Use Git for collaboration; Git is required when a release candidate must be traceable. A Plugin-capable AI client improves discovery, but the generated repository also includes a standalone Skill and Markdown fallback. Go 1.26 or newer is needed only when building from source.
codex plugin marketplace add kcrmin/Stackcord --ref v1.0.0
codex plugin add stackcord@stackcordStart a new Codex conversation after installation so the six Stackcord Skills and lifecycle hooks are loaded from the installed snapshot.
Download the Plugin zip for the current platform together with checksums.txt, verify its SHA-256, and unpack it. The bundle contains .agents/plugins/marketplace.json, the six Skills, lifecycle hooks, project templates, both bootstrap scripts, and distribution/platform.json. That platform record binds the Plugin version to the matching CLI asset and checksum URL.
Ask the AI “Install this verified bundle locally.” It can inspect the platform record and run the matching checksum-first bootstrap. To install the unpacked Plugin through Codex CLI, add its directory as a local marketplace and install the listed Plugin from that marketplace:
codex plugin marketplace add /absolute/path/to/unpacked/stackcord
codex plugin add stackcord@stackcordThe bootstrap accepts only HTTPS release URLs, except loopback HTTP used by tests; it verifies the checksum and a doctor smoke test before atomically replacing the CLI. Hooks never download or install software.
From the product repository:
cd cli
go test ./...
go build -o ../bin/stackcord ./cmd/stackcordWindows PowerShell uses go build -o ..\bin\stackcord.exe .\cmd\stackcord. Put the resulting binary on PATH or tell the AI its absolute path. Run stackcord doctor --json to inspect Git and optional capabilities. This source-build path is for contributors; ordinary users should prefer the verified bundle.
For source-tree development, add this repository as a local marketplace and install it from Plugins or Codex CLI:
codex plugin marketplace add /absolute/path/to/stackcordIn Codex CLI, open /plugins after adding the marketplace. Plugin installation is optional; generated projects retain repo-local behavior.
Say “Start a new service with me” in an empty parent directory or “Adopt this existing project without overwriting my files” in an existing repository. The AI inspects the filesystem and Git first, loads the relevant Skill, and groups independent routine questions with recommended answers. Recommendations remain proposals until submitted; sensitive decisions require explicit answers. It saves normalized checkpoints and reports the current section, remaining sections and approximate question count. Saved decisions are recovered rather than asked again.
After initialization, use ordinary requests such as “What should I do next?”, “Build this feature”, “Check the contract and database impact”, or “Prepare a production candidate”. You should not need to manage internal IDs or command arguments.
Confirm that README.md, AGENTS.md, .agents/skills/use-project-harness/, .harness/, specs/, contracts/, and docs/ exist. Ask the AI to run a context audit and Git inspection. The audit must use repository files as evidence and must report unknown or stale items instead of inventing answers.
Read Core concepts, then choose New project or Existing project. See Task management and work reservation before coordinated parallel work. Use Troubleshooting when clone, context, Git, or optional-tool state is unclear.
When the product needs editable external mockups or an independent UI repository, use UI workspace and external mockups to choose the appropriate directory or submodule boundary.
Ask “Continue the saved questions.” Stackcord shows accepted decisions, the current section, completed/remaining sections and an approximate remaining-question range. Independent routine questions appear together with recommendations; sensitive questions and prerequisites stay explicit. A recommendation is not a submitted answer.
For direct inspection, use stackcord project discovery --draft <draft-root> --json before initialization, or stackcord project discovery --root <project-root> --json afterward. Omit --json for a readable summary; --locale en or --locale ko overrides the saved language. The command does not change files. Old checkpoints without progress metadata report unknown progress.
The complete optional discovery input is shown by stackcord project checkpoint --help. Save accepted decisions and remove answered questions together, then update section estimates. After initialization, question/decision text remains in specs/product/ and section planning lives in .harness/discovery.yaml, so a clone can recover without the original draft. Harness readiness means that scope is known and no marked blocking question remains; it neither grants policy approval nor automatically creates the harness.
Run stackcord dashboard --root . and open the printed local URL. The bundled browser UI needs no Node runtime or hosted account. It shows discovery, live GitHub Issues, PR links, your assigned issues and requested reviews, settings, and diagnostics. Stop the command to close the session; it does not notify while closed.
Use stackcord setup --json to inspect the local UI preference, or stackcord setup --ui enable --apply (also disable or ask) to record it. Plugins offer this choice during first use; host installation hooks never install software or open windows. Codex and Claude use the same project files and CLI, with separate host manifests and hook adapters. Claude package validation is exercised; executable fixtures are not proof of every host version's session behavior.
Settings require a preview before applying. Shared edits become working-tree proposals; personal language/UI choices remain local. A stale revision from another session is rejected. Interrupted operations leave receipts/locks for inspection rather than silently overwriting newer state. Commit and review shared changes through the normal feature-branch PR flow.