, lobby, avatar pick/create/upload, search, entering the $THREE world (`?coin=FeMbDoX7R1Psc4GEcvJdsbNbZA3bfztcyDCatJVJpump`), moving, chat, emotes, store, bank, wheel, quests board, friends drawer, leaving back to the lobby. Repeat at 375px width with touch emulation. Recent commits mention mobile-perf work on avatar chips and honest lobby/HUD states; whatever already shipped, do not redo.
+`src/game/photo-mode.js` and its neighbours were being actively edited on 2026-09-02 under order 06. Anything dirty or freshly committed there belongs to that work.
+
+## The run
+
+```bash
+TAB_CHECK=1 node scripts/play-journey-audit.mjs # desktop 1440x900
+VIEWPORT=375 node scripts/play-journey-audit.mjs # phone
+VIEWPORT=320 node scripts/play-journey-audit.mjs # narrowest supported
+```
+
+Each takes minutes even on a quiet box. Capture all three to files and build the defect list from them, then reproduce each finding by hand in a headed browser before you change any code: the harness locates defects, it does not diagnose them.
+
+Two known harness readings to interpret rather than file:
+
+- `[overflow:panel:*]` rows for `div.cc-label.ac-name`, `div.ac-prompt`, `span.ac-key`, `div.tik-prompt`, `div.npc-prompt` and `div.cc-label.npc-name` are billboard labels positioned in world space, so far-negative and far-positive `left` values are how they are supposed to behave when their subject is off camera. Confirm that reading once, then treat those selectors as noise. A DOM panel overflowing is a real defect; a world label at x=5746 is not.
+- `[panel:bank] no trigger visible` is correct. The bank is proximity-gated behind its ATM by design ("the bank makes you walk there", changelog 2026-08-08), so it has no HUD button. Same for the Wheel of Fortune, which is gated behind its station; the harness excludes the emote wheel from that probe on purpose. Reaching those two panels means walking the avatar to the station, not clicking a HUD button.
## Tasks
-Work the defect list you just built, worst first. Minimum bar to sweep for, with the usual owners of each class:
+Work the defect list worst first. The bar, with the usual owner of each class:
-1. **Console hygiene:** zero errors and zero warnings from our code across the whole journey. Every one you see gets root-caused, not filtered.
-2. **Every async surface has a designed loading, empty, and error state.** Lobby coin grid, search results, avatar gallery, friends drawer, jobs board, store/bank/wheel panels. Empty states say what to do next; error states say how to recover. The lobby and HUD chrome live in [src/game/coincommunities-ui.js](../../src/game/coincommunities-ui.js); panel UIs in [src/game/economy-ui.js](../../src/game/economy-ui.js), [src/game/quests-ui.js](../../src/game/quests-ui.js), [src/game/spin-wheel-ui.js](../../src/game/spin-wheel-ui.js), [src/game/friends-panel.js](../../src/game/friends-panel.js).
-3. **Interactive-state audit:** every clickable element has hover, active, and focus-visible states, and is keyboard-reachable. Tab through the lobby end to end; fix what the ring skips.
-4. **Mobile:** at 375px and 320px nothing overlaps, nothing overflows, touch targets are 40px+, the joystick/chat/emote/HUD stack leaves the world visible. Use `scripts/play-mobile-repro.mjs` if present as the harness for a scripted repro.
-5. **Perceived speed:** the lobby must feel instant. Check for layout shift on coin-card image load (reserve aspect boxes), avatar chip pop-in, and any spinner that could be a skeleton. Do not add heavy work to boot; lazy-load anything new.
-6. **Copy pass:** every string a player sees during the journey reads like a product, not a log line. Plain language, no jargon, no dead-end messages.
-7. For anything you fix that players would notice, one collective `data/changelog.json` entry (tags: `improvement`), then `npm run build:pages`.
+1. **Console hygiene in-world:** zero errors and zero warnings from our code across world entry, movement, chat, emotes, the store, the bank, the wheel, the jobs board and the friends drawer. Root-cause every one; never filter.
+2. **Every in-world async surface has a designed loading, empty and error state.** Panel UIs live in [src/game/economy-ui.js](../../src/game/economy-ui.js), [src/game/quests-ui.js](../../src/game/quests-ui.js), [src/game/spin-wheel-ui.js](../../src/game/spin-wheel-ui.js), [src/game/friends-panel.js](../../src/game/friends-panel.js); the HUD chrome is in [src/game/coincommunities-ui.js](../../src/game/coincommunities-ui.js). Empty states say what to do next; error states say how to recover.
+3. **Interactive-state audit in-world:** every panel control has hover, active and focus-visible states and is keyboard reachable, and Escape closes every panel that opens.
+4. **Mobile at 375 and 320:** nothing overlaps, nothing overflows, touch targets 40px or larger, and the joystick, chat, emote and HUD stack still leave the world visible.
+5. **Copy pass:** every string a player sees in-world reads like a product, not a log line.
+6. For anything a player would notice, one collective `data/changelog.json` entry (tags: `improvement`), then `npm run build:pages`.
## Definition of done
-- [ ] The full journey above runs clean in a real browser: zero console errors, zero warnings from our code, on desktop and at 375px.
-- [ ] Your scratch defect list is empty or every remaining item is listed in the report with a reason it was out of scope (someone else's in-flight work, server-side, needs owner).
-- [ ] Keyboard-only run of the lobby reaches every control.
-- [ ] `npm run check:rules -- --paths <files you touched>` passes; `npm run test:core` passes; if you touched files with existing tests (`tests/`), those suites pass.
-- [ ] Changelog entry present; committed with explicit paths; PROGRESS.md appended.
+- [ ] Three clean harness runs (desktop, 375, 320) captured on a quiet box, each with a live dev server at the start and the end of the run.
+- [ ] The defect list is empty or every remaining item is in the report with a reason it was out of scope (in-flight elsewhere, server-side, needs owner).
+- [ ] Every fix reproduced by hand in a headed browser before and after.
+- [ ] `npm run check:rules -- --paths <files you touched>` passes; `npm run test:core` passes; any existing suite covering a file you touched passes.
+- [ ] Changelog entry present; committed with explicit paths; [PROGRESS.md](event-PROGRESS.md) appended.
## Never blocked
| Blocker | Resolution |
|---|---|
-| Multiplayer server absent locally | `npm run dev:walk-all` boots web + colyseus together; solo-world flows also work offline by design, so verify what you can and say which lane you used. |
-| A defect lives in a file another agent has dirty in `git status` | Fix only if trivial and non-overlapping; otherwise record it in the report as in-flight elsewhere and move on. |
-| A fix needs a server change in `multiplayer/` | That is in scope; the room handlers are in `multiplayer/src/rooms/WalkRoom.js` and the module tests under `tests/` show the patterns. |
+| Load average too high to measure | Wait, or run it later in the session. Reporting a contended run as a defect list is worse than reporting nothing; the 2026-09-02 attempt is the cautionary example. |
+| Dev server dies mid-run | It is shared with every other agent on this worktree. Restart with `npm run dev:walk-all`, discard the poisoned run, rerun. |
+| A defect lives in a file another agent has dirty in `git status` | Fix only if trivial and non-overlapping; otherwise record it as in-flight elsewhere and move on. |
+| A fix needs a server change in `multiplayer/` | In scope. Room handlers are in `multiplayer/src/rooms/WalkRoom.js`; the module tests under `tests/` show the patterns. |
| Playwright missing browsers | `npx playwright install chromium`. |
## Report format
-The defect list with each item marked fixed / out-of-scope-and-why, the journey verification (desktop + 375px, console state), test output verbatim, files committed.
+The three run files, the defect list with each item marked fixed / out-of-scope-and-why, the load average at the start and end of each run, test output verbatim, files committed.
diff --git a/prompts/finish/event-08-event-closeout.md b/prompts/finish/event-08-event-closeout.md
deleted file mode 100644
index e6b292087..000000000
--- a/prompts/finish/event-08-event-closeout.md
+++ /dev/null
@@ -1,92 +0,0 @@
-# EV-08: Event closeout: record what ran, count the souvenirs, recap, retire the pack
-
-**How to run this:** paste this whole file into a fresh Claude Code chat opened in
-`/workspaces/three.ws`, or say "execute `prompts/finish/event-08-event-closeout.md`". Read
-[00-CONTEXT.md](event-00-CONTEXT.md) and `CLAUDE.md` first.
-
-Rewritten 2026-09-01 to its remainder. The original order's first task (freeze the leaderboard
-standings before their one-week Redis TTL) can no longer be done: nobody ran it, the board key
-`event:lb:three-first-meetup` carried `BOARD_TTL_S = 7d` from its last write
-(`api/_lib/event-leaderboard-store.js`), so it expired no later than 2026-08-16 19:30 UTC, and
-`app_settings` holds no event, standing, or leaderboard key (verified by a read-only query).
-The standings are gone. Say so; do not reconstruct them.
-
-## Binding operating clause
-
-1. Finish 100%. Never end with a question or an unexecuted plan. Delete this file when the
- definition of done is verified and log the outcome in [PROGRESS.md](event-PROGRESS.md); any
- remainder gets a follow-up file or an OWNER-ACTIONS row in
- [../production-100/OWNER-ACTIONS.md](production-100-OWNER-ACTIONS.md) first.
-2. **No prize is ever paid by code.** Winner settlement is the owner's decision, and with the
- board gone it is a decision, not a lookup.
-3. Hard rules: no mocks, explicit-path commits, no em-dash characters.
-4. The log reads below have a deadline: Cloud Run keeps logs 30 days by default, so the
- multiplayer log lines from the 2026-08-09 window disappear around 2026-09-08. Run them
- first.
-
-## Step 0: re-derive the state
-
-```bash
-export PATH="$HOME/google-cloud-sdk/bin:$PATH"
-curl -s https://three.ws/api/version # what is live now
-curl -s https://three.ws/event.json | head -c 400 # the explicit no-event state
-git show 5616ff9b8^:public/event.json # the window that ran: 2026-08-09 17:00 to 19:30 UTC
-gcloud run revisions list --service three-ws-api --region us-central1 --project aerial-vehicle-466722-p5 --format='table(name,metadata.creationTimestamp)' | head -20
-gcloud run revisions list --service hyperfy-world --region us-central1 --project aerial-vehicle-466722-p5 --format='table(name,metadata.creationTimestamp)' | head -20
-```
-
-If `gcloud` auth is dead, revive it in-session (`gcloud auth login --no-launch-browser` fed
-through a fifo; the CLAUDE.md never-blocked table names this) before anything else. This
-order is the one case where the cloud read is the deliverable, so route around a dead auth
-by reviving it, not by skipping the read.
-
-Decide which world the event happened in from the revision timestamps: (a) the event build
-(`2bae5d8c2` or later, per the preflight entry in [PROGRESS.md](event-PROGRESS.md)) was serving
-both the API and the multiplayer server before 2026-08-09 17:00 UTC, so the quests, drop,
-and leaderboard were live for visitors; or (b) it was not, and visitors saw the pre-event
-world. Both get an honest closeout.
-
-## Tasks
-
-1. **Count what was granted.** The souvenir grant logs one line per attendee in the
- multiplayer service (`souvenir laurel-meetup -> <account>`, format documented in
- `docs/event-souvenirs.md`). Read the window with
- `gcloud logging read 'resource.type="cloud_run_revision" resource.labels.service_name="hyperfy-world" textPayload:"souvenir laurel-meetup"' --freshness=30d --project aerial-vehicle-466722-p5`
- and record the distinct-account count. Read `/population` and event-job completions the
- same way if the lines exist. In world (b) the count is zero by construction; record that.
-2. **Write the closeout entry** in [PROGRESS.md](event-PROGRESS.md): which world, the revision
- evidence, the grant count, the fact that the standings expired unexported and when, and
- the one-line lesson (the export must be a cron or a post-window hook, never a manual
- order). Add one OWNER-ACTIONS row: winners cannot be settled from the board; the owner
- decides whether to settle from the souvenir list, re-run the event, or close it out.
-3. **Recap for the community, only if the event ran.** In world (a), one `data/changelog.json`
- entry in plain language (what happened, how many attended, what was dropped), tag
- `feature`. In world (b), no entry; a recap of an event nobody could join would be fake.
-4. **Retire the pack.** Verify [02-play-polish-sweep.md](event-02-play-polish-sweep.md) and
- [06-photo-mode-share.md](event-06-photo-mode-share.md) against their deliverables (the
- retirement policy in [../README.md](../README.md)); run whichever still has agent-doable
- work, then delete them and this file. Delete the directory when nothing numbered is left
- and record the retirement in [../README.md](../README.md).
-
-## Definition of done
-
-- [ ] World (a) or (b) determined from revision timestamps, not from memory, and recorded.
-- [ ] Souvenir grant count (distinct accounts) recorded with the exact log query used.
-- [ ] PROGRESS.md closeout entry written; OWNER-ACTIONS row added.
-- [ ] Changelog entry present in world (a) and absent in world (b).
-- [ ] 02 and 06 either shipped and retired or rewritten to their remainder.
-- [ ] `npm run check:rules -- --paths <files you touched>` exits 0.
-
-## Never blocked
-
-| Blocker | Resolution (act, do not ask) |
-|---|---|
-| `gcloud` auth dead | Revive it in-session per the CLAUDE.md playbook. If the Workspace reauth policy refuses non-interactively, write the exact commands into the PROGRESS entry, add an OWNER-ACTIONS row for the read, and finish tasks 2 through 4 with the facts you have. |
-| Logs already rotated (after about 2026-09-08) | Record that the count is unrecoverable and why, with the retention math. That is a complete answer. |
-| `app_settings` schema question | Read-only queries only; this order writes nothing to the database. |
-
-## Report format
-
-World determination with revision names and timestamps; the log query and its distinct
-count; the PROGRESS entry text; the OWNER-ACTIONS row text; which of 02 and 06 were run,
-retired, or rewritten; the commit SHAs. No trailing questions.
diff --git a/prompts/finish/event-PROGRESS.md b/prompts/finish/event-PROGRESS.md
index e606bbcc9..92c8d07e7 100644
--- a/prompts/finish/event-PROGRESS.md
+++ b/prompts/finish/event-PROGRESS.md
@@ -322,3 +322,61 @@ even though `9cff4cb2a` had already committed it. Reset with
in-world browser walkthrough it has always needed, and the box is at load average 62 on 16 cores,
the same contention that defeated the order-04 and order-07 agents; a walkthrough run under that
would not be trustworthy evidence even if it completed. The pack therefore cannot retire yet.
+
+## 2026-09-02 · Order 02 · The in-world walk was driven for the first time, and the box refused to let it be measured
+
+**The residual that has blocked this order since 2026-08-08 is retired: the /play world DOES boot
+headless on this machine.** Both the order-04 and order-07 agents recorded that it would not (`page.goto`
+timeout, `Page crashed`, "the 3D boot OOMs the browser under load") and verified their work against
+harnesses instead. It boots. `scripts/play-journey-audit.mjs` under headless Chromium with
+`--use-gl=swiftshader --enable-unsafe-swiftshader` drove a real cold load, the lobby, search, the create
+modal, the gallery, entry into the $THREE world, the HUD, and the store / bank / wheel / emote panel
+probes. There is no need for anyone to build a new harness for this order; the one in `scripts/` reaches
+every surface the order asks about.
+
+**What it could not do is measure, and the reason is the box, not the code.** Two runs:
+
+1. **Load average 43 on 16 cores.** Reached the world at +332s. The shared Vite dev server on port 3000
+ died at +67s and never came back inside the run, so from there every line is
+ `net::ERR_CONNECTION_REFUSED`: the store panel's lazy chunk, `/avatars/default.glb`, the animation
+ clips, `/api/auth/me`, the world save. A `LOADER NEVER CLEARED` and a "failed to load dynamically
+ imported module" in that run are artifacts of a dead server, not defects, and it would have been easy
+ and wrong to file them.
+2. **Load average 220 on 16 cores.** The harness printed `LOBBY NEVER BECAME VISIBLE` at +125s and then
+ `[state:lobby-first-paint] {"cards":21,"skeletons":0}` at +211s, and printed
+ `GRID NEVER RESOLVED (no cards, no empty state, no error state)` about a grid holding 21 cards. Every
+ verdict in that run is a false negative. Stopped it rather than let it add load and produce a defect
+ list made of contention.
+
+The harness measures with wall-clock waits, so contention does not slow it down, it makes it lie. That
+is now written into the order as a precondition with the numbers above, so the next agent checks
+`uptime` before believing anything it prints.
+
+**What was verified anyway, from the uncontended first minutes:** `[focus:lobby] 114 tabbable, 0 kinds
+without a focus ring, 0 that refuse focus` at 1440x900, 21 coin cards, 0 skeletons, no empty state. That
+agrees with the preflight's independent lobby pass (130 desktop / 91 mobile tab stops, 0 missing rings).
+
+**Two harness readings documented as noise rather than filed as defects,** because a later agent would
+otherwise chase them: the `[overflow:panel:*]` rows for `div.cc-label.ac-name`, `div.ac-prompt`,
+`span.ac-key`, `div.tik-prompt`, `div.npc-prompt` and `div.cc-label.npc-name` are world-space billboard
+labels, so an `x` of 5746 or -287 is them behaving correctly when their subject is off camera; and
+`[panel:bank] no trigger visible` is right, because the bank is proximity-gated behind its ATM by design
+and has no HUD button, as is the Wheel of Fortune behind its station.
+
+**Shipped:** `prompts/finish/event-02-play-polish-sweep.md` rewritten to its remainder. It now carries
+the quiet-box precondition with the measured failure modes, the three exact runs to make (desktop, 375,
+320), the two readings to ignore, and the same task list scoped to the in-world half only, with the
+lobby half marked verified and off limits.
+
+**Remains:** the in-world defect list itself, one quiet-box session. Nothing about it is blocked on a
+credential, a deploy, or an owner.
+
+**Not mine, in flight.** Order 06 is being worked by a concurrent agent in this same worktree right now
+(`1c7410eef` 19:01Z, `37f60a291` 19:07Z, `ad129473d` 19:19Z, `ac68136d4` 19:25Z: the retake fix, unit
+tests, a photo-mode check script, and an e2e driving the real capture path). Left entirely alone. Its
+changelog entry was still absent at 19:26Z and belongs to that agent's run, not to a second hand.
+
+**Order 08 is complete and its file is deleted.** Its closeout is the entry above this one; its two
+blocked cloud reads are OWNER-ACTIONS rows 13 and 15, and the production defect it turned up is row 18.
+Its last task, retiring the pack, is now a line in [README.md](event-README.md), since it can only happen
+after 02 and 06 land.
diff --git a/prompts/finish/event-README.md b/prompts/finish/event-README.md
index 0a13c4a44..533e2cf9e 100644
--- a/prompts/finish/event-README.md
+++ b/prompts/finish/event-README.md
@@ -2,17 +2,19 @@
Work-order pack for the first live community event held in the `/play` world. Each numbered file is a self-contained paste-and-run prompt per the [prompts/ standard](../README.md); [00-CONTEXT.md](event-00-CONTEXT.md) carries the shared facts and MUST be read first by every order.
-The event is over and `public/event.json` is back to its explicit no-event state. Five of the eight orders were verified shipped on 2026-09-01 and retired (their files remain readable in git history); what is left is the closeout and two polish orders that were never formally run.
+The event is over and `public/event.json` is back to its explicit no-event state. Six of the eight orders are retired (their files remain readable in git history). Two polish orders are left, and neither is blocked on anything but a quiet machine.
-| Order | What it ships | State (measured 2026-09-01) |
+**What actually happened, established 2026-09-02 and written up in [PROGRESS.md](event-PROGRESS.md):** the API and frontend shipped the event build ahead of the window, so the countdown, the `/event` page and the in-world meetup layer (agenda, go-live moments, fireworks) ran. The `three-ws-multiplayer` service was never redeployed, so the quest line, the leaderboard and the souvenir grant were absent from the running process. Nothing was granted, no standing ever existed, and there is nothing to settle. All of that code is live on production today and is waiting on the next `public/event.json` window.
+
+| Order | What it ships | State (measured 2026-09-02) |
|---|---|---|
| 01 event countdown | Real event times, countdown on `/play` and the home page | Retired: `src/game/event-countdown.js`, `src/home-event-banner.js`, changelog 2026-08-07 |
-| [02-play-polish-sweep.md](event-02-play-polish-sweep.md) | Full UI/UX defect sweep of the `/play` journey: states, a11y, mobile, copy, perceived speed | Open. The lobby half was verified by the preflight (0 focus rings missing at 130/91 tab stops); the in-world half (store, bank, wheel, jobs, friends, emotes) was never walked and no defect list was produced. |
+| [02-play-polish-sweep.md](event-02-play-polish-sweep.md) | In-world UI/UX defect sweep of `/play`: states, a11y, mobile, copy | Open, rewritten to its remainder. The lobby half is verified twice over and off limits. The world boots headless here after all, and `scripts/play-journey-audit.mjs` reaches every surface, so what is left is one run on a quiet box: the order now carries the load-average precondition and the two harness readings that are noise. |
| 03 event landing page | `/event` with countdown, schedule, .ics, live headcount | Retired: `pages/event.html`, `src/event-page.js`, `api/play/population.js`, changelog 2026-08-08 |
| 04 event quests + leaderboard | Server-authoritative event jobs and a live ranking | Retired: `multiplayer/src/event-window.js`, `event-leaderboard.js`, `api/play/event-leaderboard.js`, 42 tests, changelog 2026-08-08 |
-| 05 event cosmetic drop | Free commemorative wearable granted during the window | Retired: `multiplayer/src/event-drop.js`, `public/accessories/laurel-meetup.glb`, `docs/event-souvenirs.md`, changelog 2026-08-08 |
-| [06-photo-mode-share.md](event-06-photo-mode-share.md) | Photo mode with a branded share card: capture, preview, download, copy | Open. Built (`src/game/photo-mode.js`, HUD button, P key) but never verified on both render engines and never announced: no `data/changelog.json` entry exists. |
+| 05 event cosmetic drop | Free commemorative wearable granted during the window | Retired: `multiplayer/src/event-drop.js`, `public/accessories/laurel-meetup.glb`, `docs/event-souvenirs.md`, changelog 2026-08-08. Shipped as code; never reached a player, see above. |
+| [06-photo-mode-share.md](event-06-photo-mode-share.md) | Photo mode with a branded share card: capture, preview, download, copy | Open and under active work by another agent as of 2026-09-02 19:25Z (retake fix, unit tests, a check script, an e2e through the real capture path). What remains is that agent's: the cross-engine verification and the `data/changelog.json` entry, which is still absent. |
| 07 preflight verification | The event-eve go/no-go sweep | Retired: ran 2026-08-08, recorded NO-GO with per-stage evidence in [PROGRESS.md](event-PROGRESS.md) and `docs/event-readiness/LIVE-OPS.md` |
-| [08-event-closeout.md](event-08-event-closeout.md) | The honest closeout: what ran, what was granted, recap, retire the pack | Open, rewritten to its remainder. The leaderboard's Redis record expired about 2026-08-16, so the standings can no longer be exported; the log-derived facts still can, until roughly 2026-09-08. |
+| 08 event closeout | What ran, what was granted, recap, retire the pack | Retired 2026-09-02. The closeout is the 2026-09-02 entry in [PROGRESS.md](event-PROGRESS.md); no community recap was written and that is deliberate, since nothing was granted and attendance is unknowable. Its two blocked cloud reads are OWNER-ACTIONS rows 13 and 15, and the live defect it found (`MULTIPLAYER_INTERNAL_URL` unset, so the live headcount shows no number) is row 18. |
-Run order now: 08 first (its log reads have a deadline), then 02 and 06 in either order. When all three retire, the pack retires with them: delete this directory and record it in [../README.md](../README.md) the way retired campaigns are recorded there.
+Run order now: 02 and 06 in either order, though 06 has an agent on it. **When both retire, the pack retires with them:** delete this directory and record it in [../README.md](../README.md) the way retired campaigns are recorded there. Nothing else in the pack is waiting on anyone.
diff --git a/prompts/finish/okx-ai-PROGRESS.md b/prompts/finish/okx-ai-PROGRESS.md
index eb5e469e8..843626ea9 100644
--- a/prompts/finish/okx-ai-PROGRESS.md
+++ b/prompts/finish/okx-ai-PROGRESS.md
@@ -6,6 +6,124 @@ Work Order 04 session, no earlier entries existed because no earlier work order
---
+## 2026-09-02, Work Order 04: gauntlet rebuilt for the forge listing, one false promise removed, funding is the only gate
+
+Fourth session on this stream today. WO-04 is the funded end-to-end payment test, so the
+paid half stays gated; everything that does not move money is finished, and the harness that
+runs the moment funding lands was rebuilt, because it would not have worked.
+
+### The gauntlet was buying a listing that no longer exists
+
+`scripts/okx-e2e-gauntlet.mjs` was written against the pre-2026-08-22 line-up: plain-JSON
+REST POSTs to `text-to-3d`, `avatar` and `fbx-export`, reading a `model_url` straight out of
+a 200. The listed product has been seven A2MCP rows since then. A funded run would have
+signed a real authorization, sent an MCP JSON-RPC call the old code never built, and failed
+on the first case with the money already committed. Rewritten around what OKX actually
+sells:
+
+- Every paid case now does the real buyer loop: unpaid `tools/call forge_3d` to 402, sign
+ through the TEE wallet, replay, take the job id, poll the FREE `forge_status` tool to
+ `done`, then verify the GLB bytes. Generation is asynchronous, so a case that stopped at
+ the 200 would have proved nothing.
+- Four paid rows covered individually (`forge-draft` 0.01, `forge-standard` 0.05, `forge-hd`
+ 0.25, `forge-image` 0.25), none "covered by" another.
+- **Case 3r keeps the rigged-artifact assertion alive.** The listed forge rows deliver a
+ static mesh by design, so the work order's "flagship ships a skeleton and skin weights"
+ check would have quietly evaporated with the relisting. It now buys the back-burner
+ `avatar` row ($0.50), which is the row that actually rigs, and verifies bones and non-empty
+ skin weights with `scripts/okx-verify-glb.mjs --rigged`.
+- **Case 1d is new and it FAILS against production**, on all four paid rows: a spec-compliant
+ MCP client (`Accept: text/event-stream` + `MCP-Protocol-Version`) is answered **402 on
+ `initialize` and `tools/list`**, so it can never read a tool description or a parameter
+ schema. That is the sibling session's discovery-paywall fix, already in the tree with unit
+ tests, not yet deployed. The unit tests prove the fix; case 1d is the production-side proof
+ that it has not shipped, and it will flip to PASS on the deploy. Payment is unaffected:
+ `tools/call` is never a discovery method.
+- **A budget preflight that refuses to start.** `--budget` prices a run off the catalog
+ module, and a paid run reads `balanceOf(buyer)` first and exits rather than half-running.
+ A half-funded run burns the cheap cases and then fails the dear ones on balance, which
+ reads exactly like a rail defect and is not one.
+- Case 5d's assertion was rewritten. It judged "no tool ran" by string-matching the response
+ body, and the 402 body legitimately carries the bazaar discovery block whose worked example
+ contains a specimen job id. It now reads the structured result and the receipt header.
+
+Dry run against production: **1, 5d and 7 PASS, 1d FAILS as described**, everything else
+reports "dry run" as designed. Evidence in `prompts/okx-ai/e2e-evidence/`.
+
+### Defect found and FIXED: we told paying buyers a failed job was free to retry
+
+`shapePoll()` in `api/_mcp-studio/studio-shape.js` is shared by the free ChatGPT surface and
+the paid OKX rows. Its failed-job message was written for the free lane and served verbatim
+to both:
+
+> 3D generation hit a snag upstream, it costs nothing to try again.
+
+On the OKX rows that is false about the buyer's money. Settlement happens when the lane
+accepts the job (`anySuccess` in `api/okx/3d/[service].js` judges the submit, which returns
+`pending` with a job id), so a generation failure after acceptance **has been charged**, and
+a retry is a new paid call. Measured, not assumed: over the 30 days to today, **255 of 10,278
+`forge_creations` rows (2.5%) ended `failed`** after acceptance, so this was reaching roughly
+one in forty paying buyers.
+
+`shapePoll` now takes `paid`, the OKX surface passes it, and the free lane's wording and
+response shape are byte-for-byte unchanged (its shape is the published custom-GPT Action
+contract, byte-guarded against its submission source, so a new key there belongs to that work
+stream, not to this fix). The paid message states the settlement model instead, and carries
+`retryBackends`, the alternate engines `/api/gpt-forge` already names on a terminal failure,
+so "an error the buyer can act on" now means something the buyer can act on. Two regression
+tests in `tests/api/okx-forge.test.js`; the free-surface tests in `tests/api/3d-studio.test.js`
+still pass untouched, which is the point.
+
+`docs/okx-marketplace.md` carried the same false claim ("Generation failures are free to
+retry"). Corrected in both places it appears, with the measured failure rate stated. The
+on-chain listing copy and every catalog description already said the accurate thing ("You pay
+only when a job is accepted"), so the doc and the runtime message were the outliers, not the
+promise. **This is the work order's "fix the code or fix the promise" release blocker, and
+both halves are done.** Ships on the same owner-gated deploy.
+
+Also fixed: `scripts/okx-listing-payload.mjs --delta` died inside `JSON.parse` when run
+without its stdin pipe, which reads as a broken script rather than a missing pipe. It now
+names the pipe it wants.
+
+### Verified live today (not re-deriving the sibling sessions, adding to them)
+
+- Wallet **logged in** as `[email protected]`, buyer address confirmed unchanged at
+ `0x75d00a2713565171f33216e5aa2a375e076ecf69`. The OTP owner action from every earlier
+ version of the funding request is discharged.
+- Three-copy check PASS (module 16 rows / 7 listed == live == listing submission), and
+ `--delta` against the live `service-list` is **7 updates, 0 creates, 0 deletes**: the
+ on-chain listing is exactly our catalog.
+- Balances at block 69607441: buyer **0 USD₮0, 0 OKB**; payTo 2.427731 USD₮0, 0.839596 OKB;
+ relayer 0 USD₮0, 0.02 OKB. Gas priced at 0.02 gwei, so one settle costs 0.000002 OKB and
+ the relayer covers roughly 10,000 of them. The buyer needs no OKB.
+- The buyer's Solana account is `9PirGw9wVLLNFgVyjgAt5jvuFQwJ3pYUBWt9n3vZfnyc` and its Base
+ account is the same EVM address; both are empty. The Solana accept carries a `feePayer`, so
+ case 7's real paid legacy leg needs the fee and no SOL.
+
+### Blocked, and it is exactly one thing
+
+**Funding the buyer.** `prompts/okx-ai/e2e-evidence/FUNDING-REQUEST.md` is rewritten against
+the current catalog: **5.0 USD₮0 on X Layer (196) to `0x75d0…cf69`**. A clean run settles
+$1.07 and needs a starting float of **$1.32** (the floor, not the sum: verify refuses any
+authorization above `balanceOf`, including the ones designed to be rejected); the rest is the
+fix loop. The money largely returns, because the buyer pays our own merchant wallet, so the
+net platform cost of a full run is the gas. `payTo` holds 2.427731 USD₮0 and could fund it
+without an external transfer, but its key is in Secret Manager and `gcloud` auth is expired
+on this box, so either route is an owner action.
+
+### Next
+
+1. Owner funds the buyer. Then `node scripts/okx-e2e-gauntlet.mjs --yes` runs cases 1, 1d, 2,
+ 2b, 3, 3i, 3r, 5a, 5b, 5c, 5d, 6, 7 and finally case 4 (on-chain verification of every
+ settlement the run produced), with the fix loop per phase 3.
+2. Case 1d stays red until the pending deploy ships. That deploy now carries three fixes from
+ three sessions today: the discovery paywall, the x402 v1 header string, and this session's
+ failed-job wording.
+3. `docs/okx-marketplace.md` gets its "verified behavior" section, and its "Not yet
+ demonstrated end to end" note comes out, when the first tx hash exists. Not before.
+
+---
+
## 2026-09-02, Work Order 07: independent audit, docs closure, memory; the 402 was still speaking x402 v1
Third session on this stream today. Ran WO-07 against production without reading the other
diff --git a/prompts/finish/production-100-00-INDEX.md b/prompts/finish/production-100-00-INDEX.md
index 448c4c5da..702d178e2 100644
--- a/prompts/finish/production-100-00-INDEX.md
+++ b/prompts/finish/production-100-00-INDEX.md
@@ -109,12 +109,13 @@ Strategy for what to run next inside E: [../roadmap/fable-playbook.md](roadmap-f
|---|---|
| [04-fact-check-mixed-verdicts.md](production-100-04-fact-check-mixed-verdicts.md) | Open; live run 2026-08-10 scores `mixed` 0/10 and 40% overall; verdict thresholds untouched since 2026-07-28. |
-### G. Event (closeout only)
+### G. Event (two polish orders)
| Order | State |
|---|---|
-| [../event/08-event-closeout.md](event-08-event-closeout.md) | Run first in this pack: the souvenir grant count is readable from the multiplayer logs until about 2026-09-08; the standings board expired unexported. |
-| [../event/02-play-polish-sweep.md](event-02-play-polish-sweep.md), [../event/06-photo-mode-share.md](event-06-photo-mode-share.md) | The in-world walkthrough was never done; photo mode was built but never verified on both engines or announced. |
+| Closeout | Done 2026-09-02, order file retired. The world server never shipped the event build, so nothing was granted and no standing ever existed: [../event/PROGRESS.md](event-PROGRESS.md). Its remainder is OWNER-ACTIONS rows 13, 15 and 18. |
+| [../event/02-play-polish-sweep.md](event-02-play-polish-sweep.md) | Open, rewritten to its remainder. The world does boot headless here and the existing harness reaches every surface; what is left is one run on a quiet box. |
+| [../event/06-photo-mode-share.md](event-06-photo-mode-share.md) | Open, with another agent actively on it as of 2026-09-02. Cross-engine verification and the changelog entry remain. |
### H. Distribution and listings (owner- or commit-gated)
diff --git a/prompts/finish/production-100-04b-fact-check-publish-run.md b/prompts/finish/production-100-04b-fact-check-publish-run.md
new file mode 100644
index 000000000..30d929d48
--- /dev/null
+++ b/prompts/finish/production-100-04b-fact-check-publish-run.md
@@ -0,0 +1,104 @@
+# P100-04b: The `mixed` fix is written and tested. Nothing has measured it yet.
+
+**How to run this:** paste this whole file into a fresh Claude Code chat opened in
+`/workspaces/three.ws`, or say "execute `prompts/finish/production-100-04b-fact-check-publish-run.md`".
+Read [00-INDEX.md](production-100-00-INDEX.md) and `CLAUDE.md` first.
+
+This is the remainder of P100-04, whose code half landed on 2026-09-02. That order's
+diagnosis and fix are done and unit-tested; what is left needs credentials this workspace
+does not have and a deploy only the owner can approve.
+
+## Binding operating clause
+
+1. Finish 100%. Never end with a question or an unexecuted plan. Delete this file when the
+ definition of done is verified.
+2. **Never publish a number you would not defend.** Respect the degraded-check guard and
+ the 10% error-rate ceiling in `api/_lib/fact-check-benchmark.js`. Never weaken either to
+ make a run publishable. A run that regresses gets diagnosed, not published.
+3. Hard rules: no mocks, no fixture-fitting, explicit-path commits, no em-dashes.
+
+## What already shipped (do not redo it)
+
+- **`partial` is a fourth stance.** `agents/fact-checker/src/llm-verdict.js` exports
+ `STANCES` and its rubric asks for `partial` when a source finds the claim true in one
+ respect and wrong, overstated, or only conditionally true in another. Before this, the
+ rubric explicitly told the model to collapse such a source onto one side, which made the
+ `mixed` verdict unreachable from evidence that agrees with itself.
+- **`computeVerdict` counts partial weight** as stance-bearing evidence that takes neither
+ side (`api/x402/fact-check.js`). With zero `partial` sources its output is identical to
+ the old function, which is the anti-seesaw guarantee.
+- **The LLM response reader no longer fails silently.** `extractJsonArray` replaces a
+ non-greedy regex that stopped at the first `]`; both stages now report
+ `stance extraction unreadable` / `query generation unreadable` instead of returning a
+ fabricated all-neutral `insufficient`.
+- Tests: `tests/api/fact-check-verdict.test.js` (synthetic stance distributions, clear-cut
+ pins) and `tests/fact-check-degradation.test.js` (extraction shapes, degradation
+ contract). Docs: `docs/fact-check.md`, `agents/fact-checker/README.md`. Changelog: two
+ 2026-09-02 entries.
+
+## Step 0: re-derive the state
+
+```bash
+curl -s https://three.ws/api/version # is the fix live yet?
+curl -s https://three.ws/api/fact-check-benchmark | head -c 1200
+git log --oneline -5 -- api/x402/fact-check.js agents/fact-checker/src/
+npx vitest run tests/api/fact-check-verdict.test.js tests/fact-check-degradation.test.js
+```
+
+If `/api/fact-check-benchmark` already reports a run generated after the fix landed with
+`mixed` above zero, this order is spent: log the outcome in
+[PROGRESS.md](production-100-PROGRESS.md) and delete this file.
+
+## Tasks
+
+1. **Get an LLM lane.** The benchmark's in-process runner needs one, and on 2026-09-02
+ this workspace had none: no `GROQ_API_KEY` / `OPENROUTER_API_KEY` / `NVIDIA_API_KEY` in
+ `.env` or `.env.local`, and `gcloud` refused every call with "Reauthentication failed.
+ cannot prompt during non-interactive execution", which also takes out the Vertex Gemini
+ anchor (`GOOGLE_CLOUD_PROJECT` plus ADC). Restore `gcloud` auth first; that alone
+ unlocks Vertex, and it unlocks reading the rest off the service env
+ (`gcloud run services describe three-ws-api --region us-central1 --project
+ aerial-vehicle-466722-p5 --format=yaml`). Search needs no key: Wikipedia and DuckDuckGo
+ are keyless rungs, though a `BRAVE_API_KEY` / `TAVILY_API_KEY` materially improves the
+ evidence the verdict sees.
+2. **Check the production chain is not itself degraded before spending a run.** On
+ 2026-09-02 a live free-lane check returned
+ `query generation unavailable: openai 429 billing_not_active`, meaning every free rung
+ had already failed and the paid backstop is dead. A benchmark run in that state measures
+ provider availability, and the guard will refuse it. Fix the lane first
+ (`docs/ops/llm-lanes.md`).
+3. **Run the suite in-process, cache disabled, and capture per-claim detail:**
+ ```bash
+ FACT_CHECK_DETAIL_FILE=/tmp/fc-detail.json \
+ node --env-file=.env scripts/fact-check-benchmark.mjs --in-process
+ ```
+ That writes `data/_generated/fact-check-benchmark.json` without publishing. Compare its
+ `by_class` against the last published run (`mixed` must leave zero, no class may
+ regress) and read `/tmp/fc-detail.json` for the per-source stances behind any class that
+ moved the wrong way.
+4. **Publish only an improved, non-degraded run:** re-run with `--publish`. The live
+ `/api/fact-check-benchmark` reads the DB row first, so this reaches the public page
+ without a deploy, but the CHAIN itself only improves once the code is deployed, so
+ sequence it after the ship.
+5. **Verify the page in a real browser:** `/fact-check` renders the run, denominators, and
+ the per-class and per-difficulty tables.
+6. **Log the outcome** in [PROGRESS.md](production-100-PROGRESS.md) with the before/after
+ per-class table, and add a `data/changelog.json` entry (tag `improvement`) naming the
+ new headline number.
+
+## Definition of done
+
+- [ ] A real, non-degraded in-process run exists with a before/after per-class table.
+- [ ] `mixed` is above zero and no class regressed; if one did, the diagnosis is written
+ down and the run is NOT published.
+- [ ] An improved run published to the DB and rendered on `/fact-check` in a browser.
+- [ ] Changelog entry written, outcome logged in PROGRESS.md, this file deleted.
+
+## Never blocked
+
+| Blocker | Resolution |
+|---|---|
+| No LLM key anywhere on the machine | Restore `gcloud` auth for the Vertex anchor, then read the service env. If a key genuinely exists nowhere, that is one line for OWNER-ACTIONS naming the single missing variable, not a reason to leave the rest undone. |
+| The deploy is owner-gated | Prepare it so the ship is one command (`npm run deploy:gcp:full` after `npm run prep:worktree -- --apply`) and say so. Publishing the run does NOT need the deploy, but running it against the OLD chain measures the old code, so do not publish that as the fix's number. |
+| Accuracy improves on `mixed` but a neighbour regresses | That is the seesaw the calculus was designed against. Read `/tmp/fc-detail.json` for the regressed claims' stances: a `partial` where the source plainly affirms or plainly refutes is a rubric problem, not a calculus one. |
+| The run is refused as degraded | Working as designed. Fix the lane, then re-run. Never lower `MAX_ERROR_RATE`. |
diff --git a/prompts/store-submissions/_generated/TRACKER.md b/prompts/store-submissions/_generated/TRACKER.md
index a1abcbae9..949e1c4cd 100644
--- a/prompts/store-submissions/_generated/TRACKER.md
+++ b/prompts/store-submissions/_generated/TRACKER.md
@@ -17,8 +17,8 @@ Legend: **not-started** | **blocked** | **ready-to-submit** | **submitted** | **
| **Claude Connectors Directory** | blocked | 02, 03 | Prompt 02 deliverables complete (`claude-reviewer-guide.md`, `claude-tool-call-evidence.md`, real review-mode entitlement in `mcp-server/src/payments.js`) — 16/16 paid tools return clean `PaymentRequired` unpaid. **New (2026-07-08 re-audit):** live re-test found the production free lane (`forge_free`/`POST /api/forge`) hanging 90–100s+ with no response on 4/4 consecutive attempts — re-verify green immediately before filing; see evidence file §"2026-07-08 re-verification" for the diagnosis (NIM `SUBMIT_TIMEOUT_MS`/fallback-chain latency, not a code defect in the reviewer path itself). `[HUMAN: Team/Enterprise org + Owner role required for in-app portal; OR use public form]` · Privacy policy must be deployed with MCP section (§9 of answer sheet) · `[HUMAN: generate a real `MCP_REVIEW_SECRET` value + decide which vendor credentials (e.g. `REPLICATE_API_TOKEN`) to hand a reviewer for the funded path, then include both in the submission's private reviewer notes — mechanism is built and working, only the actual secret provisioning + credential-sharing decision remains]` | — |
| **Claude plugin marketplace** | ready-to-submit | 10 | Verified 2026-07-08: 4 plugins (`three-ws-core`, `three-ws-developer`, `three-ws-pump-fun`, `three-ws-3d`) all validate `--strict` clean (fixed a real `category`-field-in-plugin.json defect found by the validator) and were installed end-to-end via the real Claude Code CLI (`claude plugin marketplace add ./` → `claude plugin install <id>@three-ws --scope project`, all 4 succeeded, all `enabled:true`). The 3 bundled MCP servers (`@three-ws/mcp-server`, `@three-ws/scene-mcp`, `@three-ws/avatar-mcp`) were driven through a real MCP `initialize`→`tools/list` handshake against their published npm packages and returned their real tool surfaces (19, 3, 3 tools). Evidence: `_generated/claude-plugin-marketplace-e2e.md`. `[HUMAN: submit `three-ws-3d` (and optionally the other 3) via `https://platform.claude.com/plugins/submit` or the claude.ai admin directory submissions page per `marketplace/plugins/three-ws-3d/SUBMISSION.md` — no API for this step]` | — |
| **Claude Agent Skills pack** | ready-to-submit | 11 | Pack complete and independently re-audited 2026-07-08: 40 skills in `.agents/skills/`, all 3 3D-creation skills present with endpoint shapes verified against `api/_mcp-studio/tools.js` + `mcp-server/src/tools/*` (real code, not assumed), manifest regenerates clean (`node scripts/build-skills-pack.mjs --check`), 3d/creative subset greps zero coin/wallet/x402 strings, install-path docs at `docs/agent-skills.md`, `STRUCTURE.md` row present, `data/changelog.json` entry present, `npm run build:pages` passes. **Live E2E still blocked** as of this re-audit (2026-07-08 01:45 UTC): `forge_free` MCP calls timed out at 90s (×2) and a raw `POST /api/mcp-studio` timed out at 60s — same root-cause P0 `readBody()` hang noted at the top of this file; code fix has landed (commit `ba37182f3`), prod re-verification against the current live revision remains. Nothing left to build; re-run the smoke test against the live revision. | — |
-| **OpenAI App Directory** | ready-to-submit | 04–06 | Package + audit done (`_generated/openai-submission.md`, 7/7 policy PASS; 3 real-model screenshots). **B1 RESOLVED (verified live 2026-07-14):** real `POST /api/mcp-studio tools/call forge_free` completed a full generation against production (1.6 MB GLB on R2, `model/gltf-binary`); the `readBody()` fix (commit `ba37182f3`) is serving. Note: `forge_free` input schema now rejects extra properties, send `{"prompt": "..."}` only. **B2 RESOLVED (verified live 2026-07-14):** `https://three.ws/viewer?src=<glb>` returns 200 and serves the studio viewer; emitted `viewerUrl` links are live. **Partner Network: ACCEPTED 2026-07-14** (welcome email to [email protected]). **Identity verification: COMPLETE** (verified organization on platform.openai.com, owner-confirmed 2026-07-14). Submission answer sheet finalized 2026-07-14: support contact `[email protected]` + `/support` (200), privacy `/legal/privacy` (200), tool list/annotations/schemas refreshed to the live 9-tool surface, fresh E2E (1.45 MB GLB + viewer 200). `[HUMAN: final submit in the partner portal / App Directory flow]` | — |
-| **OpenAI GPT Store** | live | 12 | **PUBLISHED publicly 2026-07-14** (Share → Everyone): `https://chatgpt.com/g/g-6a563a3b49a88191abf346245491a444-three-ws-3d-studio` returns HTTP 200 to a logged-out visitor. Builder profile Name attribution set to "three.ws"; domain verification for the clickable byline link was completed via the Manage Domain flow (the modal initially rendered blank; reopening it fixed that). NOTE: an earlier draft GPT exists at `g-6a5672fbf3f48191b559e482c7fcbf51` (was 404 logged-out / Only me): `[HUMAN: delete the draft duplicate in My GPTs to avoid confusion]`. Actions lane E2E-verified same day (real GLBs both contract branches, safety gate live, compliance grep clean). Tier story (re-verified live 2026-08-06, `_generated/forge-free-tier-evidence.json`): the lane **defaults to the standard tier** on both the Actions and MCP surfaces, and the platform-funded high tier is an opt-in `tier:"high"` the caller can request (a `d97e8e94d` attempt to make high the default was rolled back in `b39d6e2f7`). | https://chatgpt.com/g/g-6a563a3b49a88191abf346245491a444-three-ws-3d-studio |
+| **OpenAI App Directory** | held (B3 output quality) | 04-06 | Package + audit done (`_generated/openai-submission.md`, 7/7 policy PASS; 3 real-model screenshots). **B1 RESOLVED (verified live 2026-07-14):** real `POST /api/mcp-studio tools/call forge_free` completed a full generation against production (1.6 MB GLB on R2, `model/gltf-binary`); the `readBody()` fix (commit `ba37182f3`) is serving. Note: `forge_free` input schema now rejects extra properties, send `{"prompt": "..."}` only. **B2 RESOLVED (verified live 2026-07-14):** `https://three.ws/viewer?src=<glb>` returns 200 and serves the studio viewer; emitted `viewerUrl` links are live. **Partner Network: ACCEPTED 2026-07-14** (welcome email to [email protected]). **Identity verification: COMPLETE** (verified organization on platform.openai.com, owner-confirmed 2026-07-14). Submission answer sheet finalized 2026-07-14: support contact `[email protected]` + `/support` (200), privacy `/legal/privacy` (200), tool list/annotations/schemas refreshed to the live 9-tool surface, fresh E2E (1.45 MB GLB + viewer 200). **HELD 2026-09-02 (task 07 final verification):** every protocol surface re-verified live and green (`initialize`, `tools/list` = 11 tools, `resources/list` = 2 widget resources, a real keyless `forge_free` returning a 2,256,144-byte GLB, the served OpenAPI lint-clean, `/api/ar` correct per device class, the widget's four states in a `window.openai`-less browser, and `POST`+`GET /api/3d/studio` validating against the served schema). Submission is held on ONE new product blocker, B3 in `openai-submission.md` §0: the free lane intermittently returns a degenerate slab (6 of 40 consecutive live generations) that `scoreGlbQuality` rated 0.976 / `ok` and shipped without vision QA. Scorer fix landed in `api/_lib/glb-quality.js`; it needs a deploy plus one clean re-verification generation. `[HUMAN: final submit in the partner portal / App Directory flow]` | |
+| **OpenAI GPT Store** | live | 12 | **PUBLISHED publicly 2026-07-14** (Share → Everyone): `https://chatgpt.com/g/g-6a563a3b49a88191abf346245491a444-three-ws-3d-studio` returns HTTP 200 to a logged-out visitor. Builder profile Name attribution set to "three.ws"; domain verification for the clickable byline link was completed via the Manage Domain flow (the modal initially rendered blank; reopening it fixed that). NOTE: an earlier draft GPT exists at `g-6a5672fbf3f48191b559e482c7fcbf51`. **Escalated 2026-09-02: it is no longer private.** It was 404 logged-out / Only me when this note was written; a logged-out fetch now returns HTTP 200 and renders the same title, "ChatGPT - three.ws 3D Studio", as the canonical listing. Two publicly reachable GPTs under one name is a live confusion risk to resolve BEFORE the App Directory submission: `[HUMAN: delete the draft duplicate in My GPTs to avoid confusion]` (needs the owner's ChatGPT My GPTs login; there is no API for it). Actions lane E2E-verified same day (real GLBs both contract branches, safety gate live, compliance grep clean). Tier story (re-verified live 2026-08-06, `_generated/forge-free-tier-evidence.json`): the lane **defaults to the standard tier** on both the Actions and MCP surfaces, and the platform-funded high tier is an opt-in `tier:"high"` the caller can request (a `d97e8e94d` attempt to make high the default was rolled back in `b39d6e2f7`). | https://chatgpt.com/g/g-6a563a3b49a88191abf346245491a444-three-ws-3d-studio |
| **Official MCP Registry** | ready-to-submit | 13 | `server.json` manifests verified clean; `[HUMAN: Run mcp-publisher commands (staged in prompt 13 output) — do not publish without explicit operator approval]` | https://registry.modelcontextprotocol.io/?q=io.github.nirholas |
| **Smithery** | ready-to-submit | 13 | `_generated/mcp-directories/smithery.md` written (214 lines, real per-server listing copy) | — |
| **Glama** | blocked | 13 | `_generated/mcp-directories/glama.md` written; repo-root `glama.json` and `public/.well-known/glama.json` now committed (Glama auto-ingests `io.github.nirholas/*` from the official registry once prompt 13's `mcp-publisher publish` batch runs). `[HUMAN: sign in to Glama with the `nirholas` GitHub account to claim/verify the listing]` | — |
diff --git a/prompts/store-submissions/_generated/openai-submission.md b/prompts/store-submissions/_generated/openai-submission.md
index 8985970c2..7491e59c6 100644
--- a/prompts/store-submissions/_generated/openai-submission.md
+++ b/prompts/store-submissions/_generated/openai-submission.md
@@ -368,11 +368,25 @@ The landscape shots must show the **shipped ChatGPT inline widget** (`api/_mcp-s
resource `ui://widget/three-studio-model.html`) rendering the GLB produced by a real `forge_free` call
— not a standalone page and not a mockup.
-`[HUMAN: re-capture the landscape widget shots against the shipped inline widget (render a real
-forge_free GLB in the ChatGPT-app sandbox, or the widget resource loaded standalone), and confirm the
-App Directory form's exact required screenshot dimensions/aspect. Earlier captures were taken against a
-now-removed standalone viewer bundle and must not be reused. The widget renders any GLB at any
-viewport.]`
+**Status 2026-09-02 (task 07).** Two of the three sub-items are now resolved, one is blocked:
+
+- **Resolved: no stale viewer references.** Nothing in the kit points at the retired
+ `https://three.ws/apps-sdk/` viewer as "the widget". The only `apps-sdk/` strings left in
+ `_generated/` are OpenAI's own developer-docs URLs and a repo source path, neither of which is a
+ capture target.
+- **Resolved: the capture harness works.** The shipped widget resource was driven end to end in a
+ `window.openai`-less Chromium off a real live tool payload and reached its ready state with
+ `<model-viewer>.loaded === true`, so a re-capture is a mechanical step, not an open question.
+- **Blocked on B3 (§0): no submission-grade model to photograph.** The three generations run for this
+ checklist returned, respectively, a flat slab, an untextured blob, and a recognizable cat carrying a
+ baked-in ground plane and a floating backdrop panel. A screenshot is the single most scrutinized
+ artifact in the listing and must not advertise the defect. **Re-capture once the B3 fix is deployed
+ and a generation comes back clean.**
+
+`[HUMAN: confirm the App Directory form's exact required screenshot dimensions and aspect ratio. The
+current files are 1440x1520 portrait, 1600x1000 and 1280x800 landscape; only the portal states the
+requirement, so this cannot be checked from here. The widget renders any GLB at any viewport, so
+re-capturing to whatever the form asks for is trivial once B3 clears.]`
---
@@ -439,9 +453,10 @@ curl -s -X POST https://three.ws/api/mcp-studio -H 'content-type: application/js
- [x] Served discovery schema is live: `https://three.ws/.well-known/3d-studio-openapi.yaml` returns
200 in production, as do `/.well-known/ai-plugin.json`, `/legal/privacy`, `/legal/tos` and
`/support` (re-verified 2026-09-02).
-- [ ] Screenshots — **re-capture the landscape widget shots against the shipped inline widget**
- (`api/_mcp-studio/component.js`); the earlier standalone-viewer captures were removed as part of
- collapsing the duplicate viewers. Confirm the form's required dimensions. `[HUMAN]`
+- [ ] Screenshots: re-capture against the shipped inline widget (`api/_mcp-studio/component.js`).
+ Verified 2026-09-02 that no evidence file still cites the retired `apps-sdk/` viewer, and that the
+ widget drives correctly for capture. Held on B3: no generation clean enough to photograph yet (§4).
+ Required dimensions still need the portal. `[HUMAN: confirm required dimensions]`
- [x] Inline widget re-verified live 2026-09-02 in a `window.openai`-less Chromium: the skybridge
resource paints its empty state ("No model yet"), accepts a real tool payload over `postMessage`,
shows the generating state, and reaches ready with `<model-viewer>.loaded === true` on the real