8f75e27554
* feat(#3045): deny an executor dispatch that drops its isolation flag Every isolation gate already resolved correctly. The resolved value then reached the executor through a prose instruction telling the model to substitute it into a call the model composes itself, and nothing verified the substitution. When it was dropped, the executor edited and committed in the user's primary checkout with no consent and no warning. A prose backstop would be the same class of artifact as the defect, so this is a shipped PreToolUse hook on the Agent tool. It fires at the instant of the call rather than being read once at the top of a workflow, which is the only placement the model cannot skip. The guard is inert unless it can positively establish that this is a GSD project, that the project resolves to harness isolation, and that the dispatch targets an executor. A non-GSD repo has no invariant to enforce. Where it cannot read the configuration at all, it denies rather than assuming, with its own reason -- a guard that cannot verify must not answer safe. A malformed payload allows rather than throwing. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(#3045): extend the isolation guard to Cursor Cursor is the second of only two runtimes that resolve harness isolation, so shipping the guard for Claude alone left half the exposed surface unguarded while the changeset implied it was covered. The two runtimes fail differently. On Claude the harness flag is a per-dispatch kwarg the model must copy into a call it composes, and the defect is that it can be dropped. On Cursor the flag is --worktree, which applies to the whole session, and the subagent-start payload carries no isolation field at all. There is no flag to check, so the guard verifies the effective state instead: whether the workspace is genuinely running outside the user's primary checkout. That is a stronger check than the Claude one because it tests reality rather than intent, and it is commented so nobody later rewrites it into a flag check. Isolation is established two ways, either sufficient: the workspace resolves to a linked git worktree, or it sits under the worktree root Cursor manages. The second matters because a directory Cursor placed there is a legitimate isolated session even before it becomes a distinct git worktree, where linkage alone would report no repository. Detecting linkage required a new primitive rather than the existing context resolver. That resolver short-circuits on finding a local .planning directory before it ever compares the git directory to the common one -- and an isolation worktree normally has its own checked-out .planning. Reusing it would have read a correctly isolated session as unisolated and denied it, which is the failure direction that gets a guard switched off. The comparison is now its own shortcut-free function that the resolver delegates to after its own shortcut, so existing behavior is unchanged, and the case that would have broken is pinned. The subagent type is checked before any configuration is read, so an unreadable config cannot deny a dispatch this guard would never have enforced against. The input-schema comment on the Cursor hook documented only the fields common to every event and omitted the ones specific to this one. That omission cost a halt during this work; it now documents both. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(#3045): enforce the resolved dispatch decision, not the host capability The guard keyed on the registry's dispatch.isolation, which says only that a runtime is CAPABLE of harness worktrees. The decision that actually governs a dispatch is the one the workflow resolves after gating, and that legitimately comes out as sequential in three documented cases: a project setting use_worktrees false, a per-plan submodule intersection, and the base-check auto-degrade. The workflow tells the model to omit the flag in exactly those cases, and the guard was denying every one of them. The third case matters most. The preceding fix made the base-check degrade on git timeouts and a missing git binary, where it had previously answered "safe". That correction is right, and it means a transient hang now degrades to sequential far more often than before -- so the two changes composed into a trap where the workflow behaved exactly as designed and the guard blocked it. The workflow already resolves isolation in shell, deterministically, which is what makes it a trustworthy source in a way the model-authored call is not. It now records that resolved value through a dedicated verb, and both guards read it first. A fresh record is authoritative, so sequential dispatches pass untouched. Absent or stale, the guards fall back to the capability check combined with the project's use_worktrees setting, which still covers the case that never reaches the workflow. Also widened the matcher to accept Task alongside Agent, since a host that names the tool Task would otherwise leave the guard silently inert while implying coverage; stopped assuming Claude when no runtime is declared, which is the shipped default and would have demanded a Claude-only argument elsewhere; and made a non-git project inert rather than denied, since advising a worktree session is not actionable without a repository. The original diagnosis never modeled sequential mode as legitimate. That omission is what let this through, and it is now recorded there. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(#3045): record at resolution and bind the record to its dispatch Two independent reviews converged on the same failure: the guard was fail-open in a default install, so it did not catch the defect it exists to catch. A shipped project carries no runtime key, which made "runtime not confidently known" the common case rather than a corner one. A record asserting that isolation was required but carrying no flag then fell through to a capability lookup that answered "none", and the dispatch was allowed. The flag itself only arrived from a second shell block -- the same block a model dropping the argument would also skip. A test had pinned that behavior as intended. The record is now written by the resolver, as an unavoidable consequence of asking for the value, rather than by a step the model is told in prose to go and run. A guard against a prose-carried value cannot itself depend on prose. Mode, flag and identifiers are written together and atomically, so the flagless window is gone, and a record asserting isolation with no resolvable flag now denies instead of degrading. Runtime is also resolved from the installer's own recorded default, which makes confident resolution the normal case. The per-plan submodule gate degrades after the phase-level decision and never re-recorded, so a plan that legitimately ran sequentially was denied against a still-fresh phase record. It now records its own, scoped to the plan. A record also authorized any dispatch for four hours. One phase degrading to sequential could silently license an unisolated dispatch in the next. Records now carry phase and plan, the guards require them to match, and the window is minutes rather than hours -- the resolver rewrites it before every dispatch, so a long window bought nothing and only widened the hole. The flag validator rejected any value beginning with two dashes, which is exactly the form Cursor and Windsurf declare, so their real value could never have been stored. Writer and reader also derived the record path differently and diverged inside a linked worktree without local planning state. The predictable path remains a way to silence the control without leaving a trace in the diff. It grants no access an agent with shell does not already have, so it is documented as accepted rather than redesigned around. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(#3045): correct the staleness boundary and unmask a vacuous parity test The remote runner returned twenty failures. One was a real production defect the boundary case existed to catch: a record whose age exactly equalled the staleness window was treated as fresh, so it stayed authoritative for one tick past its own expiry. Freshness is now strictly inside the window. The parity test meant to stop the two guards' executor lists from drifting could never have failed. Its project fixture was a bare directory rather than a repository, so the non-git inert branch answered before the executor list was ever consulted. It asserted agreement it never actually measured. The fixture is now a real repository, like every sibling in the file. A test also asserted that Windsurf declares the worktree flag. It does not -- Windsurf resolves to no isolation by design, having no named concurrent dispatch to isolate. The test claimed a registry fact that was never true, and a comment in the resolver repeated it. Both corrected, and the test now proves what it should have all along: that the parser accepts any bare flag value, rather than one runtime's supposed value. The new guard was missing from the bundled-hook whitelist, which is the surface that decides what actually ships, and the per-plan gate had gained calls to the launcher without the preamble those calls require. The changeset carried parenthetical product descriptions the purity rule forbids. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * chore(#3045): backfill changeset pr number Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * test(#3045): make the guard tests hold on Windows Two tests redirect HOME to control where the installer-persisted runtime default is read from. Node resolves the home directory from USERPROFILE on Windows and never consults HOME, so both silently read the real runner profile, found no recorded runtime, and asserted against a project the hook had not recognised. The production code was already correct in asking the platform rather than the variable; only the tests were wrong to assume one variable answers everywhere. The helpers now mirror the override onto both. The symlink spoofing test also created a directory symlink unconditionally, which needs elevated privileges on Windows. It survived on this runner, but it would fail on any host without them, so the creation is now attempted and the test skips explicitly when it cannot be done -- a bare return would have counted as a pass and hidden the gap. Skipping alone would have left the platform uncovered, so the behaviour it proves is now also driven in-process through an injected realpath, following the seam already used for the clock. That case no longer depends on privileges at all, and the end-to-end test keeps its original assertions wherever symlinks work. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
33 KiB
How to install GSD Core on your runtime
Install GSD Core (@opengsd/gsd-core) into the AI coding runtime you use every day. This guide gives you the standard installer path for each supported runtime, then covers the manual path for machines without Node.js.
What you need: Node.js 18+ and npm (or npx). If you do not have Node.js, jump to Installing without Node.js.
Why the installer is required
GSD Core ships agent and command files in Claude Code's native frontmatter format. Each supported runtime expects a different schema, directory layout, and command-invocation syntax. The installer performs the necessary transformations — for example, converting tool lists and colour values for OpenCode and writing TOML agent entries for Codex.
Do not copy files from agents/ or commands/ directly. Doing so bypasses the transformations and produces schema-validation errors or missing commands.
Standard install
Run the installer from any directory. It prompts for your runtime and whether to install globally (all projects) or locally (this project only).
npx @opengsd/gsd-core@latest
That is the only command you need for a fresh install or to re-run the installer after switching runtimes.
Per-runtime instructions
Claude Code
npx @opengsd/gsd-core@latest --claude --global
Skills land in ~/.claude/. Commands appear as /gsd-* slash commands in your next Claude Code session. Restart Claude Code to pick them up.
Override the install directory:
CLAUDE_CONFIG_DIR=~/.claude-alt npx @opengsd/gsd-core@latest --claude --global
Hook coverage
GSD registers the following Claude Code hook events automatically on install:
| Event | Hook | Purpose |
|---|---|---|
SessionStart |
gsd-check-update.js, gsd-session-state.sh |
Update check, session orientation |
PostToolUse |
gsd-context-monitor.js, gsd-read-injection-scanner.js, gsd-phase-boundary.sh, gsd-graphify-update.sh |
Context monitoring, read-time scan, phase boundary detection |
PreToolUse |
gsd-prompt-guard.js, gsd-read-guard.js, gsd-workflow-guard.js, gsd-worktree-path-guard.js, gsd-agent-isolation-guard.js, gsd-validate-commit.sh |
Prompt guard, read-before-edit, workflow + worktree safety, agent-dispatch isolation, commit validation |
SubagentStop |
gsd-context-monitor.js |
Context headroom tracking after subagent completion |
Stop |
gsd-context-monitor.js |
Context headroom tracking before model stop |
PreCompact |
gsd-context-monitor.js |
Context awareness before conversation compaction |
FileChanged (matcher: config.json) |
gsd-config-reload.js |
Hot-reloads .planning/config.json context mid-session when you edit your GSD config — no session restart required |
The FileChanged hook is always-on and a no-op when .planning/config.json does not exist in the project. Editing that file while a session is running injects an additionalContext summary of the new configuration so the agent picks up model overrides, workflow toggles, and hook settings immediately.
Claude Code — native plugin install
GSD Core ships a .claude-plugin/plugin.json manifest, which enables installation and lifecycle management through the Claude Code plugin system. This path is additive — the npm installer above remains fully supported, and the two approaches differ in namespace and lifecycle only.
Install paths
Option A — marketplace or git install (once listed):
claude plugin install gsd-core
Option B — zero-friction skills-dir load: Claude Code automatically discovers any directory under ~/.claude/skills/ that contains a .claude-plugin/plugin.json as a plugin. To use gsd-core this way, place (or symlink) the gsd-core package directory there:
# Example: place the package under ~/.claude/skills/gsd-core/
# Claude Code loads it as gsd-core@skills-dir on the next session start.
# No explicit install step required.
Command namespace
Plugin commands are namespaced as /gsd-core:<command> — for example, /gsd-core:plan-phase. This is distinct from the classic npm/file-copy installer, which exposes commands as /gsd:<command>. Use whichever namespace corresponds to your install method.
Lifecycle
claude plugin enable gsd-core
claude plugin disable gsd-core
claude plugin update gsd-core
Hooks
The plugin wires gsd-core's always-on guard and update hooks automatically via hooks/hooks.json. No manual hook registration is required.
Prerequisites
The gsd-tools binary (installed as part of the @opengsd/gsd-core npm package) must be available on your PATH for gsd commands to execute their backing logic. The plugin delivers the command, agent, and hook surface; the npm package delivers the runtime CLI.
Node.js (node) must also be available on your PATH. The plugin's always-on guard hooks (wired in hooks/hooks.json) are invoked as node "${CLAUDE_PLUGIN_ROOT}/hooks/<script>". Some Claude Code distributions ship as a standalone binary and do not expose a node executable on PATH; in those environments the plugin's hooks will not run. Verify with node --version before relying on the plugin hooks.
Runtime build (self-healing). The runtime CLI's compiled modules under gsd-core/bin/lib/*.cjs are build artifacts (ADR-457): they are compiled from src/*.cts by npm run build:lib and shipped prebuilt in the npm tarball. A plugin-marketplace or git-clone install materializes the repository tree directly and never runs that build step, so those files are initially absent. The CLI heals this automatically: the first gsd-tools invocation detects the missing output and compiles it once (using the bundled typescript devDependency), then proceeds normally. You may see a one-time gsd: runtime library not built — compiling once… notice on stderr; subsequent commands are unaffected. If auto-build cannot run (for example node_modules was pruned to production-only and typescript is unavailable), the CLI prints an actionable message telling you to run npm install && npm run build:lib in the plugin directory.
Claude plugin marketplace discovery (ZCODE and compatible runtimes)
GSD Core also ships a .claude-plugin/marketplace.json marketplace manifest (sibling to plugin.json). Runtimes that implement the Claude plugin marketplace contract — such as ZCODE — can discover and install GSD Core from a custom marketplace source without a manual clone:
- In your runtime's plugin UI, add a custom marketplace source pointing at
open-gsd/gsd-core(GitHubowner/repoform). - GSD Core appears in the catalog and can be installed directly from the UI.
This path is additive and changes nothing about the Claude Code plugin install above (.claude-plugin/plugin.json is unchanged). The marketplace entry's source is ./, so it reuses plugin.json's commands / skills / hooks mapping. The catalog version tracks package.json (it lives at plugins[0].version and is stamped by the release version-sync), so the version you see in the marketplace matches the npm release.
OpenCode
npx @opengsd/gsd-core@latest --opencode --global
The installer writes four surfaces under ~/.config/opencode/ (XDG) or ~/.opencode/: flat slash commands in commands/ (plural — the directory OpenCode discovers slash commands from, #2329), file-based subagents in agents/, on-demand skills in skills/<name>/SKILL.md, and a native plugin in plugins/gsd-core.js. It converts agent frontmatter to OpenCode's schema — removing the tools: field and converting colour values to hex — and emits each skill with spec-compliant frontmatter (name matching the skill directory plus a description). Skills are loaded on demand via OpenCode's native skill tool; commands remain invokable as /gsd-*. See Installing without Node.js — OpenCode transformations if you need to understand what changes.
GSD safety hooks on OpenCode. OpenCode does not register lifecycle hooks the way Claude Code does (its hooksSurface is none), so GSD's prompt-injection guard, read-before-edit guard, injection scanner, and context monitor would otherwise be inert. The bundled plugin (plugins/gsd-core.js) closes that gap: OpenCode auto-discovers plugins/*.{ts,js} files under its config directory at startup and the adapter bridges OpenCode's event bus (tool.execute.before/after, session.created, file.edited) onto GSD's existing hook scripts, spawning them as subprocesses. No opencode.json entry is needed — the plugin is loaded by directory auto-discovery (the config plugin array is for npm packages only). A blocking hook aborts the tool call; an advisory hook surfaces its message without blocking.
Your plugin directory is pinned to CommonJS (accepted trade-off, #2544). GSD's adapter is a CommonJS .js file, and Node decides a .js file's module type by walking up for the nearest package.json. So the installer writes a minimal {"type":"commonjs"} marker into the plugin directory itself — plugins/package.json on OpenCode and Kilo, extensions/package.json on pi. It is written only when GSD actually stages its adapter there, it never overwrites a package.json GSD did not write, and uninstall removes only its own.
The trade-off: that marker shadows your config root for every .js file in that directory, not just GSD's. If you author your own plugins as ESM .js and rely on a "type": "module" at the config root, they will stop resolving as ESM. This is deliberate — it is strictly narrower than the pre-#2544 behavior, which wrote the marker over <configRoot>/package.json itself and destroyed whatever was there — but it is a real constraint rather than a pure improvement, which is why it is stated here.
Mitigation: author your own plugins as .ts. OpenCode and Kilo compile plugin TypeScript with Bun, and a package.json type field does not affect .ts resolution — so a .ts plugin is unaffected by the marker. Failing that, keep ESM plugins outside the auto-discovered directory and load them as npm packages via the config plugin array.
Override the install directory:
OPENCODE_CONFIG_DIR=~/.config/opencode-alt npx @opengsd/gsd-core@latest --opencode --global
Kilo
npx @opengsd/gsd-core@latest --kilo --global
The installer writes the same three surfaces under ~/.config/kilo/ (XDG) or ~/.kilo/ as for OpenCode — flat commands in command/, subagents in agents/, and skills in skills/<name>/SKILL.md — since Kilo derives from OpenCode and shares its config schema and skill layout.
Override the install directory:
KILO_CONFIG_DIR=~/.config/kilo-alt npx @opengsd/gsd-core@latest --kilo --global
Codex
npx @opengsd/gsd-core@latest --codex --global
Skills land in ~/.codex/skills/gsd-*/SKILL.md. Agents are written as standalone ~/.codex/agents/gsd-*.toml files, which Codex auto-discovers — that is the sole registration source for each role; config.toml only carries the shared [agents] dispatch-tuning scalar (max_depth), not a per-role table (#2406). Restart Codex (or run codex --reload) after install.
Minimum supported version: Codex CLI 0.130.0. Earlier versions had additional skill-root scanning that can produce duplicate listings.
Hook coverage
GSD registers the following Codex hook events automatically on install (requires Codex CLI 0.137.0+ for the stable hook-event schema):
| Event | Hook | Purpose |
|---|---|---|
SessionStart |
gsd-check-update.js |
Update check at session open; Windows installs also emit a commandWindows field pointing to the .cmd shim so Codex picks the correct executor on Windows without requiring per-OS config regeneration |
SubagentStart |
gsd-context-monitor.js |
Inject context / GSD_AGENT_NAME awareness at subagent open |
Stop |
gsd-context-monitor.js |
Context headroom tracking before model stop |
PostToolUse |
gsd-context-monitor.js |
Mirror the context-monitor coverage available in Claude Code |
All registered hooks are managed by GSD and are removed cleanly on --uninstall.
Kimi CLI
Support boundary — legacy
kimi-clivs Kimi Code. This integration targets the legacy/Pythonkimi-clicustom-agent contract. Thekimi --agent-file <configRoot>/agents/gsd.yamllaunch shown below is accepted bykimi-cli. The newer npm Kimi Code (@moonshot-ai/kimi-code, e.g.0.11.0) does not accept--agent-file; it discovers skills through fixed skill roots and--skills-dir. The generated/skill:gsd-*skills work in both, but the custom-agent (--agent-file) surface is specific to legacykimi-cli. For Kimi Code, point it at the installed skills root with--skills-dir <configRoot>/skillsinstead of using--agent-file.
npx @opengsd/gsd-core@latest --kimi --global
Skills land in Kimi's first existing generic user skills root:
~/.config/agents/skills/gsd-*/SKILL.mdwhen~/.config/agents/skillsalready exists, or when neither generic root exists yet~/.agents/skills/gsd-*/SKILL.mdwhen~/.agents/skillsalready exists and~/.config/agents/skillsdoes not
Start a new Kimi CLI session after install, then invoke GSD skills with /skill:gsd-*, for example:
/skill:gsd-new-project
The installer also writes the GSD custom agent definition to the same selected config root: <configRoot>/agents/gsd.yaml with its prompt at <configRoot>/agents/gsd.md; subagents land under <configRoot>/agents/subagents/gsd-*.yaml and <configRoot>/agents/subagents/gsd-*.md.
Kimi custom agents do not auto-activate just because the files exist. Launch Kimi with the generated agent file when you want the GSD agent surface:
kimi --agent-file ~/.config/agents/agents/gsd.yaml
If your machine already uses ~/.agents/skills and does not have ~/.config/agents/skills, GSD installs there instead and the launch command becomes:
kimi --agent-file ~/.agents/agents/gsd.yaml
Kimi also discovers user skills from the brand-specific ~/.kimi-code directory. If your Kimi setup is already centered on ~/.kimi-code, install there explicitly:
npx @opengsd/gsd-core@latest --kimi --global --config-dir ~/.kimi-code
Then launch the generated agent from that directory:
kimi --agent-file ~/.kimi-code/agents/gsd.yaml
For brand-specific scripted installs, use:
KIMI_CONFIG_DIR=~/.kimi-code npx @opengsd/gsd-core@latest --kimi --global
Avoid arbitrary KIMI_CONFIG_DIR roots unless your Kimi configuration also adds the matching skills/ directory to Kimi's extra skill directories. GSD can write files there, but Kimi will not auto-discover skills outside its documented generic and brand-specific roots without that Kimi-side configuration.
--kimi --local is intentionally deferred and guarded in v1; use the global install path above for Kimi CLI.
Hook coverage
GSD wires its lifecycle hooks into Kimi's native [[hooks]] array in config.toml — by default ~/.kimi/config.toml (overridable via Kimi's own KIMI_SHARE_DIR environment variable, a directory deliberately separate from the ~/.config/agents skills root above). Kimi CLI's hooks system is documented as Beta. GSD's entries are wrapped in # GSD Hooks BEGIN/# GSD Hooks END marker comments, so reinstalling only ever rewrites GSD's own block and never touches hand-written [[hooks]] entries around it.
| Event | Hook | Purpose |
|---|---|---|
SessionStart |
gsd-check-update.js, gsd-session-state.sh |
Update check and session-state bootstrap at session open |
PreToolUse |
gsd-prompt-guard.js, gsd-read-guard.js, gsd-worktree-path-guard.js, gsd-workflow-guard.js, gsd-validate-commit.sh |
Prompt-injection guard, read-before-edit guidance, worktree path safety, workflow guard, and commit validation before tool calls |
PostToolUse |
gsd-context-monitor.js, gsd-phase-boundary.sh, gsd-read-injection-scanner.js, gsd-graphify-update.sh |
Context window tracking, phase-boundary detection, read-time injection scanning, and graph updates after tool calls |
Stop |
gsd-context-monitor.js |
Context headroom tracking before the model stops |
PreCompact |
gsd-context-monitor.js |
Context headroom tracking before compaction |
SubagentStart |
gsd-context-monitor.js |
Inject context / GSD_AGENT_NAME awareness at subagent open |
SubagentStop |
gsd-context-monitor.js |
Context headroom tracking at subagent stop |
All registered hooks are managed by GSD and are removed cleanly on --uninstall.
GitHub Copilot
npx @opengsd/gsd-core@latest --copilot --global
Skills land in ~/.copilot/. GSD installs as agent .md files and repository instruction files.
GSD also wires Copilot's lifecycle hooks and instruction files:
AGENTS.md(local installs) — written at the repository root, which GitHub Copilot CLI reads as primary instructions, alongsidecopilot-instructions.md.- Lifecycle hook — a
sessionStarthook config is written to.github/hooks/gsd-session.json(local) or~/.copilot/hooks/gsd-session.json(global). It is a self-contained inlinecommandhook (no separate hook script to install), so it can never reference a missing script. The hook is advisory-only: at session start it surfaces whether the project has a.planning/workflow.
Both are removed (and any user-authored content preserved) on --uninstall.
Override the install directory:
COPILOT_CONFIG_DIR=~/.copilot-alt npx @opengsd/gsd-core@latest --copilot --global
Cursor
npx @opengsd/gsd-core@latest --cursor --global
Artifacts land in ~/.cursor/. GSD installs skills (~/.cursor/skills/gsd-*/SKILL.md), agents, and rule references. Cursor exposes each skill once in the / menu while keeping it available for contextual model invocation. Upgrading removes manifest-managed legacy ~/.cursor/commands/gsd-*.md copies that previously duplicated those menu entries; unknown user-authored command files are preserved.
Override the install directory:
CURSOR_CONFIG_DIR=~/.cursor-alt npx @opengsd/gsd-core@latest --cursor --global
Windsurf / Devin Desktop
Windsurf has rebranded to Devin Desktop. Both runtime names are accepted — use either --windsurf or --devin-desktop.
npx @opengsd/gsd-core@latest --windsurf --global
# or equivalently:
npx @opengsd/gsd-core@latest --devin-desktop --global
Use a workspace install for Windsurf slash commands. Workspace installs write /gsd-* commands as Windsurf workflow files under .windsurf/workflows/. Windsurf discovers those .md workflow files in Cascade and exposes them through the / menu. Global-scope Windsurf workflow installation is intentionally a no-op for now because global workflow locations are outside GSD's normal user-owned runtime config directory.
Override the install directory:
WINDSURF_CONFIG_DIR=~/.codeium/windsurf-alt npx @opengsd/gsd-core@latest --windsurf --global
Cline
GSD gives Cline both skills (≥ v3.48.0) and the .clinerules/ directory integration — no custom slash commands are registered.
# Global install (all projects — skills + rules directory)
npx @opengsd/gsd-core@latest --cline --global
# Local install (this project only — rules directory only)
npx @opengsd/gsd-core@latest --cline --local
GSD writes the .clinerules/ directory form:
.clinerules/gsd.md— the GSD rule file. Cline loads every.md/.txtfile in the.clinerules/directory automatically; no custom slash commands are registered..clinerules/hooks/PreToolUse— a lifecycle hook (Cline v3.36+). It is an executable script that receives the tool-call context as JSON on stdin and returns a JSON decision (cancel/errorMessage/contextModification). The GSD hook guards.planning/artifacts from direct edits and otherwise allows the operation; it fails open, so a hook error never blocks you. Cline runs hooks on macOS and Linux only.
Global install additionally:
- Emits each GSD command as
~/.cline/skills/<name>/SKILL.md. Cline ≥ v3.48.0 loads skills from~/.cline/skills/automatically — no configuration needed. - Merges GSD instructions into
~/.agents/AGENTS.md, the cross-tool global instruction file Cline reads. The block is marker-delimited, so your ownAGENTS.mdcontent (and other tools' entries) is preserved, and--uninstallstrips only the GSD block.
Local install writes the .clinerules/ directory into the current project only. No skills
directory is created for local scope.
Cline's global hook directory (
~/Documents/Cline/Rules/Hooks/) is not yet populated by the installer — project-scope hooks (.clinerules/hooks/) and the globalAGENTS.mdinstruction target cover the common cases.
CodeBuddy
npx @opengsd/gsd-core@latest --codebuddy --global
GSD installs four surfaces. Slash command definitions land in ~/.codebuddy/commands/gsd-*.md and appear as /gsd-help, /gsd-phase, /gsd-ship, etc. in the / menu. Subagents land in ~/.codebuddy/agents/gsd-*.md. Skills land in ~/.codebuddy/skills/gsd-*/SKILL.md — emitted with user-invocable: false so they stay out of the / menu (the commands surface is the sole / entry point) and remain available for model invocation. CodeBuddy hooks are written to settings.json. No mcp.json is written: GSD ships no MCP server.
Hook coverage
GSD registers the following events automatically on install (Claude hook event dialect):
| Event | Hook | Purpose |
|---|---|---|
SessionStart |
gsd-check-update.js, gsd-session-state.sh |
Update check, session orientation |
PreToolUse |
gsd-prompt-guard.js, gsd-read-guard.js, gsd-workflow-guard.js, gsd-worktree-path-guard.js, gsd-agent-isolation-guard.js, gsd-validate-commit.sh |
Prompt guard, read-before-edit, workflow + worktree safety, agent-dispatch isolation, commit validation |
PostToolUse |
gsd-context-monitor.js, gsd-read-injection-scanner.js, gsd-phase-boundary.sh, gsd-graphify-update.sh |
Context monitoring, read-time scan, phase boundary detection |
SubagentStop |
gsd-context-monitor.js |
Context headroom tracking after subagent completion |
SubagentStart |
gsd-context-monitor.js |
Context headroom tracking at subagent start |
Stop |
gsd-context-monitor.js |
Context headroom tracking before model stop |
PreCompact |
gsd-context-monitor.js |
Context awareness before conversation compaction |
CodeBuddy's own background sub-agent dispatch (run_in_background: true) is a caller-side invocation parameter, not something GSD's installed agent files control — there is no frontmatter field to set on GSD's agent artifacts to request it.
Qwen Code
Qwen Code uses the same open skills standard as Claude Code 2.1.88+.
npx @opengsd/gsd-core@latest --qwen --global
Skills land in ~/.qwen/skills/gsd-*/SKILL.md.
GSD's main-loop skills are emitted with Qwen's optional numeric priority frontmatter field so the most-used workflows surface first in the /skills TUI list. Higher values sort earlier (per Qwen's skills spec), so core commands such as /skills for new-project (100), plan-phase (90), and execute-phase (85) appear above utility skills, which are left unset (default 0). This affects only the /skills list order — slash-command completion and /help remain alphabetical.
Subagents land in ~/.qwen/agents/gsd-*.md as native Qwen subagents, converted to Qwen's own name:/description:/tools: (YAML block list) frontmatter schema rather than Claude Code's.
Override the install directory:
QWEN_CONFIG_DIR=~/.qwen-alt npx @opengsd/gsd-core@latest --qwen --global
Hook coverage
Qwen Code supports 15 hook events. GSD registers the following events automatically on install:
| Event | Hook | Purpose |
|---|---|---|
SessionStart |
gsd-check-update.js, gsd-session-state.sh |
Update check, session orientation |
PostToolUse |
gsd-context-monitor.js, gsd-read-injection-scanner.js, gsd-phase-boundary.sh, gsd-graphify-update.sh |
Context monitoring, read-time scan, phase boundary detection |
PreToolUse |
gsd-prompt-guard.js, gsd-read-guard.js, gsd-workflow-guard.js, gsd-worktree-path-guard.js, gsd-agent-isolation-guard.js, gsd-validate-commit.sh |
Prompt guard, read-before-edit, workflow + worktree safety, agent-dispatch isolation, commit validation |
SubagentStop |
gsd-context-monitor.js |
Context headroom tracking after subagent completion |
SubagentStart |
gsd-context-monitor.js |
Context headroom tracking at subagent start |
Stop |
gsd-context-monitor.js |
Context headroom tracking before model stop |
PreCompact |
gsd-context-monitor.js |
Context awareness before conversation compaction |
Augment Code
npx @opengsd/gsd-core@latest --augment --global
Skills land in ~/.augment/skills/ and slash command definitions land in ~/.augment/commands/. GSD installs skills, agents, and commands (/gsd-phase, /gsd-ship, etc.). GSD's managed lifecycle hooks are registered into Augment's own settings.json hooks block (Claude hook event dialect, covering session-start, tool-use, and phase-boundary events) — no statusline ownership. #2097 also registers the GSD companion MCP server under settings.json's mcpServers.gsd (see Connect a host to the GSD MCP server).
Antigravity
npx @opengsd/gsd-core@latest --antigravity --global
The installer auto-detects the Antigravity config directory (~/.gemini/antigravity, ~/.gemini/antigravity-ide, or ~/.gemini/antigravity-cli). Uses Gemini-compatible settings policy.
Override the install directory:
ANTIGRAVITY_CONFIG_DIR=~/.gemini/antigravity-alt npx @opengsd/gsd-core@latest --antigravity --global
Trae
npx @opengsd/gsd-core@latest --trae --global
Skills land in ~/.trae/. GSD installs skills, agents, and rule references.
ZCode
npx @opengsd/gsd-core@latest --zcode --global
ZCode is Z.ai's desktop Agentic Development Environment for the GLM-5.2 model. GSD installs skills (nested SKILL.md bundles), slash commands, and subagents under ~/.zcode/:
- Skills →
~/.zcode/skills/gsd-<name>/SKILL.md(invoke with$gsd-<name>in chat) - Commands →
~/.zcode/commands/gsd-<name>.md(invoke with/gsd-<name>) - Subagents →
~/.zcode/agents/gsd-<name>.md
ZCode's skill format is identical to Claude Code's, so no runtime-specific converter is required — GSD lands as a pure declarative descriptor with no hardcoded installer branches. ZCode also natively imports skills and MCP config from ~/.claude; if you install GSD for both Claude and ZCode, you may see duplicate GSD skills inside ZCode, which is expected. To connect ZCode's MCP servers to GSD's companion server, see how to connect the GSD MCP server.
GSD's hook-automation and native-MCP-registration integrations are not yet wired for ZCode — both are blocked on ZCode not yet publishing the on-disk config format for its plugin Hook component or the settings filename/schema for its MCP store. See the ## zcode section of the host-integration capability matrix for the cited source URLs.
pi
npx @opengsd/gsd-core@latest --pi --global
pi is a bun-runtime programmatic CLI whose extensions implement pi's own ExtensionAPI (registerCommand/registerTool/registerProvider/pi.on) rather than a settings-file or slash-markdown surface. GSD ships a single native-extension file:
- Extension →
~/.pi/agent/extensions/gsd.js(global) or.pi/extensions/gsd.js(local)
The .js suffix is load-bearing: pi auto-discovers extensions by scanning that directory and keeping only names ending in .ts or .js, and it skips anything else silently — no error, no log line. GSD shipped the file as gsd.cjs through 1.7.0, which pi therefore never loaded, so /gsd never appeared (#2470). Upgrading removes the stale gsd.cjs; if you had added a manual extensions entry in ~/.pi/agent/settings.json as a workaround, you can drop it.
The extension registers a /gsd command and a gsd_invoke tool that dispatch GSD commands via a bounded subprocess call to gsd-core/bin/gsd-tools.cjs (no fully-populated in-process command-routing hub exists — see the matrix's Stage 2 note). This is a plugin-only install: pi has no shared-settings hook surface (hooksSurface: none) and, unlike Claude/OpenCode/Kilo, no host-read markdown surface at all — pi's /gsd command is registered programmatically by the extension, not discovered from files, so GSD installs the extension plus its universal gsd-core/ engine payload and the shared hooks//hooks/lib/ bundle (spawned by the extension itself, not by any config-file hook bus), and does not write any commands/, agents/, or skills/ directory for pi. The extension bridges GSD's session_start/before_agent_start/session_before_compact/tool_call lifecycle events to those staged hooks/ scripts as bounded, fail-open subprocesses, and steers pi's active model (modelMode: active) to a tier-resolved bare anthropic id via pi.on('before_provider_request', ...). See the ## pi section of the host-integration capability matrix for the negotiated axes and citations.
Local vs global install
All examples above use --global, which installs GSD once for your user account. To scope an install to a single project, replace --global with --local:
npx @opengsd/gsd-core@latest --claude --local
A local install writes into the .claude/ directory at your project root. Local install settings take precedence over global ones when both exist.
Installing prerelease editions (Next / Nightly / Insiders / Preview)
Prerelease editions of runtimes (Windsurf Next / Devin Desktop Next, Cursor Nightly, VS Code Insiders, Codex preview channels, etc.) read from a sibling config directory. Set the matching *_CONFIG_DIR env var before running the installer:
WINDSURF_CONFIG_DIR=~/.codeium/windsurf-next npx @opengsd/gsd-core@latest --windsurf --global
Select the corresponding stable runtime in the installer prompt. GSD does not enumerate prerelease editions as separate named runtimes — they are best-effort via this env-var mechanism and are not separately tested in release CI.
Installing without Node.js
If you cannot run npx (for example, on a Windows machine without Node.js), you have two options.
Option A — Use a machine that has Node.js. Any machine with Node.js will do: WSL, a Linux VM, a CI runner, or a Docker container. Run the installer there, then copy the output directory to your target machine. For OpenCode:
npx @opengsd/gsd-core@latest --opencode --global
# Then copy ~/.config/opencode/agents/ to the Windows machine
Option B — Manually transform the source files. The agent source files live in agents/ in the GSD Core repository and are in Claude Code's native frontmatter format. Each runtime expects a different shape. For the exact field transformations per runtime, see Manual install / no-Node.js setup in the User Guide, which covers the OpenCode transformations in full detail and points to the installer's convert*Frontmatter functions for other runtimes.
After install
Restart your runtime to pick up new commands and agents. Then start a new project or onboard an existing repo:
/gsd-new-project # greenfield project
/gsd-onboard # existing codebase
If the command is not found after restart, verify the install directory matches the runtime's expected config path. The prerelease-editions section above covers the most common mismatch.
"… is not on your PATH" after install
If the installer's global bin directory is not on your PATH, it prints a one-time warning with a copy-paste command for your shell. The suggestion list covers zsh, bash, and fish (plus PowerShell, cmd.exe, and Git Bash on Windows). For fish, run the line it prints:
fish_add_path '/path/to/global/bin'
If the directory is already on your PATH but the installer still warns, open a new fish session (exec fish) to pick up the change.