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(changelog): regenerate the public feeds from the current entry list

nirholasd869ed716dae
diff --git a/CHANGELOG.md b/CHANGELOG.md
index f1e6186c6..6f52e6d98 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -19,6 +19,7 @@ Public history for [three.ws](https://three.ws), newest first. New pages come fr
- **A database blip during a scheduled job now reads as a database blip, not a healthy run**: Two of the scheduled jobs behind the trading fleet could fail without anything looking wrong. The job that reallocates budget across strategies talks to the database over a socket, and when that socket never opened the failure arrived as an object carrying no message at all. The job caught it, answered with an empty error string, and the scheduler recorded a success, so the health record stayed green while the loop did nothing. Failures of that kind are now recognised as what they are, a database that could not be reached, so the run is marked unhealthy and retried on the next tick instead of being logged as fine. Every page and endpoint that reads from the database inherits the same correction: this class of outage now returns a retryable service-unavailable response rather than an internal error. The fleet watchdog also gained a seventh check, over the calibration pass that measures whether a high-conviction coin actually wins as often as its score claims, so that loop can no longer go quiet unnoticed. `[fix, infra]`
- **A dedicated channel for first-time pump.fun creator fee claims**: The first time a coin creator pulls their accrued fees out of the vault is an irreversible on-chain signal that they are actually engaged with what they launched, and it was only visible by polling an API. There is now a delivery lane that posts each of those first claims to its own Telegram channel as it lands, one message per claim, with the creator, the amount in SOL, and links to the transaction and the coin. It keeps its own state so switching it on never dumps a backlog into the channel, it drains a backlog in order rather than skipping ahead if it falls behind, and it recognises claims by transaction rather than by timestamp, so two claims landing in the same second both get posted instead of one being silently dropped. The feed stays quiet until a claims indexer is configured: the direct on-chain fallback samples under a second of activity on a program this busy, which is not enough to tell a first claim from a repeat one. `[feature, infra]`
- **A failed database migration can no longer freeze the schema in silence**: Every deploy carries database migrations that must run before the new code does. When one of them failed, the runner printed a single line with the database's own error and stopped, naming neither the file that failed nor the migrations queued behind it. Those stayed unapplied and unmentioned, so the site ended up serving code that reads columns the database never got, and the first visible symptom was an unrelated-looking error on a live page. The runner now names the failing file, how many ran before it, and every migration still waiting, and the deploy step that catches a database running behind the code is now part of the documented deploy path instead of something you had to remember to run. `[infra, fix]`
+- **A generation that quietly switched engines no longer counts as a failure**: When a 3D generation lane loses your job, the platform resubmits it to another lane and you still get your model. Until now the abandoned attempt was recorded as a plain failure, with nothing tying it to the run that actually succeeded, so the platform's own health readout treated its recovery machinery as an outage: over the past week, 13 of the 14 lost jobs were rescued within seconds and all 14 still counted against generation health. Each recovered attempt is now linked to the run that replaced it, so the status page and our failure reports show what you experienced instead of what the machinery did. (`/status`) `[improvement, infra]`
- **A paid-call share link now previews on Slack, Telegram and LinkedIn**: Every settled x402 call has a permalink whose social preview image is drawn on demand. Slack, Telegram and LinkedIn ask an image URL for its type and size before they fetch it, and this one answered "method not allowed" to that probe, so the card silently never appeared on those platforms. It answers them now. Three more fixes to the same card: a receipt is captioned with the network it actually settled on rather than always reading mainnet; a link shared in the seconds before its receipt finishes saving, or during a cache outage, is retried a minute later instead of being pinned to the generic card for an hour; and a damaged record shows a plain "unknown" rather than a "NaNd ago" timestamp or an invented amount. (`/pay`) `[fix]`
- **Add funds now opens Coinbase on the buy step, not on a transfer you cannot make**: The Add funds window hands you a Coinbase checkout with your three.ws wallet already filled in as the destination. That checkout was not told which step to open on, so Coinbase could land you on "transfer from your Coinbase balance", which is no help at all if you reached Add funds precisely because you hold nothing yet. It now always opens on the buy step for the asset you picked (USDC or SOL, both settling on Solana). The suggested dollar amount is also rounded to real dollars and cents before it reaches Coinbase, so an amount worked out from a shortfall can no longer show up as something like $25.777777 on the checkout screen or in the window itself. `[fix, improvement]`
- **AgenC task lists no longer hang when a Solana provider goes quiet**: Browsing tasks on the AgenC coordination protocol reads directly from Solana, and the platform keeps a chain of providers so a bad one gets skipped. That skipping only worked when a provider answered with an error. A provider that accepted the request and then went silent held the page open for minutes with nothing to react to, which is what happened to the task list, the heaviest read of the set. Each provider now gets a fixed time to answer before the next one is tried, so a silent provider costs seconds instead of stalling the page. (`/agenc`) `[fix]`
@@ -27,6 +28,7 @@ Public history for [three.ws](https://three.ws), newest first. New pages come fr
- **Agents that discover us the documented way can now reach the wallet tools**: An AI client that finds three.ws on its own follows a trail: it calls a tool, gets told it needs to sign in, and reads a small public document that lists the permissions it is allowed to ask for. That document had fallen behind. It advertised eight permissions while the sign-up endpoint actually grants twelve, and the three it left out were exactly the ones that unlock the agent wallet tools, so a client doing everything correctly could never ask to check a balance, pay for a service, provision a wallet, or put an endpoint up for sale. The published list and the list the server honors are now read from one place, with a test that fails the build if they ever drift apart again. Separately, the paid-call example an agent receives with the payment request pointed at a sample model file on a domain that does not exist, so the very first call a paying agent copied failed on a dead address; it now points at a real model served from three.ws. Every discovery endpoint was re-checked end to end afterwards. (`/docs/mcp-agent`) `[fix, security]`
- **An Oracle coin page now tells you what happened after the call, and stops letting a score be read as a percentage**: A conviction of 100 sitting above a chart that had already spiked and collapsed looked like a broken engine. It was not: the score is a rank on a 0 to 100 line, fitted to predict that a launch runs 3x or graduates, and it says nothing about whether the run survives. Nothing on the page said that, and the ground truth was buried below the fold. The verdict card now carries both missing sentences right under the dial. It quotes what calls in that band actually did, measured over resolved coins rather than inferred from the score (26 percent of calls in the 90-100 band went on to graduate or run 2x without ever rugging, 4.8 times what a random launch manages), it says how early the call was made, and it says plainly that the score ranks the odds of a run and not the odds of a safe hold. Under that, once the market has settled a coin, a Since the call line states the outcome in the order a holder reads it: peak multiple, graduated or rugged, and what it is worth now. Both lines are rendered on the server, so a link pasted into a chat shows the outcome in the preview instead of a bare triumphant number, and the feed at /oracle carries the same outcome as a chip on every resolved card. The accuracy pages were reporting three numbers wrong in the same way. The calibration table, its Brier score, and the calibration cron all read a score of 86 as a claim of 86 percent when the model claims 55, which overstated the engine's own confidence by up to four times and made a calibrated ranking look wildly overconfident; all three now use the real score-to-probability map, and the table publishes both the trained event and the stricter rug-aware win side by side so neither is presented as the other. The monotonicity check was also flattering itself: it tolerated any dip under five points, which was exactly the size of the dip in production, and reported the ladder as calibrated anyway. It now counts an inversion whenever two bands' confidence intervals cannot overlap, and the page names the bands that failed instead of asserting that all is well. (`/oracle`) `[fix, improvement]`
- **Another site can no longer swap your agent's body or burn your avatar credits behind your back**: The two avatar onboarding endpoints (the one that opens the photo editor, and the one that attaches a finished avatar to an agent) accepted a request that arrived with your login cookie but no proof you meant to send it. A page you had merely visited could therefore repoint your agent to a different body, or spend your hourly avatar-editor allowance, without you doing anything. Both now require the same one-time token every other account-changing action on three.ws already asks for, and both refuse the request outright without it. API-key callers are unaffected. A second change: when the photo editor is not configured in an environment, the request is turned away before it counts against your hourly allowance, so a temporary outage no longer locks you out of the flow for an hour once it is back. `[security, fix]`
+- **Anyone can run an inference node: the operator client now ships CPU and GPU images, and proves itself end to end**: The open inference network is the part of the roadmap where the compute stops being ours. The node operator client is how you join it: your machine generates a Solana keypair, registers itself with the platform, claims inference jobs, runs them locally with a real open model, and returns every answer with a signature that binds it to your key. There is no shared password anywhere in that flow, so nothing can be leaked and reused to impersonate your node, and the platform re-derives each signature from its own record of the job before it accepts the work. This release makes the graphics-card story honest. The old image claimed it would use an NVIDIA card automatically when one was present, and it never could: the runtime it shipped had no GPU support compiled in at all. There are now two real images, one for CPU and one for NVIDIA cards, and the client tells you which device actually ran each job. Ask for the GPU on a machine that cannot reach one and it stops immediately with the reason, instead of quietly earning at CPU speed for a month. Three faults in the job pipeline were also fixed: results submitted by a node hit a server error every time, registered nodes read back as unregistered, and every finished job leaked two silent background errors. A single command (npm run e2e) now registers a node, sends it a job, runs the model, and verifies the signed answer against the real platform code, and it fails loudly if any step regresses. (`/docs/inference-node-operator`) `[feature, fix, infra]`
- **Apps that sign in with a client secret can now revoke and inspect their own tokens**: If you connect an app to three.ws over OAuth and registered it with a client secret sent in the Authorization header (client_secret_basic, the standard choice for server-side apps), only the token endpoint accepted those credentials. Revoking a token or checking whether one is still valid answered that the client_id was missing, and answered that the credentials were bad the moment you added it, so a confidential app had no way to sign a user out or verify a token it owned. All three endpoints now take the same credentials, in the header or in the form body, exactly as the OAuth revocation and introspection standards require. Dynamic client registration also now returns client_secret_expires_at alongside a new secret, so your client can tell that a three.ws secret never expires rather than guessing from a missing field. (`/docs/authentication`) `[fix, security]`
- **Avatar likeness is now measured against your photos, not assumed**: The promise behind the selfie flow is that the avatar looks like you, and until now nothing on the platform checked that. Two quality systems already existed and neither answered the question: one scores whether a generated model looks photorealistic, which a beautiful model of a stranger passes easily, and the other scores the head's shape offline against a synthetic test set and kept no record per avatar. Every finished reconstruction is now rendered from three angles and compared to the photos it was built from using the same kind of face-recognition model a phone uses to decide a face unlocks it. That produces a score out of 5 filed against the generation itself, and the scale is anchored to the model's own published cut-off for "same person" rather than to anyone's opinion, so 4 out of 5 means confidently you and not arguably you. A weekly sweep keeps the numbers current, and one of the three angles is a turned head on purpose: an avatar built by painting a photo onto a template can look right face-on and fall apart the moment it turns, and that is now a number somebody can see. The score is a measurement, never a guess: when the photos have no usable face, or the finished model has no recognisable face at all, it says which, instead of quietly reporting a bad score. `[infra, improvement, docs]`
- **Buying a skill from your phone no longer needs SOL in your wallet**: The QR code in the skill payment modal now hands your wallet a transaction we build for you, instead of asking the wallet to compose the transfer itself. Three things change for the buyer: three.ws pays the Solana network fee, so a wallet holding only USDC can complete the purchase; a creator who has never received this token before no longer causes the payment to fail, because the missing token account is created in the same transaction; and the whole payment still settles under one signature. Wallets that do not support this newer Solana Pay flow are one click away from the classic direct-transfer QR, which stays available under the code. (`/marketplace`) `[improvement, fix]`
@@ -36,16 +38,18 @@ Public history for [three.ws](https://three.ws), newest first. New pages come fr
- **Comments on your models now read as a sentence in the bell and file under the Social tab**: When someone comments on a model you made, you get a notification. It was arriving half-built. The notification centre sorts every alert into a tab (Sales, Social, Creations, and so on) using a list it keeps in the browser, and comments were missing from that list, so a comment landed in Account and Security instead of Social, next to withdrawals and security warnings. The bell had the same gap from the other side: with no wording written for the comment alert, it fell back to printing the raw internal name of the event, so the row simply read comment, with no hint of who wrote it or what they said. Comments now sort into Social, matching the preference switches in settings that were already treating them as social, and the row reads as a real sentence naming the commenter and quoting the opening of what they wrote. Three documentation pages were corrected against the code in the same pass: the Notifications reference was missing comments from its category table, the Agent Symphony page was missing five Agora arena and guild events from its sound map, and the Character Library page described its API as callable from any website when it has always answered three.ws and its partner sites only. (`/notifications`) `[fix, docs]`
- **Connecting an app to your account now fails loudly instead of handing it a token that works nowhere**: The OAuth server three.ws uses to connect outside apps (Claude Desktop, Claude Code, any MCP client) had three ways of telling a client the wrong story. An app that named a different resource when trading its authorization code for a token was quietly handed a three.ws token anyway, so it thought it was connected and then hit an error on every single request it made. An app whose client secret was rejected got a bare 401 with no challenge header, which several OAuth libraries retry forever instead of reporting as a wrong secret. And an app that registered a homepage, a logo, or its software version got a confirmation that omitted all of them, which reads as "rejected" and pushes it to register a second time under a new identity. All three now answer plainly: a wrong resource is refused up front with invalid_target, a rejected secret comes back with the challenge header the spec requires, and the registration confirmation lists everything that was actually saved. (`/docs/authentication`) `[fix, security]`
- **Controls that promised an action and did nothing now actually do it**: Several buttons across the site were links pointing at nothing, so clicking them moved the page to the top and nothing else happened. On a coin or category page whose market data failed to load, "Reload the page" was one of them: the one control offered for recovery had no handler at all, so a reader who hit a data outage had no way out of it. It now re-runs the real request in place, skeletons and all, so a recovered feed repaints without a full page load. The "+ Link wallet" button on your account page was the same kind of dead end, pointing back at the section it already sat in; it now opens your Solana wallet, signs the ownership message, and links the wallet to your account, including moving a wallet that is currently linked somewhere else once your signature has proved it is yours. Clear filter on derivatives, Show all on stablecoins, Retry on search and on the world generator, skip for now in the dashboard checklist, Create one in the agent registry, and View more in Related Agents were all links pretending to be buttons; they are real buttons now, which means they answer the space bar, sit in the tab order where you expect, and show a visible focus ring. Picking a coin in the compare search is announced properly to a screen reader as you arrow through the list. `[fix, improvement]`
-- **Cookbook posters came back, and a docs link you copy now reopens the page you were reading**: Eight developer pages were read against the code they describe and exercised in a real browser: the MCP 3D tool and CI asset gate recipes, and the payment sessions, developer platform, media API, agent-forged gallery, generation watch, and KOL tracker guides. Two things were broken for readers. Every cookbook recipe that shows a poster image was rendering without it: the figure system accepted three kinds of illustration and silently deleted the fourth, which is the kind the recipe posters use, with nothing in the page or the console to say a picture had gone missing. All of them render again. Separately, the previous/next buttons and the sidebar at the bottom of a documentation page produced a link that only worked while you stayed on the page: copy it, share it, or reload it, and you landed back on the page you started from instead of the one the link named. Those links now open the page they promise, correct the address bar to that page's own permanent link, and rename the browser tab to match, while in-page section links still jump to the section as before. The pages themselves were corrected where they had drifted from the code: the gallery's sample receipt quoted an old price for the tier it describes, the media API's two library samples reported counts no live response would return, and the KOL tracker described all four of its endpoints as sharing one dispatcher when the trade feed is served by its own file. Every command and code sample printed on the eight pages was run, and every link on them resolves. `[fix, docs]`
+- **Cookbook posters came back, and a docs link you copy now reopens the page you were reading**: Eight developer pages were read against the code they describe and exercised in a real browser: the MCP 3D tool and CI asset gate recipes, and the payment sessions, developer platform, media API, agent-forged gallery, generation watch, and KOL tracker guides. Two things were broken for readers. Every cookbook recipe that shows a poster image was rendering without it: the figure system accepted three kinds of illustration and silently deleted the fourth, which is the kind the recipe posters use, with nothing in the page or the console to say a picture had gone missing. All of them render again. Separately, the previous and next buttons and the sidebar on a documentation page produced a link that only worked while you stayed on the page: copy it, share it, or reload it, and you landed back on the page you started from instead of the one the link named. Those links now open the page they promise, correct the address bar to that page's own permanent link, and rename the browser tab to match, while in-page section links still jump to the section as before. The pages themselves were corrected where they had drifted from the code: the gallery's sample receipt quoted an old price for the tier it describes, the media API's two library samples reported counts no live response would return, and the KOL tracker described all four of its endpoints as sharing one dispatcher when the trade feed is served by its own file. Every command and code sample printed on the eight pages was run, and every link on them resolves. `[fix, docs]`
- **Creator subscriptions renew again instead of failing every billing run in silence**: No subscription on the platform has ever renewed. When a period came up for renewal, the billing code read two fields off the subscription record, the chain to bill on and the coin to bill in, that had never been created in the database. The charge failed instantly with a database error, and because the renewal job catches each subscriber separately it kept reporting a healthy run: subscribers processed, none charged. Signing up hit the same wall from the other side, creating the subscription and then erroring out. Both fields now exist, default to Solana as the home chain, and are filled in from the checkout that started the subscription, so a renewal reaches the subscriber as a payment request they can approve. A second fault behind it is fixed too: the renewal request was being attached to the creator's account instead of their agent, which could never have been paid. A plan with no agent attached now says so plainly rather than failing as an unexplained error, and the underlying database reason is written to the logs instead of being thrown away. `[fix]`
- **Crypto news stops leading with gold prices and Indian equities**: The filter that decides whether a story from a broad outlet belongs in a crypto feed matched its keyword list as raw text, so "defi" fired inside the ordinary word "defied" and three Economic Times headlines about Indian share prices led /markets/news and the top of the daily digest. The ticker detector had the same flaw: it read the English word "near" as NEAR Protocol, "optimism" as OP, and "maker" as MKR, which pinned live price widgets to unrelated stories and pushed NEAR to third place in the archive trending strip on the strength of phrases like "near-term profit booking". Both now match whole words, and those three projects count only when the story actually names them. Separately, the FXStreet feed we pull was the publisher site-wide one (gold, EUR/JPY, central banks) rather than its crypto desk, so "India Gold price today" arrived pre-approved; it now points at the crypto feed. (`/markets/news`) `[fix, improvement]`
- **Draft with AI actually returns drafts, and the X panel stops 500ing on a bad id**: The Draft and Draft x3 buttons in the agent's X panel asked the server for tweet options and got back a differently shaped answer, so the panel cleared itself and showed nothing every time. Drafting now returns the options it was asked for: one per click of Draft x3, each written from its own angle, in the tone picked in the panel and in the voice of the agent it belongs to, with Thread mode splitting a completion into real thread parts. Drafting is also gated like every other write, so another site cannot spend a signed-in account's model budget. Elsewhere on the same panel, a malformed id or agent filter now comes back as a plain validation error instead of a server error, approving a queued post can no longer be published twice by a double click, and scheduling has a sane ceiling on how many posts and rules one account can queue. `[fix, improvement]`
- **Eight docs pages checked line by line, and four stale claims corrected**: A sweep of the viability, press kit, partners, agent identities, OKX marketplace, avatar engines, stage, and world lines pages, with every command in them actually run and every link followed. Four things were wrong and are now right. The partners page pointed the NVIDIA row at a docs path when the card on the live page goes to /nvidia. The agent identity docs claimed a job handle expires after about thirty days, which nothing in the code enforces; the page now states what is true, that the signed handle carries no expiry and the job record stays readable. The marketplace page showed a create call replying with a queued status and a generation stage, neither of which the endpoint has ever returned, so an agent coding against the sample would have waited on a state that never arrives; the page and the endpoint's own published example now show the running status and generate stage the code emits. The same page read as though our marketplace listing were live when it is still in review, and now says so plainly. The world lines anti-cheat table was also missing one of the per-IP limits it lists. (`/docs/partners`) `[docs, fix]`
+- **Eleven links that led nowhere now land where they promise**: The audit that guards every page against dead ends had a hole in it: the rule that serves the not-found page matched every address on the site, so the audit read every link as reachable and reported a clean bill of health it could not vouch for. With that fixed, eleven links turned out to lead to pages that were never built. Sign in to save on the choreography and gesture pages pointed at an address that does not exist, so the one thing those pages ask you to do could not be done. The extension terms page linked its own terms of service to a missing address, twice. The motion studio offered to show you how to capture an avatar from a selfie and led nowhere, as did the pricing page's link to your account, the memory tutorial's link to your library, and two entries in the dashboard trading menu. Pressing Message on a friend in the 3D world walked you into a blank page instead of the conversation, and now opens the thread with that person. Three outbound links to other projects' documentation had rotted and now point at the pages that replaced them. The audit itself was rebuilt to be trustworthy: it no longer calls a live site dead because the site answers a robot differently than a browser, and it no longer treats the blog and docs it ships, or the app manifest it generates, as missing files. `[fix, improvement]`
- **Every agent now has a verifiable on-chain history you can read on its page**: Agents do things on-chain after they are registered: they launch a token, collect reputation attestations, change hands, delegate execution rights. Until now the platform recorded only who an agent currently was, so all of that history was dropped, and anyone who wanted it had to scan a chain themselves. Open any agent under the Discover directory and there is now an On-chain history panel: a timeline of everything the platform has indexed for that agent, each entry stamped with the real time the chain says it happened and linked to the transaction on a block explorer so you can check it yourself. The same data is available to developers as one public endpoint. Three failures were fixed to make it real. The Solana half of the indexer was crashing on every single agent it looked at, so despite tracking more than fifteen hundred agents it had recorded zero events; it now records them, and a token launch names the agent's actual coin instead of the wrapped SOL that paid for it. On the EVM side, five of the twenty-two supported chains were producing blocks faster than the indexer could read them, so their backlog grew every hour while the status page called them healthy: the indexer now widens its reading window until it catches up, and the status page publishes how many blocks behind the worst chain is, as a number. Reputation was the third: the platform was already asking every EVM chain for agent feedback and vouches on every pass, then throwing the answers away, so not a single reputation event had ever reached the index. Those now land alongside the rest of an agent's history, with the score, its tags, and a link to the attestation. (`/docs/agent-index`) `[feature, fix, infra]`
- **Every avatar you scan now arrives with an on-chain identity of its own, minted for you in the background**: When a scan or a prompt finishes building your avatar, the finished model is written to permanent storage and the agent behind it is given a real on-chain identity at the same moment. It is a Metaplex Core asset on Solana, minted into the platform's own agent collection with the royalty terms and registry enrolment every other platform mint carries, and it is owned by the agent's own wallet rather than by us. There is no wallet prompt, no fee, and no step to remember: by the time the avatar shows up in your library the identity already exists and anyone can look it up. It is created as a draft, which means it says the agent exists and this is its avatar, not that it is live and taking work. Going live stays a deliberate choice you make later. If the mint cannot run for any reason your avatar is completely unaffected, because it is already delivered and saved before the mint is ever attempted. Agents that want an Ethereum-side identity as well can have one registered under the ERC-8004 standard, and the whole path is now documented, including the single setting an operator flips to arm it for the main Solana network and the script that proves the whole chain end to end on the test network. (`/create/selfie`) `[feature, docs]`
- **Every Oracle card now says why that coin scored, and what a score like it has actually been worth**: The live conviction board had a copy problem that made a working engine look like a broken one. Each card carried a one-line read of the launch, and that line was assembled from whichever of the four pillars scored highest. Buy-side momentum is the highest pillar on about nine in ten of the coins at the top of the feed, so nine in ten cards printed the same sentence, word for word, down the whole page. Two hundred cards, one take, nothing learned. Meanwhile the engine had already written a specific, quantified reason for every single coin, and it was only visible if you opened the coin. Those reasons now ship with the feed, so each card quotes its own evidence and the historical lift behind it: 40+ early buyers at 6.8x the base rate, top holder at 5-15%, more sellers than buyers. Each reason is phrased in the units a trader thinks in rather than the model's internal labels. Every card also carries a second, quieter number: how often a coin scored into that band went on to graduate or run 2x or more without ever rugging, measured across the 61,916 scored coins the market has already resolved. That is deliberately a stricter question than the score answers, because the score ranks the odds of a run and says nothing about surviving it, and a holder is entitled to both numbers rather than being left to read a conviction of 99 as a 99 percent chance of anything. Two badges were cleaned up in the same pass: the prime pill that only ever repeated the tier printed beside it is gone, and the momentum badge, which was firing on 93 percent of the live feed, now appears only when buy pressure alone would carry a launch to prime. (`/oracle`) `[improvement, fix]`
- **Keyboard users stop tabbing past the same Skip to content link twice, and the homepage hint is readable again**: Two accessibility defects on high-traffic pages are fixed. The shared navigation injects a Skip to content link as the first thing you reach with the Tab key, so keyboard and screen-reader users can jump straight to the page body instead of walking the whole header. It only checked for its own copy of that link before injecting, but roughly three dozen pages write the anchor into their own markup under a different class, so those pages ended up with two identical skip links and you had to tab past the same control twice to get anywhere. The check now recognises every variant, so there is exactly one skip link per page. On the homepage, the hint under the interactive dragon that reads move cursor, hold to breathe fire was drawn in white at 18 percent opacity over a near-black panel, about 1.6 to 1 against the background where the accessibility standard asks for at least 4.5 to 1; anyone in bright light or with reduced contrast sensitivity simply could not see the instruction for the thing it was explaining. It now sits at 5.7 to 1 and stays legible without changing the look of the scene. The Genesis and Walk pages also had a skip link with no styling rule attached, which left it as stray visible text above the header; both now use the shared style and hide until focused like every other page. `[fix]`
- **Links inside the tutorials now go where they say, and the AWS page names the right region**: Every tutorial on the site is written as a markdown file that lives beside the other documentation, and the links inside it were written against that location. The page that renders them was reading those links as if they pointed at site addresses instead, so a tutorial that sent you to a related guide, or cited the exact source file a step describes, sent you to an address that does not exist. Links that reach across to another document, that point at a specific section of one, or that cite a file in the source tree are now each resolved to the right destination, with source files opening on GitHub. The scene tutorial's own list of related tutorials pointed at a secondary address that rendered them under a generic title, and now uses the real one. Separately, the AWS page claimed the platform's production deployment runs in AWS us-east-1. It does not: production runs on Google Cloud Run, and us-east-1 is where the AWS Marketplace billing and entitlement APIs are called. The page now says that instead. `[docs, fix]`
+- **One cookbook recipe stopped introducing itself as a generic page in search results and shared links**: All five cookbook recipes are drawn by one shared page template, and the server fills in each recipe's own title, description, and social card as the page is served. One recipe was skipping that step. The template carried a leftover marker naming a specific recipe as its own address, and the server reads that marker as this page already knows who it is, so it left that one page alone: text to 3D from the command line went out with the placeholder title every recipe shares and a one-line description that said nothing about it. Search results, and the preview card in any chat where the link was pasted, showed the placeholder instead of the recipe. The template now names the cookbook index rather than any single recipe, matching how the tutorial and walkthrough templates have always been set up, so all five recipes present themselves properly. A test now walks the real page templates and fails if any of them ever claims one of its own recipes again, which is what let this sit unnoticed. (`/cookbook/text-to-3d-cli`) `[fix]`
- **Oracle pages now say the data is briefly unavailable instead of showing you a blank market**: When the database had a short connectivity blip, the Oracle read endpoints answered with an empty result and a 200 OK, which reads as fact rather than as an outage. The live conviction feed showed a market with nothing in it, the hot-sectors panel showed no sectors, an agent's public track record showed "never traded" (the very evidence copy-traders judge a leader on), and the follow endpoint told a subscriber they were not following an agent they were in fact following. Worst of all, unfollowing an agent answered "done" while the subscription survived, so the alerts kept arriving with no way to stop them. Some of those answers were then cached for up to five minutes, outliving the blip that caused them. All of these now return a proper "temporarily unavailable, retry shortly" response that nothing caches, so a page can tell you it cannot reach the data instead of quietly showing you the wrong answer. `[fix, improvement]`
- **Paid 3D downloads work from a browser again, and the marketplace report stops ignoring every Solana price**: Buying a 3D asset or an animation clip from a web page was impossible: those two endpoints answered the browser's permission check with a payment demand and none of the cross-origin headers a wallet needs, so the page could never even read what it was being asked to pay. Both now answer that check the same way every other paid endpoint does, on the challenge, on the errors, and on a free listing. The paid marketplace report had a second, quieter problem: it recognised prices set in USDC on Base but not the identical prices set in USDC on Solana, so every Solana-priced agent was counted as unpriced and left out of the average, lowest and highest figures buyers were paying for. Solana prices now count. The auth health check no longer charges for a call it refuses, its request reader now has the same size limit as its neighbours, and the avatar optimization batch can finally be repriced at deploy time like every other endpoint. (`/docs/x402-endpoints`) `[fix]`
- **Paying to resolve a name that does not exist no longer costs anything**: The paid name-resolution call on /api/x402/pay-by-name sells one thing: the on-chain address behind a @username, a .sol name, or a raw wallet. When the name resolved to nothing it still took the $0.001 toll and handed back an empty answer. It now returns a plain 404 and leaves the payment unspent, so a buyer only ever pays for an address they actually received. Two smaller corrections shipped alongside: the pump launch feed's headline fields (newest mint, name and symbol) kept describing the newest launch when the list is ranked by liquidity instead of by age, rather than whichever coin happened to sit at the top; and internal and subscription callers can once again reach the Permit2 gas-sponsoring demo on deployments that have no card-facilitator credentials, since those callers never sign a payment in the first place. (`/docs/x402-endpoints`) `[fix]`
@@ -59,6 +63,7 @@ Public history for [three.ws](https://three.ws), newest first. New pages come fr
- **Sharing a build now checks that the coin world you are sharing to is real**: When you screenshot a voxel build in a coin's /play world and share it, the build is filed under that coin's address. The endpoint only checked that the address was the right length, not that it was an address at all, so anything between 32 and 64 characters was accepted and filed away under its own storage slot. It now checks the format the way every other coin-world surface does: a Solana mint or an EVM contract address gets through, anything else is turned away before it reaches storage. Real worlds are unaffected, on both Solana and the EVM chains. The live event board picked up its own test coverage in the same pass, including a check that a player's account key never appears in what the board sends to the browser. (`/play`) `[security, fix]`
- **The $THREE holder count is back on the token page and its share card**: The holder count on the /three-token page, and on the badge image that previews wherever a $THREE link is shared, had been showing a dash instead of a number. Our own holder snapshot is rebuilt from a Solana indexer that has been refusing the scan, and once the snapshot went stale the page correctly declined to show an out-of-date figure, so the count simply disappeared rather than being briefly missing. Holder count now falls back to a second, independent source when our snapshot cannot refresh, so the figure stays real and current through an indexer outage. The share card was also reading a narrower set of sources than the page it mirrors, which is why the two could disagree; both now read the same ones. Separately, a shared agent link unfurls noticeably faster: the card was fetching the agent's portrait only after every other lookup had finished, and now fetches it alongside them. (`/three-token`) `[fix, improvement]`
- **The 3D generation SDK can finally pay per call, and its bring-your-own-key option actually does something**: The npm package that turns a prompt into a 3D model documented two things it did not do. It listed a per-call option for supplying your own Meshy or Tripo key, and that option was read from nowhere: the key was silently dropped, so anyone following the README's own example was told to add a key that the generator then said was missing. It also documented paying per call in USDC as the default billing lane, when the flag it described was ignored by the endpoint it was sent to. Both are now real. A key passed on the call is used, overrides the one set on the client, and is resent on every status check, which the generator requires and the old code never did, so a job started with your own key could never be polled to completion. The pay-per-call lane now routes to the endpoint that actually accepts payment, quotes Solana USDC, and hands back a job you poll for free, and it refuses upfront the options it cannot honour rather than accepting them and ignoring them. The package now also exposes a way to read one job's state without waiting for it, so a generation survives a restart. Alongside that, the reference pages listed a price and a set of engines that had both moved on, named an error the service never returns, and pointed at three addresses that no longer exist: the sign-up link for API keys in the copy-trading and LLM router packages went to a dead page, as did the link to the 3D studio tool server. All of them now resolve. `[sdk, fix, docs]`
+- **The 3D viewer no longer starts up blank when its box has not been measured yet**: Every page that shows a 3D avatar builds its viewer the moment the page's script runs. On pages where the area holding the avatar has not been given its size yet, because the layout lands a frame later or the panel starts hidden, that meant the viewer was built at zero pixels wide and zero pixels tall. Graphics drivers reject an image with no width or height, so the viewer came up against a surface the driver had refused to create: the browser console filled with rejected-texture and incomplete-surface errors, and on the mocap studio the avatar stage was the one that hit it. The viewer now starts at a minimum size and resizes itself to the real box on the first frame that has one, which it was already set up to do for every later resize. Nothing changes for a page whose avatar area is measured up front, which is most of them. (`/mocap-studio`) `[fix]`
- **The activity floor stops reporting a flat week when nothing has closed yet**: The header on the live agent activity page counted a week of wins, losses and profit from the moment each position was opened, not the moment it resolved. A conviction position routinely takes days or weeks to settle, so by the time a trade had a result it had already aged out of the window that was supposed to count it. The tiles were therefore pinned at zero wins, zero losses and a confident PnL of plus zero forever, which reads as a floor that traded all week and broke exactly even. Entries are now counted on when the agent acted and results on when they settled, and the endpoint also reports how many closed positions carry a realized figure, so a genuinely flat week is told apart from a week with nothing closed yet: the tile says not available instead of inventing a zero. The win rate is graded against the market's own base rate rather than a hardcoded fifty percent, so a hit rate several times better than a random launch stops being painted red. Every tile now explains what it measures on hover, the stats row no longer re-reads all nine numbers to a screen reader on every incoming trade, and a stats outage shows the real values it has instead of shimmering as though it were still loading. Two lists on the page, the movers and winners rails and the insight strip, had their two-line rows collapsing onto a single run-on line; both stack properly again and truncate long names instead of overflowing. (`/activity`) `[fix, improvement]`
- **The agent directory describes itself correctly in all 84 languages again**: The English description Google and every social preview show for the agent directory was updated a while back to say the directory spans Solana and every supported EVM chain. The 84 translated versions were never updated with it, so a reader in Spanish, Japanese, Arabic or any other language got the older, vaguer line that named no chain at all. All 84 now carry the current wording, with Solana and three.ws left in their real spelling rather than bent to fit each language's grammar. Retired copy was cleared out at the same time: eight strings that no page shows anymore had been sitting in every one of the 84 catalogs, which is 672 dead entries the site was shipping to readers for no reason. (`/agents`) `[fix, docs]`
- **The agent interaction lab is reachable again from the demos index**: The demos index has always carried a card for the agent interaction lab, the set of small pages where an avatar high-fives you, climbs the page title, bounces on a trampoline, or falls asleep when you leave it alone. Clicking that card gave you a 404. The lab hub lives in a folder rather than as a single page, and the rule that turns a demo name into a page file was sending the request to a file that never existed, so the one link into the lab was the only link on the index that could not be followed. The sub-demos were fine the whole time; only the door was missing. It is fixed, and the demo route map now records the whole lab, including the lost-avatar 404 page that has its own card on the index but was never listed. (`/demos/`) `[fix, docs]`
@@ -77,6 +82,7 @@ Public history for [three.ws](https://three.ws), newest first. New pages come fr
- **The free fact-check lane is callable from a browser again, and two paid endpoints answer to the price you set**: A site calling the fact checker from a browser got a confusing failure: the preflight check said the call was allowed, then the browser threw away the answer. The free daily lane was not sending the cross-origin headers the paid lane sends, so a perfectly good verdict never reached the page. Both lanes now answer identically, whether the check succeeds or the claim is rejected. Separately, the fact checker and the paid LLM proxy were the last two paid endpoints whose price was fixed in code: every other one can be repriced at deploy time, and now these two can too, with no code change. The IBM Granite health feed also stops reporting itself healthy when its own database read fails, so a dashboard cannot show green through an outage. Exploit-history and gas-oracle outages are now written to the server log rather than disappearing silently, which is the difference between an alert and a mystery. (`/docs/x402-endpoints`) `[fix, improvement]`
- **The free object library can no longer be emptied by one bad upload**: The CC0 prop library at /objects reads a single index file from the CDN. If that file were ever uploaded truncated or in the wrong shape, the gallery treated it as a library with nothing in it and told the CDN to remember that answer for the next five minutes, so a bad upload kept the shelves bare long after it was replaced. A damaged index is now handled the same way as the CDN being briefly unreachable: the page stays quiet, nothing gets cached, and the props come back on the very next request. Two smaller fixes went with it: a request with a broken page cursor (a limit of "abc", a negative offset) is rejected outright instead of first fetching an index it was never going to use, and the developer documentation no longer claims the library API can be called from any website, which it cannot; it answers three.ws and its partner sites, and everything it points at is still yours to download and reuse anywhere. (`/objects`) `[fix, docs]`
- **The internal money and platform health boards tell the truth again**: Three internal dashboards watch the platform for us: one for site and cron health, one for every money subsystem, one for payment outcomes. Two of them had gone quietly wrong. The platform board judged each scheduled job against a hand-written idea of how often it should run, and three jobs that run once a day were being judged against a window of minutes, so they reported as broken for roughly 23 hours out of every day and the board could never come back all-clear; every job is now measured against the schedule it genuinely runs on. One of its probes was also posting to a payments endpoint without declaring it was sending JSON, so it had been failing on its own mistake rather than on anything real. The money board was reading the wrong timestamp column on two payment logs and showing no activity at all for a payment ring that settles thousands of payments a day. It also accepted the credential our fund-moving scheduled jobs carry as an ops password; it now takes the same admin sign-in as the other boards, so an ops credential can never be escalated into triggering a payment job. `[fix, infra, security]`
+- **The legal pages now say what is in them, and call the terms by their real name**: The six pages under three.ws/legal (the index, privacy policy, terms of service, end-user license agreement, risk disclosure, and content policy) were audited in a real browser at phone and desktop widths. Every page loads, every link on them resolves, and the layout holds at 320 pixels wide. Three things were wrong. The terms of service was filed under the name Terms of Use in the site's page index, so a search result or a shared link showed one name while the document itself showed another; it is now Terms of Service everywhere. The privacy policy, the terms, and the EULA each carried a search-result description that only restated their title, so a search engine or a chat preview told you nothing about what was inside; each now describes what the document actually covers. And several of the policies linked to each other by their raw file names rather than their real addresses, which split the same document across two web addresses; they now all use the address shown in the browser bar. The content policy was also missing its publication date in the page index, which is what the changelog and sitemap read. (`/legal`) `[fix, docs]`
- **The limits you set on an agent's spending can no longer be skipped**: An agent that pays other agents on its own is governed by the policy you set on its wallet: a ceiling per payment, a rolling 24-hour ceiling, a separate ceiling per payee so one peer cannot quietly absorb the whole day's budget, and a kill switch that halts every kind of autonomous spending the moment you flip it. Those rules were enforced wherever the platform asked for them, but the enforcement itself had a soft spot: code that asked the guard to check a payment without first loading the agent's settings was handed an empty policy, meaning no ceilings and no kill switch, and nothing anywhere said so. The guard now reads the agent's real policy itself in that case, and refuses outright to price a payment for an agent that no longer exists. Every ceiling and the kill switch are now proven, not asserted: one command spins up a real database and a real server, then shows each limit turning away an over-limit payment, eight simultaneous payments competing for the last dollar of headroom with exactly one getting through, a frozen agent halted before the other agent is even contacted while your own withdrawals stay open, and every settled payment coming back as a receipt with the payee, the amount and the transaction signature on your agent's ledger. (`/docs/a2a-payments`) `[security, improvement, docs]`
- **The link checker now covers the docs, blog, and chat pages it had been skipping, and three dead links are fixed**: Every build runs a link audit that walks the site and fails if any link points at a page that does not exist. It was only ever reading three directories, and the docs reader, the blog, and the chat app live in three more that the router serves just as publicly. So those pages shipped unchecked while the audit still reported a clean zero, which is the worst way for a safety check to fail: green with the gap intact. The audit now walks every directory the router can serve a page out of, and a new test derives that list from the routing table itself, so the next section added to the site cannot quietly slip past it. Two smaller blind spots went with it: a code sample printed on a documentation page was being read as if it were a real link on that page, and a link built up in JavaScript from several pieces was being reported as a dead one. Three genuinely dead links found along the way are fixed, and the audit is now part of the pre-build gate rather than something someone has to remember to run. `[fix, infra]`
- **The machine-readable API spec stops telling agents they need a login for a lane they can just pay for**: Agents and crawlers read /openapi.json to work out how to call this platform without a human present, and two of its entries were sending them the wrong way. The MCP endpoint was listed as requiring a bearer token or an API key, when in fact an unauthenticated call answers with a payment challenge: an agent can settle in USDC and get to work with no key, no token, and no account at all, and a credential is accepted but never required. Any agent that trusted the spec went off to obtain a login it did not need. The spec also described only the paid half of the Forge generation lane, hiding the free price lookup on the same address that lets an agent read the per-tier pricing and the accepted request shape before it commits any money. Both are corrected, the session-teardown call the MCP route already advertises in its own headers is documented, and the human API reference now matches. The sitemap, robots.txt, llms.txt, and llms-full.txt feeds were swept in the same pass and every address they publish was confirmed to resolve. (`/docs/api-reference`) `[fix, docs]`
@@ -90,6 +96,7 @@ Public history for [three.ws](https://three.ws), newest first. New pages come fr
- **The payment-outcome board no longer reports itself healthy while it is blind**: One of the internal boards we watch payments with is built from three independent panels: payments arriving, settlements going out, and the fee wallet that pays for them. If a panel could not be read, the board swapped in an error for that panel but still answered with a plain success, so a read where all three had failed looked identical to a read where everything was fine. It now names every panel it could not see and answers with a partial-success status instead, which is what our alerting keys on. Separately, all three ops boards were spending the same strict per-address budget that guards sign-in attempts, so simply watching a dashboard could lock an operator out of logging in from that address. The boards now use the generous read budget the rest of the signed-in read surface uses. `[fix, infra]`
- **The public custody page is switched off until its on-chain anchor works again**: The /integrity page invited anyone to verify our custodial wallets against a Merkle root anchored on Solana. The snapshot half of that has been running correctly all along: every six hours the platform reads each wallet's live mainnet balance, builds the tree, and records the root. The anchoring half has not. No epoch has been committed on-chain since the feature launched on 23 June, so the page was headlining custody you can verify above a history table where every row read anchor_failed, and telling readers the latest root would be anchored shortly when in fact nothing had been. A trust page that overstates what it can prove is worse than no trust page, so we have taken it down rather than leave it up making a promise the data does not support. Nothing else changes: snapshots keep running and keep accruing, the public API behind the page (/api/custody/integrity and /api/custody/anchor) is untouched and still reports every epoch along with an honest epochs_anchored count, and wallet owners can still recompute their own inclusion proof. The page comes back when anchoring is fixed, not before. (`/proof`) `[fix, security]`
- **The Reputation Staking Market is live: back an agent and earn from what it actually did**: Stake SOL behind an agent you believe in and earn a share of a daily reward pool, sized by that agent's signed on-chain record: accepted tasks, passed validations and real reviews pay, while failed validations, disputes and revocations cost double what work earns. The market page had shipped without the code that drives it, so it rendered empty no matter what was staked; it now loads live, ranks agents by the conviction escrowed behind them, shows any wallet's positions with earnings that update as they accrue, and settles a withdrawal back to the staker who signed the stake. Your principal is never slashed and can be withdrawn at any time: bad conviction earns nothing, it never costs you your stake. Every number is re-derivable from the chain, the exact formula runs identically in your browser and on our servers, and the whole cycle is provable on devnet with one command. Devnet is open to everyone; mainnet stays owner-gated. (`/reputation/market`) `[feature, fix]`
+- **The security headers every three.ws page serves are now checked against the live site, not just the config**: Every page three.ws serves carries a set of browser security rules: which scripts are allowed to run, that the connection must stay encrypted, that browsers must not second-guess a file's type, how much of the current address is passed on when you follow a link, and who is allowed to put a three.ws page inside a frame. Those rules are written in one config file, but until now nothing checked that the site actually answered with them. The sweep that loads real pages in a real browser now also compares each response against that config, and it can be pointed at the live site to check the whole catalogue in about a minute. Running it against production found no gap: all 25 of the most script-heavy pages, and the rest of the catalogue with them, serve exactly the rules the config declares. Two fixes came out of building it. Files that are not web pages, like robots.txt and the machine-readable service descriptions, were being reported as wrong when they were correct. And a page whose file changed while the server was running could be sent with the script rules of the previous version, which would have blocked its own scripts and shown a blank page; the server now notices the change and recomputes them. `[security, infra]`
- **The selfie camera now catches a blurry or blown-out shot before it costs you a build**: The capture screen showed a green "Sharp" badge on frames that were smeared past recognition. Measured against six real portrait photographs, the blur floor sat below every reading a real photo can produce, so the gate never fired once: a shot ruined by camera shake sailed through and you found out a minute later, after the build. The floor is now calibrated from those measurements and sits in the gap between blurred and sharp faces, so "Hold steady. The image is blurry." appears while you can still act on it. A face half-blown by window glare also used to pass, because averaging bright and shadowed halves lands back in the acceptable range; the camera now measures how much of your face is washed out and tells you to turn away from the window. Two capture bugs went with it: on a phone, the helper widgets pinned to the bottom of the screen sat on top of the camera and swallowed taps meant for the shutter and Cancel, and the Retake and Use this buttons stayed on screen during live capture. The shot check after the shutter is faster too. (`/create/selfie`) `[improvement, fix]`
- **The server-side agent stops falling through its own free lanes, and the cookbook command you copy now runs**: The tool-using agent behind the server tools model in chat, and the trading copilot beside it, both run on one shared list of language-model providers tried in order until one answers. Two rungs of that list were broken in ways nobody could see from the outside. The first OpenRouter rung named a paid model, so it either billed the platform key against our own free-first policy or was refused outright, and every OpenRouter rung after it named a model id that had since been retired, so those answered with a not-found error rather than a result. With that whole tier effectively dead, one busy moment on the remaining fast lane dropped the agent straight onto the last resort and the request failed. All OpenRouter rungs now ride the same live free model the rest of the platform uses, and two more genuinely free lanes with full tool-calling support were added to the chain, so the loop has independent quota to fall back on instead of one shared pool. Four documentation pages were corrected against the code they describe in the same pass: the agent runtime page listed a guard response shape that did not match what the endpoint returns, the Nemotron spotlight quoted an image-check prompt with a paragraph missing and a vision model list that skipped one of its own lanes, and the two command-line cookbook recipes named a flag the script does not accept and a file length it had outgrown. The recipe file you download is the same file that was wrong, so that flag works now when you paste the line straight out of the page. (`/docs/agent-runtime`) `[fix, improvement, docs]`
- **The Smart Money Radar stops going blank when pump.fun rate-limits us**: Every coin the platform watches gets a verdict recorded after the fact (did it graduate, pump, go flat, or rug), and that verdict is what the Smart Money Radar and the Pre-Launch Radar are built from. The verdict came from a single live lookup against pump.fun, and any hiccup there (a rate limit, a timeout, a coin aged out of their index) was written down as "unknown" and never looked at again. Unknown coins are skipped by the wallet-reputation graph and by the proven-creator list, so a few throttled minutes could permanently erase whole batches of coins from both radars. Two changes fix it. We now cross-check against our own graduation index, which we build independently, so a coin we already watched graduate is scored as graduated no matter what the third-party lookup says. And an unknown is now treated as "not yet known": it is retried on a backoff for the first week of a coin's life instead of standing as a permanent answer. `[fix, improvement]`
diff --git a/data/changelog.json b/data/changelog.json
index 0ca6a65d5..1ae5f8d03 100644
--- a/data/changelog.json
+++ b/data/changelog.json
@@ -1,6 +1,45 @@
{
"$comment": "Curated changelog entries for three.ws, the editorial layer of the public changelog. New PAGE launches are tracked automatically via the `added` field in data/pages.json; this file is for everything else holders care about: improvements, fixes, SDK releases, trading/payments capabilities, security work. scripts/build-page-index.mjs merges both sources into CHANGELOG.md, public/changelog.json, and public/changelog.xml. Newest first. Allowed tags: feature, improvement, fix, sdk, infra, docs, security. `link` is optional and must be a live three.ws path.",
"entries": [
+ {
+ "date": "2026-08-14",
+ "title": "Eleven links that led nowhere now land where they promise",
+ "summary": "The audit that guards every page against dead ends had a hole in it: the rule that serves the not-found page matched every address on the site, so the audit read every link as reachable and reported a clean bill of health it could not vouch for. With that fixed, eleven links turned out to lead to pages that were never built. Sign in to save on the choreography and gesture pages pointed at an address that does not exist, so the one thing those pages ask you to do could not be done. The extension terms page linked its own terms of service to a missing address, twice. The motion studio offered to show you how to capture an avatar from a selfie and led nowhere, as did the pricing page's link to your account, the memory tutorial's link to your library, and two entries in the dashboard trading menu. Pressing Message on a friend in the 3D world walked you into a blank page instead of the conversation, and now opens the thread with that person. Three outbound links to other projects' documentation had rotted and now point at the pages that replaced them. The audit itself was rebuilt to be trustworthy: it no longer calls a live site dead because the site answers a robot differently than a browser, and it no longer treats the blog and docs it ships, or the app manifest it generates, as missing files.",
+ "tags": [
+ "fix",
+ "improvement"
+ ]
+ },
+ {
+ "date": "2026-08-14",
+ "title": "A generation that quietly switched engines no longer counts as a failure",
+ "summary": "When a 3D generation lane loses your job, the platform resubmits it to another lane and you still get your model. Until now the abandoned attempt was recorded as a plain failure, with nothing tying it to the run that actually succeeded, so the platform's own health readout treated its recovery machinery as an outage: over the past week, 13 of the 14 lost jobs were rescued within seconds and all 14 still counted against generation health. Each recovered attempt is now linked to the run that replaced it, so the status page and our failure reports show what you experienced instead of what the machinery did.",
+ "tags": [
+ "improvement",
+ "infra"
+ ],
+ "link": "/status"
+ },
+ {
+ "date": "2026-08-14",
+ "title": "Cookbook posters came back, and a docs link you copy now reopens the page you were reading",
+ "summary": "Eight developer pages were read against the code they describe and exercised in a real browser: the MCP 3D tool and CI asset gate recipes, and the payment sessions, developer platform, media API, agent-forged gallery, generation watch, and KOL tracker guides. Two things were broken for readers. Every cookbook recipe that shows a poster image was rendering without it: the figure system accepted three kinds of illustration and silently deleted the fourth, which is the kind the recipe posters use, with nothing in the page or the console to say a picture had gone missing. All of them render again. Separately, the previous and next buttons and the sidebar on a documentation page produced a link that only worked while you stayed on the page: copy it, share it, or reload it, and you landed back on the page you started from instead of the one the link named. Those links now open the page they promise, correct the address bar to that page's own permanent link, and rename the browser tab to match, while in-page section links still jump to the section as before. The pages themselves were corrected where they had drifted from the code: the gallery's sample receipt quoted an old price for the tier it describes, the media API's two library samples reported counts no live response would return, and the KOL tracker described all four of its endpoints as sharing one dispatcher when the trade feed is served by its own file. Every command and code sample printed on the eight pages was run, and every link on them resolves.",
+ "tags": [
+ "fix",
+ "docs"
+ ]
+ },
+ {
+ "date": "2026-08-14",
+ "title": "Anyone can run an inference node: the operator client now ships CPU and GPU images, and proves itself end to end",
+ "summary": "The open inference network is the part of the roadmap where the compute stops being ours. The node operator client is how you join it: your machine generates a Solana keypair, registers itself with the platform, claims inference jobs, runs them locally with a real open model, and returns every answer with a signature that binds it to your key. There is no shared password anywhere in that flow, so nothing can be leaked and reused to impersonate your node, and the platform re-derives each signature from its own record of the job before it accepts the work. This release makes the graphics-card story honest. The old image claimed it would use an NVIDIA card automatically when one was present, and it never could: the runtime it shipped had no GPU support compiled in at all. There are now two real images, one for CPU and one for NVIDIA cards, and the client tells you which device actually ran each job. Ask for the GPU on a machine that cannot reach one and it stops immediately with the reason, instead of quietly earning at CPU speed for a month. Three faults in the job pipeline were also fixed: results submitted by a node hit a server error every time, registered nodes read back as unregistered, and every finished job leaked two silent background errors. A single command (npm run e2e) now registers a node, sends it a job, runs the model, and verifies the signed answer against the real platform code, and it fails loudly if any step regresses.",
+ "tags": [
+ "feature",
+ "fix",
+ "infra"
+ ],
+ "link": "/docs/inference-node-operator"
+ },
{
"date": "2026-08-14",
"title": "The legal pages now say what is in them, and call the terms by their real name",
@@ -48,15 +87,6 @@
],
"link": "/forge"
},
- {
- "date": "2026-08-14",
- "title": "Cookbook posters came back, and a docs link you copy now reopens the page you were reading",
- "summary": "Eight developer pages were read against the code they describe and exercised in a real browser: the MCP 3D tool and CI asset gate recipes, and the payment sessions, developer platform, media API, agent-forged gallery, generation watch, and KOL tracker guides. Two things were broken for readers. Every cookbook recipe that shows a poster image was rendering without it: the figure system accepted three kinds of illustration and silently deleted the fourth, which is the kind the recipe posters use, with nothing in the page or the console to say a picture had gone missing. All of them render again. Separately, the previous/next buttons and the sidebar at the bottom of a documentation page produced a link that only worked while you stayed on the page: copy it, share it, or reload it, and you landed back on the page you started from instead of the one the link named. Those links now open the page they promise, correct the address bar to that page's own permanent link, and rename the browser tab to match, while in-page section links still jump to the section as before. The pages themselves were corrected where they had drifted from the code: the gallery's sample receipt quoted an old price for the tier it describes, the media API's two library samples reported counts no live response would return, and the KOL tracker described all four of its endpoints as sharing one dispatcher when the trade feed is served by its own file. Every command and code sample printed on the eight pages was run, and every link on them resolves.",
- "tags": [
- "fix",
- "docs"
- ]
- },
{
"date": "2026-08-14",
"title": "The 3D generation SDK can finally pay per call, and its bring-your-own-key option actually does something",
diff --git a/public/changelog.json b/public/changelog.json
index 1630ea86e..944087407 100644
--- a/public/changelog.json
+++ b/public/changelog.json
@@ -1,5 +1,5 @@
{
- "generated_at": "2026-08-14T20:26:48.282Z",
+ "generated_at": "2026-08-14T20:45:59.062Z",
"generated_by": "scripts/build-page-index.mjs from data/pages.json + data/changelog.json",
"site": {
"name": "three.ws",
@@ -141,6 +141,17 @@
"fix"
]
},
+ {
+ "date": "2026-08-14",
+ "type": "update",
+ "title": "A generation that quietly switched engines no longer counts as a failure",
+ "summary": "When a 3D generation lane loses your job, the platform resubmits it to another lane and you still get your model. Until now the abandoned attempt was recorded as a plain failure, with nothing tying it to the run that actually succeeded, so the platform's own health readout treated its recovery machinery as an outage: over the past week, 13 of the 14 lost jobs were rescued within seconds and all 14 still counted against generation health. Each recovered attempt is now linked to the run that replaced it, so the status page and our failure reports show what you experienced instead of what the machinery did.",
+ "link": "/status",
+ "tags": [
+ "improvement",
+ "infra"
+ ]
+ },
{
"date": "2026-08-14",
"type": "update",
@@ -226,6 +237,18 @@
"fix"
]
},
+ {
+ "date": "2026-08-14",
+ "type": "update",
+ "title": "Anyone can run an inference node: the operator client now ships CPU and GPU images, and proves itself end to end",
+ "summary": "The open inference network is the part of the roadmap where the compute stops being ours. The node operator client is how you join it: your machine generates a Solana keypair, registers itself with the platform, claims inference jobs, runs them locally with a real open model, and returns every answer with a signature that binds it to your key. There is no shared password anywhere in that flow, so nothing can be leaked and reused to impersonate your node, and the platform re-derives each signature from its own record of the job before it accepts the work. This release makes the graphics-card story honest. The old image claimed it would use an NVIDIA card automatically when one was present, and it never could: the runtime it shipped had no GPU support compiled in at all. There are now two real images, one for CPU and one for NVIDIA cards, and the client tells you which device actually ran each job. Ask for the GPU on a machine that cannot reach one and it stops immediately with the reason, instead of quietly earning at CPU speed for a month. Three faults in the job pipeline were also fixed: results submitted by a node hit a server error every time, registered nodes read back as unregistered, and every finished job leaked two silent background errors. A single command (npm run e2e) now registers a node, sends it a job, runs the model, and verifies the signed answer against the real platform code, and it fails loudly if any step regresses.",
+ "link": "/docs/inference-node-operator",
+ "tags": [
+ "feature",
+ "fix",
+ "infra"
+ ]
+ },
{
"date": "2026-08-14",
"type": "update",
@@ -329,7 +352,7 @@
"date": "2026-08-14",
"type": "update",
"title": "Cookbook posters came back, and a docs link you copy now reopens the page you were reading",
- "summary": "Eight developer pages were read against the code they describe and exercised in a real browser: the MCP 3D tool and CI asset gate recipes, and the payment sessions, developer platform, media API, agent-forged gallery, generation watch, and KOL tracker guides. Two things were broken for readers. Every cookbook recipe that shows a poster image was rendering without it: the figure system accepted three kinds of illustration and silently deleted the fourth, which is the kind the recipe posters use, with nothing in the page or the console to say a picture had gone missing. All of them render again. Separately, the previous/next buttons and the sidebar at the bottom of a documentation page produced a link that only worked while you stayed on the page: copy it, share it, or reload it, and you landed back on the page you started from instead of the one the link named. Those links now open the page they promise, correct the address bar to that page's own permanent link, and rename the browser tab to match, while in-page section links still jump to the section as before. The pages themselves were corrected where they had drifted from the code: the gallery's sample receipt quoted an old price for the tier it describes, the media API's two library samples reported counts no live response would return, and the KOL tracker described all four of its endpoints as sharing one dispatcher when the trade feed is served by its own file. Every command and code sample printed on the eight pages was run, and every link on them resolves.",
+ "summary": "Eight developer pages were read against the code they describe and exercised in a real browser: the MCP 3D tool and CI asset gate recipes, and the payment sessions, developer platform, media API, agent-forged gallery, generation watch, and KOL tracker guides. Two things were broken for readers. Every cookbook recipe that shows a poster image was rendering without it: the figure system accepted three kinds of illustration and silently deleted the fourth, which is the kind the recipe posters use, with nothing in the page or the console to say a picture had gone missing. All of them render again. Separately, the previous and next buttons and the sidebar on a documentation page produced a link that only worked while you stayed on the page: copy it, share it, or reload it, and you landed back on the page you started from instead of the one the link named. Those links now open the page they promise, correct the address bar to that page's own permanent link, and rename the browser tab to match, while in-page section links still jump to the section as before. The pages themselves were corrected where they had drifted from the code: the gallery's sample receipt quoted an old price for the tier it describes, the media API's two library samples reported counts no live response would return, and the KOL tracker described all four of its endpoints as sharing one dispatcher when the trade feed is served by its own file. Every command and code sample printed on the eight pages was run, and every link on them resolves.",
"link": null,
"tags": [
"fix",
@@ -379,6 +402,17 @@
"fix"
]
},
+ {
+ "date": "2026-08-14",
+ "type": "update",
+ "title": "Eleven links that led nowhere now land where they promise",
+ "summary": "The audit that guards every page against dead ends had a hole in it: the rule that serves the not-found page matched every address on the site, so the audit read every link as reachable and reported a clean bill of health it could not vouch for. With that fixed, eleven links turned out to lead to pages that were never built. Sign in to save on the choreography and gesture pages pointed at an address that does not exist, so the one thing those pages ask you to do could not be done. The extension terms page linked its own terms of service to a missing address, twice. The motion studio offered to show you how to capture an avatar from a selfie and led nowhere, as did the pricing page's link to your account, the memory tutorial's link to your library, and two entries in the dashboard trading menu. Pressing Message on a friend in the 3D world walked you into a blank page instead of the conversation, and now opens the thread with that person. Three outbound links to other projects' documentation had rotted and now point at the pages that replaced them. The audit itself was rebuilt to be trustworthy: it no longer calls a live site dead because the site answers a robot differently than a browser, and it no longer treats the blog and docs it ships, or the app manifest it generates, as missing files.",
+ "link": null,
+ "tags": [
+ "fix",
+ "improvement"
+ ]
+ },
{
"date": "2026-08-14",
"type": "update",
@@ -434,6 +468,16 @@
"fix"
]
},
+ {
+ "date": "2026-08-14",
+ "type": "update",
+ "title": "One cookbook recipe stopped introducing itself as a generic page in search results and shared links",
+ "summary": "All five cookbook recipes are drawn by one shared page template, and the server fills in each recipe's own title, description, and social card as the page is served. One recipe was skipping that step. The template carried a leftover marker naming a specific recipe as its own address, and the server reads that marker as this page already knows who it is, so it left that one page alone: text to 3D from the command line went out with the placeholder title every recipe shares and a one-line description that said nothing about it. Search results, and the preview card in any chat where the link was pasted, showed the placeholder instead of the recipe. The template now names the cookbook index rather than any single recipe, matching how the tutorial and walkthrough templates have always been set up, so all five recipes present themselves properly. A test now walks the real page templates and fails if any of them ever claims one of its own recipes again, which is what let this sit unnoticed.",
+ "link": "/cookbook/text-to-3d-cli",
+ "tags": [
+ "fix"
+ ]
+ },
{
"date": "2026-08-14",
"type": "update",
@@ -573,6 +617,16 @@
"docs"
]
},
+ {
+ "date": "2026-08-14",
+ "type": "update",
+ "title": "The 3D viewer no longer starts up blank when its box has not been measured yet",
+ "summary": "Every page that shows a 3D avatar builds its viewer the moment the page's script runs. On pages where the area holding the avatar has not been given its size yet, because the layout lands a frame later or the panel starts hidden, that meant the viewer was built at zero pixels wide and zero pixels tall. Graphics drivers reject an image with no width or height, so the viewer came up against a surface the driver had refused to create: the browser console filled with rejected-texture and incomplete-surface errors, and on the mocap studio the avatar stage was the one that hit it. The viewer now starts at a minimum size and resizes itself to the real box on the first frame that has one, which it was already set up to do for every later resize. Nothing changes for a page whose avatar area is measured up front, which is most of them.",
+ "link": "/mocap-studio",
+ "tags": [
+ "fix"
+ ]
+ },
{
"date": "2026-08-14",
"type": "update",
@@ -771,6 +825,17 @@
"security"
]
},
+ {
+ "date": "2026-08-14",
+ "type": "update",
+ "title": "The legal pages now say what is in them, and call the terms by their real name",
+ "summary": "The six pages under three.ws/legal (the index, privacy policy, terms of service, end-user license agreement, risk disclosure, and content policy) were audited in a real browser at phone and desktop widths. Every page loads, every link on them resolves, and the layout holds at 320 pixels wide. Three things were wrong. The terms of service was filed under the name Terms of Use in the site's page index, so a search result or a shared link showed one name while the document itself showed another; it is now Terms of Service everywhere. The privacy policy, the terms, and the EULA each carried a search-result description that only restated their title, so a search engine or a chat preview told you nothing about what was inside; each now describes what the document actually covers. And several of the policies linked to each other by their raw file names rather than their real addresses, which split the same document across two web addresses; they now all use the address shown in the browser bar. The content policy was also missing its publication date in the page index, which is what the changelog and sitemap read.",
+ "link": "/legal",
+ "tags": [
+ "fix",
+ "docs"
+ ]
+ },
{
"date": "2026-08-14",
"type": "update",
@@ -915,6 +980,17 @@
"fix"
]
},
+ {
+ "date": "2026-08-14",
+ "type": "update",
+ "title": "The security headers every three.ws page serves are now checked against the live site, not just the config",
+ "summary": "Every page three.ws serves carries a set of browser security rules: which scripts are allowed to run, that the connection must stay encrypted, that browsers must not second-guess a file's type, how much of the current address is passed on when you follow a link, and who is allowed to put a three.ws page inside a frame. Those rules are written in one config file, but until now nothing checked that the site actually answered with them. The sweep that loads real pages in a real browser now also compares each response against that config, and it can be pointed at the live site to check the whole catalogue in about a minute. Running it against production found no gap: all 25 of the most script-heavy pages, and the rest of the catalogue with them, serve exactly the rules the config declares. Two fixes came out of building it. Files that are not web pages, like robots.txt and the machine-readable service descriptions, were being reported as wrong when they were correct. And a page whose file changed while the server was running could be sent with the script rules of the previous version, which would have blocked its own scripts and shown a blank page; the server now notices the change and recomputes them.",
+ "link": null,
+ "tags": [
+ "security",
+ "infra"
+ ]
+ },
{
"date": "2026-08-14",
"type": "update",
diff --git a/public/changelog.xml b/public/changelog.xml
index 808d0e4fa..e5d903591 100644
--- a/public/changelog.xml
+++ b/public/changelog.xml
@@ -110,6 +110,14 @@
<category>infra</category>
<description>Every deploy carries database migrations that must run before the new code does. When one of them failed, the runner printed a single line with the database's own error and stopped, naming neither the file that failed nor the migrations queued behind it. Those stayed unapplied and unmentioned, so the site ended up serving code that reads columns the database never got, and the first visible symptom was an unrelated-looking error on a live page. The runner now names the failing file, how many ran before it, and every migration still waiting, and the deploy step that catches a database running behind the code is now part of the documented deploy path instead of something you had to remember to run.</description>
</item>
+ <item>
+ <title>A generation that quietly switched engines no longer counts as a failure</title>
+ <link>https://three.ws/status</link>
+ <guid isPermaLink="false">2026-08-14:A generation that quietly switched engines no longer counts as a failure</guid>
+ <pubDate>Fri, 14 Aug 2026 12:00:00 GMT</pubDate>
+ <category>improvement</category>
+ <description>When a 3D generation lane loses your job, the platform resubmits it to another lane and you still get your model. Until now the abandoned attempt was recorded as a plain failure, with nothing tying it to the run that actually succeeded, so the platform's own health readout treated its recovery machinery as an outage: over the past week, 13 of the 14 lost jobs were rescued within seconds and all 14 still counted against generation health. Each recovered attempt is now linked to the run that replaced it, so the status page and our failure reports show what you experienced instead of what the machinery did.</description>
+ </item>
<item>
<title>A paid-call share link now previews on Slack, Telegram and LinkedIn</title>
<link>https://three.ws/pay</link>
@@ -174,6 +182,14 @@
<category>security</category>
<description>The two avatar onboarding endpoints (the one that opens the photo editor, and the one that attaches a finished avatar to an agent) accepted a request that arrived with your login cookie but no proof you meant to send it. A page you had merely visited could therefore repoint your agent to a different body, or spend your hourly avatar-editor allowance, without you doing anything. Both now require the same one-time token every other account-changing action on three.ws already asks for, and both refuse the request outright without it. API-key callers are unaffected. A second change: when the photo editor is not configured in an environment, the request is turned away before it counts against your hourly allowance, so a temporary outage no longer locks you out of the flow for an hour once it is back.</description>
</item>
+ <item>
+ <title>Anyone can run an inference node: the operator client now ships CPU and GPU images, and proves itself end to end</title>
+ <link>https://three.ws/docs/inference-node-operator</link>
+ <guid isPermaLink="false">2026-08-14:Anyone can run an inference node: the operator client now ships CPU and GPU images, and proves itself end to end</guid>
+ <pubDate>Fri, 14 Aug 2026 12:00:00 GMT</pubDate>
+ <category>feature</category>
+ <description>The open inference network is the part of the roadmap where the compute stops being ours. The node operator client is how you join it: your machine generates a Solana keypair, registers itself with the platform, claims inference jobs, runs them locally with a real open model, and returns every answer with a signature that binds it to your key. There is no shared password anywhere in that flow, so nothing can be leaked and reused to impersonate your node, and the platform re-derives each signature from its own record of the job before it accepts the work. This release makes the graphics-card story honest. The old image claimed it would use an NVIDIA card automatically when one was present, and it never could: the runtime it shipped had no GPU support compiled in at all. There are now two real images, one for CPU and one for NVIDIA cards, and the client tells you which device actually ran each job. Ask for the GPU on a machine that cannot reach one and it stops immediately with the reason, instead of quietly earning at CPU speed for a month. Three faults in the job pipeline were also fixed: results submitted by a node hit a server error every time, registered nodes read back as unregistered, and every finished job leaked two silent background errors. A single command (npm run e2e) now registers a node, sends it a job, runs the model, and verifies the signed answer against the real platform code, and it fails loudly if any step regresses.</description>
+ </item>
<item>
<title>Apps that sign in with a client secret can now revoke and inspect their own tokens</title>
<link>https://three.ws/docs/authentication</link>
@@ -252,7 +268,7 @@
<guid isPermaLink="false">2026-08-14:Cookbook posters came back, and a docs link you copy now reopens the page you were reading</guid>
<pubDate>Fri, 14 Aug 2026 12:00:00 GMT</pubDate>
<category>fix</category>
- <description>Eight developer pages were read against the code they describe and exercised in a real browser: the MCP 3D tool and CI asset gate recipes, and the payment sessions, developer platform, media API, agent-forged gallery, generation watch, and KOL tracker guides. Two things were broken for readers. Every cookbook recipe that shows a poster image was rendering without it: the figure system accepted three kinds of illustration and silently deleted the fourth, which is the kind the recipe posters use, with nothing in the page or the console to say a picture had gone missing. All of them render again. Separately, the previous/next buttons and the sidebar at the bottom of a documentation page produced a link that only worked while you stayed on the page: copy it, share it, or reload it, and you landed back on the page you started from instead of the one the link named. Those links now open the page they promise, correct the address bar to that page's own permanent link, and rename the browser tab to match, while in-page section links still jump to the section as before. The pages themselves were corrected where they had drifted from the code: the gallery's sample receipt quoted an old price for the tier it describes, the media API's two library samples reported counts no live response would return, and the KOL tracker described all four of its endpoints as sharing one dispatcher when the trade feed is served by its own file. Every command and code sample printed on the eight pages was run, and every link on them resolves.</description>
+ <description>Eight developer pages were read against the code they describe and exercised in a real browser: the MCP 3D tool and CI asset gate recipes, and the payment sessions, developer platform, media API, agent-forged gallery, generation watch, and KOL tracker guides. Two things were broken for readers. Every cookbook recipe that shows a poster image was rendering without it: the figure system accepted three kinds of illustration and silently deleted the fourth, which is the kind the recipe posters use, with nothing in the page or the console to say a picture had gone missing. All of them render again. Separately, the previous and next buttons and the sidebar on a documentation page produced a link that only worked while you stayed on the page: copy it, share it, or reload it, and you landed back on the page you started from instead of the one the link named. Those links now open the page they promise, correct the address bar to that page's own permanent link, and rename the browser tab to match, while in-page section links still jump to the section as before. The pages themselves were corrected where they had drifted from the code: the gallery's sample receipt quoted an old price for the tier it describes, the media API's two library samples reported counts no live response would return, and the KOL tracker described all four of its endpoints as sharing one dispatcher when the trade feed is served by its own file. Every command and code sample printed on the eight pages was run, and every link on them resolves.</description>
</item>
<item>
<title>Creator subscriptions renew again instead of failing every billing run in silence</title>
@@ -286,6 +302,14 @@
<category>docs</category>
<description>A sweep of the viability, press kit, partners, agent identities, OKX marketplace, avatar engines, stage, and world lines pages, with every command in them actually run and every link followed. Four things were wrong and are now right. The partners page pointed the NVIDIA row at a docs path when the card on the live page goes to /nvidia. The agent identity docs claimed a job handle expires after about thirty days, which nothing in the code enforces; the page now states what is true, that the signed handle carries no expiry and the job record stays readable. The marketplace page showed a create call replying with a queued status and a generation stage, neither of which the endpoint has ever returned, so an agent coding against the sample would have waited on a state that never arrives; the page and the endpoint's own published example now show the running status and generate stage the code emits. The same page read as though our marketplace listing were live when it is still in review, and now says so plainly. The world lines anti-cheat table was also missing one of the per-IP limits it lists.</description>
</item>
+ <item>
+ <title>Eleven links that led nowhere now land where they promise</title>
+ <link>https://three.ws/changelog</link>
+ <guid isPermaLink="false">2026-08-14:Eleven links that led nowhere now land where they promise</guid>
+ <pubDate>Fri, 14 Aug 2026 12:00:00 GMT</pubDate>
+ <category>fix</category>
+ <description>The audit that guards every page against dead ends had a hole in it: the rule that serves the not-found page matched every address on the site, so the audit read every link as reachable and reported a clean bill of health it could not vouch for. With that fixed, eleven links turned out to lead to pages that were never built. Sign in to save on the choreography and gesture pages pointed at an address that does not exist, so the one thing those pages ask you to do could not be done. The extension terms page linked its own terms of service to a missing address, twice. The motion studio offered to show you how to capture an avatar from a selfie and led nowhere, as did the pricing page's link to your account, the memory tutorial's link to your library, and two entries in the dashboard trading menu. Pressing Message on a friend in the 3D world walked you into a blank page instead of the conversation, and now opens the thread with that person. Three outbound links to other projects' documentation had rotted and now point at the pages that replaced them. The audit itself was rebuilt to be trustworthy: it no longer calls a live site dead because the site answers a robot differently than a browser, and it no longer treats the blog and docs it ships, or the app manifest it generates, as missing files.</description>
+ </item>
<item>
<title>Every agent now has a verifiable on-chain history you can read on its page</title>
<link>https://three.ws/docs/agent-index</link>
@@ -326,6 +350,14 @@
<category>docs</category>
<description>Every tutorial on the site is written as a markdown file that lives beside the other documentation, and the links inside it were written against that location. The page that renders them was reading those links as if they pointed at site addresses instead, so a tutorial that sent you to a related guide, or cited the exact source file a step describes, sent you to an address that does not exist. Links that reach across to another document, that point at a specific section of one, or that cite a file in the source tree are now each resolved to the right destination, with source files opening on GitHub. The scene tutorial's own list of related tutorials pointed at a secondary address that rendered them under a generic title, and now uses the real one. Separately, the AWS page claimed the platform's production deployment runs in AWS us-east-1. It does not: production runs on Google Cloud Run, and us-east-1 is where the AWS Marketplace billing and entitlement APIs are called. The page now says that instead.</description>
</item>
+ <item>
+ <title>One cookbook recipe stopped introducing itself as a generic page in search results and shared links</title>
+ <link>https://three.ws/cookbook/text-to-3d-cli</link>
+ <guid isPermaLink="false">2026-08-14:One cookbook recipe stopped introducing itself as a generic page in search results and shared links</guid>
+ <pubDate>Fri, 14 Aug 2026 12:00:00 GMT</pubDate>
+ <category>fix</category>
+ <description>All five cookbook recipes are drawn by one shared page template, and the server fills in each recipe's own title, description, and social card as the page is served. One recipe was skipping that step. The template carried a leftover marker naming a specific recipe as its own address, and the server reads that marker as this page already knows who it is, so it left that one page alone: text to 3D from the command line went out with the placeholder title every recipe shares and a one-line description that said nothing about it. Search results, and the preview card in any chat where the link was pasted, showed the placeholder instead of the recipe. The template now names the cookbook index rather than any single recipe, matching how the tutorial and walkthrough templates have always been set up, so all five recipes present themselves properly. A test now walks the real page templates and fails if any of them ever claims one of its own recipes again, which is what let this sit unnoticed.</description>
+ </item>
<item>
<title>Oracle pages now say the data is briefly unavailable instead of showing you a blank market</title>
<link>https://three.ws/changelog</link>
@@ -430,6 +462,14 @@
<category>sdk</category>
<description>The npm package that turns a prompt into a 3D model documented two things it did not do. It listed a per-call option for supplying your own Meshy or Tripo key, and that option was read from nowhere: the key was silently dropped, so anyone following the README's own example was told to add a key that the generator then said was missing. It also documented paying per call in USDC as the default billing lane, when the flag it described was ignored by the endpoint it was sent to. Both are now real. A key passed on the call is used, overrides the one set on the client, and is resent on every status check, which the generator requires and the old code never did, so a job started with your own key could never be polled to completion. The pay-per-call lane now routes to the endpoint that actually accepts payment, quotes Solana USDC, and hands back a job you poll for free, and it refuses upfront the options it cannot honour rather than accepting them and ignoring them. The package now also exposes a way to read one job's state without waiting for it, so a generation survives a restart. Alongside that, the reference pages listed a price and a set of engines that had both moved on, named an error the service never returns, and pointed at three addresses that no longer exist: the sign-up link for API keys in the copy-trading and LLM router packages went to a dead page, as did the link to the 3D studio tool server. All of them now resolve.</description>
</item>
+ <item>
+ <title>The 3D viewer no longer starts up blank when its box has not been measured yet</title>
+ <link>https://three.ws/mocap-studio</link>
+ <guid isPermaLink="false">2026-08-14:The 3D viewer no longer starts up blank when its box has not been measured yet</guid>
+ <pubDate>Fri, 14 Aug 2026 12:00:00 GMT</pubDate>
+ <category>fix</category>
+ <description>Every page that shows a 3D avatar builds its viewer the moment the page's script runs. On pages where the area holding the avatar has not been given its size yet, because the layout lands a frame later or the panel starts hidden, that meant the viewer was built at zero pixels wide and zero pixels tall. Graphics drivers reject an image with no width or height, so the viewer came up against a surface the driver had refused to create: the browser console filled with rejected-texture and incomplete-surface errors, and on the mocap studio the avatar stage was the one that hit it. The viewer now starts at a minimum size and resizes itself to the real box on the first frame that has one, which it was already set up to do for every later resize. Nothing changes for a page whose avatar area is measured up front, which is most of them.</description>
+ </item>
<item>
<title>The activity floor stops reporting a flat week when nothing has closed yet</title>
<link>https://three.ws/activity</link>
@@ -574,6 +614,14 @@
<category>fix</category>
<description>Three internal dashboards watch the platform for us: one for site and cron health, one for every money subsystem, one for payment outcomes. Two of them had gone quietly wrong. The platform board judged each scheduled job against a hand-written idea of how often it should run, and three jobs that run once a day were being judged against a window of minutes, so they reported as broken for roughly 23 hours out of every day and the board could never come back all-clear; every job is now measured against the schedule it genuinely runs on. One of its probes was also posting to a payments endpoint without declaring it was sending JSON, so it had been failing on its own mistake rather than on anything real. The money board was reading the wrong timestamp column on two payment logs and showing no activity at all for a payment ring that settles thousands of payments a day. It also accepted the credential our fund-moving scheduled jobs carry as an ops password; it now takes the same admin sign-in as the other boards, so an ops credential can never be escalated into triggering a payment job.</description>
</item>
+ <item>
+ <title>The legal pages now say what is in them, and call the terms by their real name</title>
+ <link>https://three.ws/legal</link>
+ <guid isPermaLink="false">2026-08-14:The legal pages now say what is in them, and call the terms by their real name</guid>
+ <pubDate>Fri, 14 Aug 2026 12:00:00 GMT</pubDate>
+ <category>fix</category>
+ <description>The six pages under three.ws/legal (the index, privacy policy, terms of service, end-user license agreement, risk disclosure, and content policy) were audited in a real browser at phone and desktop widths. Every page loads, every link on them resolves, and the layout holds at 320 pixels wide. Three things were wrong. The terms of service was filed under the name Terms of Use in the site's page index, so a search result or a shared link showed one name while the document itself showed another; it is now Terms of Service everywhere. The privacy policy, the terms, and the EULA each carried a search-result description that only restated their title, so a search engine or a chat preview told you nothing about what was inside; each now describes what the document actually covers. And several of the policies linked to each other by their raw file names rather than their real addresses, which split the same document across two web addresses; they now all use the address shown in the browser bar. The content policy was also missing its publication date in the page index, which is what the changelog and sitemap read.</description>
+ </item>
<item>
<title>The limits you set on an agent's spending can no longer be skipped</title>
<link>https://three.ws/docs/a2a-payments</link>
@@ -678,6 +726,14 @@
<category>feature</category>
<description>Stake SOL behind an agent you believe in and earn a share of a daily reward pool, sized by that agent's signed on-chain record: accepted tasks, passed validations and real reviews pay, while failed validations, disputes and revocations cost double what work earns. The market page had shipped without the code that drives it, so it rendered empty no matter what was staked; it now loads live, ranks agents by the conviction escrowed behind them, shows any wallet's positions with earnings that update as they accrue, and settles a withdrawal back to the staker who signed the stake. Your principal is never slashed and can be withdrawn at any time: bad conviction earns nothing, it never costs you your stake. Every number is re-derivable from the chain, the exact formula runs identically in your browser and on our servers, and the whole cycle is provable on devnet with one command. Devnet is open to everyone; mainnet stays owner-gated.</description>
</item>
+ <item>
+ <title>The security headers every three.ws page serves are now checked against the live site, not just the config</title>
+ <link>https://three.ws/changelog</link>
+ <guid isPermaLink="false">2026-08-14:The security headers every three.ws page serves are now checked against the live site, not just the config</guid>
+ <pubDate>Fri, 14 Aug 2026 12:00:00 GMT</pubDate>
+ <category>security</category>
+ <description>Every page three.ws serves carries a set of browser security rules: which scripts are allowed to run, that the connection must stay encrypted, that browsers must not second-guess a file's type, how much of the current address is passed on when you follow a link, and who is allowed to put a three.ws page inside a frame. Those rules are written in one config file, but until now nothing checked that the site actually answered with them. The sweep that loads real pages in a real browser now also compares each response against that config, and it can be pointed at the live site to check the whole catalogue in about a minute. Running it against production found no gap: all 25 of the most script-heavy pages, and the rest of the catalogue with them, serve exactly the rules the config declares. Two fixes came out of building it. Files that are not web pages, like robots.txt and the machine-readable service descriptions, were being reported as wrong when they were correct. And a page whose file changed while the server was running could be sent with the script rules of the previous version, which would have blocked its own scripts and shown a blank page; the server now notices the change and recomputes them.</description>
+ </item>
<item>
<title>The selfie camera now catches a blurry or blown-out shot before it costs you a build</title>
<link>https://three.ws/create/selfie</link>
diff --git a/public/features.json b/public/features.json
index 90c596320..1f2a0cae9 100644
--- a/public/features.json
+++ b/public/features.json
@@ -1,5 +1,5 @@
{
- "generated_at": "2026-08-14T20:26:47.957Z",
+ "generated_at": "2026-08-14T20:45:58.830Z",
"generated_by": "scripts/build-page-index.mjs from data/pages.json",
"site": {
"name": "three.ws",