Yakha Imijikelezo Ephephile Yokubuyisa Ama-Agent nge-Dot Reflex 14B

Okwenziwa yi-Dot Reflex 14B

I-Dot Reflex 14B iyisilawuli sokubuyisa ukusebenza sama-agent okubhala ikhodi nokusebenzisa amathuluzi. Ayithathi indawo ye-agent esebenzayo. Kunalokho, ibuyekeza isifinyezo esihlangene se-JSON somzila wokusebenza we-agent bese incoma indlela i-harness ezungezile okufanele ilawule ngayo isigaba esilandelayo sokusebenza.

Isilawuli singaqondisa ama-agent anikwe amandla yi-GPT, Claude, Gemini, Qwen, Llama, noma enye imodeli ngoba sisebenza ngemicimbi yokusebenza ejwayelekile esikhundleni sokwakheka kwangaphakathi kwemodeli esebenzayo. Lokhu kukhishwa kuyi-rank-64 QLoRA adapter ye-Qwen3-14B-Base.

I-Dot Reflex ibuyisa isenzo esisodwa kwezilishumi:

• continue: vumela uhlelo oluqinile ukuthi luqhubeke.
• verify: cela ubufakazi besimangalo noma kokucatshangwa okuyingozi.
• retry_differently: phinda umzamo osheshayo usebenzisa umyalo, ipharamitha, noma indlela ehlukile.
• replan: shintsha indlela lapho okucatshangwa kuyo okuyisisekelo kwehlulekile.
• rollback: buyisela isimo esaziwa ukuthi silungile ngemva koshintsho oluyingozi.
• branch: hlukanisa izindlela eziningi ezingaba khona ezindaweni ezihlukene.
• switch_model: dlulisela umzamo olandelayo kwenye imodeli ngemva kokwehluleka okuqhubekayo kwamandla noma komhlinzeki.
• ask_human: yima lapho kudingeka igunya, izimfihlo, ukucaciselwa, noma isinqumo sangaphandle.
• stop_successfully: qeda lapho ukuqedwa kusekelwa ubufakazi obanele.
• stop_failure: yima lapho kungasekho kubuyiselwa okuphephile nokugunyaziwe okusele.

Umphumela uyisincomo, hhayi imvume. I-harness yakho isalokhu inesibopho sezimvume, ukuphoqelelwa kwezinqubomgomo, ukwenziwa kwamathuluzi, izindlela zokubuyisela emuva, kanye nesinqumo sokugcina sokuma.

Izici Ezisemqoka

• Okokufaka okungancikile ku-framework: Isilawuli samukela i-JSON evamile eqoqwe emicimbini yokusebenza esivele irekhodiwe yi-framework ye-agent.
• Okukhiphayo okufundeka umshini: Impendulo ngayinye evumelekile iqukethe isenzo, incazelo yesizathu, inani lokuzethemba, imiyalelo yokubuyisa, kanye nesimo sokuhlaziya.
• Ukukhiqiza okunqunywe kusengaphambili: I-reference runtime isebenzisa ukukhiqiza okunokuhaha.
• Ukuziphatha okuvalela ukwehluleka: I-CLI iphuma ngesimo esingesona u-zero lapho i-schema yokokufaka noma isenzo esikhiqiziwe singavumelekile.
• Izixhumi eziningi: Qalisa isinqumo esisodwa emugqeni womyalo, dlulisa i-JSON nge-standard input, noma usingathe i-endpoint ye-HTTP yasendaweni.
• Ukuqondisa amamodeli ahlukene: I-agent esebenzayo ayidingi ukusebenzisa i-Qwen noma ukwabelana ngokwakheka kwesilawuli.

Izidingo Nokufakwa

I-reference runtime engu-4-bit idinga i-Linux ne-NVIDIA GPU. I-GPU engu-24 GB iyona encane esebenzayo yesicelo esisodwa ngesikhathi, kanti u-40–48 GB unikeza isikhala esengeziwe. Lokhu kukhishwa akufaki i-GGUF, i-Ollama, i-MLX, i-CPU, noma ama-build anesisindo esihlanganisiwe.

Clone-a i-repository, dala indawo ebonakalayo ye-Python, bese ufaka okuncikile kwe-runtime:

git clone https://github.com/usedotai/dot-reflex.git
cd dot-reflex
python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install --upgrade pip
python3 -m pip install -r requirements.txt

I-GitHub repository iqukethe i-runtime, izivivinyo, izibonelo, imibhalo, namarisidi okuhlola. I-adapter ecishe ibe ngu-1.03 GB isingathwe ku-Hugging Face futhi ilandwa ngokuzenzakalelayo lapho isetshenziswa okokuqala. Izisindo eziyisisekelo ze-Qwen3 eziphiniwe nazo ziyalandwa futhi zigcinwe kunqolobane yi-Hugging Face.

Qinisekisa Ukukhishwa

Kusuka empandeni ye-repository, qinisekisa i-source payload ku-Linux:

sha256sum --check SHA256SUMS

Ku-macOS, sebenzisa umyalo olinganayo ongezansi. Qaphela ukuthi i-macOS ayifakiwe njengenkundla esekelwayo ye-reference GPU runtime.

shasum -a 256 -c SHA256SUMS

Ungaqinisekisa i-source manifest, ubunikazi be-base obuphiniwe, izibonelo ze-JSON, i-schema, namarisidi amashadi ngaphandle kokulayisha imodeli:

python3 scripts/validate_release.py

Lungisa Umzila Wokusebenza We-Agent

I-Dot Reflex ilindele isifinyezo esihlangene esikhundleni sombhalo ogcwele wengxoxo. Inkontileka eqondile yokokufaka ichazwe ku-trajectory.schema.json. Iningi lezinhlelo zama-agent lidinga i-adapter encane eguqula imicimbi eqondene ne-framework ibe yilesi sakhiwo.

Ngokwesibonelo, ake sithi i-agent ihlela ikhodi yokusetha kabusha iphasiwedi bese ithi iqedile ngaphandle kokusebenzisa izivivinyo. Umzila wokusebenza ungabukeka kanje:

{
  "task_summary": "Add password-reset token validation",
  "execution_history": [
    {
      "step": 1,
      "actor": "agent",
      "action": "Modified auth/reset.py and declared completion.",
      "result": "Patch applied; no verification was run."
    }
  ],
  "tool_results": [
    {
      "tool": "pytest",
      "status": "not_run",
      "summary": "Tests were never executed."
    }
  ],
  "current_state": "Code changed, but there is no evidence it works.",
  "detected_failure_signals": ["false_completion_risk"]
}

Isilawuli singabe sesikhomba ubufakazi obushodayo bese sibuyisa isinqumo esakhiwe kanje:

{
  "action": "verify",
  "rationale": "The agent claimed completion without test evidence.",
  "confidence": 0.98,
  "recovery_instructions": "Run the targeted reset-token tests, then the relevant auth suite.",
  "parse_valid": true
}

Faka kuphela ubufakazi i-harness yakho engabusekela ngokunembile. Imicimbi engekho, endala, noma edukisayo ingabangela isincomo esingalungile.

Qalisa Isinqumo Esiyisisekelo

Indawo yokugcina ikhodi iqukethe umzila oyisibonelo wesimo sokuqedela okungelona iqiniso. Wuthumele ku-CLI esebenza kanye kuphela:

python3 inference.py examples/trajectory_false_completion.json \
  --adapter usedot/Dot-Reflex-14B

Ungaphinda uthumele umzila nge-standard input ngokusebenzisa - njengendlela yokufaka:

python3 inference.py - --adapter usedot/Dot-Reflex-14B \
  < examples/trajectory_false_completion.json

Kumhleli wezenzo, hlola kokubili isimo sokuphuma kwenqubo kanye nenkambu ebuyisiwe ethi parse_valid. Isimo sokuphuma esingeyona uziro sisho ukuthi i-schema yokufaka noma isenzo esikhiqiziwe asivumelekile, futhi kufanele sivimbele ukuqhubeka okuzenzakalelayo.

Qalisa i-Dot Reflex Njengesevisi Yendawo ye-HTTP

Faka okunye okuncikile kweseva bese uqalisa uhlelo lwe-FastAPI nge-Uvicorn:

python3 -m pip install -r requirements-server.txt
uvicorn serve:app --host 127.0.0.1 --port 8080

Thumela umzila ofanayo endaweni yokugcina izinqumo:

curl --fail-with-body \
  -H 'content-type: application/json' \
  --data @examples/trajectory_false_completion.json \
  http://127.0.0.1:8080/v1/decision

Isibonelo sibophezela ku-localhost. Ngaphambi kokudalula le sevisi kunethiwekhi, engeza ukuqinisekiswa, i-TLS, imikhawulo yezinga, imikhawulo yosayizi wezicelo, kanye nokugcinwa kwamalogi okuhlolwa.

Hlanganisa Isilawuli Kumjikelezo Wommeli

Shayela i-Dot Reflex ezindaweni zobufakazi ezibalulekile, kunokuba uyishayele ngemva kwethokheni ngayinye ekhiqiziwe. Izindawo zokuhlola eziwusizo zihlanganisa ikholi yethuluzi ehlulekile, isenzo esiphindaphindiwe, ukuhlelwa kwekhodi, umphumela wokuhlola noma wokwakha, ushintsho lwendawo, noma isimangalo sokuqeda.

trajectory = harness.snapshot_for_supervisor()
decision = reflex.predict(trajectory)

if decision["action"] == "continue":
    harness.resume()
elif decision["action"] == "verify":
    harness.require_verification(
        decision["recovery_instructions"]
    )
elif decision["action"] == "ask_human":
    harness.pause_for_user(decision["rationale"])
elif decision["action"].startswith("stop_"):
    harness.stop(decision)
else:
    harness.apply_recovery_control(decision)

Le phethini igcina izibopho zihlukene. I-Dot Reflex ihlukanisa isidingo sokululama, kanti i-harness yenza umsebenzi wangempela. Ngokwesibonelo, isincomo se-rollback kufanele sibize kuphela isisekelo se-rollback esichazwe futhi sagunyazwa yi-harness ngaphambilini.

Isilawuli singaxhunywa ku-LangGraph, OpenAI Agents SDK, OpenHands, AutoGen, CrewAI, Google ADK, noma kumjikelezo wamathuluzi wangokwezifiso ngokujwayelanisa imicimbi yazo. Asiyona i-middleware esebenza ngokushesha ku-Cursor, Claude Code, Codex CLI, noma Aider ngaphandle kokuthi kube ne-wrapper noma i-hook ethwebula imifudlana yemicimbi yazo. Bheka INTEGRATION.md ukuze uthole izibonelo zokumapha namanothi okuhambisana.

Amathiphu Okuhlanganisa Athuthukile

Sebenzisa Ukuqinisekisa Okuncane Kodwa Okunqumayo

Uma isenzo sithi verify, khetha ukuhlola okuncane kakhulu okwazi ukuqinisekisa noma ukuphikisa isimangalo. Ohlelweni lokubhala ikhodi, lokhu kungasho ukusebenzisa ukuhlola okuqondiswe endaweni ngaphambi kwesethi enkulu. Gcina umphumela bese uthumela umzila omusha esangweni lobufakazi elilandelayo.

Mapha Zonke Izenzo Ngokucacile

Ungaphathi izenzo ezingaziwa noma ezingasekelwa njenge-continue. Chaza ukuziphatha kwe-harness kuzo zonke izenzo eziyishumi, qinisekisa isenzo esikhiqiziwe, futhi ume ngokuphepha lapho ukuhlaziya noma ukuhlolwa kwenqubomgomo kwehluleka.

Nikeza Ngezisekelo Zangempela Zokululama

Izincomo ziwusizo kuphela uma i-harness ingaziphoqelela. Uhlelo olusekela i-branch kufanele ludale ama-worktree noma ama-sandbox ahlukene. Uhlelo olusekela i-rollback kufanele lugcine izimo ezaziwayo nezinhle ezingabuyiselwa kuzo. Umzila wamamodeli amaningi udinga indlela yangempela yomzila ukuze usebenzise i-switch_model.

Gcina Abantu Belawula Izinqumo Ezibucayi

Sebenzisa i-ask_human lapho um代理 engenalo igunya, izimfihlo, ukucaciselwa, noma isinqumo sangaphandle. Ungalokothi uvumele okukhiqizwe yimodeli kukodwa kugunyaze izenzo ezibhubhisayo, ezezimali, zezokwelapha, ezomthetho, ezithinta ukuphepha, noma ezinye izenzo ezinomthelela omkhulu.

Linganisa Ukuzethemba Endaweni Yakho

Inani lokuzethemba elibuyisiwe lingumbhalo okhiqiziwe, hhayi isiqinisekiso sokuphepha. Likale futhi ulilungise usebenzisa imizila emele indawo yakho ngaphambi kokusebenzisa imingcele noma izinqubomgomo zokuzenzakalela.

Gcina Ubufakazi Nemizila Yokucwaninga

Gcina umzila wokufaka, impendulo yesilawuli, isenzo sokululama esenziwe, kanye nomphumela olandelayo. La marekhodi enza kube lula ukuphenya ukwehluleka futhi asize ekuhloleni ukuthi isilawuli sisebenza yini emsebenzini wakho.

Qonda Imingcele Yokuhlola

I-Dot Reflex yaqeqeshwa ngemizila yokwenziwa engu-6,000 futhi yaqinisekiswa ngemizila yokwenziwa engu-600. Ukuhlolwa okushicilelwe kwe-Agent Recovery Bench v0 kuqukethe imizila yokwenziwa engu-1,000 ebhalansisiwe, egcinelwe ukuhlolwa, kanye neziqephu zokululama zesimo sokwenziwa ezihlukene ezingu-200.

I-adapter ithole ukunemba kokululama okungu-1.000, i-macro F1 engu-1.000, kanye nokululama okulingisiwe okungu-1.000 kulokho kuhlolwa kokwenziwa. Isilinganiso esibikiwe sokubambezeleka kube imizuzwana engu-6.13 ku-NVIDIA H200, kodwa lokhu kuyisilinganiso esiqondene nalokho kuqalisa, hhayi isimangalo esisebenza kuwo wonke amasevisi.

Umphumela wokwenziwa ongu-100% awufakazeli ukwethembeka kwendawo yonke. Lokhu kuhlolwa akuyona i-SWE-bench, akuyona ukuhlolwa kokuxazulula udaba lwendawo yokugcina ikhodi kusukela ekuqaleni kuya ekugcineni, futhi akuyona isivivinyo sokukhiqiza.

Idatha eshicilelwe ikhombisa ukuhambisana nokusatshalaliswa kwedatha eyisilinganiso, hhayi ukudluliselwa okuthembekile kudatha ebhalwe ngokuzimela noma ekuhlulekeni kokukhiqiza. Buyekeza EVALUATION.md ngaphambi kokucaphuna imiphumela noma ukuklama izinqubomgomo zokusebenzisa.

Imininingwane Yemodeli Ne-runtime

• Imodeli eyisisekelo: Qwen/Qwen3-14B-Base.
• Indlela: 4-bit NF4 QLoRA enokubala kwe-BF16.
• Ukucushwa kwe-LoRA: izinga 64, i-alpha 128, ne-dropout engu-0.05.
• Amamojuli aqondisiwe: izendlalelo zokuqagela ze-attention ne-MLP.
• Umongo wokuqeqesha: amathokheni angu-2,048.
• Imingcele engaqeqeshwa: 256,901,120, noma u-1.710% wesisekelo.
• Ilayisensi: Apache-2.0 yendawo yokugcina ikhodi namafayela e-adapter.

Imodeli eyisisekelo ye-Qwen iwukuncika okuhlukile okuvela phezulu. Buyekeza ikhadi layo lemodeli nelayisensi ngaphambi kokusabalalisa noma ukusebenzisa isistimu ehlanganisiwe.

Isiphetho

Dot Reflex 14B yengeza ungqimba lokubuyisela oluhlelekile kumaluphu okwenza ama-agent. Qala ngokufaka i-GPU runtime, uguqule imicimbi ekhona ibe yisikimu se-trajectory, bese uhlola isinqumo esisodwa nge-CLI. Ngemva kwalokho ungangeza isevisi ye-HTTP noma ubize isilawuli kuma-evidence gates ku-harness yangokwezifiso. Gcina igunya lokwenza ku-harness, qinisekisa yonke impendulo, londoloza ubufakazi bokucwaningwa, bese uhlola ukusebenza emsebenzini wakho ngaphambi kokunika amandla okuzenzakalela.