Sebenzisa i-jev-chat-jarvis ku-macOS ukubona izinhloso ze-WeChat, uhlole ubungozi futhi udale izimpendulo

Isingeniso Sephrojekthi

jev-chat-jarvis iyithuluzi losizo lemiyalezo ye-WeChat elenzelwe i-macOS. Lapho kuvela umlayezo wombhalo ku-WeChat ongatholakala ngokucacile ukuthi uvela komunye umuntu, ifunda okuqukethwe kwesikrini futhi isebenzise imodeli yasendaweni ukuhlaziya umlayezo, bese ibonisa inhloso yangempela yalo musho, izinga lobungozi ukusuka ku-0 kuye ku-9, iziphakamiso zesenzo kanye nezimpendulo ezingaba khona.

Leli thuluzi lisebenza ngendlela yokufunda kuphela; alijovi ikhodi ku-WeChat, alisebenzisi i-hook, futhi alisusi ukubethela kwesizindalwazi. Futhi alithumeli imiyalezo ngokuzenzakalelayo. Umsebenzisi angakopisha okuqukethwe kwesiphakamiso kuphaneli entantayo, noma achofoze u-“Faka” ukuze abeke umbhalo ebhokisini lokufaka le-WeChat; ukuthumela ekugcineni kusenziwa ngumsebenzisi mathupha.

Imisebenzi Eyinhloko

• Ukuhlaziya inhloso nobungozi: Isekela ukuhlonza ngaphandle kwezibonelo kwezinhlobo zenhloso ezingu-8, futhi inikeza ukuhlukaniswa kobungozi neziphakamiso zesenzo. I-README yephrojekthi isebenzisa izibonelo zokuhlola zase-Chinese ezingu-22, ngenani lokuhlonza inhloso elingu-86.4%.
• Ukwakhiwa kwezimpendulo ezingaba khona: Inezinhlobo zamazwi ezingu-10 ezakhelwe ngaphakathi. Uhlobo ngalunye lukhiqiza izimpendulo ezimbili, ezizinzile nezikhululekile, bese imodeli yasendaweni izihlela.
• Ukucubungula okuhambisanayo: Ukuhlola umlayezo nokukhiqiza izimpendulo kungaqalwa ngesikhathi esisodwa. Isikhathi esiyisibonelo esinikezwe yi-README ku-M1 Pro cishe imizuzwana engu-1.5 yenhloso kanye nemizuzwana engu-1.5 kuya kwengu-2 yezimpendulo; isivinini sangempela sincike kudivayisi nasensizakalweni.
• Amabhokisi okuthola e-YOLO: Ngokuzikhethela, ingamaka imiyalezo etholwe yi-OCR ewindini le-WeChat, ihlukanise imiyalezo yomunye umuntu kweyakho, futhi ibonise izinga lokuzethemba.
• Iphaneli entantayo yomdabu: Iphaneli ibonisa ingxoxo yamanje, isimo sokuhlaziya, imiyalezo esalindile, umongo, inhloso, izinga lokuhlonza, izinga lobungozi, iziphakamiso zesenzo kanye nezimpendulo ezingaba khona.

Ukufaka Nokuqalisa Okokuqala

Indlela yokuqala: Landa uhlelo lwe-macOS

Uma ufuna ukulusebenzisa ngqo, landa iphakheji ecindezelwe ye-macOS ku-Releases, uyikhiphe bese uhudulela i-.app kokuthi “Applications”. Njengoba uhlelo lungakaqinisekiswa, ekuqaliseni kokuqala chofoza kwesokudla kuhlelo ku-Finder bese ukhetha “Open”; ungachofozi kabili ngqo.

Uma i-macOS ibonisa ukuthi “Yonakele futhi ayikwazi ukuvulwa; kufanele uyihambise ku-Trash”, ngokuvamile iphakheji ecindezelwe elandwe esipheqululini iba nesibaluli sokuhlukaniswa. Ungalisusi ngokushesha uhlelo; sebenzisa lo myalo ku-Terminal:

sudo xattr -r -d com.apple.quarantine /Applications/jev-jarvis.app

Uma igama lohlelo lishintshiwe, isibonelo jev-jarvis 2.app, shintsha indlela ekumyalo ibe yindlela yangempela.

Indlela yesibili: Qalisa kusuka kukhodi yomthombo

Ngaphambi kokuqalisa kusuka kukhodi yomthombo, qinisekisa ukuthi i-WeChat isiqalisiwe nokuthi i-Terminal noma uhlelo oluqalisayo lunemvume yokuqopha isikrini. Ngemva kokungena enkombeni yephrojekthi, sebenzisa:

./start.command

Iphrojekthi isebenzisa uv ukuphatha indawo ye-Python. Uma ingekho i-uv esebenzayo ohlelweni, izindlela ezimbili zokuqalisa zizozama ukulanda nokusebenzisa iskripthi esisemthethweni sokufaka; uma kwehluleka, zizozama nokusebenzisa i-Homebrew ekhona. Amalogi okufaka nawokuqalisa atholakala ku:

~/Library/Logs/jev-jarvis.log

Ukunikeza i-macOS Izimvume

Ekuqaliseni kokuqala, landela imiyalelo ukuze unikeze i-jev-chat-jarvis imvume “Screen Recording” ngaphansi kokuthi “System Settings > Privacy & Security”, bese uphuma uphinde uvule uhlelo ukuze imvume isebenze.

Umsebenzi othi “Faka” udinga nemvume ye-“Accessibility”. Lapho uchofoza okokuqala u-“Faka”, uhlelo luvame ukubonisa isicelo semvume. I-v0.3.1 nangaphambilini kudinga nokuthi unikeze python3.12 imvume ehambisanayo.

Lawa mavume asetshenziselwa ukufunda okuqukethwe kwewindi le-WeChat, ukuthola isilawuli sokufaka nokwenza umsebenzi wokufaka; awasho ukuthi ithuluzi lizothumela imiyalezo ngokuzenzakalelayo.

Ukulungisa Amamodeli Nezikhiye

Iphrojekthi inengqimba yokwahlulela nengqimba yokukhiqiza, futhi izikhiye zazo zombili zingashiywa zingenalutho:

• Ingqimba yokwahlulela: Uma kungalungiselelwanga ukhiye we-TypeSafe Jev, kusetshenziswa imodeli yasendaweni ye-decider-2b. Ukulanda kokuqala kudinga cishe u-7 GB.
• Ingqimba yokukhiqiza: Uhlelo oluhlanganisiwe lunokhiye owabiwe owakhelwe ngaphakathi, ngakho ungazama ukukhiqiza izimpendulo ngaphandle kokulungisa ukhiye wakho.

Ukusebenzisa Izilungiselelo Zezithombe

Chofoza isithonjana segiya ekhoneni eliphezulu kwesokudla lewindi elintantayo, noma uvule iwindi lokulungisa ngokukhetha “J > Model Settings…” kubha yemenyu. Isikrini sezilungiselelo sisekela ukuhlela amakheli, izikhiye namamodeli ezinsizakalo ezintathu: i-Jev, ezihambisana ne-OpenAI kanye nezihambisana ne-Anthropic.

Ungachofoza u-“Get Model List” ukuze ufunde amamodeli kusixhumi se-/models sensizakalo, noma ufake imodeli ngesandla uma isixhumi singasekelwa noma isicelo sehluleka. Ngemva kokukhetha imodeli, ungachofoza u-“Test Connection”; ukuhlola kuthumela ukubingelela okungaguquki kusetshenziswa izilungiselelo ezingakagcinwa, akufundi okuqukethwe kwe-WeChat, kodwa kungase kudale izindleko ezincane zensizakalo.

Ngemva kokugcina izilungiselelo, kufanele uphume uphinde uvule uhlelo; izilungiselelo ngeke zishintshe khona manjalo kulokhu kuqalisa. Ifayela elizenzakalelayo elihlelwa iwindi lezilungiselelo lithi:

$XDG_CONFIG_HOME/jev-jarvis/env

Uma i-XDG_CONFIG_HOME ingasethwanga, indlela yangempela ithi:

~/.config/jev-jarvis/env

Ukwakha Ifayela Lendawo Ngesandla

Ungakha nefayela lokulungisa ku-Terminal:

mkdir -p ~/.config/jev-jarvis
cat > ~/.config/jev-jarvis/env <<'ENV'
# Ingqimba yokwahlulela (ngokuzikhethela): uma ishiywe ingenalutho kusetshenziswa i-decider-2b yasendaweni
export TYPESAFE_API_KEY=""

# Ingqimba yokukhiqiza: noma yisiphi isiphetho esihambisana ne-OpenAI
export OPENAI_API_KEY="sk-ukhiye_wakho"
export OPENAI_BASE_URL="https://api.deepseek.com"
export OPENAI_MODEL="deepseek-chat"
# Amanye amamodeli adinga ukuvalwa kwemodi yokucabanga
# export OPENAI_EXTRA_BODY='{"enable_thinking":false}'
ENV
chmod 600 ~/.config/jev-jarvis/env

Ukuhlelwa kokubaluleka kokulungisa kungokulandelayo: okuguquguqukayo kwendawo kuphakeme kunefayela le-env lomsebenzisi, lona liphakeme kune-.env yephrojekthi. Engxenyeni yokukhiqiza, ukulungiswa kwe-OpenAI kuza kuqala kunokwe-Anthropic; ukhiye owabiwe owakhelwe ngaphakathi usetshenziswa kuphela uma kungalungiselelwanga lutho kuzo zombili.

Izinto Okufanele Uziqaphele Lapho Ukhetha Insizakalo

• I-DeepSeek ingasebenzisa deepseek-chat, i-README eyibala njengenketho esheshayo ehloliwe.
• I-Zhipu ingasebenzisa glm-4-flash, futhi kufanele ugcwalise okuguquguqukayo okuhambisanayo ngokweqembu lokulungisa elisekelwa yiphrojekthi.
• I-Ollama yasendaweni ingasebenzisa http://localhost:11434/v1, ukhiye ube ollama, imodeli ibe isibonelo qwen2.5:7b.
• Gwema ukusebenzisa amamodeli okucabanga. I-README ixwayisa ngokuthi inqubo yokucabanga ingase iqede max_tokens, ibangele ukuthi izimpendulo eziphakanyisiwe zingabi nalutho; uma usebenzisa i-DeepSeek, khetha deepseek-chat.

Ukusetshenziswa Kokuqala

Qalisa uhlelo bese uvula i-WeChat, ugcine iwindi lengxoxo le-WeChat litholakala. Ithuluzi lifunda indawo yengxoxo ngokuqukethwe kwewindi nange-Vision OCR; alidingi ukuthatha isikrini sonke sesistimu, ngakho iphaneli entantayo ayingcolisi i-OCR.

Lapho kutholwa umlayezo wombhalo oqinisekile ukuthi uvela komunye umuntu, iphaneli izobonisa ngokulandelana ingxoxo yamanje nesimo sokuhlaziya, umlayezo ocutshungulwayo nomongo, inhloso, izinga lokuhlonza, izinga lobungozi, iziphakamiso zesenzo kanye nezimpendulo ezihlelwe ngokwezinhlobo zamazwi.

Ngakwesokunxele kwempendulo ngayinye kuboniswa amathuba okuhlelwa kwasendaweni, kanti ngakwesokudla kunenkinobho ethi “Kopisha” nethi “Faka”. “Kopisha” yenza ukunamathisela ngesandla kube lula; “Faka” ibhala umbhalo ebhokisini lokufaka le-WeChat, kodwa ayichofozi ngokuzenzakalelayo inkinobho yokuthumela.

Isibonelo Sokusetshenziswa

Isibonelo, uma omunye umuntu ethumela ukuthi “Landela le mfuneko namuhla”, ithuluzi linikeza umlayezo engqimbeni yokwahlulela ukuze liwuhlonze, bese linikeza isiphakamiso sesenzo ngokusekelwe kumphumela wobungozi. Ngemuva kwalokho ingqimba yokukhiqiza ikhiqiza izimpendulo eziningi ngamathoni ahlukahlukene ngesikhathi esisodwa, bese izihlela kabusha ngokuhleleka kwasendaweni. Ungakhetha eyodwa ngokomongo, uchofoze u-“Kopisha” noma u-“Faka”, bese uyihlola futhi uyithumele wena.

Uma indlela yomlayezo ingenakuqinisekiswa, isibonelo umbhalo udlula kwesokunxele nakwesokudla noma usendaweni emaphakathi, iphaneli izowumaka ngokuthi “Direction unconfirmed” futhi ngeke iwusebenzise njengom目标 wempendulo. Imiyalezo ehlonzwe ngokucacile ukuthi ivela komunye umuntu kuphela ezovula ukuhlola nokukhiqiza.

Amathiphu Okusetshenziswa Okuthuthukile

Ukuvula Amabhokisi Okuthola e-YOLO

Ungavula amabhokisi okuthola ngesikhathi sokuqalisa ngokusebenzisa okuguquguqukayo kwendawo:

JEV_BOXES=1

Ngemva kokuqalisa, ungashintsha futhi u-“YOLO Detection Boxes” kubha yemenyu. Amabhokisi avuselelwa cishe njalo ngomzuzwana: umugqa oqinile oluhlaza okwesibhakabhaka umele isilawuli sokufaka esitholwe yisixhumi se-Accessibility, kanti umugqa onamachashazi owolintshi umele indawo yokufaka eqagelwe ngokwemikhawulo yesithombe-skrini. Umugqa onamachashazi awusho ukuthi isilawuli esibhalekayo sitholakele ngokuqinisekile.

Ukwenza Ngokwezifiso Izinhlobo Zamazwi

Sebenzisa i-JEV_TONES ukwengeza noma ukumboza izinhlobo zamazwi. Uhlobo ngalunye lwehlukaniswa ngo-|, ngefomethi ethi “igama=incazelo”, isibonelo:

export JEV_TONES='Inkosi Yokubalekela Umsebenzi=njengochwepheshe onolwazi lokugwema umsebenzi, dlulisela umsebenzi ngobuhle nangokuhlonipha'

Incazelo kufanele icacise ithoni edingekayo nokuthi akufanele iphenduke ibe yiluphi uhlobo lwesitayela. Ngemva kokushintsha, kufanele uqale kabusha uhlelo. Iwindi lezilungiselelo okwamanje aliguquli i-JEV_TONES, i-JEV_BOXES noma i-OPENAI_EXTRA_BODY.

Ukuzihlola Nokuthola Izinkinga Ngezingqimba

Uma usebenzisa ikhodi yomthombo, ungahlola ngokwehlukana izinqubo zokubona, ukwahlulela, ukukhiqiza nokuhlola ukubuyela emuva:

uv run python src/perception.py
uv run python src/judge.py "Landela le mfuneko namuhla"
uv run python src/judge_zh_test.py
uv run python src/generate.py --check
uv run python -B -m unittest discover -s tests
uv run python probe/bootstrap_regression.py

I---check yengqimba yokukhiqiza ihlola kuphela ukuhlaziywa kwemininingwane yokufakazela ubunikazi; ayilingani nokuhlolwa kokuxhumana kwangempela. Ukuqinisekisa ukuthi insizakalo ingakhiqiza umbhalo, sebenzisa u-“Test Connection” ewindini lezilungiselelo.

Ukubuka Amalogi Nokukhulula Isikhala Sediski

Uma ukucutshungulwa komlayezo wokuqala kuhamba kancane ngemva kokuqalisa, ngokuvamile kubangelwa ukufudumezwa kwemodeli yasendaweni. Ungabuka ilogi yakamuva:

tail -40 ~/Library/Logs/jev-jarvis.log

Ilogi iqukethe izikhathi zezigaba ezahlukene, kodwa ayiqukethe okuqukethwe kwemiyalezo, ngakho ilungele ukwabelwana ngayo nonjiniyela lapho ubika inkinga.

Inqolobane yemodeli yokwahlulela yasendaweni ithatha cishe u-7 GB futhi itholakala ku:

~/.cache/huggingface/hub/models--Mapika--decider-2b

Ngemva kokuyisusa, izolanda futhi lapho kusetshenziswa ukwahlulela kwasendaweni ngokuzayo. Indawo ye-Python virtual environment ithi:

~/Library/Application Support/jev-jarvis/venv

Ukususa uhlelo akususi le ndawo ngokuzenzakalelayo; ungayihlanza ngesandla uma kudingeka.

Imikhawulo Yomsebenzi Othi “Faka”

Ithuluzi liqala ngokusebenzisa isixhumi se-macOS Accessibility ukubhala nokufunda emuva ukuqinisekisa. Uma ezinye izinguqulo ze-WeChat zinganikeli ngesilawuli sokufaka, lizozama indlela ehambisana nombono: liqinisekise iwindi, indawo yokufaka nesihloko, livule i-WeChat, lichofoze indawo yokufaka, lifake umbhalo bese liwuhlola nge-OCR.

Uma kusetshenziswa indlela ehambisana nombono, ithuluzi linganyakazisa igundane; ngakho ungasebenzisi ikhibhodi, igundane noma ushintshe ingxoxo ngesikhathi sokufaka. Alisebenzisi ibhodi lokunamathisela noma i-Cmd+V, alicindezeli ngokuzenzakalelayo inkinobho yokuthumela, futhi liguqula imigqa emisha namathebhu kube izikhala.

Uma kutholwa okusalungiswa okukhona, indlela ehambisanayo izoma bese ikutshela ukuthi usebenzise “Kopisha” ngesandla; izoma futhi uma iwindi, ukugxila noma iseshini kushintsha. Njengoba imikhawulo yokubona ne-OCR kungase kuhlulele kabi, hlola ingxoxo yamanje nokuqukethwe kwebhokisi lokufaka ngaphambi kokusebenzisa.

Imikhawulo Eyaziwayo

• Okuqukethwe kwezithombe nezitika akukwazi ukufundwa; izimpendulo ezicashuniwe ziphathwa njengombhalo ojwayelekile; amakhadi ama-akhawunti asemthethweni angase aphathwe njengemiyalezo okufanele ihlaziywe.
• Isakhiwo se-WeChat sesikrini esigcwele noma izinguqulo ezintsha ze-WeChat kungabangela ukuthi ukuhlonza kwehluleke; izakhiwo ezihlobene zitholakala phezulu kwe-src/perception.py.
• Ezimweni zamawindi amaningi, kuqala kuhlonzwa iwindi eliyinhloko elithi “WeChat” noma “微信”.
• Imiyalezo yomunye umuntu kuphela eqinisekiswe ngokucacile ukuthi inendlela efanele evula inqubo yempendulo; imiyalezo engenakuqinisekiswa iyakweqiwa.
• Indlela yokufaka ehambisana nombono isadinga ukuhlolwa okwengeziwe kumatimu amnyama, izikrini eziningi kanye nezinhlobo ezahlukene ze-WeChat.

Ubumfihlo Nezikhumbuzo Zokusebenzisa

Inhloso yephrojekthi ukufunda okuqukethwe kwesikrini kudivayisi yakho naku-akhawunti yakho kuphela; ayifaki lutho, ayisebenzisi i-hook, ayisusi ukubethela kusizindalwazi se-WeChat, futhi ayithumeli imilayezo ngokuzenzakalelayo. Sicela uyisebenzise kuphela kudivayisi yakho futhi uhambisane nesivumelwano selayisensi yesofthiwe ye-WeChat. Izibuyekezo ze-WeChat zingabangela ukuthi ukubonwa kwesakhiwo kungasebenzi; uma uhlangabezana nenkinga, qala ubheke amalogi, bese uthumela impendulo nge-issue yephrojekthi.

Isifinyezo

i-jev-chat-jarvis ihlanganisa i-OCR yesikrini, ukunquma inhloso kwasendaweni, ukuhlukanisa izinga lobungozi, ukukhiqiza izimpendulo ezingaba khona, kanye nephaneli elintantayo lendabuko, futhi ilungele abasebenzisi be-macOS abafuna ukuqonda ngokushesha imilayezo ye-WeChat nokunciphisa ingcindezi yokuphendula. Ngesikhathi sokufaka, kubalulekile ukuphatha i-Gatekeeper, izimvume zokuqopha isikrini nezokufinyeleleka; ngesikhathi sokusebenzisa, khetha imodeli yasendaweni, isevisi ehambisana ne-OpenAI noma ehambisana ne-Anthropic ngokobumfihlo, isivinini nezindleko, futhi uhlale uhlola okuqukethwe mathupha ngaphambi kokukuthumela.