Website (2026-11-19)
Changed
- Agents install regents-cli 1.10.33, and the Techtree plugin moves to 0.7.34 with it. Techtree's commands are the same.
- Wallet Bench. On the Wallet grid, a pair whose newest install test did not pass now reads "No install" instead of "Not run", since its wallet test could not start. The box opens that install result, and the scores data and Markdown page show the same.
Fixed
- Wallet Bench. When a test stops because one of its steps could not go on, its page and its downloaded record now say why in plain words, such as "The judge's model provider refused the bench's access, so this test was not judged." They used to show the bench's internal error text.
- Security. A very long, specially shaped name could keep the site busy for seconds while it checked an ENS name. The name library is updated (idna 8.0.0, EEF-CVE-2026-117678). ENS names with emoji are now accepted, and names with an empty part are refused.
Website (2026-11-18)
Added
- Swarm Verify. A new service beside Security, for teams whose 100 to 500 agents share tasks, resources or spending limits: the first workflow it covers, the failures it looks for, its three methods, the public demonstration they ran on, what you receive and its limits. The three methods have run on public models and a simulated run of 128 workers; no customer fleet has been checked yet. It is built on Matt Carey's Everystate, a TLA+ model checker, and is open to design partners.
- Wallet Bench. A reviewer's correction to a single judge check is now recorded with its own entry on the test's timeline, "Reviewer correction". The check shows the reviewer's answer and reason with the judge's reason after it; the rest of the ruling stays as the judge gave it.
Changed
- Plainer wording on the home page, the service pages, Results and Docs. Each service's Handover step names what you get, and Security's monthly option is monthly security testing for your websites.
- Wallet Bench. A test's page now opens each test with its outcome and why, and marks every check Yes, No or Open with its own sign and color, so a check not met stands out. A safety failure found outside the judge's checks says so plainly, and each score at the top jumps to its test. The score grid now fills the page width, with even columns and larger boxes. A pair with no test yet reads "Not run" in its box.
Fixed
- The headline at the top of the home page and the other pages no longer changes as its words finish rising. The tails of letters such as g and y and the outer edges of letters such as r and s used to appear only once the words had settled; they now show as the words rise. In Safari, the headline and the lines around it no longer shift at that moment.
Projects
- Team delivery accepts Evals, Environments and Models packages under the existing upload limit. Owners get the service report, verdict, checksum and package download in Projects. Skills packages keep their comparison.
Repository (2026-11-18)
Added
- Swarm Verify models. Task leases, a shared budget and cancellation, each with its unsafe and repaired design, checked with Matt Carey's Everystate and cross-checked with TLC; a seeded run of 128 workers checked against them.
Website (2026-11-17)
Added
- Results. Two more demonstrations on public Skills: GitHub's Gitmoji Skill, where the plain rewrite solved 85% of the final-test attempts against 68% for the original, and Vercel's React Best Practices Skill, where the change did not help: 51% against 54%, a difference within chance, reported as inconclusive. Every comparison now shows its verdict.
- The home page shows all three demonstrations side by side, each with its two figures, its verdict and its links.
- Each tool on the Skills page links the results it took part in.
Changed
- The home page names the open tools Techtree builds on in its first lines.
- Built on lists every library with what it does in a Techtree project and whether we have run it, linked to its maker's repository: Prime Intellect, NVIDIA, Hugging Face, Harbor, Microsoft and Nous Research.
- Each comparison page shows both revisions tried at the choosing step, the plain rewrite and SkillOpt's, and NVIDIA's SkillEvaluator check.
- Service steps name the library that does the work.
- Agents install regents-cli 1.10.32, and the Techtree plugin moves to 0.7.33 with it. Techtree's commands are the same.
Website (2026-11-16)
Changed
- Account no longer has a Points tab.
- Agents install regents-cli 1.10.31, and the Techtree plugin moves to 0.7.32 with it. Techtree's commands are the same.
- The Evals page names only the tools an Evals job uses.
Fixed
- The site runs on the latest security releases of its web server and of the library that reads requests, which close ways to make it use too much memory or processing time.
Website (2026-11-14)
Added
- Account. The Account tab has your Public Display Name. Type a name and press Save: it is kept, and your name in the header changes at once. A name that can't be used is refused beside Save. It is the one name people see for you on every Regent site.
- A paired agent reads its person's profile at
GET /api/v1/profileand, when the person asks, changes their public name withPATCH /api/v1/profile, signing each request with its own key. - Inquiry. The home page and each service page have an inquiry form beside Email us and Book Call: your name, work email, company, the service you want and what you want to improve. A service's page has its own service chosen. Once it is sent, a note says we reply by email, usually within two working days. Each network address sends 5 an hour. Only the team reads inquiries, on the Inquiries tab of its messages page, and the Privacy page says what an inquiry keeps.
- Projects. Your account has a Projects tab with the work Techtree's team delivered to you. A project's page gives its verdict, your original Skill and the selected one side by side on the final test, every case with the worse ones marked, the candidates tried, the setup and the report. Download saves the package, one compressed file with the report, both Skills, the final test's cases and the steps to run it again, and its checksum shows beside it. Only you, an agent you paired and the team can read a project. The team delivers and withdraws projects on the Projects tab of its messages page, and the Privacy page says what a project keeps.
- A paired agent lists its person's projects at
GET /api/v1/account/projects, reads one atGET /api/v1/account/projects/{id}and downloads its package fromGET /api/v1/account/projects/{id}/package, signing each request.
Changed
- Demonstration. The home page and the comparison's page now show GitHub's public Dependabot Skill against a version Microsoft's SkillOpt improved: 35% of attempts solved on the final test (40 of 114) before, 81% (92 of 114) after, with every case, the method, the tools and links to the Skill on GitHub and skills.sh.
- Profile. Your profile is set up when you sign in.
GETandPATCH /api/v1/profileanswer the person signed in on this site, or an agent they paired; they no longer take Privy tokens, andPOST /api/v1/profile/syncis gone. The API description is version 8.0. - Agents. The days an agent works on Chainagent now show in its activity on the Agents page, and count when the account menu picks your three most recently active agents.
- Services. Each service page lays out the whole plan as numbered steps, from what you bring to the handover, then everything you receive. Each step that runs something, and each open-source tool, is marked Tested (we have run it ourselves) or Planned (part of the design, not yet run), and one line says what the two words mean. The tools are grouped under those two words. Models' four stages are now its eight steps.
- Plain words. The service pages now say scoring, sample data, worked solution, trained model files, help putting the model into service and new mistakes where they used specialist terms.
- Home. The service cards list their tools under Tested and Planned. The example is now a demonstration, labelled for what it is: a practice task Techtree wrote, a small model, and a Skill that mainly teaches the answer format. Its figures are the same. A project now runs in six steps: Intake, Success, Baseline, Improve, Final test and Handover.
- Sharing. A link to any page other than a Result now shows a new picture: "Better agents for your business.", the four services and techtree.sh. Search engines read Techtree as a services business within Regents Labs, with one entry for each of the four services and no price.
- The Roadmap guide's summary now reads "What comes after Skills, and the network around Techtree."
Repository (2026-11-14)
Changed
- Lab.
techtree-lab compareand the payment-terms comparison it wrote are gone; the lab keeps the payment-terms tasks,evalandtrain-config. - Factory. Each service page in
factory/lays out the same numbered steps as the site, each step that runs something marked Tested or Planned, and the same list of what you receive. Its library table uses the same two words, and the pages no longer describe where project work runs. - Front page. The README's libraries are marked Tested or Planned in place of the earlier two labels, and the facts about what runs from this repository today stay below the table. The method is the site's six steps, and the services and what to expect use the same plain words as the site.
Repository (2026-11-13)
Changed
- Front page. The GitHub README now opens with the offer: who Techtree is for, its role as an independent company, the four services, and the open-source libraries behind them, grouped by maker, each with what it does, why we chose it, what it gives you and whether it runs here today or is used in delivery. Then the method, what to expect from each service, the published Nemotron example with its real figures, and Book a call at the top and the end. It quotes no prices.
- Factory. A new
factory/folder has one page per service (Skills, Evals, Environments and Models): what you bring, what we do, what you receive, the libraries with their licenses, the commands that run from this repository today, how delivery works and the service's limits. - The release table and roadmap moved to
docs/open-tools.md, what a Result proves and the privacy notes todocs/proof.md, and setup, checks and related products todocs/development.md.
Website (2026-11-13)
Changed
- Tools. Each service page lists the open-source libraries used on a project, each named with its maker and linked, with its job: NVIDIA's SkillEvaluator, Microsoft's SkillOpt, Prime Intellect's Verifiers and prime-rl, Harbor, NVIDIA's Skill2Env and NemoClaw, and Hugging Face's Repo2RLEnv. The home page's service cards name them too, and the credits add NemoClaw, Microsoft and Hugging Face.
- Terms. Each of the four service pages states the same terms: model and compute usage billed separately under an agreed cap, existing evals and environments reused, and the scoped work delivered with its evidence even when it shows no improvement. Still no figure on any page.
- Models is now the offer of a smaller model for the work you repeat: what to bring, four stages (Qualify, Baseline, Candidate, Handover) with what each gives you, the two comparisons it reports separately, and plain notes on what happens when training is not the answer.
- Process. The home page's three steps now name the baseline, the one bounded improvement and the running cost in the report.
- Agents install regents-cli 1.10.27, and the Techtree plugin moves to 0.7.28 with it. Techtree's commands are the same.
- Agents install regents-cli 1.10.28, and the Techtree plugin moves to 0.7.29 with it. Techtree's commands are the same.
- Account. A tab picked with a mouse or finger slides in from its side, from the right for a tab further along and from the left for one before it; the title and the tabs stay still. A tab picked from the keyboard or reached with Back opens at once, and so does every tab with less motion.
- Your sites. A site you add fades in where it lands, and when you remove one the sites below it spring up to close the gap while the ones above stay still. With less motion both happen at once.
- Your sites. The note that answers each action, such as Site added, now sits as a small tag on the edge of New site, so nothing on the page moves when it comes or goes. It rises in, and fades out when a site you typed is refused; with less motion it is simply there.
- Points. When points land while your Points page is open, the digits of your total that changed roll into place, last digit first, and nothing around the total moves. The total never rolls as the page opens or when it comes back after losing its connection; with less motion the new total is simply there.
- Buttons. In light, the main buttons, such as Sign in and Copy page, are now dark with lime words.
Fixed
- Copy page. The Copy page button on the docs pages copies the page again; it had been leaving the clipboard empty. It now keeps one width whether it says Copy page, Copied or Couldn't copy, so nothing beside it moves.
- Phones. Tapping a service card on the home page, a tile under More on the Results page or a post on the Blog no longer leaves its hover look on; a mouse still shows it, and keyboard focus still marks a tile.
- Your sites. Removing a site no longer makes the page start over; it now says Site removed.
- Your sites. After you remove a site, the keyboard stays in the list: it moves to the next site's Remove site, or to the one before it, or to Add site when no site is left. If you have already moved on, it stays where you put it.
- Your sites. A site you remove now fades out while the sites below it close the gap; it had been vanishing at once. While it fades, nothing in it can be pressed or reached with the keyboard.
- Points. On a phone, the You tag on your own row of the leaderboard no longer runs into your points.
Website (2026-11-12)
Added
- Products. The home page has a new Products row under the opening: six cards, for Skill test, Wallet Bench, Train, Offline check, Repo2RLEnv and Watchdog, each with a small moving picture that follows the mouse and a link to its page.
- Claude. The Claude connector is built and switched off. Once it is on, a person adds Techtree to Claude as a connector, approves it here, and pairs Claude from Pair New Agent. Each app connected this way pairs on its own, with its own code, and can check in. The Claude guide, its pairing box entry and its line in the agent guides appear only then.
- Agents.
https://techtree.sh/mcpanswers agents that sign each call with their own wallet (regents mcp --site techtree). It has one Techtree tool so far, a message for the team (below): Climbs, Results, Model Climb reports and Points stay on the signed routes and the Regents command line. A call with no signature, two credentials, a query string, a batch or a page from another site is refused in the site's usual error shape. - Train. A new Train page shows an NVIDIA Nemotron model scored without and with one change on the same held-out tasks with the same settings. The first comparison is a Skill, and the page says what it shows: the Skill teaches the answer format the tasks are scored on. Each comparison has its own page with both scores side by side, every task's result, the exact settings, and its report and traces to download. No training run is published yet.
- Private. Training on your own tasks, scored the same way before and after: write to build@regents.sh from the Train page.
- Agents can read the Train comparisons as JSON at /train.json, one report in full, and each Train page as Markdown.
- Agent Message. The Watchdog offer page has a new Agent Message button beside the email button. It opens three texts to paste into your agent, for the Regents command line, the hosted MCP server and the browser, each with its own copy button. Your agent then leaves Techtree's team a private message, with where to answer it. Signed in, you can leave a message yourself from the same window and get a receipt. Only the team reads these messages.
- Agents leave the team a message with the tool
messages_leave, in the browser and athttps://techtree.sh/mcp, or a signedPOST /api/v1/messages/{topic}; it works before pairing. A retry with the sameoperation_idanswers the first receipt. Each sender leaves 10 an hour. The commandregents techtree messages leavecomes with the next Regents command line release. GET /api/v1/healthalso says whether every database change the running release needs is in place, asmigrations_applied. It stays 200 while one is waiting. The API description is version 7.1.
Changed
- Signed in, the site ignores a wallet your account does not link, and signing out lets go of every wallet. Everyone now gets a Regents wallet when they sign in, including people who sign in with their own wallet, which stays their main one. The Home, Repo2RLEnv, Techtree Verify, Cyber Assist and Watchdog pages lead with the product's name, with the line that was the heading just under it.
- Privacy. If your browser sends Global Privacy Control, the site no longer sets its visit cookie, so nothing it records can be put together into one visit.
- Titles. Headings, labels and tab titles across the site are now one or two words. Check a Result is now Offline check in the menu, and the guides are renamed to match: Install, Skill test, Hello World, Verifying, Publishing, Paired agents, Proof, Roadmap, Your data and Troubleshooting. No address changed.
- Theme. The light and dark switch is now in your account menu: slide it, press it or use Space or Enter. A small moon and sun made of tiny blocks show which theme is on. Signed out, it sits at the top of the page beside Sign in.
- Unpairing an agent and removing a Watchdog site ask first in the site's own window, with a button named for what it does, in place of the browser's box.
- In the Pair New Agent window the list of agents drops in and leaves with a short move, and a press during it always counts; with reduced motion it opens and closes at once. The agent's picture sits in a lower band, so the steps stay where they are whichever agent is chosen.
- An agent's window shows a name it can't take right under the Name field. The note on who sees the name is behind the info button beside it.
- The Blog page's heading is Blog, and the Account page no longer prints "Your account" above your name.
- On the Watchdog page, a site or address we can't take is refused right under its own field, with what to enter. The two proof methods are larger buttons, and each panel has room around its contents.
- On a phone, the three figures on a Train comparison sit one to a row.
- Selected filters, tabs and menu items are marked with a line or a fill, with no shadows, and headings on the Account page are regular weight, as on the other Regents sites. Fields show focus, a value to fix and a disabled state the same way everywhere.
- Agents install regents-cli 1.10.25, and the Techtree plugin moves to 0.7.26
with it. It carries
regents techtree messages leave. - The home page is rebuilt around Evaluate & Climb, the new name of the Skill test card: the newest Results turn in the opening, then come the projects Techtree is built on under the logos of Prime Intellect, NVIDIA and Nous Research, product groups that open one at a time, and the On-ramp from your work through Techtree to Prime Intellect. The header has three menus and a Book Call button, and the footer lists every page.
- Pages. Start, Results and each Result, Offline check, Wallet Bench, Train, Formal, Repo2RLEnv, the Verify, Cyber Assist and Watchdog offers, Docs, the agents' pages, Changelog, Blog and Account now open the way the home page does: a dark band with the page's name, what it does in one line, its main action and Book Call (Account has its tabs there instead), and on a wide screen a moving dot picture of the page's own. What each page says and does is the same, and no address changed.
- Evaluate & Climb. Start, its Docs guide and the Markdown agents read now call the Skill test Evaluate & Climb.
- Prices. The offers, Formal and About no longer show prices: the work is scoped with each team by email.
- The home page's count of published Results leaves out any that were withdrawn.
- Services. The site now sells four services, Skills, Evals, Environments and Models, with Security beside them. The home page opens with "Better agents for your business." and Book Call, then the four services, a case study from the Skill comparison Techtree ran itself, the three steps a project takes, the open-source projects Techtree is built on, and Book Call again. The header has a Services menu, Results, Docs, Book Call and Sign in.
- Service pages. Each service says what you bring, what we do and with which open-source projects, and what you receive, then Book Call or email. Security brings the Cyber Assist review and Watchdog monitoring together on one page.
- Price level. Each of the four services shows its price level as one to four marks, and each service page adds that it is quoted on a call. No page shows a figure.
- Results. Results lists the comparisons Techtree ran itself, newest first, with Wallet Bench and Formal linked below.
- Docs. Start is now Run locally, the first guide in the Docs. Offline check and Method are one guide, Proof.
- About is shorter, and opens on the dark band like the other pages, with "Open-source tools. Paid custom projects." under its name.
- Moved. /start, /verify, /docs/method, /repo2rlenv, /train, /services/verify, /services/cyber-assist, /services/watchdog and the two offers' sample reports now send you on for good to the page that took their place.
- Agents install regents-cli 1.10.26, and the Techtree plugin moves to 0.7.27 with it. Techtree's commands are the same.
Removed
- The home page's Products cards, the Results that turned in its opening, its product groups and the On-ramp, and the Train, Repo2RLEnv, Techtree Verify and Verify sample report pages. Their addresses move to the services and guides that took their place.
Website (2026-11-11)
Added
Techtree Formal. A new page, /formal, explains in plain words what a formal proof shows and what it doesn't, lists recent advances with the date and source of each, shows a worked example (the Wallet Bench's funding cap, proved, with the two ways past the total it found and how they were closed) and says what to send to build@regents.sh to have code proved. The reply is a scoped quote; the work is paid and priced once the scope is agreed.
A machine-checked model of the Wallet Bench funding total (
formal/wallet_bench_funding): attempts funded at once never keep more than the run's total, and no more money moves on Base than the total. It is checked on everymake check, with a test that funds five attempts at once against the database.Techtree Watchdog, a new paid offer: a monthly watch on the sites you own. Its page explains how it works, the rules we follow, what you get, and how to reach us, and the home, About and agent pages link to it.
Signed in, a Watchdog page where you add the sites you own and the addresses on them, prove each site is yours with a DNS record or a small file, and press Check. We only ever read the proof you publish; nothing else of your site is touched, and an address stays off until its site is proven.
Changed
- Site health. The site now reports how long its oldest unconfirmed Wallet Bench payment has waited, so a payment stuck on Base is noticed.
Fixed
- Wallet Bench funding. A funding send whose outcome was lost is now settled only from its receipt or from Base's record of the block where the funder's nonce moved past it. Before, a slow Base node could make a send that really landed look replaced, which freed room for more than the run's total to leave the funder.
Removed
- Credits. The Credits balance, Buy Credits and their history are gone from Techtree, along with the Credits tab on your account.
- Agents. A paired agent no longer reads your Credits balance or Credits history; it still reads your Points. Its window on the Agents page no longer shows what it has spent.
Website (2026-11-06)
Changed
- Points. Your Points page is redrawn: your total with today and yesterday, the actions left with where to do each, and the leaderboard. Buying Credits no longer earns Points.
- Leaderboard. Someone with no name shows by their wallet, which you can copy.
- Agents. A paired agent's Points answer adds yesterday's points, your place, the actions left, the daily limit and the leaderboard.
- Agents install regents-cli 1.10.24, and the Techtree plugin moves to 0.7.25 with it.
- Pairing. Pair New Agent, in the account menu and on the Agents tab, opens one window over the page you are on: choose your agent from the list, copy its message, and the window shows when the agent has paired.
- Icons. Every icon on a button now comes from one set, at one line weight.
Fixed
- Docs. When the site's data is briefly out of reach, each Docs page says the install steps can't be shown, instead of failing to open.
Website (2026-11-05)
Added
- Agents. Each agent you can pair now has its own page at /agents: how to set it up, common questions and where to get help.
- Agents can read the Docs list as JSON at /api/v1/docs.
Website (2026-11-04)
Changed
- Contact. How to reach us is now part of the About page.
- Correct data. A library that reads the site's data is updated, which closes a problem that could return the wrong results.
- Agents install regents-cli 1.10.23, and the Techtree plugin moves to 0.7.24 with it.
- Account reads. A paired agent reads its person's Credits and Points at
GET /api/v1/account/balances,POST /api/v1/account/credits/historyandGET /api/v1/account/points, as on every Regents site, with the toolsaccount_balances,account_credits_historyandaccount_points. The answers are the same as on every other site, and/tools/accountis gone. - Points. On the Points page, an entry that counted now says Counted, and an agent's entry names the agent by its wallet's newest pairing.
- Points. When Points opens, what you already have counts: your account, your linked X, GitHub and Farcaster accounts and your first paired agent each earn their one-time points at the start, once.
Website (2026-11-01)
Changed
- Agents. Your agents and Pair New Agent have their own tab, Agents, at /account/agents. Your Account tab keeps your profile and your Credits, and Pair New Agent in the account menu takes you to the Agents tab.
- Signed requests name the site they are for, so a proof made for another site is refused here.
- A page's signed browser tools take only their input: the agent's own browser
signs each request with the agent's key, and
prepare_agent_requestis gone. In a browser with no agent signer they send nothing and answeragent_auth_required. - Agents install regents-cli 1.10.22, which signs requests the new way, and the Techtree plugin moves to 0.7.23 with it.
Website (2026-10-29)
Changed
- Account menu. Your picture and name at the top right now open a small menu under them instead of a panel at the side of the page: your name and wallet, your three most recently active agents, Pair New Agent, Manage Account and Sign out. A press outside it or Escape closes it.
- Pairing. Pair New Agent, in the menu or in your agents list, takes you to the pairing box on your Account page, laid out as on every Regents site.
- On a phone, while the page list from Menu is open, the page behind it can't be pressed or reached with Tab until the list closes.
- Agents are told to pair with Pair New Agent on their person's Account page.
- Claude Duet, the connected Claude app, has its own picture wherever your agents are listed.
- The home, Wallet Bench, contact and About pages and the guides for agents now point to Techtree Verify and Techtree Cyber Assist, and each offer's email lists what to tell us.
Website (2026-10-28)
Added
- The agent skill walks an agent through testing a Skill with the command line, from checking the machine to comparing the two runs, and names the browser tools an agent can read with. No browser tool runs a test.
- Two paid offers have their own pages: Techtree Verify, which checks the workflows a team ships, and Techtree Cyber Assist, which finds and fixes security weaknesses. Each leads with the Agent Wallet Bench result and has a sample report to read.
/capabilities.jsonnames the regents-cli release agents install, 1.10.19, undercli.GET /api/v1/healthanswers whether the site is serving a catalog release as JSON,{"status": "ok"}, or 503 withunavailable: the same check as/healthz, for agents and monitors that read JSON.
Changed
- Clearer answers for agents. An agent that sends a signed request to publish a Result, keep a Model Climb report or read its person's balances and Points, but isn't paired or whose person hasn't signed in here yet, is told which, and what to do next. An agent unpaired while a request runs gets none of its person's details back. When an agent can't be signed in for a moment, the answer says agent_unavailable; send it again. The Privacy page says what Techtree keeps about an agent's wallet.
- An agent's pairing message tells it to keep its own key, in a folder named after the kind of agent it is, so two agents on one computer pair as two agents.
- Approve and Buy in the Credits window add Regents' Base builder code to the end of each transaction, so your wallet shows a few more characters in its details and the network fee is a hair higher. The amount and where it goes are the same.
- The agent skill's list of things to keep an eye on is headed "Watch list", and it names the command that reads new Results.
- The header menu calls the result checker "Check a Result", so it no longer shares a name with Techtree Verify.
- The list of browser tools moved to
/capabilities.json, and it now names the exact version of the site that is running. The health check at/healthzanswers a plainok; which catalog is served is read from/api/v1/catalogand/api/v1/bootstrap.
Fixed
- With pairing codes out for two agents at once, the agent that pairs now shows under the code it used, and the other agent keeps its code. Before, it could show under whichever agent was on screen, offering the used code again.
- Very large amounts show in full instead of stopping the page.
- The site runs on the latest security release of its web server, which closes a way to make it use too much memory.
- The Wallet Bench test machines run on the latest security release of their list of trusted certificate authorities, which stops trusting one whose keys were exposed.
- On the Results page, the harness you picked (or All harnesses) is shown in bold and underlined, like the picks in the other filter rows.
- The Buy Credits window has the Regents cut-corner frame and a full-size close button in its top right corner. Before, the button was a tiny mark in the top left.
Website (2026-10-27)
Added
- The Agent Wallet Bench tests two more agents, Muse Code and Grok Build, on the same tests, wallets and model budget as the others. Their rows on the scores fill in as their runs finish.
- A link to a Result shared on X or elsewhere shows that Result's own picture: its Climb, its verdict, how many tasks it ran on, and how many got better, stayed the same and got worse. A withdrawn Result's picture says so.
Fixed
- The site runs on the latest security releases of its web server and data libraries, which close several reported weaknesses.
- A Credits Buy pressed before the page caught up no longer opens your wallet when the chain already shows too little USDC or no approval. The button shows why instead.
- An address with an extra slash, such as /about/ or //about, shows the page-not-found message instead of an error.
- An agent asking Techtree who it is gets its answer again: whether it is paired, and what it may do here, instead of an error.
- In the files of a Wallet Bench test run from now on, each test's SHA256SUMS lists the checksum of every file as published, so you can check what you downloaded. Before, its lines showed only hidden values.
Changed
- Each paired agent's row shows its whole name and address, with See Actions and Unpair beneath. The agent's window names its parts Allowed Agent Actions, Spending and Activity, and explains the agent's address.
- Credits, Points and USDC amounts read the same way on every Regent site: from 1 up, two decimals with thousands grouped, such as 1,250.00; below 1, three significant digits, such as 0.400. Amounts are never rounded up.
- The site counts its own visits: where each one came from, such as a search engine, an AI assistant or a social site, and how far it got, such as signing in or pairing an agent. It keeps only the name of the site that sent you, never the address. The Privacy page names the cookie that groups one visit and says the site runs no outside analytics.
- Asking not to be tracked is respected. If your browser sends Global Privacy Control, the site keeps no record of where your visit came from or how far it got. The privacy policy says so.
Website (2026-10-24)
Added
- A wallet test's record at /wallet-bench//results.json, also the first file in its zip, names the agent and wallet tool by the ids the scores use, lists what happened in the test, oldest first, and names any plain file the agent left a password or key in.
Fixed
- When the site restarts, requests in progress get a few seconds to finish and background work stops cleanly, instead of both being cut off at once.
Website (2026-10-23)
Changed
- When you sign in with a wallet of your own, that wallet is your main wallet, with its ENS name. The Regents wallet made for you at sign-in no longer takes its place.
- Pair New Agent in the side panel offers IronClaw where NemoClaw was, with its own setup page at /agents/ironclaw. NemoClaw agents still pair as any agent.
- The Hermes pairing message tells your agent to keep the key it pairs with, made with the SIWA client or its own wallet.
- USDC sent to Credits some other way than Buy Credits now shows on your balance within five minutes of landing on Base, instead of within one. Buy Credits is still credited within a minute.
- An agent's window on your Account page lets you rename it. The new name shows on every Regent site; names with hidden characters are refused.
- Each agent on your Account page shows its own wallet beside its name, with a copy icon, and says when it last checked in.
- Your agents in the side panel lists the three most recently active, then those with no activity yet. Manage beside the heading opens your agents on the Account page.
Fixed
- On a shared computer, connecting a wallet after signing out can no longer sign the previous person back in. If the earlier sign-out didn't fully finish, it is finished first; if that fails, the wallet window doesn't open and you stay signed out.
- Opening or downloading a Wallet Bench test's recorded files while their storage is slow now shows a page asking you to try again in a minute, instead of an error.
- The Pair New Agent box no longer breaks the side panel when the page reconnects before you pick an agent. Each agent now keeps its own pairing code, so choosing another agent and coming back shows the same code until it expires.
Website (2026-10-22)
Fixed
- Every connection the site makes to another service now trusts Mozilla's current list of certificate authorities. Certificates from the eMudhra authorities whose keys were compromised on 20 October are refused.
- Updated a library the home page crown is drawn with to include its three security fixes.
- Updated the library that signs the site's sign-in cookies so the site's secret key can never appear in an error report.
Website (2026-10-21)
Changed
- The home page now makes one promise: improve a Skill and prove it worked. It shows the newest Result under the headline, with its verdict, and one Start button. Wallet Bench sits further down the page.
- Every Result now says Improved, Regressed or Too close to call, how many of its tasks got better, stayed the same and got worse, and how many runs back it. A higher score where most tasks got worse is never called Improved. The API gives the same verdict and run count.
- The Results page lists every Result by default and says how many it is showing out of the total. Each filter starts at All, and choosing a harness, version, model, challenge or route narrows the list from there.
- A Result made on an earlier version of its Climb says so, as its Climb's page does.
- Every public page also answers in Markdown for agents that ask for it, and says so in its page head.
- The short start at /llms.txt is now a short page: what Techtree is, the exact install line, how to check it worked, and what you pay. The agent guide at /agents.md holds the rest and no longer repeats the Skill file.
- The install line and version in the agent documents always match the release Techtree is serving.
- The home page lists all three Climbs.
- The motion lab and the empty blog are no longer part of the public site.
- The Privacy page names every service your browser connects to when you sign in.
- On a wide screen, the Results filter names sit beside their choices instead of under them.
- The home page says exactly what Prime Intellect's Verifiers does: it scores each Climb's tasks, and every Result is a signed record anyone can check.
- Docs is now one page per guide, with a list of every guide, a side menu and Previous and Next links. Each guide also answers in Markdown.
- Wallet Bench shows only the wallet tools that have results, and lists the others below the table. Every score under 10 says what cost it points, and each test's page opens with its verdict. A test names the agent with its version and the wallet tool's version.
- On a phone, the Results and Wallet Bench tables become cards, and the header fits on one row with a Menu button.
- Plainer words across Results: Agent instead of harness, Change instead of a symbol, and "Signed by the publisher · Files checked · Not rerun by anyone else" for what a Result's record shows.
- On Results, rows slide into place when you change a filter.
- The home page crown appears as a still picture first and comes alive once your browser has drawn it.
- Agents can find the Techtree skill at /.well-known/agent-skills/index.json, and each agent the side panel offers has its own skill at /agents//skill.md.
- A browser's agent can now download a Result's bundle, read the Skill it measured, and read the Wallet Bench scores and each test behind them.
Website (2026-10-20)
Changed
- Wallet Bench's recorded files hide any picture shown inside a tested agent's own output, as they already hide keys and recovery phrases.
- Wallet Bench's recorded files hide PGP private keys too.
- Wallet Bench's recorded files also hide a recovery phrase typed in capitals, numbered, or with its words joined by hyphens, slashes, dots or other marks.
- Wallet Bench refuses a tested agent's files when they would unpack to more than the bench keeps.
- Wallet Bench's scores and downloads count toward the same per-minute request allowance as the API.
- Publishing a Result is no longer, on rare occasions, refused for naming a place on your own machine when it named none.
- Browsers asking for /favicon.ico get Techtree's icon.
- Wallet Bench's recorded files also hide BMP, TIFF, AVIF, HEIC and SVG pictures and PDFs shown inside a tested agent's output, in either base64 alphabet and with indented lines.
- A hidden email address or key at the start of a line no longer breaks the recorded file it sits in, and long outputs are checked for secrets quickly.
- A shared Techtree link shows a new picture in the Canopy colours, with the Regents cut corners and their Tangerine and Powder Blue marks. Its words and techtree.sh now sit in the left part of the picture, which is all X shows on a phone, and clear of the bottom-left corner, where X writes the site's name. Before, the headline ran past that part and techtree.sh sat at the top right, out of view on a phone.
Website (2026-10-19)
Changed
- On Wallet Bench, a pass that leaves a plain password or key file on the test machine now scores at most 9.0 instead of 10.0, so only a clean pass reaches 10. The /wallet-bench.json feed carries the same scores.
- When many people open Wallet Bench's recorded files at the same moment, a few are asked to try again in a minute, so the site stays up for everyone.
- Wallet Bench no longer records pictures a tested agent saves on its test machine, and a recorded file opens as text.
- A page you opened just before the site updates keeps working while the update rolls out, instead of losing its scripts and styles.
- In the seconds after the site's database restarts, a page shows its "can't be shown right now" note for one visit instead of an error page.
- On touch screens, links that stand on their own, the side panel's Sign out and account buttons, and in Buy Credits the Purchase History link, the amount field, the links at its foot and the account page's "Manage Credits on regents.sh" link are easier to tap.
- The GitHub star count in the header arrives with the page, so it no longer goes missing when GitHub limits a shared network's visits.
- On Wallet Bench, each agent and wallet pair in the list is easier to tap on a phone.
- Techtree wears its new green Canopy colours in light and dark: a pale green page in light and a deep forest green in dark, with lime buttons and green corner marks. The home page's Skill, Harness and Env cards, the boxes to paste into your agent's chat, the section bars and the theme switch all use the new colours, and no orange or pale blue is left on any page.
- The crown on the home page sits in a deep green field with pale green beams in light and lime beams in dark, and the still picture shown where the crown can't be drawn shows the same.
- The picture shown when a Techtree link is shared uses the new colours, and the browser tab icon and app icons show the Techtree mark in Canopy green.
- When a Techtree link is shared on X, its picture now carries a written description for people who can't see it.
- In the side panel's Pair New Agent box, each agent picture sits in one frame on white in both themes, and before you choose, the box reads "Choose type of agent below to get pairing message."
- Buttons are slimmer, about 40 pixels tall, with their words on one line. On touch screens they stay at least 44 pixels each way.
- The lime Sign in and Copy page buttons have darker green corner marks and edges, so they stand out clearly on the pale page and on the lime itself.
- Wallet Bench is a score grid: a row for each agent and a column for each wallet tool, with Wallet and Install tests to switch between and a list of every test beside it. Each box scores the pair's newest result out of 10 from the judge's nine checks; a red box is a safety failure.
- A score opens its result in a window with every check and its reason, and "Download full results" gives one file with the results and every recorded log.
- The same scores are at /wallet-bench.json for other sites and agents, each linked to its result.
Website (2026-10-18)
Changed
- When your device asks for less motion, the home page headline keeps its first ending instead of changing every few seconds.
- On phones narrower than 375 pixels, the header shows the Techtree mark without the word, so the light and dark switch no longer covers it.
- The Start page's note for when there's nothing to install yet, and the Wallet Bench page before any test has run, now say what will appear and where to go next.
- The X and GitHub icons at the foot of every page, and the Techtree link on the page-not-found screen, are easier to tap on a phone.
- If the site's database stops answering, Home, Start, Docs, Results and Wallet Bench stay up for visitors who aren't signed in and say what can't be shown right now, instead of the whole page failing or wrongly saying there's nothing there.
- During a database outage those pages now answer in about half the time.
- During a database outage, signed-in visitors also see those pages, shown signed out. They're still signed in once the database is back.
- Pages other than your Account pages load less, because the wallet code now loads only on the Account pages.
Website (2026-10-17)
Changed
- The home page headline reads "Improve any agent's Harness / Environment / Skills to", and its last words turn through "use less tokens", "perform evals stronger" and "create better traces".
- While the Regent sign-in service can't say what your agents have been doing, the side panel asks again at most once a minute per agent, and activity from the Regents mobile app is shown with the other sites.
- The home page headline reads "Least-token uplift of Harness / Environment / Skill".
Website (2026-10-16)
Changed
- Every page has Sign in at the top right. Signed in, your name opens a side panel with Sign out, Manage Account, your paired agents with where each was last active, and Pair New Agent with ten agents to choose from. On a wide screen it stays open as you move between pages; on a phone it opens over the page.
- The Account page's Pair New Agent opens the side panel, and an agent that isn't paired yet is told to ask its person to open Pair New Agent there.
- The home page says what Techtree is for: "The fastest way for a business to uplift their agent Harness / Environment / Skill", built on Prime Intellect & NVIDIA Nemotron. Each wallet test starts from the same identical cloud server on Fly.io Sprites.
- Every Techtree command described here is now checked against the regents-cli 1.10.12 that people install, so the descriptions match what the commands do.
- An agent whose person hasn't signed in on Techtree yet is told exactly that, and to ask its person to sign in here once, instead of being told the account doesn't exist. Every Regent account already exists on every Regent site.
- The agent guide says what an agent's check-in carries: its person's Techtree account, with display name, ENS name and wallets.
- Main buttons take the shared Regent look: two cut corners, the hover arrow and the press motion. Each button shows one arrow, and Copy page has none.
v0.7.3 (2026-10-13)
Changed
- Techtree now names regents-cli 1.10.2, with the same Climbs, catalog,
engines and starter Skills as v0.7.2. Every answer from regents-cli 1.10
ends with what to do next: the exact command to run, or what your person
has to do first. A payment that may have gone through is never offered as
a plain rerun. Install it with
uv tool install --python 3.12 regents-cli==1.10.2, or upgrade withuv tool upgrade regents-cli. The Hermes plugin is 0.7.3.
v0.7.2 (2026-10-08)
Changed
- Techtree now names regents-cli 1.7.3, with the same Climbs, catalog,
engines and starter Skills as v0.7.1. regents-cli 1.7.3 works with an
agent key set up by the SIWA agent client's
use-wallet, and on a Mac it keeps the agent key locked with a passkey. Install it withuv tool install --python 3.12 regents-cli==1.7.3, or upgrade withuv tool upgrade regents-cli. The Hermes plugin is 0.7.2.
v0.7.1 (2026-10-07)
Changed
- Techtree now names regents-cli 1.7.2, with the same Climbs, catalog,
engines and starter Skills as v0.7.0. regents-cli 1.7 adds
regents auth accept-world-id, which accepts the person who vouched for your agent with World ID; every site then shows that person behind the agent. Install it withuv tool install --python 3.12 regents-cli==1.7.2, or upgrade withuv tool upgrade regents-cli. The Hermes plugin is 0.7.1.
v0.7.0 (2026-10-06)
Added
- A Climb can run on your ChatGPT plan as well as on your own Prime key. Sign
in with
regents techtree model login; each Climb page lists the routes it runs on, and a run uses one route for both sides. A run on your plan shows no dollar cost. Techtree sells no model calls and charges nothing. - A Climb now scores each task on its environment's own weighted total of
rewards. Today's three Climbs each still have one reward, so their scores
mean what they meant before. The Tasksmith Climb page names the Prime
Environments Hub taskset it runs,
techtree/hf-tasksmith-v1. - Published Results carry their Skill, so anyone can fetch it with
regents techtree skill fetchor rerun a Result withregents techtree climb prepare --rerun-of. Run pages show the Skill and say when a Result reruns another.
Changed
- Techtree now names regents-cli 1.6.0. The Climbs move to
hello-world-climb@3,frontier-cs-open-ended-climb@3andtasksmith-climb@2, with the same tasks, limits and starter Skills. Install it withuv tool install --python 3.12 regents-cli==1.6.0. The Hermes plugin is 0.7.0.
v0.6.1 (2026-10-04)
Changed
- Techtree now names regents-cli 1.5.0, which lets an agent add itself to
the public agent registry on Base with
regents auth register. Listing is optional. Climbs, Campaigns and starter Skills are unchanged. Install it withuv tool install --python 3.12 regents-cli==1.5.0. The Hermes plugin is 0.6.1.
v0.6.0 (2026-10-04)
Added
- A new Climb, HF Tasksmith, asks whether a Skill helps an agent make 6 real changes to Hugging Face code. Each task brings its own working and grading images, and grading happens in a fresh box that sees only the files the task names. Six more tasks are kept apart: they are run once, on the winning Skill, reported beside its result, and never decide the winner. Run pages and the Climb page say when a result comes from those held-out tasks.
Changed
- Hello World and Frontier-CS move to
@2. Only the file format changed: same tasks, model, limits and starter Skills. Results published for@1keep their pages. - Techtree now names regents-cli 1.4.0. Install it with
uv tool install --python 3.12 regents-cli==1.4.0. The Hermes plugin is 0.6.0.
v0.5.4 (2026-10-02)
Changed
- Techtree now names regents-cli 1.3.4, which adds Patchbay's commands for
known fixes and free fixes. Climbs, Campaigns and starter Skills are
unchanged. Install it with
uv tool install --python 3.12 regents-cli==1.3.4. The Hermes plugin is 0.5.4.
v0.5.3 (2026-10-01)
Changed
- A result now shows what the model provider reported the run cost, added
up over every model call on both sides, in place of a figure worked out
from a price list. If the provider did not report a cost for every call,
no total is shown and the result says why. Install regents-cli 1.3.3 with
uv tool install --python 3.12 regents-cli==1.3.3. The Hermes plugin is 0.5.3.
v0.5.2 (2026-10-01)
Changed
- A run now stops once its spend reaches the most its Campaign allows. While
it runs, Techtree adds up the cost the model provider reports for each
finished task, and when the total reaches the Campaign's maximum it stops
both sides. A stopped run has no score. Tasks still under way when it stops
can add a little to the total. If the provider does not report a cost, the
run stops too. Runs are no longer turned away before they start. Install
regents-cli 1.3.2 with
uv tool install --python 3.12 regents-cli==1.3.2. The Hermes plugin is 0.5.2.
v0.5.1 (2026-10-01)
Changed
- A new Frontier-CS starter Skill,
frontier-cs-starter-v4. It asks the agent to plan a few methods before coding, build its own test inputs and scorer, and keep improving its answer until most of the time is used. In two test runs with GPT-6 Luna it scored 0.542 on average against 0.508 without a Skill, higher on 7 of the 9 problems that could be scored. It replacesfrontier-cs-starter-v3. Install regents-cli 1.3.1 withuv tool install --python 3.12 regents-cli==1.3.1. The Hermes plugin is 0.5.1.
v0.5.0 (2026-09-30)
Changed
- Both Climbs now run GPT-6 Luna (
openai/gpt-6-luna) at high reasoning, on Hermes v2026.9.24. Install regents-cli 1.3.0 withuv tool install --python 3.12 regents-cli==1.3.0. The Hermes plugin is 0.5.0. - The Frontier-CS starter Skill is now
frontier-cs-starter-v3: it sets out GPT-6 Luna's reply and attempt limits and when to think hard. It replacesfrontier-cs-starter-v2.
v0.4.5 (2026-09-30)
Changed
- On Results, each filter row has a link that opens in a new tab: the harness's project page, the model's page on OpenRouter, and the Climb's own page.
v0.4.4 (2026-09-30)
Changed
- The Verify page now explains, in plain words, how a Result is made: Nous Research's Hermes runs the tasks, Prime Intellect's Verifiers scores every task, and the record is signed. It then says what checking a Result tells you, what it can't, and why it matters.
- The theme button in the header has slightly rounded corners.
- The agent guide at /llms.txt links Autolaunch's source at regents-ai/autolaunch.
v0.4.3 (2026-09-30)
Changed
- The Hermes plugin now installs from the
pluginfolder of regents-ai/techtree:hermes plugins install regents-ai/techtree/plugin --ref <commit> --enable, with the exact commit on the Start page. The plugin itself is unchanged (0.4.1). The separatetechtree-hermesrepository is archived.
v0.4.2 (2026-09-30)
Changed
- The About page now says what Techtree does, what makes it different, who uses it, who builds it, and its key facts. The agent guide at /llms.txt carries the same key facts.
- The theme button in the header is smaller, sits apart from the GitHub button, and shows the other theme's colours when you point at it.
- Tables on the About, Contact, Privacy and Terms pages wrap long values to fit a phone.
v0.4.1 (2026-09-29)
Added
- An agent can pair with its person's Regent account here, signing with its
own key:
POST /api/agents/v1/pair, then check in withGET /api/agents/v1/me. One pairing works on every Regent site; the agent sign-in guide shows how.
Changed
- Install regents-cli 1.2.1 with
uv tool install --python 3.12 regents-cli==1.2.1. A Result's summary now names the Skill that changed, for example "No tested Skill → frontier-cs-v2". The Hermes plugin is 0.4.1. - The Frontier-CS Open-Ended Climb page gives both runs of its starter Skill and links the second run's Result.
v0.4.0 (2026-09-29)
Added
- A second Climb, the Frontier-CS Open-Ended Climb
(
frontier-cs-open-ended-climb@1, experimental): ten open-ended programming problems from Frontier-CS, created with FrontierSmith, where no perfect answer is known. For each problem the agent writes a C++17 program, and the problem's own checker scores it from 0 to 1. - Each Climb now has its own starter Skill:
regents techtree skill starter --climb <climb>. - The home page lists every published Climb.
Changed
- Install regents-cli 1.2.0 with
uv tool install --python 3.12 regents-cli==1.2.0. The Hermes plugin (0.4.0) prepares the starter Skill of the Climb it is running.
Removed
- The source of the old
techtreecommand is gone from this repository. It is developed in regents-cli.
v0.3.1 (2026-09-29)
Changed
- Techtree's command-line tool is now part of
regents-cli. Install it with
uv tool install --python 3.12 regents-cli==1.0.0; every command that wastechtree …is nowregents techtree …. - The Hermes plugin (0.3.1) runs
regents techtreefor every step. It stops after the first comparison: it no longer proposes a revised Skill or runs a second comparison. Its 13 tools keep their names and answers. - The Hello World Climb now tests Hermes v2026.7.20 (it was 0.19.0), and was checked again on it: with the starter Skill the agent solved 23, 22 and 24 of 36 tasks across three runs, and none without it.
Removed
- Building tasks from a repository on your own computer, and running an exported Skill environment again from its folder.
v0.3.0 (2026-09-27)
This release folds in everything that was listed for 0.2.2, which was never released on its own.
Added
- Create an environment from a Skill: a set of checked tasks written from one Skill of yours, that you review, accept and can hand to someone else. Each step shows what it will do and waits for your yes before any model is called, and every step keeps its record whether it succeeded or not. No comparison is needed to finish; comparing Skills on the tasks is a separate, optional step.
techtree forge inspect-skill PATHlooks at a Skill without running any of it. It lists every file, says which ones it can carry and which it cannot and why, and follows whatSKILL.mdnames. When a file the Skill needs cannot be carried, or a link points to a missing file or outside the Skill's folder, it refuses by name before any model sees the Skill. A reduced copy you make is recorded as a new Skill that names the one it came from. Tools a Skill asks for are recorded as its request, never granted.techtree forge planprepares the planning of tasks and calls nothing;techtree forge plan-startshows exactly what leaves your computer (the Skill's files, word for word, inside Techtree's planning instructions), the provider and model, and the limits, then makes one call on your yes. The planner has no tools. It first states what the Skill claims to improve and what would show it, then proposes tasks that each test one claim as a positive case, a boundary case or a counterexample. A plan is tried once; trying again is a new plan with its own approval.techtree forge correct-proposalrecords your corrections to the claims and tasks as a new proposal; the original is kept.techtree forge constructprepares the building of the tasks and calls nothing;techtree forge construct-startmakes one call per task on your yes. Every package is then checked on your computer with Docker and the network off: its tests must fail when nothing is done, succeed for its solution and for another correct solution, and fail for a deliberately wrong one. A failed, refused or timed-out call is kept and the other tasks go on; nothing is retried on your account.techtree forge correct-taskchecks a copy of a built task that you edited by hand, the same way a written task is checked, and records it as your correction beside the original, with every file you changed.techtree forge collectandtechtree forge acceptshow every proposed task with how it went, then freeze the tasks you accept as one collection. Accepting splits the tasks into a part an improving agent may study and a held-out part it never sees; which task goes where follows from the tasks themselves, not from anyone's choice. Any later change is a new version, accepted again.techtree forge verifychecks that a collection is exactly what was accepted.techtree forge run --collection IDruns an agent on an accepted collection, with or without a Skill, and gives you each task's outcome with no comparison needed. The agent gets a shell, files, code and Skills, and no network. What the agent left in its working folder is recorded before any test runs; a link that leads out of that folder, or output past the kept limit, is a failure of its own and is not graded.techtree forge compareworks on a collection. It names the Skill the tasks were written from and the Skill on each side, even when two of them are the same file, says "Evaluation on Skill-derived tasks", and gives a verdict for the whole collection and for the held-out tasks alone. A Skill's earlier version can stand on the baseline side, so you can compare old against new.techtree upliftworks on a collection: the agent that revises the Skill sees only the studied tasks, and the revision's verdict is worked out on the held-out tasks alone.techtree forge export COLLECTION --to FOLDERwrites a private copy of an accepted collection for someone else, andtechtree forge verify-exportchecks it from its folder alone. The copy leaves out the Skill's text, the logs from writing the tasks, and everything else on your computer. It includes the tests and reference solutions, so give it only to people who check or run the tasks. The check prints the collection's fingerprint: the copy is the same collection only if that matches the one the sender gave you.techtree forge import FOLDERbrings an exported collection into another Techtree and checks every task again there, soforge runandforge comparework on it unchanged. The export's own README lists what you need and the exact commands.- The Results page answers "Keep this Skill change?". It leads with improved, regressed or not enough evidence, and on which tasks, then shows one task of each kind and the Skill change itself. Three badges say what is known: the files are verified, who reported the numbers, and that it has not yet been reproduced by anyone else. A panel shows how to run it again: what you need, the exact commands, and the limits per try and per run in model calls and tokens. Results published before this release are assessed the same way; the two already published were checked and both hold.
- Every published Result's page offers its bundle for download, and
techtree proof verifychecks the downloaded file directly, including that it is the bundle the Result names. - The tdd example shows one real
comparison made with this release: Matt Pocock's
tddSkill (MIT licence) against no Skill, on six tasks written from it. With the Skill the agent scored full marks on 6 of 6 tasks, without it on 3 of 6; on the three held-out tasks alone the Skill won 1 and tied 2. The page shows the claims, the tasks, the Skill, what limited the runs and what they cannot show, and links to the exported tasks so anyone can run them again. The home page links to it. - Every page offers five read-only tools to an agent working in a browser that supports WebMCP: the Start guide, the list of Climbs, one Climb, the list of Results and one Result. Each reads a page a person can already see. This is already live on techtree.sh.
- The home page, About, Contact and Privacy pages and
/skill.mdanswer in Markdown when an agent asks for it. The site now has/openapi.json, a sitemap, and About, Contact and Privacy pages linked from the footer. This is already live on techtree.sh.
Changed
- The home page, Start, Docs and the agent guide now lead with testing your own Skill: make tasks from it, run them without it (or with its earlier version) and with it, and compare. The home page's main action is one line to paste into your agent's chat; the install command is one click further. Start offers three paths, each with what it needs, how long it takes and where your data goes: test your Skill, take a quick look with the Hello World Climb, or build tasks from your repository. The Climb is introduced only after a controlled comparison is explained.
- Every capability on the site is marked Available now, Experimental or Planned from one list, and the only version labels come from the active release. Pages that described the v0.1 and v0.2 plans, or said the site is built on NVIDIA NeMo, no longer do.
- Hermes users get the same instruction as coding agents; the plugin's own commands are described in the Hermes section of Docs.
- The Repo2RLEnv service leaves the top menu and gains a "Tell me when it opens" email link.
techtree doctornow says whether thetechtreeHermes profile is signed in, and to which providers, checked the same waytechtree forge runchecks before it starts. A missing or signed-out profile is reported with the exact command that resolves it.- The host Hermes this release is tested on is 0.21.3, and 0.21.3 is now the minimum: experiments run in a Hermes profile of their own, and that is the Hermes it was verified on.
- The published package is built the same way every time, from a clean copy of the release commit, so what is on PyPI is exactly what was approved.
techtree forge buildandtechtree forge statusnow say in plain words what happened to every task and every candidate commit: each generated task reads "qualified" or "rejected" with the reason a person can act on, and one summary line says how many tasks qualified and why the other commits were skipped over. A build that leaves fewer than three usable tasks carries aforge_few_taskswarning, since a comparison on so few can say how one attempt went, not whether a Skill helps; the build stays usable.- Qualified for real on three public repositories with Docker and no model:
pallets/click(Python, 1 task from 10 recent commits),pallets/jinja(Python, 4 tasks from 150 recent commits; the 30 most recent yielded none, being merges and short messages) andspf13/cobra(Go, 1 task from 30 recent commits, with a Go image supplied through--dockerfile). Go test output that is not valid text stops the task generator (seen onpelletier/go-toml); that limit sits in the pinned generator and is recorded, not worked around. The click task qualified again, with the same checks, under this release's records. - A comparison now gives a verdict a person can act on: inconclusive, mixed, improved, regressed or no difference, by fixed rules applied in that order, with fewer than three graded pairs always inconclusive. The report names every task the Skill lost, even when the average went up, and says for each task whether it went both ways across attempts; with one attempt per task it says plainly that consistency was not measured. The terminal output and the machine-readable record carry the same three facts.
- Every task image, for repository tasks and Skill tasks alike, is now built with the network off. Skill tasks may only start from base images the release lists, pinned to an exact version.
- Every container Techtree starts for a task now has a cap on its number of processes, keeps only one of the special permissions a container normally has, and cannot gain new ones.
techtree publishlists every file that will leave your computer before it asks, and sends exactly those bytes. Techtree does not search what it sends for passwords or keys; it shows you every file instead.- Forge builds, runs, comparisons and revisions made with 0.2.1 or earlier stay on disk untouched but are no longer read: Techtree names the file and says its format is not supported. Build the tasks again with this release.
- Security reports go by email or GitHub private reporting, and
SECURITY.mdnames both. The Privacy page says that deleting your sign-in account does not remove Techtree's copy of your profile, and how to ask for that by email. - Buttons, the Copy page menu and the task list on a Result page move a little when pressed or changed; a keyboard press or a reader who asked for less motion gets the change at once.
- The README describes the parts of this repository and which CLI, plugin, catalog and site versions go together.
Fixed
- A Result's page used to show the verdict its report stated. The site now works the averages, the change and the decision out again from the task scores, by the same exact rule the CLI uses, and refuses a published report whose numbers disagree with its own task scores.
- Why a task run left no score is now set by Techtree, not by a file the task's tests could write. This also corrects the "verifier timed out" label on repository runs.
- A stray byte that is not valid text in a command's output no longer stops a task check.
- A declined or unanswered prompt now says how to go ahead, and the next steps
name
--homewhen you use one. - Updated the website's Ash dependency to include a security fix.
What this release does not yet show
- The guided repository walk-through planned for 0.2.2 is not in this
release. One command that would build tasks from your repository, run
both sides and compare, with a single approval, was deferred when Skill
environments became the priority. It was not shipped, and its acceptance
runs (a real model run on
pallets/click, then a second walk-through on a fresh Linux computer) never took place. The Hermes plugin tools that would have read its results back are not included either. On a repository, usetechtree forge build,forge runfor each side, andforge compare, one step at a time. - Qualification yield and rejected tasks, from the real runs.
- The first Skill environment, from a corrected copy of Techtree's own Hello World starter Skill: the planner proposed 1 task of the 3 allowed. Its first build qualified but was not collected, because its instruction restated every step of the Skill, so an agent without the Skill had all it needed. The building instructions now forbid that; no check catches it automatically. The rebuilt task qualified and was accepted as a one-task collection.
- The
tddexample, first attempt: 7 claims and 7 tasks proposed. Of 7 built tasks, 2 qualified. Three were rejected because their tests kept copies of the starting files, which cannot be told apart from leaked answers; one because its other correct solution did not score 1; one because its reference solution did not make its tests succeed. Building the 5 again gave 4 more; one was rejected again for its other correct solution. A review then found all 6 collected tasks unfair, because some gave the tested choice away and some graders refused correct answers. None was accepted, and the planning and building instructions were changed. - The
tddexample, second attempt: 5 claims and 5 tasks proposed. A person removed 2 tasks and added 4, making 7. All 7 built tasks qualified. A person's review then corrected 5 of them by hand (6 corrections in all) and left out 1,percentage-only, which had qualified; the reason it was left out was not written down. The accepted collection has 6 tasks, 3 of them held out, testing 4 of the 5 claims.
- Authoring attempts and interruptions. Every real planner and creator
call was announced and approved before it ran, on
gpt-5.6-solfromopenai-codexthrough thetechtreeHermes profile. In all: 3 calls for the first environment (one planner, two creator), 13 for thetddfirst attempt (one planner, seven creator, five creator again) and 8 for the second (one planner, seven creator). No real call was interrupted, failed or ended with an unknown outcome, so recovery from an interrupted call has been shown only with stand-ins and with a stopped walk-through that called no model. The secondtddattempt took from 20:44 UTC on 26 September (planning) to 01:17 UTC on 27 September (acceptance), including the reviews and hand corrections above; that is the project's own time, not an outside creator's. - No one outside the project has yet created or evaluated an environment with 0.3.0. The release was meant to wait for one outside creator to make an environment from their own Skill and a different outside evaluator to verify and run the export on a fresh computer. Neither has happened, so there is no outside outcome, completion time or count of the help they needed to report. Every environment so far was made by the project from a Skill it chose.
- Skill material Techtree cannot carry. Only
.md,.txt,.json,.yamland.ymlfiles that are UTF-8 text are carried, up to 256 KiB each, 64 files and 2 MiB in all. Scripts, images, other binary files and files with no extension (a plainLICENSE, for example) are not; neither are hidden files, links or special files. A Skill that needs such a file, links to a missing file or outside its own folder, or whoseSKILL.mdheader uses more than plainkey: valuelines and one level ofmetadata, is refused before any model call. Of the eight Skills the project keeps for its own work, only one could be used as it was; the other seven link outside their own folder. Tasks can only start from Python 3.12 (python:3.12-slim) and never reach the network, so a Skill whose tasks need another language, service or live download cannot become an environment yet. - Provider usage that has no price. Every real call was on a provider
subscription that Hermes reports as "included", so Techtree records no
price for it, and it never counts such a call as free. Nothing is quoted in
advance, and the site gives limits in model calls and tokens only. What the
calls used: the first environment, 14,936 tokens; the
tddfirst attempt, 165,583 tokens; its second attempt, 108,375 tokens to plan and build; thetddcomparison, 41 model calls and 338,454 tokens without the Skill and 95 model calls and 1,026,256 tokens with it. - Checks not yet made.
- No outside creator or evaluator, as above, and no run of 0.3.0 on a fresh computer.
- Skill tasks have been built and checked only on Docker running
linux/arm64 (a Mac).
techtree forge importrefuses tasks built for another platform, so thetddexample's tasks can only be imported where Docker runs linux/arm64. - The
tddexample's export was checked from its folder but was not imported into a second Techtree. Importing an export and running and comparing it there has been shown only with stand-ins in the automated checks. - Revising a Skill on a collection, with its held-out verdict, has not been run with a real model.
- The
tddexample ran each task once on each side, so consistency across attempts was not measured, and the same model wrote the tasks and was the agent tested on them. - The real Skill environment runs used Hermes 0.21.4, which is newer than the 0.21.3 this release records as tested.
- The automatic checks show that each task's tests agree with its own
sample solutions, not that they accept every correct answer; read each
task, and correct any with
techtree forge correct-task, before accepting. - Each task carries a written rubric, but nothing scores it; only the task's tests decide.
- An acceptance is a record on your own computer, not a signature; someone who rewrites a collection and its acceptance together is not caught.
- The Hermes plugin has no tools for creating an environment; in Hermes, the agent follows the steps on techtree.sh/skill.md in the terminal.
- Only Climb runs can be published. Results from a Skill environment stay on your computer.
- The browser tools need a browser with WebMCP switched on; ordinary Chrome does not offer it without changing a setting.
- Known limits.
- A try in
techtree forge runis bounded by time only: the agent gets its task's own limit (30 minutes for every task so far) and Techtree stops it 2 minutes later. Nothing limits its model calls or tokens. - A task's part is inherited forward only: a task held out in an accepted collection stays held out, even if a later collection for another Skill studies a task with the same files. Tasks whose files differ only slightly are not recognised as the same task.
- While a task runs, Techtree does not yet cap how much it prints or writes, so a task written to do harm could fill memory or disk before its time runs out. Run tasks from people you trust, and read them first.
- Checking an export shows that it agrees with its own records, not where it came from. Importing one downloads its base images from the network; after that every task is built and run with the network off.
- An export cannot include the Skill's text yet, even when you have the right to share it.
- A try in
Not in this release
- An MCP connector for other agents, hosted building and running of environments, private hosting, managed creation, payments and bounties, NVIDIA NeMo Fabric and Relay support, automatic Skill improvement, and training remain planned.
- Publishing a Skill environment, or a comparison made on one, to techtree.sh.
v0.2.1 (2026-09-19)
Added
techtree forge runruns one arm of an experiment on the tasks a forge build qualified, with your own Hermes and your own provider sign-in: the baseline arm without the Skill, the candidate arm with it. Before anything runs it shows what will run and asks. Every attempt is recorded with its patch, its test verdict, the usage Hermes reported, and — when there is no verdict — why: the agent ran out of time or did not finish, the tests ran out of time, or left nothing readable. Nothing is scored as zero for want of evidence, and no attempt is retried on your account.techtree forge statusreads a run back.- Experiments run in a Hermes profile of their own named
techtree. Create it withhermes profile create techtree --no-aliasand sign it in once withhermes -p techtree auth add PROVIDER. Techtree copies no sign-in and reads none; around every attempt it empties that profile of everything else, so each attempt starts fresh. If the profile is missing or signed out,forge runsays so and names the command before anything starts, andtechtree doctorreports whether the profile exists. Two experiments never share it at once. - An experiment records the whole version line your Hermes reports, including its build date and source commit, so a Hermes that was updated between two runs is never compared as if it were the same.
techtree forge comparepairs a baseline run with a candidate run, task by task, and writes a self-contained HTML report you can open from disk beside the machine-readable record. It says whether the Skill won, lost or tied on each task, what each arm used in time, model calls, tokens and reported cost, what the two arms were allowed to differ in, and how far the evidence carries. A pair without a verdict on both sides is shown as unresolved, never counted as zero, and the summary says "Partial" until every planned pair has one. Comparing calls no model.- A forge comparison can be revised once through
techtree uplift, the same way a Climb run can.uplift contexton a comparison writes what a reviser may read: the task instructions, both arms' results, and what the Skill is meant to improve — never the reference fix, the tests, or either arm's patch.uplift skill-sourcereads the Skill a candidate run measured back from the run's own verified copy.uplift preparetakes one revised Skill, keeps everything else about the experiment the same, refuses an unchanged Skill or a changed Hermes, and screens the revision against every task's reference fix and tests, recording each shared line rather than refusing.uplift startshows what will run, asks, measures the revision against the same baseline, and records whether it improved, regressed or matched. The revision is kept either way, and is never measured twice.techtree forge statusreads a revision back. - A candidate run keeps its own copy of the Skill it measured, and every attempt runs from that copy.
- A forge experiment is declared before it runs, and two arms are compared only when nothing but the Skill differs between them.
- A candidate Skill that names the cases a Climb scores it on is refused when it is prepared, whether for a first submission or as a revision of a measured Skill. A Skill describes the rule; it may not carry the scored inputs.
v0.2.0 (2026-09-18)
Added
- Read release notes at Changelog, available from the site header.
- Build repair tasks from a local Git repository with
techtree forge build. Each accepted task checks that the unrepaired code fails the relevant tests and that the reference repair passes them. Building tasks does not call a model. - Inspect retained build results with
techtree forge status, including failed or cancelled builds, without requiring Docker or the task generator to be available. Incomplete results are shown as incomplete, not as usable tasks. - Keep the generated task files, their content fingerprints, validation logs, and control/reference results together so a build can be inspected later.
- Browse published Results by harness, model, and exact challenge. An unknown selection does not silently show unrelated results. This is already live on techtree.sh.
Changed
- Results and objects published before this release stay available at their existing addresses. New runs must be recorded with the current CLI against the current catalog; uploads made with CLI 0.1.1 against the earlier campaign are no longer accepted. Upgrade by following the Start guide.
- CLI integrations receive structured facts, unknowns, blockers, and suggested next actions. The response format replaces the previous format; integrations must update with the CLI rather than assume old responses still apply.
- New comparisons record evaluation rules separately from execution settings. Existing signed proof bundles remain readable without rewriting their files.
- CLI, Hermes plugin, and website development now share one repository. Existing pinned installation instructions in the active Start guide remain authoritative.
- The site header shows the GitHub star button and its star count as one button.
Fixed
- Updated the website's Ash dependency to include its field-policy security fix.
- The supplied Python build image keeps the project's installed test tools available when a login shell starts.
- Task generation retains raw validation output and distinguishes missing test results from a repair that produces no failing-to-passing tests.
- Interrupted builds retain progress and failure details. Docker commands have time limits, and interrupted operations attempt to remove their owned containers.
Not in this release
- Running a coding agent on the generated repair tasks with
forge run, comparing baseline and candidate Skills withforge compare, and reporting their measured time and usage are not included. - Evaluating the default Hermes profile or a named profile with isolated state, Hermes-owned authentication, and a verified grading handoff is not included.
- A Prime reference-agent handoff, an independently authorized external-use case study, Fabric-backed Hermes/Codex comparisons, and optional Relay evidence are not included as qualified end-to-end workflows.
- Qualification has been reproduced for one selected public repair, locally and on a fresh Linux worker. This is task-pipeline evidence, not an agent repair, a demonstration of Skill benefit, or a held-out evaluation.
- The full generation and per-task deadlines were not run to expiry. Among the generator's rejection reasons, only the missing or unreadable test-output case was deliberately exercised in a real build; this is not exhaustive failure-path coverage.
- Hosted comparisons, public collaboration and payments, automatic Skill optimization, private proving, and training remain outside this milestone.