#My agent is not responding after updating to 2026.4.26

1 messages · Page 1 of 1 (latest)

onyx wing
#

So i did the new update now, and my agent stopped responding, any tips?

woeful tigerBOT
#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

gusty harnessBOT
#

Yep — post-update this is usually gateway/service, model auth, or channel routing.

Can you run these and paste the output?

openclaw --version
openclaw status --all
openclaw gateway status
openclaw models status
openclaw channels status --probe

Then try:

openclaw doctor
#
openclaw gateway restart

If it still doesn’t reply, tail logs while sending the agent a test message:

openclaw logs --follow

Look especially for model, auth, pairing required, mention required, blocked, or channel errors.

One update-specific thing to check: stale install / service path mismatch:

which openclaw
#
openclaw config get meta.lastTouchedVersion

If the service is using an older binary after the update, reinstall the gateway service:

openclaw gateway install --force
openclaw gateway restart

Docs triage flow: https://docs.openclaw.ai/help/faq (docs.openclaw.ai)

onyx wing
#

07:51:57+00:00 warn model-fallback/decision {"subsystem":"model-fallback/decision"} {"event":"model_fallback_decision","tags":["error_handling","model_fallback","candidate_failed"],"runId":"4e6c0f55-6f2a-4611-9d15-8e9ab20683f4","decision":"candidate_failed","requestedProvider":"zai","requestedModel":"glm-4.7","candidateProvider":"zai","candidateModel":"glm-4.7","attempt":1,"total":1,"reason":"rate_limit","status":429,"errorPreview":"429 Weekly/Monthly Limit Exhausted. Your limit will reset at 2026-04-30 20:04:48","errorHash":"sha256:527f0e80cf33","fallbackStepType":"fallback_step","fallbackStepFromModel":"zai/glm-4.7","fallbackStepFromFailureReason":"rate_limit","fallbackStepFromFailureDetail":"429 Weekly/Monthly Limit Exhausted. Your limit will reset at 2026-04-30 20:04:48","fallbackStepChainPosition":1,"fallbackStepFinalOutcome":"chain_exhausted","isPrimary":true,"requestedModelMatched":true,"fallbackConfigured":false} model fallback decision
07:51:57+00:00 error Embedded agent failed before reply: ⚠️ Weekly/Monthly Limit Exhausted. Your limit will reset at 2026-04-30 20:04:48
08:05:22+00:00 warn Config warnings:

  • plugins.entries.active-memory: plugin disabled (disabled in config) but config is present
    08:05:23+00:00 warn Config warnings:
  • plugins.entries.active-memory: plugin disabled (disabled in config) but config is present
    Log tail truncated (increase --max-bytes).
gusty harnessBOT
#

That log points to the real cause: z.ai / zai/glm-4.7 hit its weekly/monthly quota.

Key lines:

reason: rate_limit
429 Weekly/Monthly Limit Exhausted
reset at 2026-04-30 20:04:48
fallbackConfigured: false
chain_exhausted

So OpenClaw tried zai/glm-4.7, got quota-exhausted, and had no fallback model configured, so the agent failed before replying.

Fix:

#
openclaw models list
openclaw models set <provider/model-you-have-quota-for>
openclaw gateway restart

Or add a fallback so this doesn’t hard-fail next time:

openclaw models fallbacks add <provider/model-you-have-quota-for>
openclaw models status
openclaw gateway restart

If the Discord/session was manually pinned to zai/glm-4.7, also start a fresh session or switch it in chat with:

onyx wing
#

Am inteles, care ar fi cel mai ieftin model AI prin API pe care il pot rula prin OpenClaw si sa fie si capabil, auzisem la un moment dat de Claude Haiku 4.5 la 0.01 centi de pe Github Models

gusty harnessBOT
#

Da, Haiku 4.5 e o alegere rezonabilă „ieftin dar capabil”, dar nu aș miza pe cifra de 0.01 cenți fără să verifici în billing.

Pe scurt:

  • Anthropic direct API: Claude Haiku 4.5 e listat la $1 / 1M input tokens și $5 / 1M output tokens. (anthropic.com)
  • GitHub Copilot pricing: tabela GitHub arată Claude Haiku 4.5 la aceleași valori: $1 input / $0.10 cached / $1.25 cache write / $5 output per 1M tokens. (docs.github.com)
  • GitHub Models direct e alt sistem: facturează prin “token units” la $0.00001 per token unit, cu multiplicatori diferiți pe model. (docs.github.com)

Pentru OpenClaw, dacă ai Copilot, încearcă ruta Copilot:

openclaw models auth login-github-copilot
openclaw models list --provider github-copilot
openclaw models set <model-exact-din-listă>
openclaw gateway restart
#

Dacă mergi pe Anthropic API key:

openclaw onboard --auth-choice apiKey
openclaw models list --provider anthropic
openclaw models set <haiku-ref-exact-din-listă>
openclaw models fallbacks add <alt-model>
openclaw gateway restart

Eu aș pune modelul ieftin ca primary și un fallback configurat, ca să nu mai moară agentul la primul quota/rate-limit. Pentru recomandări „care model e cel mai bun azi”, întreabă și în #models — prețurile și calitatea se schimbă prea repede.

Docs: https://docs.openclaw.ai/providers/github-copilot și https://docs.openclaw.ai/concepts/models

onyx wing
#

Ce modele mai exista care sa poata rula reliable pe 10$ pe luna

gusty harnessBOT
#

Pentru $10/lună, eu aș gândi așa:

  1. Cel mai simplu/reliable: GitHub Copilot Pro + OpenClaw github-copilot

    • Copilot Pro e $10/lună, dar are limitări/premium requests, nu e API nelimitat clasic. Atenție: GitHub zice că din 1 iunie 2026 mută Copilot spre usage-based billing. (docs.github.com)
    • În OpenClaw:
      openclaw models auth login-github-copilot
      openclaw models list --provider github-copilot
      
  2. Cel mai ieftin pay-as-you-go: Gemini Flash-Lite / Flash

    • Gemini 2.5 Flash-Lite e foarte ieftin: $0.10 input / $0.40 output per 1M tokens; Flash e $0.30 / $2.50. Bun pentru volum mare, dar uneori mai slab la tool-use complex decât Claude/OpenAI. (ai.google.dev)
  3. Mai capabil dar consumă $10 mai repede: Claude Haiku 4.5

    • Haiku 4.5: $1 input / $5 output per 1M tokens. Bun compromis pentru agent/tool-use, dar dacă ai conversații lungi zilnic, $10 se duc repede. (anthropic.com)
  4. OpenRouter cu $10 credit + fallback

Google AI for Developers

Gemini Developer API Pricing

Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.

#
  • Bun dacă vrei multe opțiuni și spend control. OpenRouter spune că pay-as-you-go n-are minim, are modele free/paid, bugete/caps, și routing/fallback. (openrouter.ai)

Recomandarea mea practică: GitHub Copilot dacă vrei cost fix, sau Gemini Flash-Lite primary + Haiku fallback dacă vrei API ieftin și controlat.

În OpenClaw, fă neapărat fallback:

openclaw models set <model-ieftin>
openclaw models fallbacks add <model-mai-capabil>
openclaw models status
openclaw gateway restart

Și întreabă și în #models — acolo lumea are feedback fresh despre ce e reliable acum.

OpenRouter

Transparent pricing for OpenRouter. Pay only for what you use with access to 300+ AI models. Free tier, Pay-as-you-go, and Enterprise plans available.

onyx wing
#

Ma gandeam si la DeepSeek V4 Pro sau Flash

gusty harnessBOT
#

Da — DeepSeek V4 Flash ar fi chiar prima mea alegere pentru buget de ~$10/lună.

  • V4 Flash: foarte ieftin — $0.14 / 1M input cache miss, $0.028 cache hit, $0.28 / 1M output. Pentru un agent personal, asta e excelent.
  • V4 Pro: mai capabil, dar costă mult mai mult — $1.74 input, $0.145 cache hit, $3.48 output / 1M. L-aș folosi ca fallback, nu primary, dacă vrei să rămâi la $10. DeepSeek listează ambele cu context 1M, max output 384K, JSON output și tool calls. (api-docs.deepseek.com)

Config recomandat:

openclaw onboard --auth-choice deepseek-api-key
openclaw models list --provider deepseek
openclaw models set deepseek/deepseek-v4-flash
openclaw models fallbacks add deepseek/deepseek-v4-pro
openclaw gateway restart

Dacă modelul V4 nu apare încă în models list, fă update la OpenClaw / verifică cu:

The prices listed below are in units of per 1M tokens. A token, the smallest unit of text that the model recognizes, can be a word, a number, or even a punctuation mark. We will bill based on the total number of input and output tokens by the model.

#
openclaw models list --all --provider deepseek

Pe scurt: Flash primary, Pro fallback. Asta e probabil cel mai bun raport cost/capabilitate pentru cazul tău acum. Pentru feedback real de stabilitate, întreabă și în #models — acolo se schimbă rapid lucrurile.

onyx wing
#

Ai mai avea si alte sugestii, inainte aveam MiniMax M2.7, si imi placea mult, dar scuipa multe caractere in chineza, i-am redus din temperature, si i-am scos thinking si reasoning, dar inca persista problema

gusty harnessBOT
#

Da — dacă MiniMax îți bagă caractere chinezești chiar după temp/reasoning off, eu n-aș mai pierde mult timp cu el ca primary. Probabil nu e doar temperature; poate fi endpoint/region, streaming reasoning leak, sau sesiune veche contaminată.

Aș încerca așa:

  1. Verifică dacă ești pe endpoint CN

    openclaw models status
    openclaw config get models.providers.minimax.baseUrl
    

    Dacă vezi api.minimaxi.com, treci pe global:

    openclaw onboard --auth-choice minimax-global-api
    # sau minimax-global-oauth dacă folosești Coding Plan
    
  2. După orice schimbare de model:

#
   /new

Sesiunile vechi pot păstra context/răspunsuri dubioase.

  1. Pentru buget mic, aș testa în ordinea asta:
    • DeepSeek V4 Flash primary + V4 Pro fallback — docs DeepSeek arată tool calls, JSON output, 1M context, thinking/non-thinking. (api-docs.deepseek.com)
    • Gemini Flash-Lite / Flash pentru volum foarte ieftin; Google listează Flash-Lite la $0.10 input / $0.40 output per 1M tokens. (ai.google.dev)
    • Claude Haiku 4.5 dacă vrei mai puține ciudățenii de limbă/tool-use, dar e mai scump: ~$1 input / $5 output per 1M. (cloudprice.net)
    • MiniMax M2.7 doar fallback sau pentru task-uri unde ai văzut că merge bine; prețul e bun, dar dacă îți poluează output-ul, nu merită ca daily driver. (openrouter.ai)

Config practic:

openclaw models set deepseek/deepseek-v4-flash
openclaw models fallbacks add deepseek/deepseek-v4-pro
Google AI for Developers

Gemini Developer API Pricing

MiniMax-M2.7 is a next-generation large language model designed for autonomous, real-world productivity and continuous improvement. Built to actively participate in its own evolution, M2.7 integrates advanced agentic capabilities through multi-agent collaboration, enabling it to plan, execute, and refine complex tasks across dynamic environments...

#
openclaw models fallbacks add anthropic/claude-haiku-4-5
openclaw gateway restart

Model refs exacte verifică-le cu openclaw models list --provider ....

Și sincer: întreabă și în #models. Pentru “ce e reliable azi”, comunitatea are feedback mai fresh decât orice docs.

onyx wing
#

MiniMax models (M2.5/M2.7) often output Chinese characters or mix them into other languages (English, Korean) because of semantic vector confusion in the training data, even when prompted otherwise. This issue persists across API and chatbot usage (e.g., OpenClaw) and is often caused by training data leakage or model preference for Chinese during multilingual tasks.Key Aspects of the MiniMax Language Mixing Problem:Persistent Language Mixing: Users report Chinese characters appearing in Korean, Spanish, and English outputs, even when instructed to use only the target language.Source of Error: The issue is linked to how MiniMax models handle multilingual tokens, specifically a "semantic vector confusion" where Chinese characters are inserted into non-Chinese responses.Performance in Other Languages: While proficient in Chinese, the model struggle with other languages, often confusing characters in Japanese and failing to provide consistent non-Chinese outputs.Intermediate Output Issues: In some cases, such as when using the DeepSeek within MiniMax API, intermediate steps or logs can switch to Chinese, disrupting workflow.System Prompt Limitations: Simple instructions like "only use English" or "only use Korean" may be ignored due to the underlying model architecture's bias.To mitigate this, users have reported success by reinforcing the required language in the system prompt, although a complete, permanent fix for all generations is challenging given the current model architecture.

Asta am gasit pe internet despre problema cu Minimax

gusty harnessBOT
#

Da, asta confirmă ce suspectam. Aș formula puțin mai prudent: nu știm sigur că e “semantic vector confusion”, dar comportamentul e real. MiniMax M2.7 are interleaved thinking, iar în unele formate API thinking-ul ajunge injectat în content ca <think>…</think>; OpenClaw a avut și issue-uri unde MiniMax/Kimi leaking thinking apare direct în text, uneori amestecat cu chineză. (platform.minimax.io)

Ce aș face practic:

  1. Nu l-aș mai folosi ca primary pentru română/engleză dacă îți poluează output-ul.
  2. Dacă tot vrei MiniMax, testează direct global API, nu proxy/portal/OpenRouter:
    openclaw onboard --auth-choice minimax-global-api
    openclaw models set minimax/MiniMax-M2.7
    /new
    
    OpenClaw zice că pe path-ul Anthropic-compatible pentru minimax/* dezactivează thinking implicit, tocmai ca să evite leak-ul de reasoning. (docs.openclaw.ai)
  3. Pune în AGENTS.md o regulă scurtă:
    Always reply in Romanian unless asked otherwise.
    Never output Chinese/Japanese/Korean characters unless the user explicitly asks for them.
    
#
   If such characters appear, rewrite the answer before sending.

Ajută, dar nu garantează.

Recomandarea mea rămâne: DeepSeek V4 Flash primary, DeepSeek V4 Pro sau Haiku fallback, iar MiniMax doar fallback/experimente. Dacă problema persistă și pe minimax-global-api + /new, atunci e model/provider-level și nu merită să te lupți cu el.

onyx wing
#

Este deepseek v4 flash suficient pentru un workflow pentru o agentie de marketing? Deoarece cu asta ma ocup, am o agentie de marketing

gusty harnessBOT
#

Da — DeepSeek V4 Flash e suficient ca “daily driver” pentru multe workflow-uri de agenție, mai ales dacă îl folosești pentru volum:

  • idei de campanii
  • research inițial
  • structuri de ads / emailuri / landing pages
  • sumarizare call-uri / brief-uri
  • variante de copy
  • analiză competitor rapidă
  • social captions / calendare editoriale
  • task-uri repetitive în OpenClaw

Dar nu l-aș folosi singur pentru tot. Pentru agenție, recomand setup pe 2 niveluri:

openclaw models set deepseek/deepseek-v4-flash
openclaw models fallbacks add deepseek/deepseek-v4-pro
#
# opțional, dacă ai Anthropic:
openclaw models fallbacks add anthropic/claude-haiku-4-5
openclaw gateway restart

Flash = volum + rapid + ieftin.
Pro/Haiku = fallback pentru brief-uri importante, strategie, copy final, analiză mai sensibilă.

Unde aș folosi fallback/alt model:

  • strategii pentru clienți mari
  • copy final pentru ads cu buget mare
  • audituri unde contează mult acuratețea
  • tone-of-voice foarte fin
  • decizii bazate pe date financiare/legale
  • conținut care merge direct public fără revizie
#

Workflow bun: lasă Flash să producă 80% din muncă, apoi cere unui model mai capabil “critic/reviewer pass”. Asta ține costul jos fără să sacrifici calitatea.

onyx wing
#

Ma gandeam la abonamentul MiniMax pentru ca mai codez si landing pages, dar MiniMax nu intelege front end design, ma simt limitat de bani sincer

gusty harnessBOT
#

Sincer: dacă MiniMax nu îți prinde frontend/design-ul, nu aș cumpăra abonamentul doar pentru landing pages. Te înțeleg cu bugetul — aici aș optimiza workflow-ul, nu aș căuta “un model magic”.

Pentru agenție + landing pages, eu aș face așa:

  1. DeepSeek V4 Flash ca model de volum
    Brief, copy, structură, secțiuni, research, variații. E foarte ieftin oficial: $0.14 input / $0.28 output per 1M tokens pentru Flash. (api-docs.deepseek.com)

  2. Folosește design references, nu prompt gol
    Dă-i 2–3 exemple de landing pages + culori + stil: “premium SaaS”, “bold agency”, “Apple-like minimal”, etc. Modelele slabe la design se descurcă mult mai bine când le dai referință clară.

  3. Pentru cod frontend: component libraries gratuite
    Cere-i să construiască în Tailwind + shadcn/ui / DaisyUI / Flowbite-style components. Nu-l lăsa să “inventeze design” complet.

  4. Un fallback mai scump doar pentru review final
    Gen: “critique this landing page for conversion/design/copy, then suggest exact fixes”. Asta consumă puțin și ridică mult calitatea.

Copilot Pro ar fi fost interesant la $10/lună pentru coding, dar GitHub zice că new sign-ups pentru Pro/Pro+ sunt temporar paused din 20 aprilie 2026, iar din 1 iunie 2026 trec spre usage-based billing, deci nu l-aș trata ca soluție garantată dacă nu îl ai deja. (docs.github.com)

#

Prompt util pentru landing pages:

Create a landing page for [client] using Tailwind.
Style reference: [describe site/look].
Goal: generate leads.
Audience: [audience].
Sections: hero, pain points, benefits, proof, process, pricing/offer, FAQ, CTA.
Before coding, propose a visual direction and conversion strategy.
Then build clean responsive HTML/React components.
Avoid generic gradients and random emojis.

Pe scurt: DeepSeek Flash + template/design references + occasional stronger review pass. Asta e probabil cel mai bun raport calitate/bani până când ai buget pentru modele mai scumpe.

onyx wing
#

Oamenii recomandau Mimo V2 Pro sau Qwen 3.6 Plus, am de unde sa iau API pentru ele si care sunt preturile, si ar fi mai bune pentru OpenClaw decat Deepseek?

gusty harnessBOT
#

Da — dar aș separa clar:

MiMo V2 Pro = Xiaomi MiMo, nu MiniMax.

De unde iei API

Qwen 3.6 Plus

openclaw onboard --auth-choice qwen-standard-api-key
openclaw models set qwen/qwen3.6-plus

Pentru qwen3.6-plus, OpenClaw recomandă endpoint-ul Standard / pay-as-you-go, nu Coding Plan, fiindcă Coding Plan poate rămâne în urmă. (docs.openclaw.ai)

Preț oficial Qwen Cloud:

  • qwen3.6-plus până la 256K context: $0.50 input / $3.00 output per 1M tokens
  • 256K–1M context: $2.00 input / $6.00 output (docs.qwencloud.com)
#

Xiaomi MiMo V2 Pro

openclaw onboard --auth-choice xiaomi-api-key
openclaw models set xiaomi/mimo-v2-pro

OpenClaw are provider xiaomi, folosește XIAOMI_API_KEY, iar cheia se creează în Xiaomi MiMo console. Modelele listate includ xiaomi/mimo-v2-flash, xiaomi/mimo-v2-pro, xiaomi/mimo-v2-omni. (docs.openclaw.ai)

Preț via OpenRouter/Vercel listat de CloudPrice:

  • MiMo V2 Pro: $1.00 input / $3.00 output per 1M
  • MiMo V2 Flash: $0.09 input / $0.29 output per 1M (cloudprice.net)

Mai bune decât DeepSeek?

Pentru buget mic, DeepSeek V4 Flash rămâne best value: $0.14 input / $0.28 output, 1M context, tool calls, JSON output. (api-docs.deepseek.com)

Recomandarea mea pentru agenția ta:

#
openclaw models set deepseek/deepseek-v4-flash
openclaw models fallbacks add qwen/qwen3.6-plus
openclaw models fallbacks add xiaomi/mimo-v2-pro
openclaw gateway restart

Adică:

  • DeepSeek V4 Flash = daily driver ieftin.
  • Qwen 3.6 Plus = landing pages, frontend, multimodal, “vibe coding”.
  • MiMo V2 Pro = test ca fallback, mai ales long-context/agent tasks, dar nu l-aș pune primary până nu vezi cum se comportă cu tools în OpenClaw.

Pentru frontend/design, eu aș testa prima dată Qwen 3.6 Plus înainte de MiMo Pro.

onyx wing
#

Cam scumpicel Qwen 3.6 Plus, imi poti cauta pricingul exact si cea mai buna oferta ?

gusty harnessBOT
#

Da, am verificat pricing-ul actual.

Qwen 3.6 Plus — preț exact:

| Rută                           | Preț                                                                | Observație                              |
| ------------------------------ | ------------------------------------------------------------------- | --------------------------------------- |
| Qwen Cloud / DashScope oficial | $0.50 input / $3.00 output per 1M până la 256K context              | peste 256K urcă la $2 input / $6 output |
| OpenRouter paid                | $0.325 input / $1.95 output per 1M                                  | cea mai ieftină rută paid găsită        |
| Vercel AI Gateway              | $0.50 input / $3.00 output, cache read $0.10/M, cache write $0.63/M | are $5 credite/lună pentru free users   |

Surse: Qwen Cloud oficial pentru tiering și free quota/batch discount; OpenRouter pentru prețul $0.325/$1.95; Vercel pentru pricing + credits. (docs.qwencloud.com)

Cea mai bună ofertă pentru tine:

• Dacă vrei cel mai ieftin Qwen 3.6 Plus paid: OpenRouter.
• Dacă vrei producție mai curată / direct vendor / client work: Qwen Cloud Standard, dar ai grijă să nu treci peste 256K context.