ContractorKeith/zigagent

最近记录的来源活动: 2026年11月2日 15:26

SkillsMP 收录数据更新: 2026年11月5日 19:57

已收集 skills: 1

GitHub 星标: 0
GitHub Forks: 0

用 AI 助手安装

复制这段 Prompt，发给你正在使用的 AI 助手。

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/ContractorKeith/zigagent.

仓库 README

仓库 README 来源：README.md

查看 README 源文件

展开完整 README

# zigagent A small native Zig CLI that supervises an installed Codex for repository fixes. Version **0.1.0** implements the v1 supervisor. Native Linux and Mac acceptance, packaged-binary subscription checks, and independent release skeptic reviews have passed. See [the plan](docs/PLAN.md) and [release checklist](docs/RELEASE.md); published artifacts appear on [GitHub Releases](https://github.com/ContractorKeith/zigagent/releases). ## Install (after a release is published) Choose `zigagent-0.1.0-linux-x86_64.tar.gz` (Linux glibc 2.28+) or `zigagent-0.1.0-macos-arm64.tar.gz`. Download that archive and `SHA256SUMS` from the same trusted release. Verify its SHA256 before extracting (`sha256sum` on Linux, `shasum -a 256` on macOS), inspect the expected contents in [release docs](docs/RELEASE.md), then copy its executable `zigagent` to a directory on PATH, such as `~/.local/bin`. Python and Zig are not runtime requirements; OS libc and external Codex are. Install official Codex separately, pinned to **0.166.0** (`>=0.166.0,<0.166.1`; no prereleases). For an existing Node/npm setup, `npm install -g @openai/codex@0.166.0`, then `codex --version`. Authenticate separately with official `codex login` and confirm `codex login status` reports ChatGPT. An eligible ChatGPT subscription and a model available to that account are required; zigagent does not provide Codex or model access. See [backend controls](docs/CODEX-BACKEND.md) for the tested invocation and configuration limits. ## Build and test On newer macOS SDKs, pinned Zig may need an installed compatible SDK selected for the build process; see [SDK setup](docs/TESTING.md#macos-sdk-compatibility). No system configuration change is needed. Requires Python 3 for tests and release tooling, Git for scratch-repository assertions, and Zig **0.15.2** for builds. The native binary has no Python dependency. ```bash # Official archives, pinned SHA256, Linux x86_64 or macOS arm64: toolchain=$(bash scripts/install-zig.sh) export PATH="$toolchain:$PATH" bash scripts/accept.sh ./zig-out/bin/zigagent --help ./zig-out/bin/zigagent --version ``` The installer writes only to `~/.local/share/zigagent-toolchains` (or its explicit directory argument). The gate runs format checking, unit tests, build and compiled-process black-box tests, and is the same entrypoint CI runs on both target hosts. See [testing](docs/TESTING.md). ## Run ```bash ./zig-out/bin/zigagent run --help ./zig-out/bin/zigagent login --help ./zig-out/bin/zigagent run --model gpt-6.1-sol --json 'Fix the parser' # One schema-version-1 result; without a check, verification is not_requested. printf '%s' 'Fix the parser' | ./zig-out/bin/zigagent run --model gpt-6.1-sol --json --no-input - ``` ```bash ./zig-out/bin/zigagent run --model gpt-6.1-sol --workspace ./repo \ --check 'zig build test' --json --no-input 'Fix the parser' ``` `--check` runs `/bin/sh -c COMMAND` in the original workspace directory with EOF stdin, only after valid Codex completion and zero backend exit. It shares the execution deadline; an exhausted budget never starts it. Check output retains a bounded, UTF-8-safe combined stdout/stderr tail (read order across separate pipes). Only the explicit check command is shell syntax; prompts, models and paths stay literal. Failure preserves edits and never starts a repair or retry. An open directory handle preserves check location if the backend renames or replaces the workspace path. | Outcome | Run / verification | Exit | |---|---|---| | Completion without check | completed / not_requested | 0 | | Check exits zero | completed / passed | 0 | | Check exits nonzero or is signaled | failed / failed (`check_failed`) | 5 | | Failure before check launch | failed / not_run | 1–4 by cause | | Deadline before/during check | timed_out / not_run or failed | 124 | | SIGINT/SIGTERM before/during check | cancelled / not_run or failed | 130/143 | A started but interrupted check is `failed`; `exit_code` is the observed normal exit status, or null when signal-terminated/unobserved. Its retained output is still reported. Plain mode prints `verification: STATUS (exit N)` when a check was requested, followed by its output and an explicit truncation marker when needed. Errors go to stderr; successful summaries/checks go to stdout. JSON mode emits exactly one schema-version-1 envelope and no backend/check chatter. The complete error/exit matrix is in the [spec](docs/features/supervisor/SPEC.md#result-schema-and-exits). `--model` overrides `ZIGAGENT_MODEL`; missing/empty models fail usage. Values are separate arguments, not `--flag=value`. `--` ends run option parsing. Workspace and additional-directory paths resolve from the caller's current directory. Invalid arguments exit 2, with one JSON envelope when requested. Help/version need no backend or credentials. Omitted/`-` prompts read nonterminal stdin under the preflight deadline, capped at 1 MiB. An argument prompt leaves caller stdin untouched. All prompts reach Codex through a separate stdin pipe; model/path values are literal argv data. Invalid UTF-8 and NUL prompt bytes fail usage. Preflight shares one `--timeout` budget; authenticated execution receives a fresh full budget. Summaries retain the last completed agent message's prefix; diagnostics retain a bounded stderr tail. Truncation flags identify loss; invalid/cut UTF-8 bytes become `?`. Missing login starts official `codex login` only when a controlling terminal exists and `--no-input` is absent. Login output goes to that terminal, never JSON stdout; `--login-timeout` separately bounds login and the following ChatGPT auth verification. `zigagent login` uses that same bounded flow. Unattended callers receive `authentication_required`; API-key login and unrelated status failures return `authentication_failed` without launching login. A mutating run is never retried. SIGINT/SIGTERM cancel with exits 130/143; deadline expiration returns 124. Owned groups receive TERM, then KILL after one second (100ms following ordinary leader exit), before the leader is reaped. Cleanup drains/reaps for at most three seconds total. Descendants retaining pipes or closing them cannot hold normal completion indefinitely. Closed caller stdout stops owned work; final output writes each have a two-second bound. Deliberately detached processes and uninterruptible OS waits remain outside cleanup guarantees. All edits remain on failure. [Backend controls and configuration limits](docs/CODEX-BACKEND.md) describe the exact invocation. Subcommand `--help` stops parsing immediately; trailing arguments are ignored, while errors before it still fail. Option values beginning with `--` are treated as missing values: use `./--context` or an absolute path for a directory named `--context`. Every repeated `--add-dir` is validated. The backend is a separately installed Codex with ChatGPT subscription login. Codex owns the agent loop. The caller supplies OS isolation; working directories and added context paths do not enforce an access boundary. See the [glossary](CONTEXT.md).

正在加载仓库 README…

这个仓库中的 skills

职业分类待补全

skill

职业分类

描述

更新

skeptic-env

未分类

Run zigagent review against disposable repositories and fake or explicitly authorized real Codex backends.

原文语言：英语

2026-11-02

已展示 1 / 1 个已收集 Skill。