Repository code change proposal for automated software development by generative response engines

The automated software development system using a generative response engine within controlled environments addresses security risks by generating diff-formatted proposals and requiring human oversight, ensuring secure and accountable software development automation.

WO2026241000A1PCT designated stage Publication Date: 2026-11-19OPENAI OPCO LLC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/US2026/025855
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-09-10
Filing Date
2026-04-29
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

Generative response engines, powered by advanced deep learning techniques, face challenges in automating software development due to potential security risks and the need for human oversight in managing repository changes, as compromised models can execute unauthorized operations or exfiltrate sensitive data.

Method used

An automated software development system using a generative response engine that operates within a containerized environment with controlled network access, generates diff-formatted change proposals, and includes a review mechanism to ensure human oversight before integrating changes into a repository, thereby ensuring security and accountability.

Benefits of technology

The system provides secure and accountable automation of software development tasks by restricting model access, capturing changes within controlled environments, and requiring human approval, thus enhancing security and maintaining repository integrity.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure US2026025855_19112026_PF_FP_ABST
    Figure US2026025855_19112026_PF_FP_ABST
Patent Text Reader

Abstract

Systems and techniques are described herein for automating software development. For example, an apparatus is configured to: obtaining, from a machine learning model, an indication that a software development task provided to the machine learning model is completed, wherein the software development task comprises instructions in natural language requesting to modify code in a repository; generating a diff, using a deterministic application, illustrating changes generated by the machine learning model to the code in a cloned repository, wherein the diff comprises at least one change to a text file associated with the repository; and generating a change proposal to apply the changes generated by the machine learning model to the repository.
Need to check novelty before this filing date? Find Prior Art

Description

CHANGE PROPOSAL FO Al I O M ED SOFT Vi ARE DEA ELOFMENT R¥ GENERATIVE EbPOME ENGINESCROSS REFF RI-. NCH TO RELATE© APPLICATIONS(000 H This application claims pnnnty to: S, n0n->piovlsloMl appiimititm number 'i 9 32 I filed on September!0;2025. entitled: CHANGE PROPOSAL FOR AUTOMATED son A ARP DI Vn OPMl VI RY C U P M ivr RI'SPOXSl ENGIM’S. w hichmtn elaima the benefit of ILS. provisional egptic io-H number 63 / 806^37, filed <t® May 15, 2025, emitted Al i OM V l-l) sop jw \RF Iti M OP l- S l W H H ObNLR A U - KFSPONSb ENGINE, % htvh ov expre^ly foim-pura d h foteitee herafo fo then entiretyBACKGROUND(0 0 1 Generative response engines such as large to tiage todels teprcsesti: a significant uuk- cnt: m the field nl amficuil telhgvnutf. uifobomzuig emaptimr~h1jj<ed mmtml bra iuite undemanding anti rnerattom Generative rasgte e engines, powered by advanced deep learning technfouex, have demt tiated asLmNtem capabilities in tasks sadt as test genemtin, translation, su marizatfon, and even sods gmterati m Generative response engines can sifi through vast amounts of text data, extract context, and pro vide caberent responses to n wide array of.quedes.BRIEF DESCRIPTIO: OF TEIE SEVERAL VIEWS OF THE DRAWINGS10003 Details nf o e ar more aspects pfth subject matter described m this disclosure are set forth in the: accompnHying ttaiwfog arid the tiesmgtfon below. However, tite accompanymg: drawings illn ra:te only »me typical tispeets of fob efisefosura and are therafora rati fo be w idere limiting of ite scope. Odter feteu, aspects, and advanfeiges will become apparent item the descriptitin, tlte duxungx. 0tid the claims.(00041 FIG. I illustrates an example system supporting a generative response engine during inference operaifo s m accordance with some aspects of the disclosure,(00051 FIG, 2: i a conceptual btock diagram of att automated software dewh tn»f system using a gene tive response engine in accordance -with m aspects of the -diaelGsiue, iw««j 1 Kt 2 dlu. Rates,.1 il« stl!' rt,d' si oscw-i- Fn gemrm g wirnme u-u ig.i gene^ioc response engine ipape ndance with sorae aspects of the disclosure.? s tw foFIGS, A-4© Mmtrate vanows screens and conceptnal diagrams «f an automated software development system tssmg a generative response enmrte m aceo danc with some aspects of the disclosure.pM08 MG. 5 is another ermcepuml bkjch diagram of an automated software devefo meni syst em rising a generativerespontteengine m nceor ance with some aspects of thedisclosura,FIG, ft iiiwtmtes a flowchart of a process r generating change proposals using a generative response engine in accordance with some.aspects of the disclosure.MG 7 s a b:< k d i ar' Ulustratmi: an ex-im fe a nne-tearmra' perform accoformce with some aspects of the foscfosnm.|Ml 1| FIGS. SA, I-IG. SB, and FIG, 8C illuatrdfo sn example transformer aral ■mccturc in j^cordr’cc wn > ' c a<.]w ot the d, kwif e|Wi2 FIG, ft shows art example of a system for implementing some aspects of the present technology.DETAILED DESCRIPTIONfCMHSi Generative response engines sneh as large language modeis represent a significant milestone in the field of arliftdrf mielligeiioe, revohntanieing compther-hase natural language undemanding and gewmtiom Generative response ongmes, powered by advanced deep learning leehniques, have 4em»wed astonishing eapafetl I ties, ing such as text genemtiou, translation, anmman tfon, and even code generation,|0M M soy generative respmise engines pjfov i e a epn vem&t iomd user interfoe-e powered by a hatbtit wlwmby the user account interacts with the gerserative response engi ne through natural hnguage ct ersatioo svith the chatbot, Such a user Interlace provider an intuftiw format to provide prom foor instrueiiaus to the generative response wgme. In IhcL the cnwe tional iw interface powered by the chatbot can be so effective, that users can feet as if they are taeravtmg with a person,. Some user acemmls End the generative response -engine ef etive enough that they utilise the conversational userinterface powered by the chatbot as they would an assistant,{WI$ One use of generative response eagine is automated application development and software dewloptnent ftmetibbs to automate complex tasks, enhance productive and reduce manual e-rmrs. For example, generative response engines are driven by advanced artificial intelligence i Al} mmkd and can interpret mtrural language specifications or high-level7te‘H65AWfo.y t s twos and the repository uses eirtefirf verteoti tracking and.a ptoces to revolldaie cede after it Iwss the eontemeriked environment.I& MQ Fmafly, al lowing an Ml, model to write code with access to native filesystems or unrestricted ftei tok access intrisditetes sabstefitlal security risks, induding the potential for unintonded side sffets or msltetems behavior. A compromised or misal ign d model eonld execute unauthorized file oputa to s, corrupt repositories, or exfiltrate ejedentiafe if granted access: to local aecteto or thtemet connectivity.Systems -and. techniques are described hereto tor aatomatfog software development using^n ML mode! configtwd to receive natural language inatotetfons, generate source code modifications, and prepare a; -structured change proposal tor Integtatton into a versfon-eon tolled repository, to some - imple: eatafionx, the automating software development system obtains a diff aoterated I the i model bm st its completion «»t a iftn jo. rn dnprmmt task <tn im rhe diff to assemble a change proposal that includes the updated file content? repository tree st etore^ and Commit lineage. The eh cage proposal esm then be tendered to "a user via a ftonteend interface tor inspection and approval prior to integration into the terget branch bf the repository, This ss pOs. K'h supptVk st uclravd Oidrmste mn t i c eduttrtg tmk\ w hi pte cs mg the iMhshls and travc bihty to changes made by ife ML model.IM2O| to some as s, the automating softwata development system spawns a enuteiner d execution environment eon figured according -to uSeWdfefincd parameters. The container may include a preloaded development environment matching the asi r ’s tmttehain, package dependencies, or rwdime -ennfigutatiom Before providing the software development te.sk to the Ml, model, the containers network in tertbee can be disabled, ensuring the model cannot exfiltrate date, downed external packages, or otitet %vi perform unauth rizod bparatfons. This design '-.e po »C!! W, mprratobte nsfctehec fomra iiiirudiMng uracss i itefo fifoss stems ra the repository ctedeatials,. Follow me task oxeemmn, the ML* bnetoted changes arc: captured to produce a deten sinisfic difo. which is e nuted bos® the container tor downstream processing. The m work disabled container tlteteby enforces terted boundaries betos'eco ML -activity and fossome eases, as an alternative t disabling the external hvork aece si the c ntemer may ra jm> sdeteo ei conlrof extendi ne tek acckws ba^cdsoftware de opttomt te.-T slate. For example, a wft am developmem tosh ma be associated w-ith a pecific API, and back end may configtoe the mxv to allow the comamer townueet to a domain assu iuied with the API so develop l rk.toq»K $ tc c 11 1! P verb commands such as get, posh etej. In tehe e s, pr yF 85 Ttoto may allcw' interaction with other ackage repushori^, API sbrii s, shd so forth. The ptrixy can. also compl fK blockTo ensure accountability and EOirectness, the system includes a h'iima»-in>*to»lo^ review mechanism wltereby the dilT d a summary of the model’s pro mie changesar&displayed t©"a user, I siwe eases,. a second eoto er Is spawned to apply surf wlldate the proposed changes, Thls-sue»ndaiycontainm-inay clone the repository) apply the dif’C and use a’O^ammatie interfaces such as GitHub’s API to co tmct a manifest, create a cnniinit, i d open a p-iill request:(. PR)> In some implementations, this process may Includeupdates to "text or binary tiles, or the addition and removal of repository fifes. By reparatiteg the environments where the model operates from thqse that execute repository updates, the system matatata strong security guarantees, supports enterprise deployment scenarios, and ensures that repository modillcaticos are always mediated by human judgment and structured worktops.|fth22| IG, I iitatortex an example Al assiSteui: service supporting a generative response engine during inference operations m accordance with some aspects of the present twhnelagy. Although the example system depicts particular system components and an arrangement of ich uemponenk thrs depiction ts hi utethte a divers on ot the re'-ete L hrmto”;. a td <u<stod not be Considered limiting unless spccdrtd m the appended s nnis. I or exaniplc, spine components that are illustrated as separate can be -combined with other -components, and some components -can be divided into separate eumprmeuls..(00331 At moto(s) 110 are iri l intelligence (Al) models that can generate conte in response to a prompt, The prompt can be fiom a human w a software: entity (Al or applications, such as iem application 08). The prompt Is generally in ntotiral tengusge but coa i be in code, including binary. Sr ® examples of Al toodclbri Hfi can include lan uaat models that generate hmg age. >ttds HA lGP n toh titodds, ch s* D LI 4 <<; IM VH Gb\\ '»,hsdi svrate images, and SORA, which generates Videos. ( U lGl’ l, UAH -I I. AGE EN, and S0R are all provided: by 0P11 I, but Al model(s) HO are not limited to At provided by 0MMAL Al odgl(s) HO can:a1 ro be any W® o generfitive Al and cun include: Al developed using various arehitectures such u dsH s models and t mifbwnem (e.g,, a o res^ e trensfermer arch itecture) and combsnalm of models. Some Al models ate ute* o al models- inch can receive and output. ata bi text, imayc. audio, etc.| ft24| In some instances, a language model, such as CHATQP, can receive prompts o output mages, video, code, applicator, etc., which it an provide by interfacing With one nr more ©tor models, as will be addressed further herein.l247^^Wt3. P85tW< H10O2 Al modeUs) I i 0 may be part <rt a p« tieiaiive msponstt engine Ml), where generative ponse engine Ml) may tKfe mto or stwte,ompdtietrts for to porthig interaetitm with Al mndel(sj 110. tteerscan interact with generative response engine 140 through a front end 102 rand appheatians can iateraet with generahve response engbte 140 thr.myh \Ms H. Front end 102 •>c?\c.3 the hitmlto raid ntk'iim dia between llx- u.'j;rad gen tm> it pim-.e cr-piii. MO I* UKMmp^^ a pi iphical ikei mted.^e ioi wbicb PWS n >,.i to frovate input fort t tu! -..nw- wmmunicatln with: generative:res ratss:M0 and Al mrtdsl(s) 1 10. Generally, users interact tlwiigh a graphical use? imertbcc 104 that olk-n m ttdcs a croi ra'sehmtai mkraftiee. and a plinatinns I ntmet through the API 106, but thi k nut» i-equiremem.|W White some components are shewn as be mg pun ©f generative response engine 140 tmd some are oimmfo <>1 g ramitivc response engine 1 Hi, this for ifororatmn puipraww and some entoponenis su'cb as tool 1.30 might be part of generative response engine 14, and some component such as nuxteratton system DP might be mit-side of generative response engine 14. As used heic tL gcitaatiw W'-praiwparty serviees that prspsre prompts to fee Sent t© Al nkxtete, that are primiti ve level services that A I kkiel(s) 1 10 eim ntili / e to generate a n pra c ro a prompt Services thin arc utKrtdc nf gcra-rati tegxnoe engine 1 0 arc fogwfoh separate. Fw cv mipfo. tsvnt end 102 might he a llrrt- party front and (provided by the rome party that pro virtes generative response engine 140) but femt end 102 ed t mleri.rae x sth ge cra c se-pmise engine I 40 lather than aid m iLveb mg » ros row m a prompt.(003-7 $t e cumppnent tof generative tsspn se engine Mn might be execrated by etmiptiiirig programming w (CPU), wdwiw»snbi might bu cxvaukd bj gniphural processirt tt its mg., Al uwdel(g) 110).bi ime miihwjimtm chvni spphemism I3h might pnn xk a graphical nwi mmrtaw that is custom to eltent tippltemtor 13d tor can display graphical user totorface 104 provided by generative response erigine 140 provider{002fr| Graphical user lnterfoee 104 is the platform throurh winch itsera interact with the generative.response engine 140. It can: be a wefobased chai window, ms applteaiici, or any interface that Mipp-tm data tnpnt si d output Graphical roer mu-riace Ht4 mhutira a cutncrwhno between the user and Al mydel(s) 110 as the user provide prompts m graphical user interface: 104 to which Al iwdei(s) 110 respmid, and generative res m. engine DO presents these im. i mses m rmphwai n-w- rater dice OM, In some aspeets. graphical u\ei infed ce 104 presents <i4765-S " WtoJ* 851 W! eratvera tiunal interfere, which has Mribws of aennveraatinh thread: bet wees a uw account and Al m<x fet HO|M3(8| Graphical USES' iferefere 804 i$ retfeguted to perform in ut handling aii output presentation. The typeoflnputs that can be received can be relative to the specifies of generative res ortse engine 140 wdfor Al m fel(s) 1111 For example, a language model is generally-configured to accept texts but When the Al model isa multi -modal At models feat end 10 t accept oice and hnages vide.[W3I | In -some aspects* front end 102 can be a layer to accept any inputs, provide: the inputs to generative response engine WO, receive outputs fe generative response engine 140 and display theoutpras, Graphical wer interface 104 can -accept an input typesas part ofdie prompt* and front end Al service 142 can determine which Al model or collection of Al models are best suited to respond u> the prompt* Front cad A.i service 14 can be one or more Al models that cun ba "trained to determine which Al model should be ed to respond to a prompt, sod in some examples, can be used to improve pwstopte or provide additional eortfigura ion: details; such a h ittog a. reasoning model should spend processing a. prompt la general, frbnt end Al service 142 phtete u-ter tnkri e 104 by peHomntm m!erfc nee npcrfh.m hteou.a prompt is sent to generative response engine 140- Generally graphical user interface 104 is present on a client device, while front: end Laervice 1.2 'might be present on the same client deviceor no a server in a data venter(0032j Graphical t r interface 104 is also configured to maintain the context of the convcraafion, which allows for coherent and relevant responses. For example, graphical user interface 104 is responsible for providing die- conversation thread and other relevant context accessible th fi«mt end '1- 2 to the Al models along with the -specific prompt In an example, a eonveteutton between the user;«;<:re.m d th® Al reodte(s): 110 rem have -aken several Arms 1 prompt, response, prompt, ms ot. etc.). When the osar account provides a further prompt, the urapbual user inierfhee1B4can provide that prompt to Al mo ei( ) 110 in lire ermtext of the entire ceuverxfoon.|hfoH| In another example, graphical user interfere 1(M might be configured to provide a.system prompt along -with a user-provided prompt.. A system prompt fe hidden from the: users fin she Al model J he ss^tei® jnstsopt s n be used to define dte AFspdifSona, style, and constraints. There, can be levels of system prompts* A hiylte tev of a system promps might be prouded by die M nttnleks) Ht prmidcr and is meant to establish poliefes for the behavior of Al mbdel(s) 110; This highest level of system4765-S " to. F 851 W! prompt should be rohibited too being edited. A cnstomieteion system prompt han be used to m mmi' - the bcteu ror to she M mra l and h often pn>v ded through:m XPi call, or prtw i.led b a user ascmrai when creating a: customized version o f AT mndel I 10; A sit ft tower level of system prompt might include hidden information about a task. This can include chaimof hou n to>m;s h>, H!:' m,»dcl, o> Mtoi;skmr JB u gtomton sh,f \l mo>.l (s) 1 h» o wraknm wth to complete a task Accordingly, graphical user interface 104 dues not display ail of the output of the Al m fel.jW | phical user interface UM is also configured to dsspl o the responses iro he AI model, which might include text code snippets. Images, or interact K demem,jW$| tn some aspects, Al odei(s 110 trau provide inswcttons-to front end 102 that instruct graphical user interface 104 about how io display some of the output from the Al model. For esumpfe, the Al model andtor gene tive response engine can di met graphicalinterfile© 10-4 to present code to a eodtospecide format, or to present interactive graphics, r static images. In Other exam ples, the A I model and or generative response engine ©an direct graphical user interface 104 to present an interactive document editor where graphtai user interface 104 ©an be presentedtoe use) 4, >'ii)l sad the \l m>*. wto eutotoorak on 'he document,IMM in some aspects* Al model (s) 1 10 can provide Inslntetions to pe lmtion stale 12fi to record facts in a pe ualixation notepad. ersonaiixation state I® can store Information received directly from the user account or intoned from prompts provided by the user account, trad can be used to provide com to the Al model such that output is personal bed for the user account in some examples, front end K12 can be configured to notify the user account that s memory -A; ereteed. Persdnalizatfoo slate 12$ eao slso in l summaries of past nve tipn ihto stK hctowfi toe W'.ei txtomrt sod she gcncramc re&pimw engttiv I W a a ^a hahte index of past conve tkm threads. In this vay, peraonaiteattos te 126 "aid generative response engine 140 and / or Al raodel(s) 110 to learn specific facts abdut a user account and recall past totoradtom. The past teraetibris W be used to guide futuro tes onsex by generative response muFte I to(md m X! mu telni 1 10 irh th.u genu-mto o <prawe enprae HO nd r Al muddrai 1 10 becomes pum al d tn a pafiiciilds user account.|W37| Generative po ©righto MO erm provide' one dr tpbte application progmtrnnthg interfaces ( f’i(s ) 106. APIs enable deveto ers to Integrate the Al rabdeFs capabilities into c acmal dpphc.bu m and wrr hx, 4 PI -! 06 pros ide groeraramtora ack m.s to eeneras ice response gagin© 140, allowing tor customised intetactiteis and &nc iosahtKw A lulc AP1» l”t are shown ftas part generative res^bnw engine 140, st some "examples, APIs 1G ate part -of tom end 102, API 1 6 am refer tn points of totoss to Al mode (a) 110.. In same examples API Ito en phte a ® locatedat context reauagemeni service I|W3R| APIs 1 to can accept at etured rerfeusts containing prompts. emrtex and conttg»thn psremetera, For example, an API can be ed to provide prompt and 4n1Jc the prompt into system pomipi-.- '-.mi uwi pTcH-i is hi Mibie spec’s APfo!ton <. U? piro wc specific mpm» u>i whub M nw ei(s) 110are eonrigured tn respond with a specificbehavior; Forexample,an API can be ured to speedy ass output in a partic ular format or st rue m red output. For example, let fee ctot completion.-V'i. fee \Pl ral. n sp edy parameters tor thv output..reJt 'is to n kmgfe to the detofo wtpre, ami s.xxT\ aspects to the lone to' fee language used la the res se, The o pui ears alw be st etored otrtpo end in a data interchange fbnnat s»eh as JavaScript object notation (JSQfeh For example, the input tpre y may me lude a scheme that Ideuilftes a speei fi otop-ut stractare. Some common APfc are for participatingInaconversation(CbatCb pietion API), forpfovidmg asingle sc g’tsu^e (Com k-sto MM r to cum- sine tex5. hfto embeddings, s Lutoddh- s API}, ck. 7 to API.re also be cred u> redtoste' specific decision btiUiitories that. Al t delfo) 110 might be trained to Interpret For example. the mmkmtion API.ut rake ads.mwge of M d-tomt ’tetskre 509k rronrefo uunkiratkre detotowm stomp. In the ease of the mutomfem M’l and totesx, the API togfe give access jo services, torer than the Almodel toraxampia, the ntod ion API might be m interface m motomtoa system fob, addressed tokm,|00»l Some toher coinniw APIs mdads ilte Fme- miug API, which aliws itev fojKre to iwtomize Al w (sj i 10 of generative response engine MO using their wathe Audm■red Sto^Ii APIs, which cause Al mtoel ) H O to feut speech or eudire and the Image Gmteretito API, which mires Al mtoeife I Ip to ont iit images (which may request ii eratims of cilier models).(IMMO'I There mi itoo he APIs that direct Al mtoei(s) 110 asitor generative res reiM enme 140 to interface with other appHcatiuns of other generative Al engines, In such eases, the »pcs.if application hr Al engine might he sgeeifefe nr Al nre-delfe) 110 might to allowed to elwre another appli tor ofgene iw response engine 140 to me in response t prempt imu As described herein, a Responses API can be an Improved API for interacting with: l mo el(s) 110. Respire res API -si ta e ns a paremu tor « mp»w 1 D, wh idt ere ho -^vd • c retam Al models) 110 to a state itoilto model was in at a partteular response thereby fe towfog a user account <u A eloper to rmfe 4 prauniw pretest, wifeoui toreae m re-prompl A! modetot I bi er having to rtphcaK a previous series of prompts. Further, Restores®® API can faeilitete4765-S "to-to. P S51 W! fimetiotetopi chaining to enable execution of bpe tfons Psqmrhtg several fofteticthfeol calls vtotom she added overhead sto responding to an API fefe io e; JI r e Him tom or ovsl Instead, deterministic code associated: with the AH cso be directed to tali luucltons er is (iheW g other Al models) behalf ef the Al model, and return data resulting f m the function or tool call: to the A l model.|W«| In short graphical user terfaee 104 and APIs 106 can be used to provide prompts -.<■< Al odel(s) 110 of generative response etigine 140. ifem ts are sometimes di h tiated into prompt types, For example, a system prompt can he a hidden prompt that sets the behavior an gsodc Ines to; the Hcner me response cngi-m A use’' prompt the e toctt input pro ued to the user, which may include qnesttoto, conwnds, or wtonnatton.H»43i Generative resp-oase engine 1411 may also Include a context management service 1211 The frncttoa of context management service 120 ts tomanage and ocgmxta the few of data a aag key subsystems, enabling Al mo el(s) 110 to generate responses foi are contextually relevant accurate, end enriched with additional int n liun as required,10044] Action 122 facilitates auxiliary tasks that extend beyond basic text generation, In Otoe aspects, aettan 122 c be actions that correspond to an API 106, In some aspects, aefe s 1 2 can h« agmtoe actions that Almod-ehs) 1 10 decide to take tirom out a msteA meto as described io the prompt For example, an a i on e bs to- call tool 130 or even other Al mottelfo) 110, Tools 130 can include internet browsers, editors such as code editors, other Al tools, etc, While tools are shown outside uf nertatw response engine 140. some tools might be p of generati response engine 140 and some tools might beexternal tools.Prompt 124 is the request or command provided by the user acceont through front end 102, in some aspects, prompt 124 can he further supplemented by a system prompt and other fofcmummiihte might be included by graphical awr interface 104 or API I ( or w-oe te with a custom Al model. In some aspects, prompt 124 can be modified or enhanced by Al mpdelfs) 110 as addressed further below. Additionally, its the user a e mai provides prompts ami Al modeks) 110 provide responses,,a eotwemafims thread forms. As the user account provides a new prompt, ibis is appended to the uvetall cuuw ttou mfo added to prompt 124, Thus; a user aceoiml might think of a first tf®gr*provided message as a fct prompt and a second user- rovided message a® « w<ond prompt, ami so »n, but prompt 1 4 pwveat- to, Al moc h 1 10 can include a Ihtuail of uset-p v fded messagesand resportses fem Al mode If s) 110 i»-g tnu i-turncouversatiou. The actors In the conscrsaiioa thread cart be laKfel s<s thm Al ntoilJt Utt can review Hie turns of the eoto'-er tom Gcmfely, prompt 124 will include an csttro corwcrsation thread, but in some1.24765-8 fotort, F 851 WO! instance's, prompt 124 dy fee shortened if it exceeds 'a maximum accepted. ietigth (generally me wired by a number rtf lotos).|6IM6) fo some examples, context management service 126 can also route prompts:an response through moderation system h. In some aspects prompts are provided a* prompt surety astern 1 12 before bu,' pro^’J^d to \i jn dJ^j!ID Ifomp: -areis '-restore 1 '2 rellyi re-4 to use one or toure t cimiqws to evaluate prompts to ensure a prompt is not tequ.! hip A 8 mudch * s 1 10 to generate moderated content; In some aspects prompt safely system 132 can utilfeg text pattern matching, classifiers, and / or other Al techniques.1(9 471 Since prompts can evolveover time through the course of a conve tion, consisting of rom and responses, prompts east be repeatedly evaluated at each I HI the eon versation. W48 Pe mali tion stele 12 can litete continuity and personalities In conversations. It allows the •: system to maintain itser-s eeifle context, pmfercnees, or details ibai may inform tutors* foerate io « PeratoUi) L ist ton state 126 can include a memory fife of persisted data fom previous interact u mt or sessions that provide tetekground toformation to maintain continuity. In some aspects, information can be recorded io pemonalmatron state 126 at the insiwfina of Al nrndeKs I W when A) mifotofo) 1 10 fo ity a fact or date that it dete m i nes. should he ed in memory hecatise the firei are ’ m < ait & might be useful j« biter conversations or sessions. tn some aspects,.perwnalistaiiPn stem no can also include s ntesteed concepts extracted fro past conversation threads, and personalisation state 1 6 can tffo encompass the ability of Al mode!($) H to search through past interactions to find relevant information to a current etwersat n!brefo(0049) Conversation metadata 128 can aggregate -data points relevant to the conversation, including user prompt 124, action 122, and pursonatation state 126. This consolidated iniormatton package serves: as / ftc input for Al models) I IP. ■Conversation metadata 8 can label parts of a prompt as u -p i ed, Al mod *prevlded, a system prompt, pstounali fon state 126, date foui action I 2 or toot 130 (addressed below).I0050| Al utodel(s) 110 are the care engines dial process.inputs ( tn context management \er! Cc l?i;>,m»l genctate uulpuh In inw jtv ex l untfkh I ID?resy he a g n ifo t oSibrmer, or autom^ressi.ve tmnsformerr but could use other -architectures, fa some examples, the transformer is mitol- edsl trttnsfomer that can ase: audio tokens (or embeddings thereof)* visual tokens (or embeddings thereof an language (or embeddingsthereof) as needed.16051 A COM fatere nf Al modelfs) 110 is to: ge nerate content to response topwapfc Th prompt e m he h mt ehtnl apphcatron I pr* xkd ' a graphical User nteribre ’0 l move <st 1 K) can be eonfigure to rsesive inputs > APIs 106 thm provide guidance <m at desired output Al mo d(s) 110 can analyse the input and Idsatd relevant pattern and associations in the data, and generate a -sequence of tokens ihn arc predicted as the nwt likely emrtumation of the -input Al ntodel(s) HO generate responses by amplmg Irom the probarilhy diAribthian of possible tokens added by the patterns observed during its Irammt, hw features f the- atoore^scssh tranrim f that result ui (hi1* toirttu uhty ate that the d toregresstee' irariuwi ought vc the decode pun »fihe trjBsioro.er.ischii eusre a1<('u t a into 'w SU'J as «?(»■' Pv._ -ro ’ she?«■©« dot pan of the transformer architecture, the transformer tocases tm predicting the tokens given the ptw k»tts uitexi tokens. The sdKit ntton wjchanwu captures tong-range dependences.amongst tokens, allow iug n v» generate amtexmrily te ni re porew.^ (in text, taho, images, and * idea s,|0052 In some; aspects, Al mtririfs) I W can generate rmiltipfe posribte: respaoses heto e presenting llse ( ul *me Al nmdeh i 1 HKurs gmerak multiple responds IMSVJ nn the input, a these responses are variations that. Al model (s 1 1 ft consider potentially raiev wt and sdherenh100531: In some as t;<,*. \l c lelt- HO va* e hutto uenurair* ■ icsponw » tossed o« certmn criteria. These criteria can trtetode relevance tn the prompt, ertoerenee, Owncy, and smoetimes adbereitoe: to spemtk sm(.lelines or rutes, depending <to the a pltoatom. Based rm this evaltiafii.m, Al mo<ito{s: J 11ft can select the most appropriate res| ense. This s l to to typically the one that scores htghc-d H thecntei in. bnkmcing tacto's like rek-sance, m mitoirestess., eohet cmre, and content moderotion: ihatoietipns / trelniug.1«054| In Marie topecto, ah strucfton provided by an API 106, a system prompt, ttr a demsimt mads h A model(ss HO van cause Al modeksJ I Ift to interpret a prompt and rewrite st or Improve the prompt tor n desired pttrproe. Fw example:, Al model(s) 11 ft can deter mine t» take prompt to make a pictore and enhance the prompt to yield a Ivetw pietare. In toese instances, Al modet($> 110 cart genE ate their ro prompts, which can be provided to tod 130 tor provided m Al nimkhto 1 to to yield a better oMput re^xm than the urignml prompt might have.|0055| Al n:to sl(s) 110 can also de more than generate content in response to a prompt in soivr u\pe> Al mratof -d 110 can nsc dckreron danev to dcl mmc the appropriate oWM' o! action Based t the prompt. In some examples, a derision bmtrobry might be nsed to cause the Al model to recognize that it is bring: asked to provide a response I n a particular format such that tire Al model will generate its response constrained by the pariieuiar fottoat. In some examples, a derision boundary ean cause the nm el to refuse to generate a responsive output if th© derision is4765-S " te. P 851 W! font foe responsive output: would violate a moderation poiiey. In some axamptes, the decision bsiiit ary might cause the Al mrofei to recognize that it foterfhcm- with another Al model or applfcaitan to jespond to the prompt. For example, when the A? model is a language anted, it might recog e that it is being ashed to output an image, and therefore, if interfaces u ifo a model that can Pulp images to provide a msptmse to the 'prompt. Ip Another example, the prompt might request. a search of the H-temet before responding,. The Al model eart use a decision boundary to recogn that ft should wnfotet a search of the Internet and use the results of font search to responding to the prompt t another example. flic prompt might icquot thet she Al model take an agemk aetto on hehulf of ’he user y interacting u sth a third-pm-t sen ice m y,, book a re' atmn tor me at.. ), and the VI model van w a Uect on boundary to recognize that it needs tu plan steps to locate the third- carte sen <c. rnuiet the shn - arty -<n set. and interact with fo third-part}' smta to complete the info. ami then report back i the user that the acttrni has been completed.Vdiui M ruudchsi 110 dtocmnne to la an auctiiic ucth on ixhalfuf the mei or to tall a tool to nd it- pivsidmg a quaht respons to tire met di anml. Al itHnielf.-t I 10 might cab <i tool 130wr cause an action 122 to he performed. As Indicated a ne. took 1 Mean include internet bro to edters stto ce n follies >. ofoer Al tools foe. Actions 122 arc actions that Al mmddf s) I 10 can ca e to be performed, perhaps using tool 130. A<> u< »i herein act ions 122 shuuld be ewsidered to cover a broad tirray of act tms that A. I models s) 110 cart perform with m wiihrmt is 1. Ah Tools 130 arc considered to cover a wide variety ef susvtees rm software that c compass touls such s a computer ijp tmg system uch that l mudclfd) 110 can ermufo the com pa ter o -er tmg system en foe user’s behalf to robotic actuators, to search browsers and speel fic appl tealions.IWS7J Additional^, mmtelfol H can ateo generate portums nf re.ponses that arc not displayed it? the u^i f-o; example, i iiwfofo 1 10 can: direst front end 102 to provide specific behaviors.;.uch ro. direction's for hero to pscscnt the response from Al mmk- y Hit to the user aecuunt, Ingnafoe es.ampte» Al modehs) 1 10 can provide respe poroons dictated be AltoTI Ofo whete poiwn sd foe avpmwc us foe \PI mieht be for the cmn>iJinptmn of the eallmg apph num but -rot for pswuttomn to the end: user. In another example, some At mmtdfo 110 are reasoning models hich w M models I lit ihm.re4.ontlgureJ m oittpin <i r.™ cbmn-of- tou hi fotore re ring a tidal respm to st prompt. The mw chaiurokfooughi might not be presented to d er accoute or uppliciitfo?: catling an API. Instead, aattoter Al rondel might summarize lite raw chain- ofoiwiight into a mure ccmsumabte and t fel outpm: for the ttw nec at tjr appilcatiomBP to I H|(W5the um ti m'iht- Xi HKHL-I L. IH he hit t e; mjfeyzedsystem 1'34, White Al txi <si 110 t n perform some of their WK modertoios, there can be instonees where ft is tendfeife u> have another service roview outputs for compliance with a moderation policy, The use of dashed lines in FIG, difewdaiss a path using output safety ■system 134 and not tisiiig (niipti safety system 134.|0 As used herein, the ptvwu th < riplion ma refer to the generative response engine or Al modcl(s) making a determinate! n reasoning prim to providing -a response to a prompt. Deterndning:can refer to a process by whfch an Al model emtpuis & response that provides- a? c el ton, In an example, art Al mote can pedor an inference opera fet to Otrtpui a sequence of tokens, Which output provides a etermiaatiors. Ip such an example, the Al model might not have made any decision other than to sample particular tokens to provide M wtput, yet the response from ths Al model might he astatement of determination. In another example, Almodels can be ttained to make particular tfefennfeutfens. In such e mtptes, the Al murid can be trained to make decision. such as m crosva deefefoa boundary to take anagenUc action or m call a tool, etc,|W60| Reasoning refer* to midaii g hsgicfe proves w’' to texx c<w,hi’>mii o pci form d-x based on provided inputs an teamed information* Although the- Al model might lack true rntderxlaodtag, the Al model uses pattern mcogftitton and linguistic associations to emulate reasoning by link g rotatedte fept s and predicting outcomes. As used herein, the term reasoning by an Al rnodri refers io the p c*.w perfoi med by Al trwfete known as febsnning wfeefe|OM11 FIG. 2 is a conceptual block diagram of an ausonuu-.'dit ware dev'cloprnsnt system 2t)0 using & generative response eng iae in accordance with some aspects of thedisclosure. In some! aspects, automated software development system 206 includes componehte of a generative rc.'p*msc engine ''ttoh as a hotst end 21 (e g * f to end Ffei, hack end 22 and Al unnivl 2. (e.g„, a trans torrhes’^based model, etc.). In tme aspect the automated scRwre development system: 200 is ebnftgme to perform d£der ihistie container setup and c'on '1«m at ion prior to providing the Al model 230access to the Container- For example,the'automate software development system 200 can eenfiguKs the container with a development envimnmenb metading code*and ten itlk the Al model 230 to access theContainer,|jtMte2 iMotoe aspects, -Al model 2 Os is n f- i taik model w hicb is targC’$e8tevpta’tta ed nufed hushfee range of dv nstremu tasks such to language undeta ndiny. code genc^ fon, image total j S. and aiore, Fotetaiicm models leverage transfer teaming to adapt eftk tenth to uvw problems with minimal additionaltrafei g. An example of afoundudtmwtiei is a reasoning: foundation model fem uses chai n-oi- feutfehl reasoning m which: fee model gewafes- ifeermediafe logical steps to wive at a hn;ri answer, 'hits process mimics human problero-soMng by str urmg the r ming path, which improves b&fe fee accuracy and i erpretabflity of the model's, ufepute. In some eases, the foundation model mm be lne-femed tbr specific iasfex, such as softvwme development, nnd fee model can be pruned to reduce weights and ns ahilto to perform i»cnemLpiirp»w reasomne, thereby using fewer computfefeufe rexcuresx and improving efScfeucy Ibr targeted applications, iw la some aspects, front end 2W can be a client application that is presented at an edge device and can he cmfemsre in various forms xuh as a uatjw appliestom. i r x-pfeltorm upplfeatfen using a bview feerer fe.g.. Efectiw, etc.), or a web browser bwmsmentoti application with structfens (e,g.sa avaSeript bundle such as React nr WebAsaembly bytecode, ete.). In' some cases, client application 212 can be retrieved from elond comparing service 204 (e,g„ as a JavaScript bmsdkri or may be rendered via ukrnd computing w-vrec A >4 via different techniques ic.g., xroref -dtk rer krmg, psubai fedratfon. cto!The:back end 330 provides art iriterlfem for front ea 310 to istoract with various citosposujfe on a MO CI or tloitul 1 or cv tmple back cud 220 us>, hides ^nuus n KCS such as debase 225, inspection services to analyze the safely of prompts, and into fece wife Al model 330, Back end 330 deludes orehestratiun engine 222 imd patch engine 224 to itriplcritenl: softwam devetopm'ent services* O ehestriife'»«t cnetoe 3*2 is cnnSgurod to receive instmerio from front end 2 KI and spuw n a ettofe n ier 240 wfedntg u dev opnwrt ot. imtitienl lit vans speefr, edge device (wg,< controlled by a user) can configure various aspects such as art executkm envit eut The exu fon apviromtot Htato a toe system <e.g», a Linux distribuli i) and ca» be iostn tenons to nnfel the base environment ( <g,, a docker tile}. The execution enyitotime t can also chtde additional instructions io tosirument ibe dcselnpmenl cnsironmcni wtfe various s fecx packu ex package managers. servers, wifigu fens, SDKs, and so forth. For sxantpfe, fee e 'cmhm emirom to can nde iswtrndferw io include a package manager te.u, pip for Python, cargo Ibr rust, toai ppm for JavaScript, etc 1 as well as urt us dependencies. v an example, a dependency can be a todopiuem tote <e.g„ a I fetor lui cheek mt» Un errors Knme Innfeiue, a ptefeer io m-fetem h formal code, le 1, a hhruty lor mtoraetfeg wife a dawfei ', -t framework fexg,. React etc t a development package. and so forth.Or hestratfea engiuc 222 retchw various informamrn hum from rod 21 it ImJudsug fee execution cm hiimumn md *paw ns container 2fe mdudtng fee todopt em envtromn rt i n»g d.pw rri arid service 342.W AWti. F to IWOl | (W66| some aspects, contatocr 2.40 has a lifecycle lh i is sulmble tor ociw srk appfteatmtw In pMtictoar, M ItMaoee of a container 2.40 is ctoated, started, opoo lh paused, stopped, sn eventually removed. t creation (c.g., docker c rte or docker nijil. container 24ft i hush in »m a specified image and includes M isolated environment, which contains the fllesysto, environment vambies, entry oim, arid network settings. When started, CMtoter 240 wn a defined process (c,g,, in the dockerfile with limited access to the host system, CMtatoer 240 is a se ed m romnent and can he accessed through services installed within the doekerfile. For example, MMSS to CMbtmer 240 iy.g- Oles, APIs, or other ftmcftMality) is provided via services instei led and configured at build time (e,p, m the docker cn iigufttoM ©r toe dMke fileX AftercotoMW 240: A Mecnted,cooteifter 240 eon be accessed v ia the defined services, and titers is implicit access to itX Conten Once the process vnd iMiuunc 240 stops. Any change made within e titeinsr 10 se.c, to {he filesystem) arc gcucinHy ephemeral and are lost ^ hen the eon ner is stopped, i removed.jtM j In » - pe v, cdntaim 24ft may clude service 2 ihaE is ccmi i utcd to interact with Al model 23<i. f ur example, Al model 2 toay M executing: hi special iwd hardpan; for floattag point ope ions <e.g., a gtophlcal processing tmii (GPU, a uitl ly mam uhte ( MAC) unit, etc,). Thc crmuuaer 24ft can expose a port to allow service 242 to teraist with Al model 23ft M< I perform fund fen p^oJated with l model 230, In wnewes, arfmts actions.by Al model 2 ft* cap. compromise the actions pci formed wlthm container 24ft. such as Eii.teni li.ng a denial of sen sec attack, purtrn n licimK social nweft (. I'MUIL and -.«'• Sorth I en llmuah wnMm,i 4<»ass isolated es t viros mifift, i t CM be stak'd himi C CHMI uwc^. arid internal MiMWds may be penrthfed. For example, c^rstalr^r -l{> ay execute shell uwiiu h te g., owl, ate.) or other applications te.g, ping, etc that can be acceded tor various mahciouv purposes.|IW8| Ip spine aspects, sw itch 245 fe:configured to permit of deny access to external network 50 by coatoinsr 240. Switch 245 may be a logical network interface rtfcontainer 2411 sod witch 245 h illu. bikd t>» 4 that back wd 22ft mmol, whether container 2 ift can MCVX CXMMI network 2 ft tnot the miental netw ork!. lai some aspects, switch 245 can be enabled tot an itiah / Mim p^n d associated with conilgursog rhe container In this tes Ckintemct 10 rn access nuwork 25ft to itv -v, packigc repository 2^2 sinrane an mtEtah / mt n period a sn aEcd with canti-tmcr 40 Package rep shory 252 can todaide xanot^ tlware packages, coa tMtor in:isagessand ftiher co paden to generate and conilciJi'e eo tmner 240 In xome aspects, ufcheshalion engine 222 is also tonllguted Eta cause milamcr 24b to access cotte lepos rx 254 to totall c tone a public or private t^poattory. For example, otohesrtatioa engine 222 CM ancWd-k lf. P 85 tWiGI HUB (either using a public tawfawot a rep sitory based on a stored secret of the ser) and retrieve m'd stere code repository: 254 within ce eta ine 240|MO|o remrei 25(l abej dl code m uekages are remlted in eonitoner 240 and the imthK km period ends. For example, switch 24S is enabled belbre Al model 2 W s prodded s242 ol court'"' or 240 and is dwatoed w en \I mode’ 2 h. iiiw s te ren- ret 242 ot'remuino 2 0 hi tin tare 4.t>eU8«cs 240 re isnabk to access e vtera.il network 250 and the data, packages., code, and other content instated on container 240 can be accessed via an Internal interface (e,g.sa network file system service), in this way container 240 cannot perform any mahe iv actions outside ot it-, env ironment and any harmful noth ilie are wreck'd re ns enurvnm n In S> HSK tw ctrt >nti i 15 c m be oaitl ned usmg an m sl interfere (s,g., igtebles in a Linns container) or as external Inferfsce (e.g„ vis a control plane associated with a ntniner service).|B7«1 In some pectS, switch 244 can be.configured based on nser oowiroL For example, a nsw as part of st project or an executed envirorsment. may deal to enable limited or b irtd network tae$s. For example, the user may rtwhgured mtoiner 240 to control the scope of etv ork avCcsv v ia n ci ititismodnm to restuct sccesv to jwtunlat nom* and p tovok Fin exampie, a mr repository may be focused on i piementittg a new service based on aa A i f v.g,» wealhertae eiCrtm aud rtsraal and external network access can be li de t network tr. Af to and frnre that fully oai tfied domai: name* In other exampie®, network access csft be time I tuted b sto sin rt evtsm d request io an cvicn^l vrsu, rer ntm a p»a?re 4 ilimi a dtlfa m {<. g sseonds), to some cases, container 4b emd todncle eiMt eonfiietwitons that identify advertent ar ma ient behuMor. such u>,» loop tliaiLOhtiiutes to execute tm idumtoto iietwork tuquest: h e on an improper <.otolma’;nkm Fable 1 below ihuxtrto g a example of roles that etro be.tpplted to a fiixwall of coniattwr 24U, ro *vommatoUme nlthrt sp 'toe \duch configures the packet Altering rales of the I MUX kernel firewalki 2 765-W to I WHjlWil The use of ipiabtes is an Example lor illustrative purposes. Ih some cases, the dxfe al uetuurk tufertee may he emupte y disabled:uid flw container e<m be ace\%<«cd s ki an ltem.»l inte tet je.g.. dusker escej.mutate w hh rhe container servicci10072| to xomo cases, as; an alternative to isab I htg the external network access, the contain® r may use a proxy to access a external network; In some cases, the back, end (e.g., otecstration engine 222) can control the proxy dynamically an sefeetively controlled g msd network access teed on softere development task site. For example, a software development task may be associated with a specific API. and back end may configure she proxy so allow the container to connect to d domain. associated with the AH to develop network requests teg., HTTP fe commands snbhns get, post, etc.). In iMtec.ea.ro, pr x tuaj. allow Internet ion v-ith other package! repositories, API swtacex, and sc lot th. I he prosy can atoc cw ■ t leteh htek i letw t k connections.73J At model 230 Is configured to intoct: with se.rvfee-.242 of container 240 to generate- eode changesandother modifications within coataher 240 based on a naturallanguage instruction, For example, service 242 can prot ide iviminal access to Al model 230 (e.g.sSTDIN, STDOUT, S ] DJ RRi and ulte Al model 2 to use the tonunnud hne tertfe.ro (Ct h so intend. road, ami modify the copy of local code repository -(e-g., cloned from code repository 2M), run fests, and usher iL elopvi lui> fa >n> I he e roc 2,? rn cl* * ro to te astunsk modtftehou* vid coinmate Hui Al sited 2 hl rx rosined with to prm ide s te’ innenonx associated wHh a sohuate development task,. For ' example, A.l model 230 te access a local G1THUB r wttory te.g., the! locally stoned repository wiihhl eteteer 240) to create, read, update, and delete code files ing ttn API. The At model 230 behaves like an wtono ow developer and reads cite u tu'a rosroto > ahnni hiteron $ln> e-r hstos. and pnteecx mime edit', new code, and fentoves code « needed. The Al pdel 230 is able to peroeive the romite eastern of she project while isolating dungc'- to ctmfemef 2.40.W74| Crnteemteial amumste development technique either we direct access to a live system tn directly modify files without netwrorfe istette (e.g„ n uhm a local dpvclopmehf eteroutenfi and risk puteniM malicious betedor, or an incomplete undemanding of the: environment based on mt if« tiohofa b mpositwy within a wetobased envirotete, For4765-8 " ri. P 851 W! example, the local eenfigutotimi can mriude various tlngs (etg., tt Un -, a configutufi ft file, etc.,) and copying the eoduten a wetob sed ftent end May require additional n dificntton to match the local developmt t cnvironmcnt For example, if the iwImplfcstAtty flag is set to false in a fecal typescript configuration file and the ernte has im 1 left -any retom hc, topping the changes.fiom a wa msed er ronmem van create additional ermrs within fiu. copied uisuu rt»ma based on ditYun ce in the fecal development eavl nmvm and the wctoha^e trom endIW751 In some aspects. Ah model 230 is configured to exeetoe in a dtfibreto hardware. -environment and can access service 342 to interact with the software development environment For example M model 230 is typically executed at inference (,g., at runtime^ on specialis'd haf ume, such m feife a m custom accelerators, such a towor processing u n, f I Pl Ah that am configured for la ge-wcafe matrix math. ML inference and training involve massive - parallel fioatiPg-ipoht. computations (e,g„ matrix mtiftiptotfens) that tire lu cr on G U hardware. For example, Al model.230 y require access to higlMhmnghpift memory fem,, high bandwidth memory (IBM)) and p lfel processing cores that: are act awlabfe in c mfenal CPU ardntevtwes.|W7h| Mnhspic sm tice fe autmnnkxl wtb ne slevcfepmcHt * ricm 200 ma< exx ufe in parallel to handle difiewnf tasks. For example, a first instance of automated software development system 2f)f) can be adding database inte rafiom a second instance of automated software development system 290 can be adding anhicjnieaifesit, and a third instance can be adding a new teafete,|(W 7| After the task by Al model 230 is cornpfefed, omhestration engine 222 is configured to use service 242 to extract the development data associated with the task, udiieh indndes i > ificat:m x,esim awds, changes, and 'any evenfe that ocetofed within co iafeer240 while bring enntroifed by Al model 230, Orehe^tmtion engine 222 sfnres: the developmetft data in database 2 3:|O078) Patch slight© 224 is corifigawd to access the development data extmetvd by whestratfen engine 222 and generate a change proposal associated with code rspnsftory 2S4, For example, jukh engine 224 can rrnt commandsto • eterm b.ijealiy poHiueM-difiofal1 the changes made w Ohm the develops «u data In some cases, patch engine 224 ma be configured to receive- th© dct pnwni data ond feeriidy the various changes fe g„ ifem> o determimstic dift) and store: he modi fixations in database 225,Uwr^WtT p to twos |(W7^1 some aspects, Al model 230 <j n genera to and mitpat the diff to database 225. For example, Al model 230 can call a determiniscc tunyiinn to some cases, Al model 23©-may he able to reason and generate the dtft’dunng inference. In some cases. Al model 230 can also review the changes to validate that its reasoning corresponds to the generated msir toas.|M | Patch ermmc 224 {or Al model 2 fib ^umcrme« a chri' h xtomifms;t<cpmsc-nttoi>,w of change* 1 c.g„ inscOfonx. ekelorss) in a paslicatar file, Al model 230 can be cc fig ed to output attwcumd mW mmum usiag varioux tochmt es, such as a data validation fi c ian ttr libmry (e,g>, cocl) that validate:te types that cross system boundaries. Table 2 below dlutsnate a iff cemmasto of two files (tog,,,to / helto.txi and,.4rihdto. ) apd eoim^bnding eu^»ut fowt that j(ktoitks' toKtoux «md uM mo Mowed by the lines to remove based on:a fea lnr " lmc$ t»» nddbased on a leading **”)<art dcvelopinait system 2(h) toemtflgumd to generate chamm ptopos a,vw<itsd w ith code r njxtux* 254, Ft a nenrity porp ^, automated -^ft ate ik '< clupnamt «i$ stent 2(MJ mas not be mfiuured io ire h access y dc repusitorv 254 and e. generate the ehtmgc pmptw for the user’s review. The change pro o l also in rdv* additional infomiation that can be used to generate all changes based on a slate of code refU fray 254. w vCampkx du Jumto-, WP< I c< n. dud, > xo.ncc I i meh cm ton ' - pc-.y<'w», te mto. <•: npc* bt beh where the eba ges are intended to be merged, a commit history that mnli es the modifications, a dlff showing line-by-line changes to fifes, and optional metadata such as a title, description, reviewers, and discussion comments for code review and ecBabor tiom For example, a commit h a snapshot ofa repository at n apeci fie point and links to a e of that snapshot The commit includes a -pointer to a tree object (e.g. a tree of files}, a list of parant commits (iw history, and other metadata (e.g. aat or, times * commit message; etc.). The commit pointe to the" new tree and previous commit and allows a graph to be constraeted te identify changes to be made out of order.|0©^2| In MW axfx ls fisc du W s pos.d H U toa p.n'\ t<i fxJun u }'»»»{X>sai wd generated and provided to back end 220 based n user input Itor example, front and 210 candisplay various: options essucieted with the change proposal* swlr s tire ability to view and generate apull req sest, Based on the user input, backend 220 accesses database 225 and generates thechange proposal|«| in Other aspects, the automated foft re development yster; • 200 east be figured o interct with. Al model 210 during before sad daring container setop. ur example, the Al mmfel 230 may receive htsinmiroas to generate multiple variants of the same software development task. In this way, the Al medei 230 may need to identify different alternatives, such as different packages, services, or other surfaces before network access is disabled or limited,|0 4| FIG, 3 illustrates a flowchart of.a process 300 for gencuting software using a ecnera c response engine in aecontmee with some as ch el the tb-drw re m some -wgect, process 300 can be executed by a computiag system to implement an automated software evetopm t system fo g., rnomted software development system 200 of FIG, )|W851 At btoek 302, the computing system receives instructions: t pufewn a software dm opm uisk fli task esn he large or Muall, and can include till k rem aspects, ot the sfot erv development task. For example, the software development tak can include a request to up^a e the pmkagev le g, wr vd uiJwhng mwerimg the piop m Jilforent IWI le g, from Co monJS to FSM), updating libraries, prmlag dtmdaut code,: ere.|WM| At block 304, the computing -system cmtftgures a container based on an execution envimnment specified » the software deveiopmesit.task. For example, block t can emtsht of retrieving a container Image (e,g,. docker image) and executing cp spondin docker in wettonSj installing packages-identified in the execution envimnment, -and so forth. In s ne:aspects* at block 304, the computing system accesses a code repository (s.g,* code repository 254) associated with the software development task (s.g,, a GiiHuh repository), hich can be cither public or private^ The competing system can clone the code repository into the container,| 7| In some aspects, as part of block MM, the M L model can be inserted into the execution en irtmmem, and eotrespondlng: libraries, scripts, and so forth for executing the ML model can- also be installed. For example, a startup script is installed to prompt the ML model to cause the Ml no’del to perforin1&0 I Alter the exaention environment of the container is configured, -at block 306, ihe iam piJtui” x cm is cmifi ued io irt lire muUmes vid vomrol oriwofo oecv^s wilhm thv container, 1 or example, a virimrl network interface within the compuimgsystem, that the eenta eraccesses can be disabled. In one example, tin? startup script fo the ML model c iftcl to a fund km to disable the network imerfaee before the ML model is initiated.in some aspects, block 3()b is associated with a docker coufigtirttimn and startup Wnfigumtidte Far example, a decker configuration phase (e.g., built ton the ttockerfite) can be static and eaehed locally within a eontetae service. In some cases, the docket image isstatic and is bruit with a standard versim? of a Linux distrifautot (e,g:>, Alpine, ete<) with instrumentation associated with the service provider. For example, a static d cker image is built to include a service.-to interface with the Al model (e.g., service 242), in other cases, the docker configuration may include building a docket image awd on dilferent environment. l'o example, different docker images can be used to develop or test software within difte em operating systems, Customizat m of' ths docker images can have security impacts and may be enabled based on heightened jutromen For example, the container may have an external network disabled at the network switch or isolated m a virtual network to prevent inadvertent maltekms actions. In » e aspects, nct’ Of. cont tmj ii ni thecudtaitwc-ai* be xvojx- to »« limited wt cl d«miain names io develop particular tetesres (e.g,, an API associated with the soft warndevelopment task},|MWJ Block Mb can also include a startup eonfiguratiou, which is c Listen, i ring of the execution environment itself For example, the startop conftguratiou can in uhe instrucrimw to Install packages, code tools, services, unit tests, mock data, packages, arid other Information, The startup configuration can also execute functions (e,g mmes.ue a seif-signed SSL certificate, etc.) to prepare the local cm. tronnseni wrthm the cotikm?ct. In thw eonliguratoi, die wntoicr hosts a fully provisioned development environment that is -aceo b m art Al model through a defined API t ierfW. The startup nfigu ion of the eoutahw can install teugttege runtime® (s. g<, Python, ode w>- uitemoners. compilers debuggeis, and integratd toolchains fhat are required to rntote or mpheto det eloper wwlfiw The startup omtotratmn cun also Cinifunre API calls to create new- fife, update uree code, Inject test cases, execute scripts, trigger ( I -style bidkl processes, or: analyse logs p duedd diking: texecutiom For example, the Al model may need to dynamically mtoii additional Python: packages, base -on acodeb T import graph, or etol ure s test datebase before running integration rests. After installation of the various aspects, the container may disable or scope network access (e:.g.,:to specific domain names) during: thestartup contigumiitoThe Al model can use various tools to ctort-e, evtoito and tefine code quality tht i mgh p < »ytai iw i te orating and validation. For exam ple, the Al model generates input scenarios, mm.t s code sto crc. simulates user ime.metk.tos. a; -ft executes periormance profiling roubnes.t2176 A? te. FtelWt;HThis may involve instrumenting n ing services, ea twiog: traces, and comparing outputs against expected baselines. The stamp cottfigufetion can also load telemetry hpofc A d synthetic date generators thte allow the ML to test edge eoridi ns and rare states withom real user d tea. The startup eooSguratioa can eonffem mod APIs, emulate network fellu, or simulate U'ABditvuciBx ixcito like wehhnok c;dlhas.ks m rtu,sx,g;e queue delays His- intecy in h>u of the coni&mer ed execution s«d ML tori ven whretratM can cause fee system to autonomous ly perform: a wide range of Software engmueriiig, and swirly assurance- I ks to feral Me the dcvclopmcte ol appl tealMs,|0»21 As block 30X. fee computing system causes the instmeti n assoc steed with the softwareconcsponding parameters and objects of the Al model are loaded mto memory, fee Al model receives fee steueftna to perform fee software development task.At block 311), the computing system Is eoni 1 gered to receive an, indication of Inference completion by fee Al fe l, Fur example, the container can teeiude a controlled signaling mechanism to indicate to the outside omltestraiion system (e.g., patch engine 224) that the Al inmk t* tirtsslkd with fee Mdi.tt fe' Jopmeid txsk, to one evunplc, uhers the M tumid is finished, the startup senpt can configure the cwtei ra process to exit wife a known status code (,g,, 0 for stKX'us't The cuu pteiog system may Include -an outside orchestration system thui v««tehv for changes to the c iiuine Maui1and nfenfiinw when fee or Kmra succe^fedv exits w ife n xt vevdfe Mains code In > mw aspects, the miptitmg system van stop p cc’-s feto m the avenl feat the wtas code Indicates Mure.M 4| At block 312.: the computing system- is esrafigured to extract fee changes made by the Al model, For uxampk-, when the cufeids orchesMion system identifies a suceerafiil ^ohtame exitsthe oatsfe <»i chc m- «»>n system -can then mount a fil e system of fee.eontaj "to read the modified contents (e.g,, based on the cdhtainer installing fee requisite services to -allow the moHuling operation). In some aspects, the container is treated as an unsafe environment to prevent any malicious actions from occitrnug wtfeta fee: host system or pn olgattug alto S actions to a code repository. The computing system can store fee changes to fee modified contents of fee code repository in a datebese (e,g,5dteabato 225),W5 in ''inne C'.pvck -d-cr-i det, <. so \tc« the dusr.et u> the te tc o to-x lot example, the c Ifent device can be presented with a diff interface feat illustrates line-by-feto Changes to th sun it, fifes tu the code repository, l l«s fell' inte'rdice presents iisera with a huwhsdtau view of the change* generated by the model and allows the user to review edits before i egration.Xt’n^’.o re -ow it Since the changes to fo repository are produced in an isolated container ( g« c tai er24Q) and t ^fmed through n see, detort isti proem fo.g,, fey mounting container 240 and g f tog diffb), users can vslldate that the modifications reflset the iutsnded changes without any risk of hidden side effects or unsmhorixed code execution. This review layer ensures transparency and control while m mhureiig the ssreiiiHv ixilmderv he wees? the 'W'uei atvd. ode an the repo toix.At block 314, tir e computing system canreceivea request to generate a change proposal associated with the code repository (e.g.scods reposiftuy 254) and may generate a corresponding patch file wi th all content For example, the change proposal can include a n m fest associated with file -dnwterc to' the repository to g., sdenfift file changes). a comran identify inc a free and a parnt associated: with the change proposal, and a diff identif nui at fe.ro one text change in at feast rare feeAt block 31, the computing system provides the change proposal to the repository. For example, to response from sc I tent device, the computing system generates die change proposal (e.g,, a PR and pushes the changes to the repository. In the case the repository is a private git server, the ce pub g system pushes the change proposal t uig an API ef the git server 1 st s»»mc cascvjhccli.ni ecan he uxhed l g puhlu wrv ice -nth ax < II I I H R huithw c sex, rhec spiifoig system can al so provide the change proposal to the user, and the user can co ppmte the change proposal tistog, standard l J commands (git am iiewh'eetui-e-. aicli) to apply the patch with a commit In seine eases, the dewlopmem. enviroaimeni ma else include graphical user mtorfeces for apply mg fte patch with a ctortorit hi this way, the user is in control of the rey towing aifo patching process, and patchescan be relied back.JM FIGS. 4A-4B ilfe straw vartous screens and conceptual diagrams df an automated software devefopment system using a generative response engine in accordance w ith some aspects of tire disclosure,0991 F1O, 4A iltetrstes a user interface displayed by a client device after the automated software development system comptetos a task. For example, FIG. 4A illustrates implementing a Jfem ipphc m with reconnection topic for a server sent event (SSE) service by a server. The n^er interface includes a summary 402 that is extracted from the data generated b the Al model ih l i k-afom$ vjurfouv cKdwns ihat occurred during die tosf meludmg xecuiimi the premier iv e, Hide fiinira ing airl s ypfocrpft franspiter ftw) to validate (he changes. The user interfile ahn ii hutes d f iritedacc 40-4 to present the changes to the user in a jnanner that allows the use? hi eusil validate the changes.4765-S fo. F' 851 W!|01M FIG, 48 illHStrates ths user forertoce options to generate change proposals based on change rifo m’c,i m Hb 4 I? thw ccw, a hn.ttm 4Uf* can be smoked l>* ibc^*’sx dtobrerr opnom to.generate a change proposal (e.., a PR, a git tsppifo nr a manual patch). In some aspects, the PR a celbborative foarnte commonly used » platfrms lake GITHUB or GITLAB to allow < elope* s to propose choages dwton© branch to ennther white enabling code review, cemmems. Cl integration, and approval wnridlnws. The FR includes lull eomrnit metadata and Is ideal for tesm-h^cd dcsetopitKnt. 1 he git apply is ctawnand-linc tool used to apply p-^toh flLs that arc tj tk' tbs ecucrated v<itn eit dtfi twtb to dtc working tree- without commitiijig, The git apply option to bi’hiwesghi and uwlul for sharing changes H ithout usin a motor repo-mory but hek, commit history and metadata rmteto manualh added. \ manual patch the generic ddl tools hkc dt ft' and patch u> apply changes,md requires cm tol handling beertuse it doesn't track toposlter: state or resolve conflicts automatically.i n In some aspects, the PR mmwtos a human re iew stop in the development v kftow, which is lmp*-rtate wheu chmiuex cnx gciicmKd automatically. Regiilthw approval fto a developer w tcMc ver before stooging ensures* d»ar code is examined for HiHctnme! eorrectoess, style, a dtitoctotofl consistency,: and potential seeuriiy « performance issues, This Itomamin- toe¬ loop model acts as a safeguard a usftst unintended msexpsences and l ^ team to matotaiu high w e gttoliiy whe» using adhine learning systems, or auio atfon u th to gestemte nr odily code. The steuetored review pmcess provides transparency, at wn bili y, and tot eppdrfomiy for before w <vh gs.s aw I’ltogt sted mto toe *'.ms ^. debav'ttes t puma!} user tnteifocs1ol the automated stoi re development system and incteite\;at option to gtototote mto ipte v s in aoomtaito t sm e aspects < the tos osu to some aspect, the user interface includes a text input 41 u fr tecew in«> a natural fengpags task from a user. I he text inpul 410 bv includes a selectable item 412 or other user c-cmrrto for selecting -an nssoebted repository and a btweh 414 of that repository to.g«, main). The mmmd language ntsk can be eluted m ditlerem modes based t»n a button input, such as the code buttes and w ask button.|0103 The primary user into -we mchtdes a ttser totetf e 416 fo seiecting a number of variants that the A l model can gen te. For ettafoplb, activating the user interface 41 x jspfov. u menu for allowing a number of variants to: try bused on the software dtovhmm<.m t sk In s»w cases, the user may want to the Al model to genemle'diiTereni va n I'm esamjfo drilucm iy w> of bmps, (e.g., a for 1m«p vetoes a while knqrited mine the mori stnbfote option. In some cases, the Al model can also try out differed libraries to allow a user to qatekly c pam dlSsmntw iauht to kletiiify a mfeired solution. to this wty, ths automated software development system c in identify fotkrem XOIIUBIB md a preferred ^UBPOB 'eil <w, due w er roterl.sce 4 i|01(M FIG. 4D lllustmtes a use interface associated with reviewing an executed task and wJa ei a t.;b interface 42b hu Sw’tru- ri.c different H.mre hu ewtemfe. the ilfo’Ur.ireJ example shews orttet ntelng a model context protocol (MC. P): service. There are -different types of MC services, such ax FastMC, Python. MCP 'se r, O bridge, etc; The Al model: can generate flte different variants that perform similar nctionality, and the user deploy the different: version with change proposals.[0105| FIG. 5 is.mother conceptual block diagram of an automated software development system 5ftft using a generativeresponse engine inaccordance with someaspects of the dtsclosmte.. Automated software development system 500 fe substantially similar to the automated software development system 200 and repetitivedescriptions to allow explanation of different featums.In-some aspects, automated software development system SOO' incht s cpm mentx of a generative response engine such as a front end (not shwn), back end 520. and Al model 530 (e.g,, s traustormsr-batsid model, etc.), an use eonh er 540 to internet with the Al model vis tod tmtion enghre 522 wd agwt. 542.|»7 Automated.software development system SOO also includes patch container 560, In some aspects, gatch container 560 Includessimilar features -to container 5411 -but ix.eonligured to perform additional softw-are development tasks. For example, pateh container 560 may be eonforered to autonomously deptoy changes to a PR via network 550, For example, pateh engine 524 of back end 520 can include instruel ns to launch p teh container 500 to perform additional deployment fimetk.( Fateh contame 560 serves as a secate^ isolated execuitoiiettvitomneht «sed to apply in K'htne win m d cede Jnmgtw bi a deu-nnssfetse and c-mtroUed mannet I lie pilch esmu cr 560 is not associated with Al model 530 and ves data generated by the Al model 530 te.g., from she database 525), The patchcontainer 360 initiates theenvironment,retrieves a -clean version of code mpo& tory 234, generates the change proposals, and chit merge the change proposal topatch eosshdmr 5( * casi perform terimg and additional post-patch validation functions. For example, because patch container 560 has access to the updated eodebas^ patch container 560 esn run:the same kinds of tests -and tools a developer might invoke total ly or in Cl without exposing that execution; context to the Al model to;aitow.an additional validation of changes beftw pushing them to a remote repository or teiggerlng:if WJ ®., F t s twos deployment or example, after the patch is applied m patch container 560:tpatch container 56-0 can execute unit tests, perform static dpalys ur linters, verify htfild, and perform security scans (e,g,, with dependency cheeks with tools like np au it or pip-audit). Because patch container- 56-0 rans additional testing in an isolated and reproducible manner, these actions are safe and can be auditefe.|01 W| The patch container 560als supports patching binary Hies;,, which is not easl ly handled through conventional cxt«b-ased patching techniques. For example. changes in binary assets (e,g.scompiled rtites. media files, or serial if ed ata) are not human-readable ami do not. have a? structured tore-h - re format like source code and standard diff took ronne-t meaningfully inteipret changes inside binary rtnaK 1 re example, Gi oiten just repsirts “binary files toffa m binary output without show mg the tracked change In this usy, the patch containerprevents retrieving the iulh original arid ntotlsfted binary bfobs, either from GITHU or a secure object store, and van tne the git cnmm'tn -lhje itocrirce to recotolnict the conwut w ilh accurate binary wment ulhtw- ittg afchiua woiklkws » to inducexmge update*, *.onipi d binaries r tohci non¬ text rc-juu e’- without tvqmnng custom drtt higie or risking data h»s\ due to hwn i m i inatehes,Pateh co abte 560 is ate stmfigittod to totthetoicM® with remote Git services to Commit awl push the updated code. This requires access to user^speciffe secrets, such as G1THUB pc -unal. K VJ tokens H’A i,-d S H keys m OAuto h'kero, rk-penrimg on ilu configured method of mOhenttcafion, In this case, toe users cat! e nfi um their secrets securely totoughthe front etid (, iron! etui 101 wlwre ioktiw '-tie eiuiApicd tml stored tn a hwkmto vault When patch container 560 is bunched. the dcnitais are injected into patch container 560 at runtime ing spctuo e wwd va s o 'mounted w:tahea$ ottering patch po ai ^6Qcm totteto to CITI U • B and create branches, commits, er pud requests as needed w ithoirt esjwst g credentials to is B trusted components.(6111'1 Patch com iuer 560 may also support automated deplo mem hooks to push changes direeih to the fergei: branch Or open a pull reqnest depending en user crihfigumtioh. For continuous intogtoltoitteom sutoiss developrnerst (CI ) enabled repositories, the automated deployment hoefe enable seamless imegmtlou with testing pi pel toes, deployment woridlo^s, or preview envimmuents. Automated software developmeiv s em *»<» can be cent; eared tu utouuitfctoly applj pxitehu- to tealurc hranclws, tag rdeww* or trqywr Cd HIl R \ch - alter ptwhin, Smee this process is handled entirely within the patch cotitaitier, which has access ip the user's secrets and operate todepend toj f the Al model, it aligns with enterprise security m^ulmmcnts and totolte iMidited, logged, or sctslnae independently.y51?4765-S " ta P S51 W!|0H2| FIG, 6 illustrates a fitrevcltart: of a process for gesterath g change proposals using: u generative response engine in occordsoce with some aspects of the dreeloxure. Ptoe 600 (ar a y other described herein) can he perforated by a computing device (pr apparatus) or a compmmt (e,g., one or sw chipsets, a SoC, one or mere processors such as one or more CPUs, SPtis, BSPs, PDs, NSPs, ownKJonwIle^ ASICs, PPG As, programmable fogfo devices, discrete gates or transistor logic somponente, discrete hardware components, ere,, an ML system such asanentai network model, shy combination thereof, and / or other component or system) of the 'compaiiag de sec, I he operations ol pjueess bW nmj he implemented as wiirear wmporans dan are executed and ran on oho or mure processors fo.g,, CPU, GPU, DSP, NPU or maod engine, SoC, processor PM of FIG, fo a d / or other ocess r(s)).J01 ) In some aspects, the computing system may receive the instructions^ to perfomt a soilware dev cis (parent t,i*k ral spawn a e< inta er h-Hed »® tm execution envtronmem identified by a br configuration, J c container procere e,sn mdude a: startup configaration which can include tcffowlrm pai mwre 'wsuciaU'd nhh a dexch’pnreta environment evxuJated wish a i e eritery (c,g„ emte rqramu' 2541, Fur example. the rept sitory may he specified by a sser (e,g,, a GITHUB repo).|$H4| After the startup eon gnration is complete, t s computing system may disable n network interface associated with: the container before tire machine learning model b provided the software development task. Forexample* this may disable a virtual network interface for external network access, but a fretwork mterfitee t r internal network (e.g;,, a private network) ray erafo c: trebled to al tow op ttoris with the ceramet.|011 Sj At block 002, the cofnpuling system is configured t provide, to a machine leading model, mslnictiuns: to perforfo a software: fetefopHrent task in natural language based on a ro xroten I ta? uun, hnie tea mug model sx coullgtired to a«.x"-o ike coutarar Ok • the d.sabbng o? the network interface (tfofo via the private network) and my send contman s, access a terminal, interface with an API, and so forth. The various mterfirees can be configmed with docker image hwumented with specific features te,g.. service 2421 that the machMie teaming model {c.g„ Al minted *o nxcsx(CH 6| In seme; aspects, the imeh e learning model uses the mstruetfesw and the container and perforra a tfomuted ds vefopmeni taskte sttefi » modifying fifes, steatmg new filessoxee dbiig imp terts. and other ope ran on x that a dtexdopet woeM peitonn I he m'tehra fearosng nuxki re a igrad to generate an output to an associated component (e, „ orchesttafion engine 222} tn iiidteate that the Uevdopmedt task is completed,2ht:te" fe fe.ft 85 iWO!(0117| At blockthe compote'jt " '-w' is configured to obtain a dfif ftom the machine ■ teaming model based on the identification " Of completion of the.software development task. In some aspects, the dift' is a change made by the machine learning model to the rep tery. For example, the diff'etm be a new fife. a removed, fife, and a change to the file.(0 I& I At block.606, the computing system b configured to- generate a change proposal for the topository baaed on the diff, wherein the diff emnp es at least one change to a text file associated unh she repository I exampie. a nont end ic g front end 210* mat display the ditT l a summary of the changes generated by the machine teaming model for human review. The front end may include various options to allow the user to generate a change p posai, such as s bisttem, ta respotbe to input at the from end togenerate thechange proposal, d emnptsting system may gepe to the change ptopoaai by accessing th« data fiw reference:m a database (e,, database 225 g In some cases, foe cferege proposal includes a manifest aa.wfeted with the file •> rue sure of the repoviton. a commit idrenily y a ime and a parnt assuented with foe chaum. prop^ul sis the dlff identifying at least one text change:W at least one file. The change ptopcbal includes a nreinfcA av'atesated with the file sterature of the ret»sfitoy, a commit idre lifyiag a tree and u parent axsod uxl u 1th the change proposal, and the dfi kl tfyfog at least one text change in at teas! tree. eft.(0110 The computing system prm tees, tv the icte device, the change pogfestiL At the client: devices, the user mny exrt’ute the change proposal to promfogate the dlbuge^ identified in the change -p opesttl to the repewi toty;(01 0| to some oases, the e rnptrtmg system may also geagfete a patch eontofoen F r example, (01211 After -g^ae tlag the change proposal, the imputing system may spawn a second coniomer associated with the repository, clone the repository in the second container (e-g., the same repository that the development container (ere:., cento i ner 240).retrieved); and generate the change request. to the second cemtatoe The second container may also apply change request to the repository based res an instruction from the efient device. Sa tore example, the change proposal whites at feast one of a -file add, fiferemove, nd a binary file update.( 122 IG. 7 is a block diagratis illustrating an example machine- teaming platform for impie e tog various -aspects of this disclosure in aeeordacee with some aspects of tir prsent tedm mgw AiWugh the example s-Astem deplete par- scufer system cmnp«men sad,m arrangement of such components, this depiction is to facilitate a discussion of the present20t2^"65-h7W3. F85 wirt ieohnulogy and should not be considered fimitmg unless specified fe the appended claims. For example, some components that are Ilhi rtoed as separate can be combined with other component$fand some components can b: dh -fei to separate components.|(if£3 -System W "may inchi data input engine 7W thatcan further include d retrieval engine 712 and data tmnsfonn engine 714. Dsns retrieval engine " H may be configured to access, interpret, request, or eivu data, which may be adjusted, reformatted, or changed (e.g„ to be interpretable by another etigintu s h as data inputengine 71 fe For example, data retrieval -engine.-712 may request data from a remote source using an AH. Data input engine 730 maybe cenfigated to access, mtcrprto, request, format, re-formal, or ceke input dads fem data ^urecsM " OL for example, data input.engine 71fi may be configured to use ata.tmoaibm engine 714 to -execute a tv-cun figuration or other change to data, such ax a data dimension reduction. In some aspects, data Uiwn 7PJ may be assoc isted with: a. single entity (e.g,, organ ization) or with aifiple eutiiies. D;da som-vcshi ’’H. include unc OE nuue u Gaining, data 'tfe,. input dute to iced >. m chfe k udrm model adata ",!12b sc g„ daks apjmst which to feto ttoc prow^or mssy compare litodyl rartpui ith, j>uch as m determine- model ou wi quality ), and / or reference. data 702m In: soma aspects, data input engine 7. I can be u pLutotiied using at kust one cump tm d KC I mwu fe c,m be tosuined through one o in« I O Jso ices and / or uetwmk interfaces. Further, the data may be stored g-g., during execution of<me or nwe operations) in suitable storage or system memory; Data input engine 710 may also be configured to "interact with a data storage, which may be impteme-nted on computing dfetx that stotes data ia storage or aystem memoryJBl j System 7fi0 may inelqde featwtssmtm engine 720- Fsamrfeatfon engine 720 may include feature- xnnCftatfog & labeiiug engine "22 (e,g.. cunSgured to nnnmate or label features fem a mode! or data, which m s 1 extracted b< ie um e rtr iRUt eugme '24 v knUim extr toum eugme 4 te-g.< configured to extract one or more feadsms fem a tmnicl or da. und or fetomx; scaling & sefcefeu cughte 726 Feuttfe scaling & selection engine 726 may be configured to detormine, select, limit, eonsWin, cUhea enate» or define features (e.g., Al fetu ss) for us© with A 1 models,|0O5| System 706 may afe include Ml irmdchne Engine 7, which may be configured to execute one or more operations on a machine femmg model feg,, model, training;, model re- configumnon, model validation, model testing), such as those described in the processesdescribed he em For ©xumpfe ML modcitog suigmc 75d toas execute,m opctution to train a machine l arattig model, such as adding, reimn m. or nitrifying a tnndel parameter. Training of a muchutoto t woi foaming model may he *>»^x’rvis«, seraLsnpervised, or o -upe se. tn some aspects, training of a machine learning mo< mty mdfofe ftndtiple epochs, ar passes of data (fog-, training data 702a) through a machine learning model precess (esa training precess), In some aspects, ill nt epochs may have different degrees of supervision (e,g,„ supervised, semUsupe sed, nr uusupei sed). Data into a model to train the model may include input data (e.g„ as described abmto and m data prtwim y output Iron a model (e g, framimg a to-cu he kammg feedback u A model parameter may include Me or o of a seed value, a model node, s odel layer, an aleoothm. a fonMom a model donnget ri (e.g Hwecp oilier model ppramators or between models) a mu s-1 e<m$trataL dr any other digital component mt enc g the output of a model x model esmneefton may include or represent a relaiton*htp between mtxiel parameters and or models, which may be epehslenf m hifefocpendent:, hfora h al, andAsr static or dynamic. The combination and cunllguratioft of the model parameters and relatwnships het'® ecu model parameters dimmed herein are cvyrativdy infeasibte for the human mind fo ifoafotomut hmrtiny tbe djscbn.ed aspects m ans WJ), a marhtnc IcanttM model max indude millions, billions, or even triil w of modd parameters. ML modeling: engine 730 stay include model selector engine 732 (e<g., cnnHgwed i select a model from among a pl rali iy of models, such as bssed on input data), parameter engine 734 (e.g,, configwed to add, remove, andfor change one or utoto parameter of a wdel) endfor model gettofedon eng p 736 (e,g., configured to generate or more machine learning models, swsh as according to model input data, mMG output data, comparison data, and / or validation daiaLIn me aspects, model >clecn«i engine?2 may be cimfigured receive ptti,andfor transmit output to ML. afo»n)hms dafoha ' 7U Sm.»ter'y, feaiurtzatipn engine 720 can utilize storage or system mctwry for storing: data and can uub / t m or more l. U JCMCOS or ne v fe im ics to? iransm juv nr totes v nig data Ml a g<*nthins databus 7M in v stme wio or utorc machine learning mudels, any of which may be hilly usbted, partially trained, w untrained< A machine few? tug: model may be or include, without hnmauon one or more i {e.g.. Mkh as tn the (v ise to a rotoumMJ t ri Mtisisc d model, an algonthm, a cental nclwurf AM. a en htuortal neural network (CNN), a generative neural network it AM, a Word2Vcc mMel bag of words model, ® term frequency-towi'se document ir queitoy {if-ldf model, a GPT (Generative Pre- trained I rami meu raMel (o? other auusreg ssixe model? a dilltbm model, i di wtois- transformer model, an encoder such as BERT (Bfo Mfoital Encoder Rep entations fo m Transformers) or LXMERT (Learning Cross- Moda ity Encoder Representations from I ran tri to. a I'raxmito Itobas uptimi / atto'i U’l’A mode, a nearer swMiMr model te g. I nearest neighbor model), a linear regression model, a forneans clustering model, a Q-Leanilngmodel, a Temporal Different (T ) model, a Deep Adreraa l Network model or y other type of model described further b in.-Some of the ML algorithms in ML algorithms database 770 can be considered generative response engines. -Generative response engines are those models me commonly referred to as Generative AL end that can receive: an input prompt and generate adfi k al content Wed r the prompt GPT$, tftu$ion models, anddtCftmtniutruttsfocTner ’-models me some mtohmrau, examples of generatke response engmes. Some spemtle csampfes of generative response vophws fhst turn be stored n the Mi sd gnrithtos database 770 include versions PALL L U \ Hin. and S A, all provided by OI’LN AL27 System 700 ess further include predictive output eneralfen engine 745 Md output vafeiatten engine 750 (e,g.sconfigured to apply-veli ation data to machine Wmingmodel ompufe 'Predictive output gcn t m engine 745,can analyze the inputand identify relevant patterns:end associations in the data it has learned to gehdrate a se^isenee of words that predictive output imerahim ci guie ’’ p dreS re the m< hkeiy evnlmuatom of Ira mput usiri* one or more model, Irom the Ml aWulhm-, dtoa e 7~U. among to pna 'dc a coherent mJ witc Wly jvk'S Jttt ans es Predreunc >smput i’e.u.v>it!o.', c 'gnw ~t> generated s '- oures h> ramphsig from the roixtodky dhtd'tostom i<f pr-relbfe words and wainc rere guided by Ora pastures Wcra d during th tnnmmc its senra aspects, predictive output gciteiatfen cu u ’’41 van gemaato multiple mssibfe responses - before presenting the final rare, Pted ietW output generation engine745 east generate multiple responses teed mi d input. and there spsrerew arc variations that predictive Output generation engine 74 esmfidera potent Lily rek' ant and coherent, Output relidattra engine 750 egn evaWd (too generated resptmrea based on certain criteria. These criteria cast i udp reivxancc to the prompt, coherence, riuemey. and wttWEW\ adherence m specific tnddhiC'c or rules, dependi ng on the appheuUnn. Based on this evaluation, output validation engine 75 se lects the t'ttosi appropriate response. This sefeditm is typically the one that scores highest on the set: eriwhg bidafretug W’ors hhe relevance, mfermofi eness,and cohererice.|01 | System 7W mm fertlter include feedltock engine 760 e*g„ conUgu dd to apply:feedback front a user and or machine to a:model) and model refinement oagiac 753 fe,g„ mlfen ed to update r re-con Injure a tnodt In some aspects toedbacl engine "kto mav receive input anfew irmtratu output te g„ output from a tramed, parJally trained, or tsritrained ttm-Jel) to ssutcorae metrics fetwhyse 765. Outcome metric <fetaba.se 765 may be configured m store utopitt from one or more modelsand may also be configured to ass oc e output wi th one or more models, In some aspects, outoome metrics database 765, er other device fe,g>, model refinement engine 755 or feedback ngi 7601 may be configured to eorreiate output, detect trends In mttpm data, and / or infer a change to input or nwfel paramerem to cause a parm ulnr model output or type ofmodel output In some aspects, model refinement wgiue 755 may receive output fem predictive output generation reure 74$ or output validation engine 750. In some aspects, model refinement engine 755 may transmit the received output to foa urira ies:? engine 721) or ML modeling engine 730 m one or more Iterative cycles.|6129 lhe e gTW *>f -y-e’em 7<jti are be twv mod fo cm sl hi duarc urw demmed lor use with other comprments or a par of a program that performs -a parfetoar hmet m (e.g., of related 1 h net tons). Any or each of these modules may be smpk meute using a computing device. In some aspects die h e ouabty of system 700 ma be split across multiple computing devices to allow- tor diunbmed processing of the data, which mat improve- output speed and reduce compittaiimial toad <m md tdu il devices 1st some aspects, system 700 spay use load -balarsc Ing tu raaiatatn stable resource load (e,g.sprocessing load, memory l d, ft bandwidth load) across multiple computing devices and to reduce the risk of a eompnilng device or connection becoming: overloaded. in these or other aspects. the different components may co nmitieate over one or more l-;O devices mxl-'or network hrterihew.1013ft) System 700 can be related te different do ams or fields of use* DescriptioBs of aspects related o sp-eetfic domains, such as natural language processing or language modeling, is not intended to limit the disclosed aspects to those specific -domains, and aspects consistent with the present disclosur. can apply to any domain that utilizes predictive: modeling based on available1dsns.accordance with some aspects of the present fehrmtogy. Examples of ML models that use a tmusfertoer neural network "(e<-.* tra sto mer orehfeeture Sfill can include, e.g., generative pretmine transformer (GPT)ISi ifeciional Irneoder Represenhriions fromTtan toOTer (BERT) mo<k-is. The mw-tenuer whitcc mu 800, which is illustrated in FIG. SA, FIG. 813, and FIG A\ includes inpu’s 802, input embedding bidek 804, pdsitionni encodings 800, eauoder &08 including encode btocto S 10, decoder S 12 including decode blocks 814, linear block 8 T6, softmas bto-AxI M, arid Output probability 820.|ftl 32| Input embedding block 804 is used to provide mpresen aifoits "for words. For example, embedding w.m be used in: text analysis. According to certain non -I. uniting examples, the reprewntm!<.•••» i « rfe- toed vector that ensixtes the meaning of the word in such a way that wrexb that m. ckwer <>’ the WGW spew a.'e expt- ied to be similar in owm^g V, to embeddings can be obtained using language modeling and feature; learning techniques, w here words or hw-c*. from the vocabulary are mapped to vectors of real numbers. According to certain rrofeimhingexamples, the input erabeddiitg block §0 can be teamed embeddings to- convert the hpst tokens ate output tokaw to ’.ix'Jots of fonw.m that V\x the sw.' IVM. M.t; the po\ u. te encodings, for example.|(il33|about dfo relative or absolute posttidn of foe tokens fo the sequent; conl ^ to e.-toma non-limiting examples portions’ cn?te‘rp« Mb, can be provided by adding positional encodings to the:input embeddings at ite inputs to the encoder 808 and -decoder 812. The poriltotifo encodings have the sains dimension as the.-embeddings, thereby enabling a summing of the embeddings with the positional encodings. There are several ways to realise the peritw! eneod gs. including leaned and fixed, ter example, fona ate eosine functions having UH let ent frequencies can be used. That is, each dimem n of the positional encoding corresponds to- a sinusoid. Other techniques of conveying positional infortnatinn can ate be used, as would be understood by a pPri u of ordinary skill in the art. For example. kr.rnte e-ritoate cmbcfo m.’s c n uiato.nl be med to obtain Motes’- ratert. te si fom-to^ ofos-.foa sinusoidal pu nt> >nal ox i ulingx radu-j than learned p.»ui«'ml citcodirsgs fo that dotitg ilh.vx J> the tnted tu cdiapfoaic t<> -<c<{iteuee teufote longer than the uncs ic unUrte during training.Encoder W can use stacked self-attention and puinhwlse, folly connected layers. Encoder 808 can be a stock: of identical tuyere (e,g., ■■■ d, ate each layer cart be an encode block, as si hast rated by encode; block 810 shown in FlO. 8B. Each encode block 810 has- tuo sub teiO < hte subhm u multi te altoidmu ute-k 822 ate tute wiffil sub Li?er fort a feed forward block 82b» white can be a pusfoomwrae folly connected toed^torward network. I he toed torv td block fo cma U? a mvtfote lit i ar nd t Rcl I }-101 | Encoder 80S usca a rcstteal ertmtocten are te each of the- hyp suMayers* followed by aat add & twin bto f -k which performs normaiiztiion. For example, the output of each -layer can be yer\nnn( Ssblaytex). To iaeilitsito these' residual connections, nil sublayers in the motto!, as -well ite ernbcd inglayers^producp output data having a same dimension.IH3& I Similar to encoder 8 8, decoder 812 uses slacked §elf*atention and poin se, fitily etomebte layers. I codcr 812 e-an ate be a slack ofM Id foal hy^-s? M;« 6), te each layer van be a debtee htote ax illustrated by decoder 81? tevmn hi 1 1<3. SB. In addllfoit to thy tv.<« tetem-i i rn.. the thknet uith mitllite te at ntert block 822 and the tes-tecs fot kte forward block 826) imtte m encode block 816. decode bh<ek 14 %. focltee a third suMsycr, which performs multi -teas I at umiion o ver thU tput of the encoder stock. Similar to encoder 808 decoder 812 ttses residual co settoPs atound each of the ste ayers, followed by layernomtalixation. Additionally, the scb-h er with nndt bead ttsniion n <« L ’•?’' ^an "be nwdifted in the deader ■.l u.m h’ pre’,. re m> MU nw h® 2L’U'»ding * ' Mrwc. cnl pcrooon - I hs-* unking, combined with the fact that the output be dings -are offset:by w position can ensure that ' the predietfom for position i can depend only ® the known output data at positions less than t.|(H17| Linear block gW ca be a teamed linear transfortnatisn, Far example, when transformer architecture SOT is bein g ured fo translate from a first language into a s md.o- g age, linear block §16 can project the outp t item the last decode soflm x block fold into word scores for the second language fo,g„ a score value for each unique word in the target / vneabuter lat each position the sentence. F r instance, if the output sentence hss seven words and the provided vocabulary for the second language has 10, WO unique words, then lOfofo) score values am generated for each of those seven mmrds. The score values Indicate the likelihood of urterte frit each word in the vocabulary in that position of the sentence.Suft ax M ek fol S then tarns the scores from linear bfoek 816 into output prohabfotiex 20 (which add up fo 1,0), In each position, the index, provides for the word with the highest prebahiliiy, anti then maps that index ta the correspond fog word in the vocabulary. Those words Siren fotm the output w trence of to tHrtM au'fotechirc MlO I Ire sottnux oprmhnn re u plred to the output ifom linear block b to convert the aw nuffl ts into output probabilities S2t) (fog,, token probabilities),|ftl39| FIG. 9 shows an example of computing system 900, which ears be, for example, any mpfoitg device making up any engine illustrated is FIG. I or any cmnp em thereof.|O. I4OJ hi Wte U^pcet rapulam yy em!HHi W a iigk» det ice trn n rshuk <^ e tn which th© fonctfo s described in this -diseto m can be distributed within s datacenter, multiple, data centers, a pesr network, etc, In some aspects, one or mure of the described system components, represents many stteh components each perfomtlng some or all of the fonnilon for which lire component is described. In some aspects, the ctw x ents can be physical or virtual devices, 411 In some aspects?, computing system 900 may comprise one or more comparing resourees pmvisfoned i ra a “dead computing” provider, for example, A. A7( ' Fl XSUC rn PI I F CU M n t"4\SA / O\ fo'Tfo prm uW by A / OS, I ‘ of Se,ifo X vIimghm. St N n ot "i C ra\lFl U R t i n t n, powioed by SI \ MH'ko 't S H. Mb, INC, l Sanm Clara, California; AZURri, provided b MICROSOFT CORPORATION of Redmond, Washington, G0OGLB CLGUO PLATFORM, provided by ALPHABET, INC. Momwm View C ifomia, tmd the like.t2<t"6 Aftmy. Ptotw0i|0142| Example cmwptrtiug system 900 includes at feast OTE processing unit fCPl’ or processor) 904 and ephneetfen 902 that cowpies various system components tncludfeg system memory 900. such as read-only memory (ROM) 910 and random access memory (RAM) 2 to processor 904, Memory 90S w be a volatile or nen-v afile memory device and ©an be a hard disk or other types of ttot»trenaitte ©emputer readable media which con store data that ure acce§ai ie by a cbmpbtefj. such as magnetic casaete^, Ras. memmy cards,. solid state memory devices, digital erasfite disks, cartridges, random access memories (RAMsJ, read-only memory (RUM), an w seme cemhmufton of these dcvsc-es.| 43| Memory 900 can Include software services, severe, logic, eto., that when thecode that defines such software is executed by the processor 904, it: causes the system to peribrm:-a.fim fofe In some aspectssa hardware service that performs a pariienlar fimetfon can include the software^ minporieni stored in a computer-readable "rhodium in e u ecifon with the necessary hardware eunfeonents. such as processor 904, connected 902, iroiput device 922, cto,, to carry out the ftmeii a.|0144 Computing system 900 can irreirete a eeehe of high-speed memory 90ft ©©mrected shs sfth ufth, us close prourmh to, or mtegsated as pert to procC’WH M,|0145| Cmtaect 902 can be a physical neeftm via a bus, or a direct ewwtiw fete processor W4, such as in a chipset atchi ture, Cenneeifou 902 can ata be a virtual connectfen, neiwerked rom-ton. Or fogfeat eomtectfon.Processor 9(M can inefede - iiy general putpnae processor and; a hardware service or software s ses stored in memory 90S, eowfigured to control processor 904 as weft as a special- jnape. processor where software instrectfo -are incorporated - Into th® actual processor design. Processor 904 may essentially be:a completely se1 «contefeed computing system, contefemg muftipie cores or processors, a bus, memory control er, cache etc. A twitrreore processor ay be s>ymmeir»c ot rosmmeftte. Processor 904 can be physfeift or x irtual10147 Tn enable user imeractfen, computing system 909 feetude tm input device 26, which CUR ropresent any number of Input mechamsms, such as a m icrophone for speech, a touefetensifive screen for gesture or graphical input, ke fesxrof, mouse, motion Input, speech, etc, Compmfog system 900 can also Include output device 922, which can be one er more of a number of output mechanisms knows to fttnse of skill the art. In some instances, multimodal system can enable a user to: provide multiple types of input / uutpui 'fo emmmmkaro mth com mftjg system 900,85 t WCH manage the user fojwfaRd system output There *s to restriction on operating on any particular hardware ngementi tout therefore the fete fetes® here may east ly be su hsti tuted for imptpvetl hardwar o -fi wm amagemmtts as they are developed.In seme aspects, putfog system 900 can refer to- a combination ©f a personal (.mmpnt s* devwe mtemteing wdh compone '. htwted m a data earner. wher h«uh foe erenpto device 'and the componerns m the date center. In such examples both the personal computing device and the -componeuta in the datacenter might have a processor, cache, memory, storage, ete,|M40 Fur clarity of explanation, in swne instances, foe present technology may be presented as including individual functional blocks including ifet te blocks comprising devices, -device components, steps or routines I n a method embodied th software, or cc b tofo of hardware and software.{M50 Any ©f the steps, o emtion&i- f etiens, « processes described herein may be performed or ‘im leme' ted fo combination of hardware and software services or service*or in combination with »ub.'i »hn tecs, I some rnnbodtments, a se e© can be software that resides in memory of & client device andfor one or more servers of a content management system: and perform one or nwte fi cifons when a prices w executes she soft w a re associated wites the service. In some emtotoiments, a service is a program or a collection of programs that carryout a apeei e foncik, to some -wibodimente, a service can bn considered a server- lite memor can be awn- transifory c ospu ter- rwufafo e med i uih.|01S i J to aome embodimems, the.eom^te^readable s age devices, medto s, and memories can toefede a cable or wireless signal contointog:a bit sh m and the like. However, when iwnfto d, non-transitory computer-readable storage media expressly exclude media such as energy, carrier signals, el^frOmagnetic waves, and signals pet sc,|P152| Methods according to th above-described examples can be implemented using eomputer-executoblC: inst tetfons that are stored w therwise avaihfe from computer- sdable media. Such Insbwtfe ©an comprise. For example, instructions and tfe wmeh cause or otherwise configure a general purpose computer, special purpose on puter, o special purpose f sre *rre d,'\ w ' oei fo, a tctem tot sun or ytou f fonutrei-, fetors of t'< ‘je c-l r resources used can be accessible over a network. The -executable computer msiruetio may be, tw example, binaries, tote ediate format tfe fons such as assembly lattgut^ge. firmware, ©r so e code. Examples of co uter- -dabk media that may be tited to Store Instructions, information used* d / cr information created during methods wording to dewib examplestide niagnetie or optical disks, wlidfState mcmo«%- devices, flash memory, USB rfevtees provided with non-vfoatile s emoty, networked storage devices, trad ro on,[0 53| Devices imptemefomg orefon s according to these iseto res Can comprise hardware, tinware s dfor software, and can take any of a variety of form iaetors. Typical examples of such ■form.fat.foeto^^rvem,.1apt»ps»xma^ t^. small form fe er peteosrel computers, peaspuai digital assistan ts, and so ot. The functionality described herein also can be embodied in peripherals or add-in cards, Such functionality can also be implemented tm a circuit board among diflkront chips or different processes executing in h single device, by way of father example.[0154| The isislwtfens, media for conveying such instructions, comparing resources for executing them, and other swe u for ippoti such computing resources arc means for providing the factions desc ribed In these disclosures.[M55| The present technology Includes Curaputerereadable storage mediums for storing inswaotions, and systems for executing any one of sire methods enfoodre in sire ia actfons addressed in the aspects of the present technology presented below'[0156[ A'pvvisJ 157 Aspect L method of automatng so arc development. comprises < fotemmg. O m a machine kv ruing mode; jn m J at urn iha*.3 si'ft ^re de lupwm reft piroidfo m the in.n hw learning: model if completed, sv IK rem the software devefopment task compritss mswuctrous tt miret <p' 'in. e rcq v line m w lift irodc 5repoapplication, i Ifostrai i ng t bangts gemmated i>y the:machi:ns isa o sag mrek I m tire crate m a drared roposifory, wherein the dift comprises ai Ica« one change to a text file associated with the repository; and generating a change proposal io apply the changes generated by the machine le ti g model to the repository.[01581 Aspssti. The method of Aspect *, further comprise based dn receiving the ipstwtfas w perform the software development u k pawning a c tiner based on an execution environment identified bya user ronfigurotlan; ^retrieving packages a s»cht«d witha dewfop ens ironmen! a^ datod with the repository; and disabling a netw rk interface associated w th the ntait before the nwhine learning model isprovided the software development task.[01591 Aspect 3 - The method of Aspect 2, forthe comp reg: wherein the efa Ifofoing niodei is cmfogured io access the container after: the disabling of foe ®twk inter& e.35|01 | Aspect 4. The method of any of Aspects J to 3, further comprising: generate, using a p g m atb interface associated with the repository, a manifest associated with a braauh, uf the: repository after the change proposal, -wherein the manifest Identifies a plurality of fl ies associated with the repository after the change proposal,|®16I| Aspects, The method: of hay rf Aspects 1 to 4, further 'comprisiag: geu ttng, using s programmatic interface associated with the repository, a commit identifying a tree and a parent: associated with the change proposal.[01 2| Aspect fe The method of any of Aspects i to Ssfurther comprising: display, using: a front end at a client device, ths diffand a summary of changes generated by the machine learning model for human review.iftl63| Aspect 7, The method of any of Aspects I re o, wherein the change proposal Includes a manifestassociated with fife sfrucln of the reposbors a dammit idenliiying a tree and: a parent associated with the change proposal, and the diff identsh > ng at least onetext change ip at least one- file.10164 Aspect The method of Aspect 7, further comprising: providing, to a fetu device, the changefij. % bent d > er, ’•uheirtu >bv s. i s< u device ex jtes the change proposal to promulgate changes identified in the change propos'd to the repps! trey.1016S| Aspect 9, The method of 'Aspect 8, wherein the change proposal includes at feast one ot a file a< d, fl ' i",o're, a, «: a tevt tk’ u <:ree|016b| peul lb. I he method of any vt Aspects I th 9, farther comprising: after gerteriftlngshe re >c n H dore g he rep<wttor -r Jre recon lonruner uni veirera: n. fire ebanyc proposal m the eto-’d routrew-f (I167| Aspect 11, The method of Aspect IO, further comprising t applying change request to the repository based on an instruction from a cheat device.|0168| A spect 12, The methodof Aspect 11:, wherein t hechange proposal includes at least one ofa file add, fife remove, and a binary fife update.|0I60| Aspect LX A computing devicefor automating software development. The computing device includes at least one memory and at least: one processor coupled to the at least one memory' ■ nd configured to: obtain, from a machine teaming model, an indication that a software4765-S " te fy S51 W! devofopmeat task provided to the machine leantiag model fe eempteted, whurofo the software development task eomprfees ifflnidiw in natu l language requesting to mixiify ebde a repositoi^; generte a diffi using a etemiimstfeapplication. il fustratingchanges generated by the machine (cemm? model to the codeiu a cfened repositoiy^w herein the diffeompri - best one change to * UM fife associated with the re ositefy; and genm^a change propntol to.apply the changes getkrnue by the machine learning model to the repositefy,[0170| Aspect I 4. The computing device of Aspect 13, wherein the at least: one processor is configured tot based onreceiving the inductions to petfbrm the soft re development task, spawn; a container based m an e eetoms environment identified by a user configuration} retrieve packages associated with s development enymennwt associated: with the repository; and disable a network interface as bte Aii the container before the machine learning model is provided thesoftware development task,[ftl7'i| Aspect 15, The computing- device of Aspect 14, wherein the machine learning model is configured to access the containerafter the disabling of the network Interface.101721 Aspect I<£ The computing device of any of Aspects 11 to 15, wherein ths at least one processor is configured to: generate, using s programmatic huerfiwv assuefeted with the repustiory, a manifest associated with a branch of the repository after the change proposal, wherein the manifest identifies a plurality of fite* associated with the tepiteitory after the change proposal.[ I73| Aspect I?. Ths computing device of any of Aspects 13 to 16, wherein the at least one processor is con fi g nred to; gstterate, using » programmatic iuterfitee associated with the epository;, a commit identifying'^ and a parent associated with toe chrmge proposal[01741 Aspect 1 H The computing device of any of Aspects 13 to 17, wherein the at least one processor is ccmfigured to. display, using a front end at n cl tent device, the diff arid a samma y of changes generated by the machine I aiming model for human review,[01751 Aspect 19. The competing: device of any to" Aspects 13 to 1, whereto the change proposal mehdtoi a manifest sxumated with fife structure of the repository, -i commit identifying: a tree and a patent assneiated with the change proposal, and ths dife identifying at feast one text change- in at least one file.[01701 Agroct 20, 1 be compt i device to Aspect in, w heroin the a? lead one pnwvwur is cnnfigiired to; provide, to a client device, the change proposal based rm a request rm a receivedl2<tto5A7Wk, P to t WOI to w tow device, wherein the client device executes the elutnge proposal to premulgare changes idenrifted in the change proposal to the repository,101771 Aspect 2 l. Tto otnpmi’ng diwto of Aspect 2.0, wherein the change sop to snciii s at least one ofofife tod, file remove, and a text file update.

[0178] Aspect 22. The computing device of any of Aspects 13 to 21, wherein the at least one processor is configured to: after generating the change proposal, spawn a second container associated with the repository; clone the repository in the second container; and generate the change proposal in the second container.

[0179] Aspect 23. The computing device of Aspect 22, wherein the at least one processor isChnftgured to; apply change respit i<< to repository bared,on an instruction from a client device*- Jftl80| Aspect 24 I IK ^emptors’ d to uf Aspect 2?, w.jercm ihc Jiaince o^to sncmces at least one of a fife. add* file remove, and a binary file opdafc

[0181] Aspect 25. A non-transitory computer-readable medium having stored thereon instructions that, when executed by at least one processor, cause the at least one processor to perform operations according toany of Aspects 1 to 12.

[0182] Aspect 26. An apparatus for automating software development comprising one or more means for performing operations according to any of Aspects 1 to 12.

Claims

124765-SClaimsWhat is chimed is:

1. A method of automating software development, comprising: obtaining, from a machine learning model, an indication that a software development task provided to the machine learning model is completed, wherein the software development task comprises instructions in natural language requesting to modify code in a repository:gcn ratsne a <hff uunc a i termmMsc applrortfon slkisUahm, change* g wjato h riw m Urou. homing model to the fe;n a drou itp itosy, nh vnn the ds il msp;s w at lw»l one change f a h'-\i til ’ a vs 'iated with she repository smgenerating a change proposal hi applj the changes generated by the machine learning model to lb® repository;2. The method of Claim I* further comprising:based on receiving the instructions to perform the software development task, spawning a container based on an execution environment identified by a user configuration;retrieving packages associated with a development environment associated with the repository; anddisabling a network interface associated with the container before the machine learning model is provided the software development task.

3. The method of claim 2, further comprising:wherein the machine learning model is configured to access the container after the disabling of the network interface.

4. The method of claim 1, further comprising:genemttng, tiring a pwgrantmatic terfwe smeiated with the mpesitoty, a manifest associated with a branch of the peeifory:afierti)echaage:pn>|X>^d, whesdjilhe manifest identifies a plurality of files associated with the repository after the change propossl5. The method of claim 1, further comprising:generating, using a programmatic interface associated with the repository, a commit identifying a tree and a parentwith the change proposal,6. The method of claim 1, further comprising:display, using a front end at a client device, the diff and a summary of the changes generated by the machine learning model for human review.

7. The method of shim I, wherein the change proposal ineludes a manifest associated with ills structure of the repository, acommit idc-niifymg a tree-and a parent associated with the change pu>pr^<d, and rhe ds fl jdertily g at lewt one test chrmgc m as ki one tile8. The method of claim 7, further comprising:providing, to the repository, the change proposal based on a request on a received from a client device.

9. The method of claim 1, further comprising:d; d.i\ I'i'j m mh'i f.n c t<» vsen s ecit'd dm Jnpo'sent '.arimv oom s plnrohU of development, vanan s generated during the snftwnre development task, whereiri the dill is generated based on die selected development variant.10, The method of claim 1, further comprising;after generating the change proposal, spawning a second container associated with the repository:cloning the repository in the second container; andgenerating the change proposal in the second container.

11. The method of claim 10, further comprising:applying change request to the repository based on an instruction from a client device.

12. The method of claim 11, wherein the change proposal includes at least one of a file add, file remove, and a binary file update.

13. A computing device for automating software development, comprising:at least one memory; andat least one processor coupled to the at least one memory and ermfigme to:obtain h«m a imrchmc learning mod. < i imhcaimn that asro development task pfr ides.! t<« rhe machine framing model is c nipfrfrd, wherein the software dfewlop ^nt task comprises instruction in nbireai tanggage reqwtidg fo modify code in a repository:generate a fefit using a deterministic application, iiteteatiu gfai ges generated by the mstfemc kammg mode! m the endc ir a cloned repo tnrx uhtrem tlx <h!i xnnpns<.s at least one change te a text fife associated with: the repository; andgenerate a change proposal to apply the changes generated by the machine learning model to the repository.14, I he txnnpuonc dw <c of ci nr ’ x wher^n she a: L.w: ’re phx.*.^ fe ennf nsed to based ett Kxc^nsg the HO( -on> to citora the software develop^' t ta^k, 4x1^ ■. a mi WBU Koed on M ex LUIH’H ctnnnnm<nt sdctmfe d hv I HHI onth’ roinnreirte'te pa^vex as><-.»tefl xyith a development v-nuucnnvnl a,oc ied with the repository; anddisable a network interface associated with the container before the machine learning model is provided the software development task.

15. The computing device of claim 14, wherein the machine learning model is configured to access the container after the disabling of the network interface.1, Th ompuhng dev s e of cfe-re ’ k wntreni she m least one pnx ’wor cnnf'gttred tn, generate, / using a programmatic interface associated with the repository, a manifest: associated with a branch sfthe repository alter the change osal, wherem the manifest identifies a pl fey of files associated with the repo U>r after the change proposaf17. The computing device of claim 13, wherein the at least one processor is configured to: generate, using a programmatic interface associated with the repository, a commit identifying a tree and a parent associated with the change proposal.

18. The computing device of claim 13, wherein the at least one processor is configured to: display, using a front end at a client device, the diff and a summary of changes generated by the machine learning model for human review.

19. The computing device of claim 13, wherein the change proposal includes aassiteiated wife: file structure of the repository, -a commit idctfefyhg a tree and a- parent associated with the dirfnL’v propel and lite difi’idcmi sng t feaM one text change m at kti one life20. A method of automating software development, comprising:4765-S " to. P S51 W! providing, to a machine learning model, Imtmetie to perform a wftwam development task, therein the trncfio define the roftwam development taskin natt l tongnage that revest changes or addmens to code a repository;obtaining a chff from the machine learning model based on idetdifieation of completion of toe: software d vcto>pme»t task; angenerating a change proposal for the repository based on the diff, wherein the diff comprises at least one change to a text file associated with the repository.