Skip to main content

aside-jun

Use Aside's signed-in browser from Codex, Claude Code or another coding agent. Route authenticated browsing, explicit Aside code-mode batches and visual artifact work between aside-codemode, Aside REPL and Aside's agent. Use when the user asks for Aside or its extensions, or needs its existing browser session; ordinary public HTTP fetching and unrelated local coding work stay with the host's tools.

소스 정보

저장소
lidge-jun/aside-skill
최근 소스 활동
2026년 9월 25일 15:04
감지된 SKILL.md 언어
영어
스타
12
포크
1

설치 방법

기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.

소스 파일 검토

설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.

파일 탐색기
18 개 파일

SKILL.md 표시 중

SKILL.md
소스 지침 · 읽기 전용 미리보기
name
aside-jun
description
Use Aside's signed-in browser from Codex, Claude Code or another coding agent. Route authenticated browsing, explicit Aside code-mode batches and visual artifact work between aside-codemode, Aside REPL and Aside's agent. Use when the user asks for Aside or its extensions, or needs its existing browser session; ordinary public HTTP fetching and unrelated local coding work stay with the host's tools.
# Aside from a coding agent Use the smallest available surface that can finish the requested work. Codex, Claude Code and other callers can invoke **aside-codemode directly**; they do not need to ask Aside's agent to run every batch. This skill is guidance, not a runtime or an installer. The user's chosen tools, scope and permissions remain decisive. ## Choose the execution surface | Work | Route | Read | |---|---|---| | Any Aside work when this caller's live tool list shows the native Aside MCP (`aside mcp`: `exec`, `repl`, often `memory_search`) | Call those tools instead of shelling out to the CLI; the REPL/exec rules below still apply | [Delegation](references/exec.md), [REPL](references/repl-api.md) | | Directory/project search within an explicitly requested Aside/code-mode workflow; 2+ independent known files, URLs, queries or captures | **Direct aside-codemode** through a tool exposed to this caller, otherwise the installed CLI | [Code-mode calls](references/codemode.md) | | First look at an unfamiliar page; one visible click; a dependent form/wizard step | Native Aside REPL, inspect → act → inspect | [REPL](references/repl-api.md) | | Current/already-open tab | Identify and attach the existing tab; never replace it with a fresh URL batch | [REPL](references/repl-api.md); verified codemode `browse.attach` is also available | | Login, MFA, account change, approval, or a browser task needing judgment | Aside `exec` — it signs in by itself; hand it the login first. Within the task's existing authorization; stop when a human gesture is required | [Delegation](references/exec.md), [credentials](references/credentials.md) | | Composed HTML/SVG report, visual explanation or PDF through Aside | `exec` using the selected Aside account's **dev-visualizer** skill; codemode may collect inputs first | [Visualizer handoff](references/visualizer.md) | | A public page that only needs HTTP, or unrelated local coding work | The host's existing reader/search/file tools, unless the user explicitly selected Aside/code mode | No Aside setup needed | Independent means one item's result is not needed to choose the next item's operation. Do not turn a login flow, cart, wizard or uncertain side effect into a parallel batch. A batch failure is a result to inspect, not an invitation to replay completed actions. Two independent reads can also run as concurrent one-shot `aside repl` processes (4 and 8 observed isolated, zero leaked tabs); see [parallel pages](references/repl-api.md#parallel-pages). ## Sign-in is exec's job `aside exec` logs in on its own with Aside Vault or a connected password manager. When a site needs a login, delegate the sign-in to exec in the task prompt; do not ask the user to sign in or unlock first. Ask the human only after exec reports a concrete blocker (locked Vault, no credential, MFA, passkey, CAPTCHA). No account switching, new accounts or policy changes. Prompt clause and details: [exec](references/exec.md#sign-in-is-delegated), [credentials](references/credentials.md). ## Direct code-mode calls from Codex and Claude 1. Look for a **currently exposed aside-codemode MCP tool** and inspect its real schema and server identity. A matching name alone is insufficient. If present, send the bounded guest program directly to it. 2. Otherwise resolve the **already installed** `codemode` command or absolute Node/CLI pair, identify its package version, and use `--code-file` with an explicit `--cwd`. [The cookbook](references/codemode.md) includes Bash and PowerShell forms, executable guest examples and result handling. 3. If neither route is available, report the missing capability and use an already available authorized native route when it can satisfy the request. Do not install, register MCP, force configuration, or update a package merely because this skill was loaded. `codemode --install-mcp` configures **Aside's** MCP client. It does not register a server in Codex or Claude Code. Host-qualified tool names vary; never invent `mcp__...` names from an example. `execute_code` accepts the schema the caller actually exposes, not arbitrary CLI flags. The guest is an async JavaScript body: `await` tool operations and **`return` the answer**. Codemode's `search`, `browse` and `fs` globals exist inside that guest. They are not globals in Codex `functions.exec`, Claude's shell, Node, or native `aside repl`. A host Code Mode tool may orchestrate a shell/MCP call if its own contract allows it; it is a different runtime from aside-codemode. For completeness-sensitive searches, identify the runtime and require the 0.9.0 result contract, including the actual metadata. Preserve `complete`, `truncated`, `partial`, `scope` and `scope.coverage`; keep per-item errors and skipped reads. An older runtime or missing metadata cannot prove exhaustive absence. Positive bounded reads may still be useful with their limits stated. Neither an exit code nor `ok:true` proves that a page contains the requested data. ## Confirm browser account and host before use For browser work, inspect `aside --version`, relevant `--help`, and `aside account list` on the execution host. Use an explicitly requested account and host; derive absolute paths from that account instead of copying `u0` from an example. Never print account credential files. For personal context (people, projects, preferences) query Aside memory before asking the user: `aside memory search "<q>" --json`, `list`, `show <path>`, or the attached `memory_search` tool. It is read-only; never edit memory files, and route "remember this" through exec. Codemode's browser runner inherits native `aside repl` defaults unless the installed build supports explicit context. npm `latest` is 0.9.2, which adds CLI `--account`/`--host` and per-call MCP `account`/`host`; 0.9.0 has neither. `codemode` has no `--version`, so identify the build by [resolving its package](references/codemode.md#identify-the-installed-version) and check the live schema. When the verified build cannot pin a required nondefault account or remote host, use native Aside's explicit `--account` / `--host` route. Keep the same workload split on that native route: REPL for known mechanical steps, exec for judgment/login. Do not silently switch global defaults to make a recipe work. For an installed build with the #44 fix, follow the [explicit context contract](references/codemode.md#context-selection-in-the-091-source-contract); verify capability rather than assuming a dev merge upgraded the installed package. Read the host instructions when running Aside `exec` or resolving a missing CLI: [macOS](references/host-macos.md) or [Windows](references/host-windows.md). CLI, browser app and background components have separate versions; current help and measured behavior take precedence over stale examples. See the dated [compatibility notes](references/compatibility.md). ## Keep execution and evidence honest - Every noninteractive `exec` uses the host deadline and a prompt that names authorized inputs, output paths and the no-question stopping rule. Read [exec.md](references/exec.md) before delegating. - Permission mode changes capability, not user authorization. Do not widen roots, enable browsing, disable biometrics or change persistence settings as a fallback to a denied operation. Use the authorized scope or report what is blocked. - Before native REPL, read `aside guide repl` and the relevant installed builtin skill via `aside skills list/show`. Do not assume the builtin catalog includes account-installed user skills such as dev-visualizer. - One-shot `aside repl "..."` commands have separate temporary sessions. Keep an inspect–act–verify flow and download verification inside one invocation. Borrow an existing tab when it must survive; use current snapshots and real refs. - Before writing session files, prepare `./tmp` and `./artifacts`. For element screenshots use the [tested source-capture and host-crop workflow](references/repl-api.md); native locator capture and clip origin are unreliable on measured builds. - A timeout or indeterminate effect may leave completed side effects behind. Inspect the destination state before retrying. Do not repeat a submission just because a CLI stopped reporting it. - Pass subcommands as separate argv words. Anything the CLI does not parse as a subcommand becomes an agent prompt: `aside version`, or `"account list"` as one quoted word under zsh, starts a paid exec session. Call agent runs explicitly as `aside exec -- "<prompt>"`. - Exit 0 is not success: REPL errors end in `[error | Nms]`, and exec printed ` • Error 403` (provider quota) with exit 0 on 1.26.906. Read the output. - Run `aside skills install` only when the user asks; it installs the official aside-browser skill into coding agents, which shadows this one. - Verify actual content and artifacts before reporting completion. A local path returned by a remote Aside host is not yet a file delivered to this caller. Additional references are conditional: [gated research](references/deep-research.md), [scheduling](references/scheduling.md), [permission diagnostics](references/permissions.md) and the [builtin catalog](references/builtin-skills.md). Historical measurements remain useful diagnostics; they are not universal promises about newer builds.
GitHub에서 보기