주제: Development

홈으로

  • AgentCLIHub Development Notes 3: Per-CLI Model, Reasoning, and Permission Settings

    AI-generated conceptual illustration of per-CLI settings. 한국어 버전 / Read in Korean This article explains the code at fixed commit 46dfca2. It does not report runtime testing or performance measurements in an actual environment. Bringing several AI CLIs into one window makes it tempting to reduce their settings to a handful of shared options. Yet the values and launch options behind familiar names such as “model” and “permissions” differ. Agent CLI Hub shares the UI structure while keeping each CLI's setting values and their meanings separate. 1. How a shared input field creates invalid combinations A model alias selected for Claude cannot be assumed to be valid when passed directly to Codex. Overwriting a shared model value every time the user switches CLIs also creates the problem of restoring the previous settings when they switch back. Settings has separate fields for ClaudeModel , CodexModel , and AntigravityModel . Reasoning effort and executable paths are also stored separ...

  • AgentCLIHub Development Notes 2: Finding Installed CLIs on Windows

    AI-generated illustration of Windows CLI path discovery. 한국어로 읽기 / Korean version This article explains the code at fixed commit 46dfca2, reviewed on October 17, 2026. It does not report runtime tests or performance measurements. codex can run in PowerShell while a GUI application on the same computer fails to find its executable. The two processes may have different PATH values and command-resolution environments. Agent CLI Hub’s discovery code determines both where a CLI is located and how to launch it. 1. Why a discovery result contains more than a path Alongside the executable path and discovery source, CliLocator.Result stores the Node executable, JavaScript entry point, PowerShell path, and whether to load a PowerShell profile. Even after locating a command, the launch method depends on whether it is a native program, an npm wrapper, or a function defined in a shell profile. For example, finding claude.cmd does not necessarily mean executing that file directly. If the asso...

  • A Beginner Trading Bot UI: Clear Stop Intent and Order Outcomes

    한국어로 읽기 TECH INSIGHT / EXPLAINED A beginner needs three answers from a trading panel: is the information current, what will this action change, and where can its outcome be checked? A screen answering those questions makes mistakes easier to notice than one simply displaying more numbers.

  • GAS PropertiesService: Scope, Validation, and Recovery for Small State

    한국어로 읽기 TECH INSIGHT / EXPLAINED A few dozen settings do not require operating a database. PropertiesService is convenient for small string-based state. Its storage contract still needs to define execution identity, input types, concurrent changes, and size limits.

  • Hosting a Manual on Cloudflare Workers: Structure and Cache Checks

    한국어로 읽기 TECH INSIGHT / EXPLAINED A manual should answer where users get stuck: does saving settings start the bot, does Stop cancel orders, and where should someone look when alerts disappear? A public document can explain these flows without crowding a small popup.

  • Minute Triggers in GAS: Runtime Budgets and Overlap Prevention

    한국어로 읽기 TECH INSIGHT / EXPLAINED A 54-second loop inside a minute trigger can look almost continuous. Its practical constraints are daily runtime and overlapping executions, more than the six-minute per-run limit. Decide the required observation frequency before filling the available time.

  • FCM Push from Apps Script: JWT Encoding and Recipient Management

    한국어로 읽기 TECH INSIGHT / EXPLAINED Push delivery involves two different tokens. An OAuth access token authorizes the server to send; an FCM registration token identifies a receiving app instance. Keeping them separate makes authentication failures and recipient failures easier to handle correctly.

  • Split Telegram Alerts: Authenticate Recipients and Track Delivery

    한국어로 읽기 TECH INSIGHT / EXPLAINED Users need to know what happened to an order; operators need to know why. Creating two messages from one event reduces noise. Splitting channels alone, however, does not establish privacy isolation.

  • Monthly-Distribution ETF Rotation: Separate Dates, Returns, and Order State

    한국어로 읽기 TECH INSIGHT / EXPLAINED Writing “rotate from the mid-month ETF to the month-end ETF” on a calendar is different from specifying executable orders. Dates, distribution eligibility, fills, and total return need separate representations. This article turns a two-instrument idea into testable states and calculations, without recommending a product or claiming profitability.

  • Chrome Web Store Submission: Reproducible Tests and Privacy Disclosures

    한국어로 읽기 TECH INSIGHT / EXPLAINED A store submission is more than an uploaded archive: reviewers need to reproduce the functionality you describe. For a trading bot panel, prepare a path that demonstrates settings, status, and errors without moving real money.

  • A Chrome Extension Bot Panel: Permissions, State, and Stop Controls

    한국어로 읽기 TECH INSIGHT / EXPLAINED Putting every bot metric in a tiny popup makes the stop control harder to find. Treat the extension as a remote control: show current state, its freshness, explicit start and stop actions, and the result of saving settings.

  • Google Apps Script Backend: Execution Budgets and Operating Boundaries

    한국어로 읽기 TECH INSIGHT / EXPLAINED Before building an order engine, build a small backend that reads prices periodically and keeps configuration. Google Apps Script lowers the cost of that experiment. A scheduled job, however, has a different operating model from a continuously running server.

  • Agent Hooks: Contracts for Blocking, Response Changes, and Notifications

    한국어로 읽기 TECH INSIGHT / EXPLAINED Hooks connect pre-tool checks, post-tool notifications, and response transformations outside the agent core. Their behavior depends less on event names than on whether the pipeline waits and what happens on failure.

  • Multi-Tenant AI Services: Isolation Across SQLite, Files, and Secrets

    한국어로 읽기 TECH INSIGHT / EXPLAINED A core multi-tenant test is: what happens when a customer A session submits customer B's resource ID? Apply that question to files, search indexes, caches, backups, and background jobs as well as list queries.

  • Knowledge Graph Retrieval: Versions, Rank Fusion, and Bounded Traversal

    한국어로 읽기 TECH INSIGHT / EXPLAINED Document search finds similar material; relationship search follows which incident a decision references. Combining them is useful, provided a link is not mistaken for proof of causation.

  • Multi-Agent Delegation: Task Contracts, Permissions, and Budgets

    한국어로 읽기 TECH INSIGHT / EXPLAINED Multiple agents help when independent tasks can produce verifiable results. If workers edit the same file or submit the same order, ownership and authorization boundaries matter before agent count.

  • Prompt Modes: What to Trim and Which Rules Must Remain

    한국어로 읽기 TECH INSIGHT / EXPLAINED A job checking whether mail exists may not need a long tone guide. It still needs permissions, privacy boundaries, and a completion contract. Use prompt modes to control optional context, not to remove essential controls.

  • LLM Provider Adapters: Contracts for Streaming, Tools, and Fallback

    한국어로 읽기 TECH INSIGHT / EXPLAINED Adding a second LLM changes more than a URL. Streaming APIs differ in tool-argument assembly, completion signals, and usage semantics. An adapter should centralize those contracts instead of scattering them across callers.

  • Three-Tier Agent Memory: Provenance, Corrections, and Retrieval Quality

    한국어로 읽기 TECH INSIGHT / EXPLAINED “Use the decision from last time” requires more than a longer recent-message window. The system needs to find who decided what, when, and in which source, then check whether it is still valid.

  • LLM Prompt Caching: Design the Boundary and Measure the Cost

    한국어로 읽기 TECH INSIGHT / EXPLAINED Prompt caching is worth testing when an application repeatedly sends the same tools and instructions. Measure reused tokens and total cost before concluding that rearranging a prompt removes most of the bill.