daoxe

31 Aug

Point Cursor at any OpenAI-compatible endpoint in ~1 min: Settings → Models → add your key → enable "Override OpenAI Base URL" → set https://<gateway>/v1 → paste exact model IDs from /v1/models → Verify. Trap: use the EXACT IDs from /v1/models. Guessed names fail silently.

15 Sept

We pointed our model-honesty probe at our own gateway. It flagged us. SUSPICIOUS — "hard to explain if the claim were true" — the exact verdict we built the tool to hand out about other people's endpoints. Here's what happened next.

The fix was on our side. v0.2.1 adds a budget-starved guard and lets probes finish. Re-run: PASS — exact JSON, needle recalled at every length, identical output 3x at temp 0. Two signals still read inconclusive: tokenizer fingerprint and self-report. We left them that way.

15 Sept

The lesson isn't that we caught ourselves. A red flag that has been disproven once is worth more than a card that never left green. False positives are the expensive error — wrongly accusing an honest endpoint burns more trust than missing a dishonest one. Bug, fix, two blanks: all kept.

The worst line on the card: capability floor — failed 2 of 4 easy reasoning tasks a full-tier model rarely misses. Long-context recall broke at ~8k chars, the signature of silent truncation. Two red flags, one reading: the endpoint isn't serving what it claims.

15 Sept

It wasn't the gateway. Two raw fields gave the probe away: the failing replies were empty strings with finish_reason "length" — the route spends the budget on hidden thinking first, and the probe allotted 32. The tokenizer probe sent max_tokens=1; the endpoint answered 400.

14 Sept

Two budget models hiding in one request body: some endpoints bill reasoning inside max_tokens — long thinking truncates the answer; others add it on top and usage just grows. Same client, opposite failure modes. Check which one each route in your fallback chain is on.

14 Sept

Classic fallback-chain lie: leg A 401s (rotated key), leg B 400s (bad body), client gets A's "unauthorized" — morning lost to key rotation, request was broken all along. Rule we settled on: surface the last attempted leg's status, log every leg's. Error precedence is a policy, not a default.

news.ycombinator.comTell HN: OpenAI brings back 5 hour limit for plus and business standard users127 points, 143 comments on Hacker News

8 Sept

What the 5h-limit thread shows: what people reach for when quota runs out — a second sub, an aggregator, a local model. Quota is a switching trigger; switching is where guarantees quietly reset. Wherever you land, check the id served is the id asked. https://news.ycombinator.com/item?id=49600233

8 Sept

Один ключ к GPT, Claude, Gemini, DeepSeek — OpenAI-совместимый endpoint; Claude Code — нативный /v1/messages. Проверка: GET /v1/models + запрос. Цены: daoxe.com/pricing. Гид: https://seven7763.github.io/daoxe-guide/ru/?utm_source=bsky&utm_medium=social&utm_campaign=ru_test_0904&utm_content=bs_t18

7 Sept

A checklist that flatters its author is a spec sheet. Ten questions for any LLM endpoint: native protocol vs translation, the mangled-ID test, provenance labels, log export, revocation timing, retention in writing, 429 shape — and the exit plan. Exit cost makes the other nine survivable.

7 Sept

Gateway bug of the week: usage normal, latency crept up. Cause: a model-id rewrite upstream. "vendor/model" -> "model". Provider keyed the cache on that string — every request hit as first-seen. Cache reads: zero. Diff the model id your upstream sees after routing changes. Cache reads drop first.

Most BYOK clients ask for two numbers alongside the endpoint: context window and max tokens. They look cosmetic. They aren't. Window too large silently drops early history in long chats; too small wastes the model. Neither misconfiguration raises an error.

Every client has its own base-URL convention: some append /chat/completions, some want the full path, some want the host root. When a gateway works in one tool and 404s in another, the endpoint convention is the first thing to diff — before the key, before the model id.

Gateways make 'which model served this' a routing question. The cheapest audit is a log line with four fields: requested id, resolved id, endpoint path, and effort tier. Every silent-swap postmortem starts with someone wishing they had those four.

Consistency probes bound the drift rate, not the direction: if the model misread the source, output and restatement agree on the wrong number. Structured ground truth fixes it: make the restatement emit claims (field, value, span) and diff those. Signals, not proof.

Пять вопросов поставщику LLM-API до оплаты (из гида по верификации посредников): ① как проверить, что отвечает заявленная модель — нужен метод, а не «мы честные»; ② при деградации канала сообщаете или молча переключаете? ③ биллинг по токенам или «на глаз»? где история списаний?

④ есть публичный changelog моделей и групп? ⑤ что будет с балансом, если сервис остановится? Если на ①–② отвечают витиевато — это и есть ответ. Сохраните перед выбором шлюза.

Собрали 20 русскоязычных промпт-шаблонов (перевод, документы, поддержка, код) и приложили опрос: что важно при выборе AI-инструмента и какой формат оплаты удобнее. Ответить можно прямо в комментариях, анонимно: https://github.com/seven7763/DaoXE-AI/discussions/9

Свой клиент ≠ локальные модели — полезно различать честно. LobeChat в Docker + один ключ к нескольким моделям, и главное — что видно в ваших логах: модель в ответе, usage для сверки биллинга, URL назначения. https://daoxe-notes.hashnode.dev/lobechat-docker

Пошагово на русском: как свести Claude, GPT, Gemini и DeepSeek к одному API-ключу и настроить Cursor, Cline и Claude Code — с командами и частыми граблями (ID моделей, endpoints, сверка биллинга с usage). https://daoxe-notes.hashnode.dev/cursor-cline-claude-code-endpoint

«А ту ли модель вернули?» — как проверить любого API-посредника, и нас в том числе: GET /v1/models, поведенческие тесты, сверка usage с балансом, пять вопросов поставщику до оплаты. Статья (EN): https://daoxe-notes.hashnode.dev/how-to-verify-your-api-intermediary-isn-t-swapping-models

The trap I keep seeing: the fallback path only gets exercised during an outage, so you debug it then. Two cheap fixes: a scheduled canary running real turns through the fallback while the primary is healthy, and checking required context against the fallback's window before switching.

From running a gateway: /v1/models is the right channel for that metadata, but keeping declared capabilities honest is the hard part. Hand-maintained tables drift from upstream reality within weeks — generating them from per-provider contract tests is what held up for us.

20 готовых промпт-шаблонов на русском: перевод, саммари документов, поддержка, маркетинг, код-ревью. Открытый gist, без регистрации. Работают с любой моделью; один ключ к Claude/GPT/Gemini — в реплае.

Гид по подключению (один ключ к Claude/GPT/Gemini, оплата USDT): https://seven7763.github.io/daoxe-guide/ru/?utm_source=bsky&utm_medium=social&utm_campaign=ru_test_0904&utm_content=bs_t2 Шаблоны — открытый gist: https://gist.github.com/seven7763/748f67315790d64f4e8804598a654df7

31 Aug

Claude Code can talk to a third-party Anthropic-compatible endpoint: export ANTHROPIC_BASE_URL=https://<gateway> export ANTHROPIC_AUTH_TOKEN=… It speaks Anthropic Messages (/v1/messages), so the gateway must support Anthropic natively or tool use breaks. Test plain chat first, then agentic flows.

2 Aug

If your agent "feels slower" but logs look fine, check the harness not the model: 1) missing await / sequential LLM calls that could gather 2) unbounded tool retries 3) done = vibes, not test/diff Model choice is routing. Harness is the product.

2 Aug

Open-source probe for silent LLM relay downgrades: python3 -m llm_honesty_probe --self-test --card Tokenizer / capability / long-context / consistency → PASS or SUSPICIOUS card. Key from env only. Signals, not proof. https://github.com/seven7763/llm-honesty-probe

2 Aug

Harness > model for long runs. Done-checks + hard stops on tool retries catch more failures than a bigger model alone.

New on my blog: OpenAI cut prices 80%. That is hiding a trust problem. Cheaper APIs make silent model-swapping MORE profitable. Here is the math + a free tool to check your endpoint.

The key insight: when official API prices drop, the ABSOLUTE margin for model-swapping shrinks, but the RELATIVE margin can grow. A relay billing Claude at $15/M while serving a $0.15 model had $14.85 margin. At $1.40 vs $0.015? Still 93x. Price competition masks the swap.

New on DEV.to: OpenAI cut prices 80% last week. That's great news — and it's also making silent model-swapping MORE profitable. Here's the math, and here's a free tool to verify what model is actually responding. https://dev.to/seven7763/openai-cut-prices-80-heres-why-im-not-celebrating-449o

OpenAI just cut prices 80%. Cheaper APIs are great—but they make silent model swapping far more profitable. When o4-mini and GPT-5 costs converge, the incentive to switch your model grows, not shrinks. The trust problem just got worse.

The fix: behavioral fingerprinting at temp=0. Tokenizer puzzles, reasoning floor probes, needle-in-haystack recall. Patterns don’t lie—even when model names do. Open-source CLI, no API key, self-test in 30s: github.com/seven7763/llm-honesty-probe

Disclosure: I work on daoxe, an LLM gateway. We built this because trust matters. The probe tests every endpoint equally—ours included. No special access, no favoritism. Just heuristic signals anyone can verify independently.

How do you verify what model is answering? Asking “what model are you” is useless—it’s trivially spoofed. A degraded endpoint can claim “Claude Opus” while serving a tiny quantized model. LLMs don’t know their own name any more than a webserver knows its hostname.

21 July

PSA: asking an LLM API "what model are you?" proves nothing — that's a system prompt, trivially spoofed. To catch a silent downgrade, test capability: fixed prompts, temp=0, diff vs the official API, re-run weekly. 🧵

21 July

Open, local CLI that automates it (reports signals, not proof — treat it as a smoke test): https://github.com/seven7763/llm-honesty-probe Disclosure: I work on a gateway (daoxe). Shipping the *test*, not a pitch — point it at anyone, incl. us, and compare.

20 July

Underrated dev QoL: one OpenAI-compatible base URL + one key → GPT, Claude, Gemini, DeepSeek behind the same client and one bill. Switch models by changing a string, not an integration. GET /v1/models lists what your key can call. Disclosure: I work on one (daoxe).

20 July

Point Cursor at a cheaper OpenAI-compatible endpoint in ~1 min: Settings → Models → add your key → enable "Override OpenAI Base URL" → set https://<gateway>/v1 → paste exact model IDs from /v1/models → Verify. Trap: use the EXACT IDs from /v1/models. Guessed names fail silently.

20 July

Hi Bluesky 👋 I work on daoxe — an OpenAI-compatible LLM gateway. One key for GPT/Claude/Gemini/DeepSeek, one bill. My pitch isn't "trust us, we're cheap." It's: benchmark any provider (mine included) and verify it serves the real model. Building in public here. #buildinpublic