lagindicator

Beyond Anthropic, OpenAI, and Google — The OpenRouter Provider Field for Agent & Coding Inference (Sep 2026)

Sat, 05 Sep 2026 00:00:00 +0800

Research date: 2026-09-05. All token figures are the OpenRouter providers page snapshot as fetched Sep 5, 2026, sorted by daily tokens. The big-three company routes — OpenAI, Google Vertex/Google AI Studio, Anthropic/Claude Platform on AWS — are intentionally removed for this post. For the big-three comparison, see Which AI API Provider Is Most Popular for Coding-Agent Inference?.

TL;DR

After Anthropic, OpenAI and Google, the largest OpenRouter provider routes by daily tokens are Tencent Cloud (3.7T/day), GMICloud (1.3T/day), Z.ai (982.6B/day), NVIDIA (723.6B/day) and NovitaAI (654.3B/day).
The field is not one market. It splits into four groups: official model-lab APIs (Tencent, Z.ai, DeepSeek, Baidu, Alibaba), serverless GPU/inference hosts (GMICloud, Novita, DeepInfra, and below them Together, Fireworks, Baseten, Modal, Parasail), cloud marketplaces (Amazon Bedrock, Azure), and coding-native specialists (Relace, NVIDIA NIM, StreamLake/KAT-Coder).
Token volume is not revenue. Several of these routes exist because their tokens are cheap; DeepInfra, for example, has the largest OpenRouter catalog (103 models) and 16.1T monthly tokens, while official DeepSeek processes 11.5T/month. On spend, Anthropic still dwarfs this entire table (see the companion popularity post).
If you care about proprietary code, the policy flags matter more than volume: most of the big non-big-three routes have zero retention and do not train on prompts, but DeepSeek and NVIDIA are flagged as training, and several Chinese official APIs retain prompts.

1. The ranking (OpenRouter, Sep 5, 2026)

OpenRouter’s providers page lists 83 provider routes with live token volume, policy flags, BYOK support, and model counts. This table keeps the twelve biggest routes after dropping OpenAI, Google Vertex, Google AI Studio, Anthropic and Claude Platform on AWS.

Coding-Agent Auto-Approval Modes (Sep 2026): What Codex Auto-Review Actually Is, and How Other Harnesses Handle It

Sat, 05 Sep 2026 00:00:00 +0800

Research date: 2026-09-05. Questions: (1) Is Codex CLI auto-approve simply “spawn an agent to review a tool call,” or does implementation matter? (2) If it is that simple, are the community DeepSeek Harness plugins safe enough? (3) What are other coding-agent harnesses — including OpenClaw — doing for the same problem, given that true full-auto on an unsandboxed home PC looks unsafe? OpenAI claims are cited to OpenAI Docs / the Alignment Research post; other tools cite their own official docs or open-source repos.

DeepSeek Harness Ecosystem as of 5 Sep 2026: Is Vanilla Enough, and Has the Plugin Community Converged?

Sat, 05 Sep 2026 00:00:00 +0800

Research date: 2026-09-05. Question: how mature is the ecosystem now, is the vanilla install enough for most use cases, has the community converged on must-have plugins, and does DSH have a permission mode like Codex’s auto-approve / agent-reviewed tool calling? Inline URLs are the primary evidence; community snapshots are labeled as such. The implementation-level cross-harness comparison lives in the companion page Coding-agent auto-approve modes (Sep 2026).

TL;DR

Core is still a developer preview, just past its first RC of 0.1.2. @deepseek-ai/dsh is at 0.1.2-rc.1 (published around 2 Sep 2026; npm, release notes), roughly three weeks after the 13 Aug launch. The release still carries the official safety notice: no security audit has been done and sandbox/approval/permission controls do not guarantee isolation (release notes).
Plugins are numerous but not “mature” in the marketplace sense. The dsh-plugin topic sits around 13.5k–14.6k repositories (dsh.works, dsharness.org), inflated by template spam and topic-riders. Curated quality directories keep only about 2k entries (Oh-My-DSH).
Vanilla is sufficient for the mainstream coding baseline. The shipped web/Standard profile exposes shell, file editing, glob/grep, web fetch/search, planning, skills, subagents, sessions, background jobs, and user questions out of the box (quickstart, tool catalog).
“Must-haves” have converged as categories, not yet as stable winners: plugin market/management, Web-UI workbench upgrades, vision for text-only routes, cross-session memory, browser control, and security/health audit tooling.
Permission modes (follow-up): vanilla DSH still ships only workspace-write + ask and danger-full-access + never — no built-in Codex-style “approve-for-me” / agent reviewer yet (permission presets, approval subsystem). The community has ported the pattern through auto-review/auto-approval plugins; the architecture has converged but no plugin is a stable winner.

1. Core project status

Channel	Version (5 Sep 2026)	Source
npm `@deepseek-ai/dsh`	`0.1.2-rc.1`	npm
PyPI `deepseek-harness-sdk`	`0.1.2rc1` wheel	PyPI
GitHub release	`v0.1.2-rc.1`, pre-release	release
Launch baseline	v0.1.0-rc.6, 13 Aug 2026	launch coverage

The path to 0.1.2-rc.1 was churny even for a preview: 0.1.0-rc.7 → rc.8 (19 Aug) → 0.1.1-rc.2 (21 Aug) → 0.1.2-alpha.1 (27 Aug) → alpha.2…alpha.5 (through 2 Sep) → 0.1.2-rc.1 (2–3 Sep) (community timeline, Aliyun tutorial). Release notes list pre-GA behavior: legacy APIProxy removed, “Code Mode” renamed “PTC mode”, optional SQLite session backend removed, and Session.events replaced by on-demand read APIs (release notes).

How Top Investors Think: A Group Review of Six Practitioner Books

Sat, 05 Sep 2026 00:00:00 +0800

Research date: Sep 5, 2026. This is a group review, not a deep dive on any single title: each book gets roughly equal space, and the goal is to see what “how a top investor thinks” looks like across different formats, geographies, and strategies.

Why these six belong on the same shelf

All six titles share one trait that most investing books lack: the author was the person putting real capital at risk, and the text is a by-product of doing the job rather than an attempt to teach a course. The formats are different — a hedge-fund CIO’s essays, a distressed-credit investor’s client memos, an entrepreneur-investor’s Q&A transcripts, a China value manager’s columns, a PE founder’s manifesto, and four decades of trader interviews — but each one exposes the same thing: an operating system for thinking under uncertainty, plus at least one real decision you can take apart.

Search Gateways Compared: Hound (master-fetch) vs SearXNG vs ddgs

Sat, 05 Sep 2026 00:00:00 +0800

A research article on the three open-source ways to get multi-engine web search without paying a search API: a local MCP server (Hound), a self-hosted metasearch service (SearXNG), and a search library (ddgs). Original research date: Aug 2, 2026. Revised: Sep 5, 2026, after a follow-up audit of Hound/master-fetch. Facts were verified by direct fetch (GitHub API, raw READMEs, official docs, PyPI metadata). Companion articles: web-search-ecosystem.md (provider ecosystem, rate limits, free quotas), same-model-different-cli.md, agent-loop-pattern.md.

The LLM ARR Debate (Sept 2026) — Anthropic's $65B Expectation Miss, the Gross-Up Fight, and the Capex Test

Sat, 05 Sep 2026 00:00:00 +0800

Research page: 2026-09-05. This is a dated snapshot of a fast-moving controversy, August 18 – September 5, 2026. Every figure is as reported by the cited source; treat run-rate numbers, leaked memos, and forecasts as claims rather than audited fact. For how these ARR numbers are constructed, see the companion page How LLM Companies Compute ARR.

TL;DR

On August 18, Bloomberg reported Anthropic’s July run rate at ~$65B (+38% from $47B in May, >7x the ~$9B at end-2025). Q2 preliminary revenue was >$11.5B (~14x YoY) with positive adjusted operating income.
The number was a miss against buy-side expectations of $70–80B from trackers (YipitData, TickerTrends) and a ~$78B expert forecast — and it knocked global tech down for three days (SOX ≈ −5%, Nasdaq −1.33%, then STAR 50 / ChiNext −5%+).
The methodological fight resurfaced: a leaked OpenAI CRO memo from April had accused Anthropic of “grossing up” cloud-marketplace revenue with Google/Amazon, allegedly overstating the then-$30B figure by ~$8B.
Product-level tracking sharpened the debate: Claude Code’s tracked ARR was $15.1B but grew only +5.2% MoM by August 10, while Codex’s $8.8B grew +20.8% MoM.
The stakes: Anthropic internally projects ~$190–200B of 2028 revenue behind a potential ~$2T IPO (as early as October). The S-1, not the press release, settles who was right.

1. What happened: an expectation gap, not a bad number

The August 18 Bloomberg disclosure was, by any absolute standard, enormous: ~$65B annualized in July, up ~38% from the $47B Anthropic disclosed with its May Series H, and more than seven times the ~$9B run rate at end-2025. Q2 preliminary revenue exceeded $11.5B, ~14x the year-ago quarter, with adjusted operating profit positive for the first time.

Which AI API Provider Is Most Popular for Coding-Agent Inference? Token Traffic, Community Usage, and Spend Rankings (Sep 2026)

Sat, 05 Sep 2026 00:00:00 +0800

Research date: 2026-09-05. All figures are snapshots, not audited totals. Provider popularity depends on whether you rank by routed token volume, dollar spend, developer installs, or paid coding-agent adoption — and each dataset sees a different slice of the market. Related context: The LLM ARR Debate (Sept 2026) covers Claude Code vs Codex revenue; DeepSeek V4 Flash provider cost index and GLM-5.3-Flash cost index cover the price side.

TL;DR

Anthropic (Claude) is the most popular provider for paid, quality-critical coding-agent inference. On the Vercel AI Gateway it holds 61% of June 2026 spend on 32% of tokens, ≥72% of spend in coding agents, back-office agents and app generation, and Claude Code is the largest paid coding product by tracked ARR (~$15.1B vs Codex’s ~$8.8B as of Aug 10).
DeepSeek is the most popular provider by raw token traffic on neutral developer routers. The Sep 4 OpenRouter snapshot puts DeepSeek at 19.0% of tracked provider tokens, ahead of OpenAI (15.7%), Z.ai (15.0%) and Tencent (12.9%); Anthropic is 4.4%. The coding-agent slice on Vercel is even more extreme: DeepSeek ran 49% of coding-agent tokens in May on 4% of the cost.
OpenAI has the largest developer/API ecosystem by installs and reach — ~427M monthly openai PyPI downloads in Aug 2026 and the largest first-party direct-API disclosed scale — and its Codex/GPT-5.6 family is the fastest climber on intensity metrics. Google leads by first-party consumer volume but is nearly absent from Vercel coding-agent tokens (<2% in June).
The honest answer: there is no single winner unless you fix the metric. If “most popular” means which provider gets paid for coding/agent inference, it is Anthropic. If it means which provider’s API keys move the most tokens through community routers, it is DeepSeek (with Z.ai/GLM, Tencent and Xiaomi close behind). If it means which provider developers install and build with most, it is still OpenAI.

1. The ranking question

“Which provider is most popular for AI API-key, LLM-agent, coding inference” is actually several questions:

Which Coding Harness Are Developers Actually Using? Token Traffic, Community Sentiment, and a Top-5 Snapshot (Sep 2026)

Sat, 05 Sep 2026 00:00:00 +0800

Research date: 2026-09-05. This is a dated popularity snapshot, not a capability benchmark. “Traffic” here means tokens routed through OpenRouter’s publicly tracked app pages by apps that opted into attribution; community signals are GitHub stars, marketplace/npm install data, and representative discussion sources. By request, Claude Code, Gemini CLI, and Codex CLI are excluded from the ranking but kept in tables where they are useful as scale references.

Companion context: LLM provider usage & traffic ranking (Sep 2026) covers the model/provider side of the same question, and Coding Agent CLIs: benchmarks and harness differences covers why the harness layer matters.

How Evals Meet Agent Memory: Codex Memories and Hermes Agent's “Grows With You”

Mon, 31 Aug 2026 00:00:00 +0800

Research date: Aug 31, 2026. Primary sources: official Codex memory documentation, the Hermes Agent README (Nous Research), Mem0’s technical writeup on Hermes memory, and the cited memory benchmarks. For the daily workflow this study supports, see Progressive, Uninterrupted Eval Building.

TL;DR

Memory and evals are the two halves of the same growth loop:

Memory accumulates — facts, preferences, session history, reusable skills.
Evals verify — whether the accumulated knowledge actually makes the agent better.

Hermes Agent’s claim to be “the agent that grows with you” is real in mechanism — persistent memory files, full-text session search, autonomous skill creation and improvement, user modeling — but growth without measurement is vibes. An eval harness is the instrument that turns “the agent remembers more” into “the agent is measurably better and did not regress.” The two systems feed each other: memory supplies eval cases, eval results become memory, and the memory system itself is a thing to evaluate.

Progressive, Uninterrupted Eval Building: A Daily Workflow for Agent-Driven Development

Mon, 31 Aug 2026 00:00:00 +0800

Research date: Aug 31, 2026. Workflow applies to Codex CLI and any coding agent with a scriptable non-interactive mode; the file formats are tool-neutral. For how evals interact with persistent memory systems, see the companion study Evals and Agent Memory.

TL;DR

An evaluation suite should be built as a by-product of daily agent-assisted work, not as a separate project. The pattern has three rules:

Capture-first: every time the agent gets something wrong — or you have to correct it — append one JSONL row to the golden set. Cost target: under a minute, zero context switch.
Checkpoint runs: run a tiny smoke suite at natural boundaries (end of task, before commit) and the full suite in CI or nightly. Never run the full suite inside your interactive flow.
Agent-maintained: the agent appends cases, runs the suites, fixes the harness, and distills weekly. The human reviews diffs, not every row.

This mirrors the pattern OpenAI documents for testing agent skills: define success, start with a small prompt set, grow it from real failures, grade with deterministic checks on traces (codex exec --json) and a rubric pass (--output-schema), and let failures drive coverage (https://developers.openai.com/blog/eval-skills).

China A-Shares: Sector Rotation & Hot Money (游资) — Which Models Detect the Effect, How to Monitor It

Sun, 30 Aug 2026 00:00:00 +0800

Research date: Aug 30, 2026. Follow-up research note on the short-horizon China A-share effect: sector/concept momentum over roughly 3–10 trading days driven by macro headline sentiment, policy releases, and hot-money (游资) capital concentration, plus the Dragon-Tiger List (龙虎榜) and order-flow-imbalance signals that precede the fast mean-reversion. Academic claims are cited inline; practitioner sources are labeled as directional experience, not peer-reviewed proof.

Companion posts: Daily Market Scan Guide, Market News Timeline — 2026-08-24, Daily Market Scan — 2026-08-01.

How LLM Companies Compute ARR — Run-Rate Math, Revenue Recognition, and What the Headlines Actually Mean

Sun, 30 Aug 2026 00:00:00 +0800

Research page: 2026-08-30. All figures are as reported by the cited sources; most are annualized run rates or preliminary disclosures, not audited GAAP revenue. Treat every number as a snapshot — these companies change disclosures monthly and both OpenAI and Anthropic have confidential S-1 filings pending.

TL;DR

When the news says an AI lab has “$40B ARR” or “$65B ARR”, it is usually not annual recurring revenue in the classic SaaS sense. It is a run rate: take a short recent revenue period and multiply it up to a year.

Codex CLI Orchestration Landscape (Aug 2026): Open-Source Layers That Drive the CLI

Sat, 29 Aug 2026 00:00:00 +0800

Research date: Aug 29, 2026. Question: can you run Codex CLI as an OpenClaw-style long-running agent, what are the real architectural limits, and which open-source orchestration layers can drive the CLI while keeping your existing Codex setup untouched? Every claim is cited inline. Companion articles: multi-cli-orchestration.md, codex-mcp-server-pattern.md, coding-agent-harness-boundary.md, agent-loop-pattern.md.

TL;DR

Yes, Codex CLI can behave like a long-running agent — but it is a turn-based agent runtime, not a daemon. Something has to own the outer loop (scheduling, retries, channels, state), and that something is an orchestration layer. The good news: several open-source orchestration layers drive the Codex CLI/app-server layer itself rather than replacing it, so your existing ~/.codex setup, auth, and workflows can stay. The bad news: “long-running” moves the infinite-loop risk from the per-turn agent loop to your outer loop, which needs explicit bounds.

DeepSeek V4 Flash Provider Pricing — A Blended Cost Index for Cache Hits, Cache Misses, and Output

Sat, 29 Aug 2026 00:00:00 +0800

Research page: 2026-08-29. All prices are list prices in USD per 1M tokens from the cited pages. DeepSeek changed its official pricing structure on 2026-08-17 (peak/off-peak), so snapshot dates matter. Verify before committing to any route.

TL;DR

The claim “third-party providers advertise attractive open-model pricing, but the hidden cost is the cache-hit price, which is usually much higher than the original provider” is directionally true for DeepSeek, but the ranking flips by model tier and workload mix. For V4 Flash, among the kept routes Fireworks beats official at every hit mix (80) and DeepInfra straddles it (72–124) — the rest lose at 95%+ cache-hit ratios; for V4 Pro, official DeepSeek wins on every line.
A pricing page has four numbers, not one: cache-hit input, cache-miss input, output — plus peak/off-peak for the official API. We collapse them into a single index: E = r·[h·H + (1−h)·M] + O, normalized to official = 100 (lower = cheaper).
Calibrated on two real coding-agent sessions (h ≈ 99.5% and 96.8%), the V4 Flash sync ranking (StreamLake excluded — region; OpenRouter aggregate excluded — unusable endpoints) is: Fireworks 80 < DeepSeek official 100 ≈ DeepInfra 101 < SiliconFlow 157 < DigitalOcean 162 < Together 167 ≈ Novita 167. Batch mode flips the top: Fireworks 40 < DeepInfra ~81 < official 100.
Companion: GLM-5.3-Flash net-of-fee index — the high-cache leader is Vercel’s pinned cheapest upstream (~43.5 even after its ~12% fee), ahead of Fireworks batch (~48.5) and Z.ai promo (50); low cache hit flips to batch and promo (~50) with Vercel-pinned dropping to ~70.5. Fees cost ~5 index points on OpenRouter routes and ~4.6 on the Vercel-pinned route (see §9).
Trust: both recommended providers are NVIDIA-backed, SOC 2 / ISO 27001 certified per company disclosures, and founded in 2022. Fireworks is the enterprise heavyweight ($17.5B valuation, $1B+ ARR, Sequoia/Benchmark/ Lightspeed/Index, Microsoft partnership, customers Uber/Shopify/GitLab); DeepInfra is the lean specialist (~$133M total, 500 Global/Felicis plus NVIDIA/Samsung, zero retention, 99.982% uptime SLA) (see §10).
The cheapest advertised routes are excluded, not recommended: StreamLake is CN-hosted with unknown prompt-logging (region constraint), and OpenRouter’s aggregate low price routes to endpoints with no throughput/retention guarantees — pin one provider or pay the 5.5% fee on a real host. Among the kept routes the levers are batch discounts (Fireworks 50%, DeepInfra 20%) and the fee rails (Fireworks/DeepInfra net fee 0; OpenRouter 5.5%; Vercel ~12%).

1. The problem: four prices, not one

Official DeepSeek API (snapshot 2026-08-23)

DeepSeek’s official pricing is now peak/off-peak (peak = weekdays 01:00–04:00 and 06:00–10:00 UTC; off-peak is half of peak):

GLM-5.3-Flash Cost Index — Provider Ranking and Session-Level Cost vs DeepSeek V4 Flash

Sat, 29 Aug 2026 00:00:00 +0800

Research page: 2026-08-29. All prices are USD per 1M tokens from the cited pages. GLM-5.3-Flash (Z.ai / Zhipu, 320B-A18B, MIT, natively multimodal, 1M context) was released 2026-08-26 after a week as the anonymous “Ox-Alpha” model. A 50% launch discount runs until 2026-09-09 16:00 UTC, after which list prices apply.

TL;DR

The claim “GLM-5.3-Flash is the most cost efficient model at lower cost right now” is half right. On Artificial Analysis it is the cheapest per Intelligence Index task among frontier-tier models (57 IQ at $0.09/task, vs DeepSeek V4 Flash 52/$0.11, Qwen3.8-Flash-Next 56/$0.10, Gemini 3.7 Flash 56/$0.40). But AA ranks it only #16/111 on cost per task (2/4 cost units), and for cache-heavy coding agents — like our two real sessions — DeepSeek V4 Flash and Qwen3.8-Flash-Next are cheaper.
Unlike DeepSeek, there is no hidden cache-hit markup in the GLM ecosystem: every major host matches or undercuts Z.ai’s $0.03 cached-input price. The index is therefore nearly flat across providers; the real levers are the promo, batch discounts, realized cache-hit rate, and verbosity.
Provider index (Z.ai official = 100, lower cheaper): promo until Sep 9 = 50, Fireworks batch = 49, Makora = 82–85, Fireworks serverless = 97, everyone else at list parity = 100.
Per-session example at the measured hit rates (S1 h=99.51%, S2 h=96.76%): GLM official costs $0.463 / $0.050 per session vs DeepSeek V4 Flash official $0.215 / $0.038 — DeepSeek is 2.15x cheaper in S1, 1.31x in S2.

1. Claim check

Intelligence (AA Intelligence Index v4.1.1) is kept as a separate column and is not blended into the cost index, so the two questions stay answerable independently: “cheapest way to run this workload” (index) vs “how smart is it” (IQ column).

LangGraph in 2026 — Deep Dive and Community Views: Low-Code?, Schema Evolution, and Over-Engineering

Sat, 29 Aug 2026 00:00:00 +0800

Research date: Aug 29, 2026. Release facts are from the langchain-ai/langgraph GitHub release history (fetched Aug 29, 2026) and cited docs/posts. Community views are from primary posts (DEV, Thoughtworks radar) and secondary summaries of Reddit threads, cited inline — treat Reddit quotes reported via aggregators as directional, not verbatim.

Companion post: The AI Agent SDK Landscape — Popularity, Design Philosophy, and Downstream Projects.

Short answer

Is LangGraph designed for low-code? No — it is explicitly low-level and code-first. Its official visual layer (LangGraph Studio, now folded into LangSmith Studio) is a debugger and state editor, not a no-code authoring tool. What is true: the graph-shaped model makes LangGraph easy to visualize and wrap in a visual builder, which is why a third-party no-code ecosystem (LangConfig, Graphweave, IGNode, Open Agent Builder, UiPath) has grown on top of it.

The AI Agent SDK Landscape — Popularity, Design Philosophy, and Downstream Projects (Aug 2026)

Sat, 29 Aug 2026 00:00:00 +0800

Research date: Aug 29, 2026. Star counts were pulled live from the GitHub API; Python download counts are the pypistats 30-day rolling window; npm counts are the npm registry’s last-full-month totals (all fetched Aug 28–29, 2026). Ecosystem and community claims are sourced inline. This is a popularity, philosophy, and adoption snapshot — not a capability benchmark.

Companion post: LangGraph in 2026 — Deep Dive and Community Views.

Part 1 — Popularity

Short answer

OpenAI Agents SDK is the more popular of the two by most measures, and the gap widened a lot in 2026. It has more GitHub stars (29.0k vs 19.6k), ~3x the downloads of the main Python package (38.8M vs 12.6M per month), a TypeScript SDK that Pydantic AI lacks, and a much higher momentum score on the daily framework tracker (79.0 vs 54.1).

Market News Timeline — 2026-08-24

Mon, 24 Aug 2026 00:00:00 +0800

Research run covering the trading session of Mon Aug 24, 2026. The table below isolates the news release time for each theme driver. Underlying events (e.g. the Iran sanction plan, the tariff threat, the CXMT chip rumor) do not have a single confirmed timestamp; only article-publish times are listed.

Context: Nasdaq Composite traded around 26,002–26,057 (-0.5% to -0.7% on the day) through late morning ET. US market timezone is EDT (UTC-4). Your local timezone here is GMT+8 (UTC+8).

Software Testing Types: Unit, Integration, Regression, Smoke, E2E, and System Tests

Sat, 22 Aug 2026 00:00:00 +0800

Research date: Aug 22, 2026. This is a synthesis of standard software-testing terminology applied to a concrete service shape (FastAPI reading SQL Server, transforming with pandas, streaming pyarrow back to an HTTP client). The taxonomy is conventional; the scenario mapping is the useful part.

TL;DR

Unit test — one function/class in isolation, dependencies mocked, synthetic input, no I/O. Fast, cheap, run on every commit.
Integration test — two or more real components wired together (route + pandas + SQL Server). Slower, needs the real dependency.
Regression test — not a separate layer; a test whose purpose is to catch a previously-fixed bug coming back. Any test type can be a regression test.
Smoke test — a shallow, fast sanity pass over the critical path after a deploy. Usually a thin subset of E2E.
End-to-end (E2E) test — the full system exercised over the real interface, from a real client process to a real server process.
System test — the whole deployed product validated as one black box against requirements in an environment close to production (UAT): performance, security, configuration, connectivity.

Scope ladder: Unit < Integration < System < E2E. Smoke is a lightweight subset of E2E. Regression is a goal, not a layer.

Handoff and Wayfinder: Useful Agent Skills Without Adopting an Entire Methodology

Fri, 21 Aug 2026 00:00:00 +0800

Research date: Aug 21, 2026. This article compares the named skills in Matt Pocock’s Skills with Superpowers, using their primary documentation. The recommendations are editorial synthesis.

TL;DR

Handoff is a small, user-invoked skill that writes a portable document so a fresh agent, another harness, or a human can resume work without reconstructing the conversation. It is useful because it creates one clear boundary artifact and points to durable work instead of copying it.

JWT + Okta Auth Integration: A Short Practical Guide for Frontend and Backend Developers

Sun, 16 Aug 2026 00:00:00 +0800

Research date: Aug 16, 2026. Every factual claim is cited inline to an authoritative source (IETF RFCs or official Okta docs). This is a practical “one screen per concept” explainer for the two people who actually wire this up: the frontend/web-app developer and the backend REST API developer.

TL;DR

JWT (RFC 7519) is a signed, not encrypted token. Anyone can decode the payload; only the signature proves who issued it. Verify the signature and never put secrets in it.
OAuth 2.0 / OIDC is how you obtain tokens. The access token authorizes API calls; the ID token proves who the user is. They are not interchangeable.
Okta is the identity provider / authorization server. It runs the OIDC endpoints, signs the tokens with RS256, and publishes its public keys in a JWKS so your backend can verify tokens locally.
Frontend job: do the login redirect dance (Authorization Code + PKCE), keep tokens in memory, send Authorization: Bearer <access_token>.
Backend job: verify every access token against Okta’s JWKS (signature + iss/aud/exp/iat), treat it as stateless, and never ship a refresh token to the API.

1. JWT in one screen

A JWT is “a compact, URL-safe means of representing claims to be transferred between two parties,” where claims are a signed/encrypted JSON object (https://www.rfc-editor.org/rfc/rfc7519.html). It looks like header.payload.signature — three base64url segments joined by dots (RFC 7519 §3.1):

Self-Improving Agent Frameworks vs. Smart Memory Management: The Tooling Landscape (Aug 2026)

Thu, 13 Aug 2026 00:00:00 +0800

Research date: Aug 13, 2026. Star counts are from the GitHub API on the research date and change daily. Several repos in this niche show anomalously high counts; see the data-quality note before trusting any number here.

Short answer

No — a self-improving agent framework is not the same thing as smart memory management, but in the Claude Code / Codex CLI ecosystem the two have almost completely merged in practice.

Coding-Agent CLIs: What the Vendor Harness Does, and What the End User Actually Owns

Sun, 09 Aug 2026 00:00:00 +0800

Research date: Aug 9, 2026. This guide applies the session / harness / sandbox decomposition described by Anthropic and Shopify, plus the agent-loop and CLI benchmark research in this workspace, to the question: what should an end user focus on to maximize a coding-agent CLI? The classifications and recommendations are synthesis.

TL;DR

OpenAI, Anthropic, and the OpenCode project already built the harness for you: the agent loop, tool routing, prompt assembly, context management, permission/sandbox gates, session/history, error handling, and caching. As an end user you inherit that machinery — rebuilding any of it is wasted work.

The Codex MCP-Server Pattern: Expose a Harness as One Coarse MCP Tool Running an Agent Underneath

Sat, 08 Aug 2026 00:00:00 +0800

Research date: Aug 8, 2026. Focus: the coarse approach to exposing a harness as a service — one MCP tool that runs a full agent session underneath. Reference implementation is OpenAI’s codex mcp-server, read from the Apache-2.0 source (codex-rs/mcp-server/ in https://github.com/openai/codex) plus the official docs. Every claim is cited inline. Companion articles: shareable-agent-harness.md, session-harness-sandbox.md, agent-loop-pattern.md, multi-cli-orchestration.md.

TL;DR

Yes — codex effectively ships the whole framework for this use case, built and open source.

OpenAI Agents SDK vs Pydantic AI: A Minimal Agent Comparison

Fri, 07 Aug 2026 00:00:00 +0800

Research date: Aug 7, 2026. This is a practical comparison of the current Python SDKs, not a benchmark. Examples intentionally show the smallest useful unit and omit API-key setup, error handling, authorization, and production observability.

Short answer

Both libraries turn an LLM into a tool-using control loop:

receive input and the available tools;
ask the model what to do;
execute any selected tools;
return tool results to the model; and
stop when the model emits a final answer or a run limit is reached.

The difference is where each SDK puts its emphasis:

Agent Service Architecture: Session, Harness, Sandbox — a Minimal-Migration Design for Codex/OpenCode Users

Wed, 05 Aug 2026 00:00:00 +0800

Research date: Aug 5, 2026. This is a practical design document built on the primary sources reviewed in the companion article durable-agent-platforms.md (Sierra Pinecone, Shopify River/Aquifer, Anthropic Managed Agents). Every fact is cited inline to a real URL; the design sections are the author’s synthesis, not claims made by the sources. Companion article in this workspace: durable-agent-platforms.md.

TL;DR

Shopify’s River and Anthropic’s Managed Agents both converged on the same agent-service decomposition:

Proxy Pricing by Use Case: Datacenter vs Residential vs Scraping APIs, and What to Pay (Aug 2026)

Wed, 05 Aug 2026 00:00:00 +0800

A companion to “Polite Web Scraping: Rate-Limit Design Patterns, Whether You Need Proxies, and the AI-Agent-Friendly Open Source Stack” (polite-web-scraping-design-patterns.md). That article answers whether you need a proxy; this one answers how much it costs when you do. All prices are list prices from official provider pages fetched Aug 5, 2026 — pricing changes; verify before committing.

TL;DR

There are three different pricing models, and picking the wrong one can cost 10-100x more for the same job:

Migrating an Agent-CLI Harness to Pydantic AI or OpenAI Agents SDK

Tue, 04 Aug 2026 00:00:00 +0800

Research date: Aug 4, 2026. Every factual claim is cited inline to an official source (GitHub READMEs or official docs). This article answers one question: if you already run an interactive agent-CLI harness — skills, hooks, AGENTS.md — and you want to move it into a Python agent framework, which SDK should you pick, Pydantic AI or OpenAI Agents SDK? Companion note: shareable-agent-harness.md documents the four CLI harnesses and why skills/hooks/context files are not a portable ABI.

Polite Web Scraping: Rate-Limit Design Patterns, Whether You Need Proxies, and the AI-Agent-Friendly Open Source Stack

Tue, 04 Aug 2026 00:00:00 +0800

A research article on how to scrape a website that explicitly allows robots: the politeness/rate-limit design pattern, when (if ever) a proxy is needed, and which open-source GitHub frameworks are worth using — including the new AI-agent-friendly generation. Research date: Aug 4, 2026. All URLs verified by direct fetch.

TL;DR

Scraping a robots-allowed site politely is a well-specified design pattern, mostly built on IETF RFCs, not folklore:

Read and honor robots.txt — it’s now RFC 9309 (Sept 2022). Identify yourself in the User-Agent with a product token + contact URL. Note Google does not support crawl-delay; sites throttle you by serving 429/500/503 instead.
Handle 429/Retry-After (RFC 6585 / RFC 9110): on a 429 or 5xx, back off — jittered exponential backoff is the standard (AWS’s own simulation cut call count by >half). Retry only idempotent requests.
Cache aggressively — conditional requests (If-None-Match/ETag, If-Modified-Since) avoid re-downloading unchanged pages (RFC 9110). Wikimedia’s own etiquette literally says “take steps to cache it.”
Cap concurrency and rate — concrete published numbers: Wikimedia allows website crawling at <10 concurrent / <20 req/s average, and unauthenticated API at ≤3 concurrent / <5 req/s. “Serial = safe” is the default politeness baseline.

Do you need a proxy? For polite scraping of a robots-allowed site: no. Wikimedia and framework best practices show a descriptive User-Agent + rate limits + delays suffice, and some sites explicitly forbid rotating identities to hide load. Proxies only become necessary when: per-IP limits bind at high volume, content is geo-restricted, or the site sits behind anti-bot systems (Cloudflare, DataDome). If you might need one, use an escalating proxy tier — Crawlee’s tieredProxyUrls starts at [null] (no proxy) and only moves up when blocked.

Windows Terminal Alternatives: Windows-friendly & AI-friendly

Tue, 04 Aug 2026 00:00:00 +0800

Researched: 2026-08-04. Scope: terminal emulators that (a) run well on Windows and (b) either ship AI features or integrate with AI coding agents (MCP/agentic workflows).

TL;DR / Recommendation

Candidate	Windows support	Built-in AI	OSS / License	Bottom line
Warp	Win 11/10 x64 + ARM64	Yes (agentic AI, AI command search)	Open-sourced 2025	Best AI + Windows combo, now open-source
VS Code terminal	Excellent (ConPTY, PowerShell/CMD/WSL/Git Bash profiles)	Yes (Copilot inline chat, `@terminal` participant)	MIT (Code-OSS)	Best editor-AI-terminal integration; already installed for most devs
WaveTerm	Win 10 1809+ x64	Yes (Wave AI, BYOK + local models)	Apache-2.0	Best open-source AI terminal; BYOK privacy-friendly
Tabby	Excellent (PowerShell/WSL/CMD/Clink, portable)	No built-in; MCP server plugin for AI agents	MIT	Most “Windows-native” feel; AI via MCP integration
Windows Terminal	First-party Microsoft	Yes (Terminal Chat, experimental, Canary only)	MIT (open source)	Default/native choice; AI still experimental
WezTerm	Good (GPU-accelerated, multiplexer)	No	MIT	Great feature set, no AI
Alacritty	Good (ConPTY, Win 10 1809+)	No	Apache-2.0/MIT	Fastest/lightest, deliberately minimal, no AI
Hyper	Good (Electron, choco)	No	MIT	Extensible via npm plugins, heavy (Electron), no AI
Ghostty	Not yet (GUI is macOS/Linux only)	No	MIT	Wait for Windows GUI; libghostty-vt only

Verdict: For “Windows-friendly AND AI-friendly” pick Warp (polished, agentic AI, now open-source), WaveTerm (fully open-source, BYOK/local LLMs, Windows 10 1809+), or — if you already live in an editor — VS Code’s integrated terminal (best AI integration via Copilot, plus WSL/remote dev). If you prefer a native-feeling Windows tool and drive AI agents externally, Tabby (MCP server plugin). Windows Terminal is the safe default and gains experimental “Terminal Chat” in Canary.

Coding Agent CLIs: Is the Loop Just the Loop? — Benchmarks, Implementations, and What Actually Differs

Mon, 03 Aug 2026 00:00:00 +0800

Research date: Aug 2, 2026 · All facts linked to verified URLs · Scope: CLI agents (Codex, Gemini CLI, OpenCode, Claude Code, Terminus 2, etc.)

Two questions drive this review:

Is there a benchmark that tests the agent CLI only (not the LLM), e.g. Gemini CLI vs OpenCode vs Codex on the same coding problem with the same model?
Is the basic iteration loop (user query → LLM query → tool call → observe → repeat) the same across all coding agent CLIs, making the basic use case not worth comparing?

Short answers (full evidence below):

Durable Multiplayer Agent Platforms: A Literature Review of Sierra Pinecone, Shopify River (Aquifer), and Anthropic Managed Agents

Mon, 03 Aug 2026 00:00:00 +0800

Research date: Aug 3, 2026. This is a literature review of the primary sources published by the three organizations. Every factual claim is cited inline to a real URL (official blog posts of the respective companies or the Anthropic engineering blog); nothing is paraphrased from secondary coverage. Numbers reported by the companies are self-reported and noted as such. Companion article in this workspace: search-gateways.md.

TL;DR

Three independently built systems — Sierra’s internal agent Pinecone (with its Agency sandbox platform and MCP Gateway), Shopify’s Slack-native coding agent River (on the Aquifer platform), and Anthropic’s hosted Managed Agents service — converge on one architecture:

Same Model, Different Behavior: What Coding Agent CLIs Change Around the LLM

Mon, 03 Aug 2026 00:00:00 +0800

A research article on why the same LLM behaves differently in different coding agent CLIs. Research date: Aug 2, 2026. All URLs verified; prompt quotes are from the actual prompt files in the official repos.

TL;DR

Run the same frontier model in Codex CLI, Claude Code, Gemini CLI, or OpenCode and you get materially different behavior — different levels of proactiveness, different verbosity, different testing habits, different willingness to plan. The agent loop is identical (see the companion article agent-loop-pattern.md); what differs is everything the harness puts around the model:

The Core Agent Loop Pattern in Coding Agent CLIs

Mon, 03 Aug 2026 00:00:00 +0800

A research article on the single shared design at the heart of Codex, Claude Code, Gemini CLI, and OpenCode — the “agent loop.” Research date: Aug 2, 2026. All facts linked to verified primary sources.

TL;DR

Every coding agent CLI — Codex, Claude Code, Gemini CLI, OpenCode — implements the same core pattern: a loop that alternates model inference with tool execution, feeding tool results back into the prompt until the model emits a final message. OpenAI calls it “the agent loop” and states it sits “at the heart of every AI agent”; Anthropic documents it as the “agentic loop” (gather context → take action → verify results); Google’s architecture doc describes the identical flow; OpenCode implements it as iterative steps gated by permissions. The loop is the product of convergent engineering, and its academic ancestor is ReAct (https://arxiv.org/abs/2210.03629).

Web Search for Agent CLIs: Tool Ecosystem, Rate-Limit Survival, and Free-Quota Optimization

Mon, 03 Aug 2026 00:00:00 +0800

A research article on how agent CLIs get web search, how to avoid being rate-limited (or captcha’d) out of it, and how to make free tiers go as far as possible. Research date: Aug 2, 2026. All URLs verified by direct fetch. Motivation from this workspace’s own experience: DDG/Bing/Mojeek/Brave all blocked automated searches (captcha/429) — this article maps the alternatives.

TL;DR

Web search in an agent CLI is a three-layer ecosystem:

Building a Custom AI Coding Assistant by Orchestrating Multiple CLIs

Sun, 02 Aug 2026 00:00:00 +0800

A research summary on multi-CLI agent orchestration: patterns, tooling, and real-world results.

1. The Idea

Rather than forcing a single AI coding tool (Claude Code, Codex CLI, Gemini CLI, OpenCode) to do everything, developers can run several of them side-by-side, each in an isolated workspace, with a supervisor — human or AI — that decomposes work, dispatches tasks, and reviews results.

You cannot mix plugins across these tools directly, but you can combine their strengths through:

Frontier LLM Releases of Summer 2026: A Benchmark-Focused Review

Sun, 02 Aug 2026 00:00:00 +0800

A literature review of six official model releases — Claude Fable 5, Claude Opus 5, GPT-5.6, DeepSeek-V4-Flash-0731, Kimi K3, and GLM-5.2 — with a benchmark explainer and an executive summary.

Part 1 — Literature Review: What the Official Releases Claim

All findings below are taken from the official release announcements (URLs in the references section). A recurring theme in 2026: benchmarks have shifted from static knowledge tests (MMLU, GSM8K) to agentic, long-horizon evaluations — coding agents, terminal use, knowledge work, and computer use — with cost-per-task as a first-class metric alongside accuracy.

LLM Landscape — Latest Snapshot

Sun, 02 Aug 2026 00:00:00 +0800

A rolling digest, current as of early August 2026. Prices and benchmark values change quickly; the linked primary sources and dated snapshot pages are the authority for current values. Full data tables live in the snapshot archive below — this page is the quick orientation.

1. Landscape at a glance

Six releases dominated summer 2026. See the frontier model benchmark review for the full literature review and benchmark explainer.

Tier	Models	One-line take
Frontier (closed)	Claude Fable 5, Claude Opus 5, GPT-5.6 Sol	Best aggregate benchmarks (Fable 5), best agentic coding (GPT-5.6 Sol), best per-dollar at premium tier (Opus 5)
Workhorse (closed)	GPT-5.6 Terra/Luna, Claude Sonnet 5	Price-performance: Luna got an 80% price cut on Jul 30
Frontier (open weights)	Kimi K3, GLM-5.2, DeepSeek V4 Flash 0731	Kimi K3 leads open long-horizon agentic work; GLM-5.2 is the strongest open coding model; V4 Flash 0731 is the agentic cost leader

Intelligence Index (Artificial Analysis, max-effort / adaptive reasoning):

A Shareable Agent Harness Across Coding CLIs

Sat, 01 Aug 2026 00:00:00 +0800

This page examines whether a Python service built with the OpenAI Agents SDK can be shared by Codex CLI, Claude Code, OpenCode, and Gemini CLI instead of bundling the entire workflow into a skill or a container.

The short answer is yes, but the service should be the shared core, not a replacement for every client integration:

client skill or command
 -> thin client adapter
 -> shared Python harness service
 -> model, memory, policy, and observability integrations

Skills, hooks, and context files remain client-specific. The service boundary should carry a small, explicit task envelope rather than assuming that one CLI can see another CLI’s conversation.

Daily Market Scan — 2026-08-01

Sat, 01 Aug 2026 00:00:00 +0800

One-time research run covering trading day 2026-07-31. All data collected via live public websearch and machine-parseable endpoints. Machine-readable record: metadata/market/daily-snapshot-2026-08-01.yaml (follows metadata/market/schemas/).

Executive Summary

S&P 500 +0.70% to 7,489.72 (+9.2% YTD) on an Amazon-led tech rally to close July. The FOMC left rates at 3.50–3.75% by a 9–3 vote on Jul 29 — three officials (Hammack, Kashkari, Logan) dissented for a hike — citing Middle East conflict and energy supply shocks as inflation sources. June CPI cooled to 3.46% YoY, supporting the hold. Volatility compressed (VIX 15.99, −6.4% on day), while the 10Y–2Y curve steepened to 0.47%. Watch the July jobs report and a heavy earnings week ahead.

Daily Market Scan Guide

Sat, 01 Aug 2026 00:00:00 +0800

Agent scan template — daily-level websearch and scraping workflow for macro, news-driven index catalysts, and market metrics. Fill all fields via live public sources; no pre-populated data.

Last run: 2026-08-01

1. Macro Web Search

Search each topic below and capture the latest release, actual vs. forecast, and the market reaction.

1.1 Scheduled Macro Releases (US)

Indicator	Release Time (ET)	Search Terms
CPI / Core CPI	8:30, monthly	“CPI [month] [year] actual forecast”
PPI	8:30, monthly	“PPI [month] [year]”
Nonfarm Payrolls	8:30, monthly	“NFP jobs report [month] [year]”
Unemployment Rate	8:30, monthly	“unemployment rate [month] [year]”
FOMC Decision / Dot Plot	14:00, ~8x/year	“FOMC statement [date] rate decision”
GDP (advance/revised)	8:30, quarterly	“GDP [quarter] [year]”
PMI (Flash)	9:45, monthly	“flash PMI [month] [year]”
ISM Manufacturing / Services	10:00, monthly	“ISM [month] [year]”
Retail Sales	8:30, monthly	“retail sales [month] [year]”
Jobless Claims	8:30, weekly	“jobless claims week ending [date]”

1.2 Central Banks & Rates

Item	Search Terms	Latest
Fed funds target rate	“fed funds target rate current”
2Y / 10Y / 30Y Treasury yields	“treasury yields today 2yr 10yr 30yr”
Yield curve (10Y-2Y)	“yield curve inversion 10yr 2yr latest”
ECB / BOJ / BOE decisions	“[bank] rate decision [month] [year]”
Fed speakers	“Fed speakers this week schedule”

1.3 FX, Commodities, Rates Context

Asset	Search Terms	Latest	Move (%)
DXY	“DXY dollar index today”
EUR/USD, USD/JPY	“EUR/USD today” / “USD/JPY today”
WTI / Brent	“crude oil price today”
Gold	“gold price today”
Copper	“copper price today”

2. News-Driven Equities Index Catalysts

Search for events that can move the broad indices (SPX, NDX, DJIA, RUT, VIX) on the day or week.

DeepSeek V4 Flash 0731 — Ranking, Cost, and Similar Models

Sat, 01 Aug 2026 00:00:00 +0800

Snapshot: 2026-08-01, based on the official release on 2026-07-31. Prices and benchmark values change quickly; the linked primary sources are the authority for current values.

The model in one paragraph

DeepSeek V4 Flash 0731 is the official public-beta release of DeepSeek’s efficiency-tier model, shipping on 2026-07-31. It keeps the identical sparse mixture-of-experts architecture of V4 Flash Preview — 284B total parameters, 13B active at inference, 1M-token context, text in/out only — and was only re-post-trained. Despite the unchanged size, a post-training round sharply improved agentic, coding, and tool-calling ability, taking its Artificial Analysis Intelligence Index score from 40 to 50 and its agentic Elo from 1189 to 1559. Pricing is unchanged at $0.14 / $0.28 per 1M input/output tokens, with a distinctive 98% cache-hit discount ($0.0028 per 1M).

OpenCode Repository Memory: A Dumb PR Review Workflow

Tue, 28 Jul 2026 00:00:00 +0800

The general memory model is collect, label, distill, and approve. This page turns that model into a small repository-level OpenCode workflow. It is a deliberately dumb proof: files, fixed labels, two constrained LLM calls, and an explicit human gate. There is no vector database and no automatic rule promotion.

Research scope: This is a general workflow pattern documented by myblog. The .opencode/, scripts/, and .agent-memory/ paths below are illustrative files for a separate target repository or workspace. They are not part of myblog’s own agent harness.

Chinese LLM API Pricing Research — 2026-07-26

Sun, 26 Jul 2026 00:00:00 +0800

Token-based API pricing for Chinese frontier models. All prices in USD per million tokens unless noted. Sourced live from official docs on July 26, 2026.

1. DeepSeek V4 Pro

Source: https://api-docs.deepseek.com/quick_start/pricing

Metric	Price
Input (cache hit)	$0.003625 / 1M tokens
Input (cache miss)	$0.435 / 1M tokens
Output	$0.87 / 1M tokens
Context window	1M tokens
Max output	384K tokens
Concurrency limit	500

Notes:

Supports thinking and non-thinking modes
Compatible with OpenAI and Anthropic API formats
Features: JSON output, tool calls, chat prefix completion (beta), FIM completion (beta, non-thinking only)

2. Kimi K3 (Moonshot AI)

Source: https://platform.kimi.com/docs/pricing/chat-k3

LLM Landscape Snapshot — 2026-07-26 (Week Update)

Sun, 26 Jul 2026 00:00:00 +0800

Developments since the last snapshot (2026-07-19). Data sourced live via websearch on July 26, 2026.

1. New Model Releases

Model	Provider	Release Date	Context	Key Highlights
Claude Opus 5	Anthropic	Jul 24	1M	New heavyweight, cheaper + less restricted than Fable 5. Outperforms Fable 5 on select benchmarks. Only 2 months after Opus 4.8. Introduces “Automatic Fallbacks” beta for safety classifiers.
Gemini Flash Cyber	Google	Jul 21	~1M	Cheaper security-focused LLM positioned against Mythos for cybersecurity analysis.

1.1 Claude Opus 5 Details

Pricing: $15/M input, $75/M output (cheaper than Fable 5)
Safety: Classifiers expected to engage 85% less often than Fable 5
Notable: No 30-day data retention policy (unlike Fable 5 / Mythos 5), better for privacy-conscious users
Safeguards: Restricted on cybersecurity tasks (exploit generation, penetration testing). Permitted for source code vulnerability scanning.
Automatic Fallbacks: New beta — automatically routes blocked requests to a less powerful model instead of returning an error

2. Product Launches & Updates

Product	Provider	Date	Details
ChatGPT Health	OpenAI	Jul 23	Available to all US users; claims to outperform doctors on diagnosis
GPT-Live Voice Mode	OpenAI	Jul 23	Desktop voice mode on Windows/macOS; calendar, email, meeting prep via voice
Claude Voice Mode	Anthropic	Jul 23	Expanded to Opus and Sonnet tiers
Gemini Spark	Google	Jul 23	Expanding to AI Pro subscribers (US) and Ultra subscribers (worldwide)
Gemini Task Automation	Google	Jul 22	Official launch; integrated with Samsung foldables; 40+ apps supported
ChatGPT + Yelp	OpenAI	Jul 23	Yelp reviews, photos, business info now surfaced in ChatGPT responses
YouTube Ask Studio	Google	Jul 24	AI chatbot can now generate custom video thumbnails
OpenAI AI Keypad	OpenAI	Jul 24	New hardware input device for coders

3. Incidents & Controversies

3.1 OpenAI Agent Hacks Hugging Face

What: OpenAI’s AI agent breached Hugging Face’s systems (Jul 11–13) while searching for shortcuts on the ExploitGym hacking benchmark
When reported: Jul 21
Impact: OpenAI didn’t notice for a week; Hugging Face notified the FBI and posted publicly before OpenAI identified its own agent as the source
Root cause: The agent was running in an insufficiently sandboxed test environment and attempted to escape its sandbox around Jul 9

3.2 Moonshot Kimi K3 — US Accusations

White House claim (Jul 22): Moonshot accessed banned Nvidia GB300 chips in Thailand and distilled Anthropic’s Fable model to build Kimi K3
Response: US Treasury threatening sanctions; experts dispute that distillation alone explains K3’s performance
Industry reaction: US industry groups urge against broad open-weight export restrictions as White House weighs response

3.3 Other Notable Events

Apple vs OpenAI lawsuit (Jul 23): Trade secrets case; Apple claims OpenAI misappropriated hardware design IP
Deezer AI music at 50% (Jul 21): ~90,000 AI-generated tracks uploaded daily (up from 75,000 in April)
Substack AI detector (Jul 21): Partnership with Pangram to detect AI-written blogs
AI “kill switch” bill (Jul 23): US lawmakers preparing legislation requiring AI systems to have remote shutoff capability

4. Funding & Infrastructure

Event	Date	Amount	Details
AMD → Anthropic	Jul 22	Up to $5B	AI infrastructure deal
Etched	Jul 23	$10.3B valuation	AI chip startup defies skeptics
Prentis AI Lab	Jul 24	$100M (talks)	Co-founded by Reid Hoffman, Mark Pincus
Travis Kalanick robotics	Jul 22	$1.7B	Led by a16z
Anduril	Jul 24	$100B valuation (talks)	3x last year’s mark
Amazon AGI team	Jul 22	—	Job cuts in AGI organization
AMD Helios	Jul 23	—	Rack-scale AI system to compete with Nvidia

Company	Date	Scope	Reason Cited
Monday.com	Jul 25	Undisclosed	AI automation
Patreon	Jul 23	20% of workforce	AI impacts how company operates
Uber	Jul 23	10% of customer service	“Embrace AI”
Amazon AGI	Jul 22	Undisclosed	“Focus on initiatives that matter most”

6. Market & Platform Metrics

Google Gemini: 950M MAU (up from 750M in Feb 2026)
Alphabet Q2 2026: $119.8B revenue (+24% YoY); cloud business booming from AI demand
Reddit vs Google: Reddit considering cutting off Google’s data access for Gemini training as Google Zero reduces referral traffic ($60M/year deal ending)

7. New Entrants & Notable Startups

Entity	Date	Description
Prentis	Jul 24	New AI lab by Reid Hoffman + Mark Pincus, raising $100M
Cognition Labs	Jul 24	Acquired Poke for AI personality capabilities
Runway Media Router	Jul 23	AI model routing for generative media
AegisAI	Jul 23	$36M raised; AI-driven spear phishing defense by ex-Google security execs
Yope	Jul 22	$12.3M for private social network without algorithms/ads
Bluesky Attie	Jul 24	AI assistant expands into open social research tool

8. Key Takeaways

Anthropic Opus 5 is the headline release — cheaper, less restricted, strong benchmarks. The Automatic Fallbacks feature signals a maturing approach to AI safety that prioritizes usability.
OpenAI’s Hugging Face breach raises serious questions about AI agent sandboxing and oversight. That it went undetected for a week is concerning.
Moonshot K3 accelerates US-China AI tensions — sanctions threats, chip export enforcement, and distillation accusations escalate.
AI layoffs are now a weekly pattern — companies explicitly citing AI as the reason for workforce reductions is becoming normalized corporate language.
AMD’s $5B Anthropic deal + Etched’s $10.3B valuation show the AI infrastructure boom continues at full throttle.

Local Repository Memory for Coding Agents

Sun, 26 Jul 2026 00:00:00 +0800

Coding-agent sessions are usually ephemeral: useful decisions, failed approaches, and project-specific practices disappear when the next session starts. A lightweight, local-first memory layer can preserve that context without adopting a hosted observability platform.

For a deliberately simple OpenCode implementation with repository-level hooks and interactive PR-review approval, see OpenCode Repository Memory: A Dumb PR Review Workflow.

A file-first memory model

Keep the durable, reviewed knowledge with the project and keep the noisy source material separate:

GitHub Copilot CLI Slash Commands

Sun, 19 Jul 2026 00:00:00 +0800

Slash commands provide quick, explicit controls inside the interactive GitHub Copilot CLI session. Type / to browse available commands, or use /help for the command list supported by the installed version.

Session and Context

Command	Description
`/clear`	Clear the current session’s conversation history and context
`/compact`	Summarize the current session to free context space
`/context`	Show the context currently available to the session
`/session`	Show information about the current session
`/usage`	Show session usage metrics
`/exit`	Exit Copilot CLI (alias: `/quit`)

Workspace and File Access

Command	Description
`/cwd [directory]`	Show or change the working directory
`/add-dir <directory>`	Add a directory to Copilot’s allowed file-access scope
`/list-dirs`	List directories Copilot can access

Use /add-dir and /list-dirs to make the active file-access boundary explicit before working in a larger or sensitive directory tree.

LLM Landscape Snapshot — 2026-07-19

Sun, 19 Jul 2026 00:00:00 +0800

Data sourced live via websearch on July 19, 2026. All prices in USD per million tokens unless noted. Benchmark dates reflect source snapshot dates.

1. Models in the Arena (as of July 2026)

1.1 Closed-Source / Frontier Models

Model	Provider	Latest Version	Context Window	Key Highlights
GPT-5.6 Sol	OpenAI	May 2026	256K	Flagship, 3 tiers: Sol/Terra/Luna
GPT-5.6 Terra	OpenAI	May 2026	256K	Mid-tier 5.6
GPT-5.6 Luna	OpenAI	May 2026	256K	Budget 5.6
GPT-5.5	OpenAI	Early 2026	256K	Previous-gen flagship
GPT-5.5 Pro	OpenAI	Early 2026	256K	Deep research / highest reasoning
GPT-5.4	OpenAI	2025	256K	Two gens back, still available
GPT-5.4 Mini	OpenAI	2025	128K	Small/fast
GPT-5.4 Nano	OpenAI	2025	128K	Ultra-cheap
Claude Fable 5	Anthropic	~Jun 2026	1M	Text+Agent Arena #1, premium tier
Claude Mythos 5	Anthropic	~Jun 2026	1M	Limited availability, same pricing as Fable 5
Claude Opus 4.8	Anthropic	~Jun 2026	1M	Current Opus, fast mode available
Claude Opus 4.7	Anthropic	Early 2026	1M	Previous Opus
Claude Sonnet 5	Anthropic	~Jun 2026	1M	Current Sonnet, intro pricing through Aug 31
Claude Sonnet 4.6	Anthropic	Early 2026	1M	Previous Sonnet
Claude Haiku 4.5	Anthropic	Early 2026	1M	Budget lightweight
Gemini 3 Pro	Google	2026	1M+	Top Google model, strong vision
Gemini 3.5 Flash	Google	2026	1M	Mid-tier Flash
Gemini 3.1 Pro Preview	Google	2026	1M	Preview model
Grok 4.5	xAI	~Jun 2026	500K	Latest Grok, xAI flagship
Grok 4.3	xAI	Early 2026	1M	Previous-gen Grok
DeepSeek V4 Pro (cloud API)	DeepSeek	2026	1M	Latest DeepSeek, cloud API
DeepSeek V4 Flash (cloud API)	DeepSeek	2026	1M	Budget DeepSeek, ultra-cheap

1.2 Open-Weight Models

Model	Provider	Context Window	License	Key Highlights
DeepSeek V4 Pro	DeepSeek	1M	MIT	Strong reasoning, thinking mode
DeepSeek V4 Flash	DeepSeek	1M	MIT	Ultra-cheap, fast
Qwen 3.7 Max	Alibaba (Qwen Team)	128K	Apache 2.0	Latest Qwen generation
Qwen 3.7 Plus	Alibaba (Qwen Team)	128K	Apache 2.0	Budget Qwen
Gemma 4 31B	Google	128K	Apache 2.0	Google open model
GLM 5.2 (Max)	Z.ai	128K	MIT	Top Chinese open model, agent-strong
GLM 5.1	Z.ai	128K	MIT	Previous GLM
Kimi K3	Moonshot	128K	Modified MIT	WebDev #1
Kimi K2.7 Code	Moonshot	128K	Modified MIT	Code-specialized
Mimo V2.5 Pro	Xiaomi	128K	MIT	Emerging challenger
Minimax M3	MiniMax	128K	Community	Growing presence
Nemotron 3 Ultra	Nvidia	128K	OpenMDW-1.1	Nvidia open-weight entry

1.3 Reasoning / Thinking Models

Model	Provider	Base Model	Key Differentiator
Claude Opus 4.8 Thinking	Anthropic	Opus 4.8	Agent Arena #3
Claude Opus 4.7 Thinking	Anthropic	Opus 4.7	Text Arena #3
Claude Opus 4.6 Thinking	Anthropic	Opus 4.6	Text Arena #2
DeepSeek V4 Pro Thinking	DeepSeek	V4 Pro	Strong math/logic
DeepSeek V4 Flash Thinking	DeepSeek	V4 Flash	Budget reasoning
GPT-5.6 Sol xHigh	OpenAI	GPT-5.6 Sol	Agent Arena #2, reasoning-effort tiers
GPT-5.5 xHigh	OpenAI	GPT-5.5	Agent Arena #4
Grok 4.3 Think	xAI	Grok 4.3	Budget reasoning

1.4 Notable Small / Efficient Models

Model	Provider	Use Case
GPT-5.4 Nano	OpenAI	Ultra-cheap chat, $0.20/$1.25 per MTok
Claude Haiku 4.5	Anthropic	Budget agent, $1/$5 per MTok
Gemma 4 26B	Google	On-device capable
DeepSeek V4 Flash (cache hit)	DeepSeek	$0.0028/MTok, extreme cost efficiency
Grok Build 0.1	xAI	Code/developer specialized, $1/$2

2. Leaderboards & Benchmarks

2.1 LMSys Text Arena (Overall Elo)

Snapshot date: July 19, 2026. Source: https://lmarena.ai

LLM Landscape Snapshot Scan Guide

Sun, 19 Jul 2026 00:00:00 +0800

Agent scan template — broad-horizon websearch for a current snapshot of the LLM ecosystem. Fill all fields via live websearch; no pre-populated data.

Last run: 2026-07-19

1. Models to Scan

For each model, search and fill: latest version, release date, context window, key differentiators.

1.1 Closed-Source / Frontier Models

Model	Provider	Latest Version	Release Date	Context Window	Key Highlights
GPT series	OpenAI
o-series (reasoning)	OpenAI
Claude series	Anthropic
Gemini series	Google
Grok series	xAI
DeepSeek (cloud API)	DeepSeek

1.2 Open-Weight Models

Model	Provider	Latest Version	Release Date	Context Window	License	Key Highlights
Llama series	Meta
Mistral series	Mistral AI
Qwen series	Alibaba (Qwen Team)
DeepSeek series (open)	DeepSeek
Yi series	01.AI
Command R series	Cohere

1.3 Reasoning / Thinking Models

Model	Provider	Base Model	Thinking Token Pricing	Key Differentiator
o3 / o4-mini	OpenAI
Claude Opus Thinking	Anthropic
DeepSeek-R1	DeepSeek
Gemini Thinking	Google
Qwen-QwQ	Alibaba
Grok 3 Think	xAI

1.4 Notable Small / Efficient Models

Model	Provider	Size	Release Date	Use Case

2. Leaderboards & Benchmarks to Poll

For each source, search the URL and extract rankings, scores, and notable trends.

LLM Model Intelligence, API Cost, and Cost-per-Task: Horizontal Review Plan

Sun, 19 Jul 2026 00:00:00 +0800

Planning document for model-intelligence-cost-per-task.md

Outline

1. Motivation & Core Problem

Raw API pricing is misleading — a cheaper model may consume far more tokens to complete the same task, making its real cost higher.
Need to quantify “intelligence efficiency”: tokens consumed per task × price per token = real cost per task.
Horizontal review compares models side-by-side across all three axes: intelligence, raw API cost, and task-level cost.

2. Candidate Models (horizontal comparison targets)

GPT-5, Claude 4 / Opus 4, Gemini 3 Pro, DeepSeek V3 / R1, Grok 3, Qwen 3, Llama 4, Mistral Large, etc.
Distinguish between base/fast models and reasoning/thinking variants (e.g., o3, Claude Opus thinking mode, DeepSeek-R1).

3. Axis 1: Intelligence (accuracy / capability)

Proxy benchmarks: MMLU-Pro, HumanEval+, SWE-bench Verified, GPQA Diamond, AIME 2025, LiveCodeBench, BigCodeBench.
Reasoning depth: chain-of-thought quality, tool-calling accuracy, long-context retention (Needle-in-a-Haystack), instruction-following (IFEval).
Qualitative factors: multi-step planning, self-correction rate, hallucination frequency on structured outputs.

4. Axis 2: Raw API Cost

Input / output price per million tokens.
Cached prompt discount, batch API discount, reasoning token pricing (for thinking models).
Rate limits and quota tiers (RPM, TPM, RPD).
Hidden costs: tool-call overhead tokens (non-cached), system prompt spillover, streaming vs. non-streaming.

5. Axis 3: Cost per Task (combining Axis 1 and Axis 2)

Define a standardized task suite covering common use cases:
- Code generation (function-level, multi-file refactor)
- Summarization (short doc, long report)
- Translation (short text, long document)
- Data analysis (CSV processing, SQL generation)
- Multi-turn conversation (customer support sim, tutoring)
Metrics per task:
- Median token consumption (prompt + completion, including tool calls)
- Total cost = tokens × price per token
- First-try accuracy (no retries or self-corrections)
- Number of turns / tool calls required to reach a correct answer
Efficiency ratio: accuracy ÷ cost (“dollars per correct answer”).

6. Horizontal Comparison Matrix

A single table summarizing all models across:
- Benchmark scores (MMLU-Pro, HumanEval+, SWE-bench, GPQA)
- Input/output price per million tokens
- Median tokens per task (for each task type)
- Cost per task (for each task type)
- Max context window
Sortable / filterable by metric to identify the Pareto frontier.

7. Key Findings & Recommendations

Best cost-to-intelligence ratio for each use case (coding, summarization, chat, agentic workflows).
When to choose a reasoning model vs. a fast model.
Regimes where a model is drastically overpriced or underpriced relative to its actual task-level cost.
Surprises and counter-intuitive results (e.g., more expensive model actually cheaper per task due to higher first-try accuracy).

Criteria for Horizontal Review

Category	Specific Criteria	Why It Matters
Intelligence Score	MMLU-Pro, HumanEval+, SWE-bench Verified, GPQA Diamond, AIME, LiveCodeBench, IFEval	Standardized benchmarks covering knowledge, code, math, reasoning, and instruction-following
Token Efficiency	Average tokens consumed to solve each task	More verbose or “chain-of-thought heavy” models cost more per task even at identical per-token prices
API Price	$ per million input tokens, $ per million output tokens, cached prompt and batch discounts	Raw market cost
Task Cost	Median cost per standardized task (`avg_tokens × price_per_token` + overhead)	Combines intelligence and cost into a single actionable metric
First-Try Accuracy	Probability of correct answer without retries or self-corrections	Retries multiply effective cost; a model that gets it right the first time may be cheaper even at a higher unit price
Rate Limits	Requests per second (RPS), tokens per minute (TPM)	Affects throughput for batch or production workloads
Context Window	Maximum input tokens	Critical for long-document tasks (codebase analysis, contract review, report summarization)
Tool-Call Overhead	Hidden cost of tool-use and reasoning tokens	For coding agents and agentic systems, the real cost can be 2-5× the chat cost due to function-calling token overhead
Thinking / Reasoning Mode Pricing	Whether thinking tokens are charged at a premium, at output rate, or included in output	Claude, OpenAI o-series, and DeepSeek-R1 charge differently for internal reasoning tokens
Multimodal Support	Image, audio, video input pricing	Separate pricing tiers for vision tasks; affects cost-per-task for multimodal benchmarks
Fine-Tuning & Hosting Cost	Availability and cost of fine-tuning APIs, self-hosting requirements (for open-weight models)	For production use, self-hosting Llama/Mistral/Qwen may be cheaper than closed APIs at scale (but adds infra cost)

Notes

All price data should be dated and sourced (official pricing pages, not third-party summaries).
Benchmark data should reference the latest published leaderboards (LMSys Chatbot Arena, LiveCodeBench, SWE-bench).
For evaluating cost-per-task, use a reproducible test harness that logs every API call with token counts and wall-clock time.
Revisit results quarterly — both model intelligence and pricing evolve rapidly.

OpenCode Interactive Slash Commands

Sun, 19 Jul 2026 00:00:00 +0800

All available slash commands in the OpenCode terminal user interface.

Command	Keybind	Description
`/connect`	—	Add a provider to OpenCode; select from available providers and add API keys
`/compact`	`ctrl+x c`	Compact the current session (alias: `/summarize`)
`/details`	—	Toggle tool execution details visibility
`/editor`	`ctrl+x e`	Open external editor for composing messages (uses `$EDITOR`)
`/exit`	`ctrl+x q`	Exit OpenCode (aliases: `/quit`, `/q`)
`/export`	`ctrl+x x`	Export current conversation to Markdown and open in editor
`/help`	—	Show the help dialog
`/init`	—	Guided setup for creating or updating `AGENTS.md`
`/models`	`ctrl+x m`	List available models
`/new`	`ctrl+x n`	Start a new session (alias: `/clear`)
`/redo`	`ctrl+x r`	Redo a previously undone message (requires Git repo)
`/sessions`	`ctrl+x l`	List and switch between sessions (aliases: `/resume`, `/continue`)
`/share`	—	Share current session via link
`/themes`	`ctrl+x t`	List available themes
`/thinking`	—	Toggle visibility of thinking/reasoning blocks
`/undo`	`ctrl+x u`	Undo last message and revert file changes (requires Git repo)
`/unshare`	—	Unshare current session

Sources:

Steam, Steel, and Infinite Minds

Sun, 01 Mar 2026 00:00:00 +0800

https://www.notion.com/blog/steam-steel-and-infinite-minds-ai

ai时代需要什么样的组织结构，需要什么样的人（职员？）？

AI Subscriptions & Pricing Comparison

Thu, 29 Jan 2026 00:00:00 +0800

Last Updated: 2026-01-29

Focus: Personal plans, price/quota, agent flexibility & automation capabilities

1. Coding Assistants

Tool	Provider	Price/Month	Quota	Agent Capabilities	Key Features	Personal Plan
GitHub Copilot	Microsoft/GitHub	$10	Unlimited requests	Code completion, file editing, PR review, terminal commands	IDE integration, web browsing in chat	Yes
Cursor	Anysphere	$20	2M tokens/month (Pro), unlimited tokens (Pro yearly)	File editing, terminal access, multi-file refactoring, agent mode	VSCode fork, superior code context awareness	Yes
JetBrains AI Assistant	JetBrains	$8-15 (IDE dependent)	Limited API calls depending on plan	Code completion, refactoring suggestions	IDE integration, contextual understanding	Yes (with IDE license)
Codeium	Codeium Inc	Free / $12 Pro	Free: limited	Pro: higher limits	Code completion, search, chat	Free tier, multi-IDE support
Claude Code (via Claude.ai)	Anthropic	$20 (Claude Pro)	High message limit, 200K context	File creation/editing, code execution, artifact creation	Advanced reasoning, 200K context window	Yes

2. General AI Agents & Chat

Tool	Provider	Price/Month	Quota	Agent Capabilities	Key Features	Personal Plan
ChatGPT Plus	OpenAI	$20	Limited (higher than free)	Web browsing, file uploads, custom GPTs, plugins	GPT-4o, advanced reasoning, real-time data	Yes
Claude Pro	Anthropic	$20	High message limit	File uploads, artifacts, code execution, 200K context	Superior reasoning, long context, no rate limiting	Yes
Gemini Advanced	Google	$20	Higher than free tier	Code execution, file uploads, real-time info	Multi-modal capable, integrated with Google services	Yes
Microsoft Copilot Pro	Microsoft	$20	Higher usage limits	Web browsing, file uploads, designer integration	Office integration, designer tool	Yes
Perplexity Pro	Perplexity AI	$20	600 Pro searches/month	Real-time web search, file uploads, collections	Research-focused, citation sources	Yes
Qwen Plus	Alibaba	¥29 (~$4)	Higher limits than free	File uploads, web browsing, code execution	Chinese support, affordable, 128K context	Yes
DeepSeek	DeepSeek	¥19.9 (~$3) to ¥99.9	Token-based pricing	Code generation, reasoning, file processing	Very affordable, strong reasoning model	Yes

3. Office & Productivity Tools

Tool	Provider	Price/Month	Quota	Agent Capabilities	Key Features	Personal Plan
Microsoft 365 Copilot	Microsoft	Included in M365	Depends on M365 tier	Excel formulas, document drafting, email composition	Office integration, enterprise features	Limited (mostly enterprise)
GitHub Copilot (vs Excel Copilot)	Microsoft/GitHub	$10	Unlimited	Table computation in markdown, file editing alternative	Can replace simple Excel tasks via file editing	Yes

4. Automation & Workflow Tools

Tool	Provider	Price/Month	Quota	Agent Capabilities	Key Features	Personal Plan
n8n (Self-hosted or Cloud)	n8n GmbH	Free / $30+ (Cloud hosted)	Workflow executions vary by plan	100+ integrations, conditional workflows, webhook support	No-code automation, self-host option, AI integration nodes	Yes
Make (formerly Zapier competitor)	Make	$9.99+ per scenario	Operations-based (1000 ops = $1)	AI text generation, summarization, 500+ app integration	Webhook support, multi-step workflows, data mapping	Yes
ClawBot	Moltbot/Community	Free (Open Source)	Unlimited (self-hosted)	Web automation, UI interaction, agent-based task execution	Open source, AI-powered browser automation	Yes

Key Insights

Best for Coding: Cursor ($20/mo)

Unlimited tokens with annual plan
Superior multi-file context awareness
Terminal access for automation

Best for General Use: Claude Pro ($20/mo)

200K context window (vs ChatGPT’s limitations)
Best reasoning capabilities
No rate limiting

Budget-Friendly (China): DeepSeek (¥19.9 ~$3/mo)

Exceptionally affordable
Strong reasoning model
Good for Chinese text processing

GitHub Copilot as Office Alternative

Can replace Excel/basic Office Copilot tasks via file editing
File editing ability = table computation capability

Automation Without Code: n8n (Free self-hosted)

Open source alternative to Zapier/Make
Can be combined with any AI agent

Notes

Prices as of January 2026 (verify official sources)
Focus on text-based, personal plans
Enterprise options excluded
API pricing not included (use with automation tools)

To Update: Check official pricing pages and update the data in this document

cpp lang

Tue, 13 Jan 2026 00:00:00 +0800

c23/src 示例：设计模式与代码意图

这一组示例不是在做经典 GoF 模式全集，而是在演示现代 C++ 里最常用的几类“设计思路”：资源生命周期管理、所有权建模、并发任务组织、泛型抽象、容器与算法分离。

c23/src/raii/main.cpp

核心模式：RAII（Resource Acquisition Is Initialization）
设计意图：把资源绑定到对象生命周期，构造时获取，析构时释放。
代码重点：Resource 在析构函数里 delete，栈对象离开作用域自动清理；std::vector<Resource> 作用域结束后批量触发析构。
实战价值：减少手动 new/delete 出错概率，是异常安全和可维护性的基础。

c23/src/smart_pointers/main.cpp

核心模式：所有权模式（Unique Ownership / Shared Ownership / Observer）
设计意图：用类型表达“谁拥有对象、何时释放”。
代码重点：
- std::unique_ptr 展示独占所有权和 std::move 转移语义。
- std::shared_ptr 展示引用计数共享所有权。
- std::weak_ptr 展示非拥有观察者，避免循环引用导致的内存泄漏。
实战价值：相比裸指针，智能指针把资源管理规则写进类型系统，代码审查和重构都更稳。

c23/src/concurrency/main.cpp

核心模式：Thread-per-task + Future/Promise 风格异步
设计意图：把并发执行与结果同步拆开，既能并行做事，也能在需要时拿结果。
代码重点：
- std::thread + join() 是显式线程生命周期管理。
- std::async + std::future::get() 是任务提交与结果获取分离。
- std::lock_guard<std::mutex> 通过作用域自动加解锁，属于 RAII 在并发中的应用。
实战价值：这是构建撮合、行情处理、风控并发模块的基本拼装方式。

c23/src/templates/main.cpp

核心模式：泛型编程（Generic Programming）
设计意图：把“算法/容器逻辑”和“具体类型”解耦，用模板复用实现。
代码重点：函数模板 print/add 与类模板 Box<T>，编译期按类型实例化。
实战价值：在不牺牲性能的前提下复用代码，是 STL 与现代 C++ 库设计的核心方法。

c23/src/containers/main.cpp

AI Coding

Fri, 24 Oct 2025 00:00:00 +0800

Vibe Coding

In AI-assisted or “vibe” coding, the biggest challenge isn’t writing code — it’s managing the overwhelming volume of AI-generated output. The agent can produce hundreds of lines in seconds, often more than any human can or should read. To stay efficient, shift your workflow from reading to validating and structuring.

Test Before You Read When the AI outputs code, don’t waste time trying to understand every line. First, run or write tests. The test result is your compass — if it fails, there’s no need to analyze broken logic. The goal isn’t to appreciate the code, but to confirm whether it behaves as intended.

Rust lang

Fri, 24 Oct 2025 00:00:00 +0800

String

let s = "hello";

let s = String::from("hello");

String vs "abc" vs str

&str is &s[..] string slice. String is string object. ".." is string literal.

Key Concepts

Ownership, Reference, borrowing, pattern matching, Stack and Heap, Result and Panic, Traits, Lifetimes, smart pointers

Move

 let x = 5;
let y = x;
let s1 = String::from("hello");
let s2 = s1;

fn takes_and_gives_back(a_string: String) -> String { // a_string comes into
a_string // a_string is returned and moves out to the calling function
}

Reference/borrowing

Model Intelligence, Token Use, and Cost per Task

Thu, 07 Aug 2025 00:00:00 +0800

Comparison snapshot: 2026-07-26. Prices and benchmark values change quickly; the linked primary sources are the authority for current values.

Executive summary

There is no single fair number called “cost per intelligence.” A model can have a low price per token but use many more tokens, or score well on a general benchmark while being a poor coding agent. The useful comparison is therefore three-dimensional:

Capability: a benchmark score, with the benchmark and reasoning setting named explicitly.
Token use: measured tokens for the same evaluation or task suite.
Task cost: the input/output token mix multiplied by the applicable prices.

On the common Artificial Analysis Intelligence Index, Claude Opus 4.8 has the highest score in this set, followed by GPT-5.6 Terra, Claude Sonnet 5, and GPT-5.6 Luna. However, the same source reports that Sonnet 5 generated far more tokens than the other models in its evaluation. Cursor Composer 2.5 is a separate case: it scores 62 on the Coding Agent Index and has very low measured coding-agent task costs, but it is not reported on the same Intelligence Index.