msd-core/docs/how-to/develop-a-capability.md

Develop a Capability for GSD 1.5+

This guide shows you how to add or change a first-party GSD Capability after the ADR-857 cutover. In GSD terms, the extension unit is a Capability. A plugin is a packaging or host-runtime term, for example a Claude Code plugin, not the unit that owns a GSD feature.

A Capability is right when the feature can be toggled as one unit and owns its own skills, agents, hooks, config keys, or command family. Keep verifier predicate contracts, the five-step loop spine, and shared infrastructure in core unless the ADRs explicitly move that boundary.

Start from the boundary

Before writing a manifest, decide whether the work is core or a Capability.

Use a Capability when the feature:

• Can be enabled or disabled without changing the meaning of the base loop.
• Owns a stable feature name such as research, ui, graphify, ai-integration, or pattern-mapper.
• Adds a step, gate, or contribution at a Loop Extension Point.
• Owns one or more feature config keys.
• Owns a command family that can route through the capability registry.

Keep the work in core when the feature:

• Defines the reliability substrate of the loop, such as the verifier predicate contract described by ADR-550 and ADR-857.
• Is required for every installation profile.
• Mutates shared host workflow state rather than adding a declared hook contribution.

Create the folder

Create one folder per Capability:

capabilities/<id>/
  capability.json
  fragments/
    plan-pre.md

The manifest path is capabilities/<id>/capability.json.

The <id> must match the id field in capability.json. Co-locate prompt fragments, owned skills, owned agents, and other owned artefacts under the Capability folder when the schema allows it. Shared host artefacts can be referenced by name, but ownership should remain clear in the manifest.

Write capability.json

Use the existing manifests as the source of truth while the schema is still first-party:

• capabilities/research/capability.json
• capabilities/ai-integration/capability.json
• capabilities/pattern-mapper/capability.json
• capabilities/ui/capability.json
• capabilities/graphify/capability.json

At minimum, a feature Capability declares:

{
  "id": "example",
  "role": "feature",
  "version": "0.1.0",
  "title": "Example",
  "description": "Adds an example planning step.",
  "tier": "standard",
  "requires": [],
  "runtimeCompat": { "supported": ["*"], "unsupported": [] },
  "skills": [],
  "agents": ["gsd-example-agent"],
  "hooks": [],
  "config": {},
  "steps": [
    {
      "point": "plan:pre",
      "ref": { "agent": "gsd-example-agent" },
      "fragment": { "path": "fragments/plan-pre.md" },
      "produces": ["EXAMPLE.md"],
      "consumes": ["CONTEXT.md"],
      "onError": "skip"
    }
  ],
  "contributions": [],
  "gates": []
}

The registry generator validates the shape, ownership, and cross-capability contracts. ref.skill must name a skill declared by the same Capability. ref.agent must name an agent declared by the same Capability.

Declare runtime compatibility

Every feature Capability must declare runtimeCompat. This is part of the GSD 1.5+ developer contract: a Capability says which runtime descriptors it can surface through, and the generator validates that declaration before the central registry is written.

Use the wildcard when the Capability is runtime-agnostic:

"runtimeCompat": {
  "supported": ["*"],
  "unsupported": []
}

Use explicit runtime ids when the Capability is intentionally narrower:

"runtimeCompat": {
  "supported": ["claude", "codex"],
  "unsupported": ["kilo"],
  "notes": {
    "kilo": "Requires a hook surface Kilo does not expose yet."
  }
}

supported is required and must be non-empty. "*" means every descriptor-backed runtime, including future first-party runtime descriptors, is compatible unless it is listed in unsupported. Explicit runtime ids and notes keys must match runtime Capability ids such as claude, codex, opencode, or kilo; typos fail node scripts/gen-capability-registry.cjs --check.

Add hooks

Loop Extension Points are the stable sites where Capabilities attach to the host loop. Phase 6 planning-time features use plan:pre so the core planner can ask the registry for active planning hooks instead of reading feature config directly.

Choose the hook kind that matches the behaviour:

• steps add a sequenced unit of work, such as running gsd-phase-researcher or gsd-pattern-mapper.
• contributions add labelled context to a host prompt.
• gates check a condition and may block when blocking is true.

Declare file artefact flow with produces and consumes. The registry uses those arrays to order hooks and to reject unsatisfied dependencies.

Ship skills — and know what you are shipping

A skill your Capability declares is not an inert asset. Its SKILL.md body is copied verbatim into the user's runtime skills directory at install, where it becomes an agent-invocable instruction file. GSD does not scan it: there is no content inspection of any kind, at install or at any later point. What you write is what reaches the agent.

That makes a skill body an instruction surface, and it carries author responsibility that an inert artifact does not:

• Write instructions you would be comfortable defending. Your reach is bounded only by what the agent will do when told. Consent, integrity pinning and reversibility are the user's protections; content review is not among them.
• Do not embed anything that tries to redirect the agent away from the user's task — reframing its role, overriding host instructions, or persisting directives past the skill's own scope. That is indistinguishable from a prompt-injection payload, and the fact that it is not scanned is not permission.
• Treat anything your skill tells the agent to read as data, not instructions. If your skill has the agent ingest a file, a URL, or tool output, say so explicitly in the body. See the untrusted-input boundary.
• The surface is disclosed. The skills your Capability contributes are named, by stem, in their own section of the pre-install consent summary (#3248). Write your skill body on the assumption that a user sees it listed before they accept. Declared agents[] are not named there today — a third-party capability's agents are never staged into the agent's instruction context, so there is nothing to disclose — but they remain classified as an instruction surface, not an inert or safe one.

The reasoning behind this posture — including why scanning was considered and rejected — is recorded in ADR-2363, and the user-facing side is the capability trust model.

Use prompt fragments

Use fragment.path for prompt text longer than a short sentence:

"fragment": { "path": "fragments/plan-pre.md" }

The generator materialises that file into fragment.inline in the generated registry. Paths must be relative to the Capability folder, must not be absolute, and must not contain ...

Use fragment.inline only for short, stable text:

"fragment": { "inline": "Add the generated example context to the planner input." }

Generate and verify the registry

After editing any Capability manifest or fragment, regenerate the committed registry:

node scripts/gen-capability-registry.cjs --write

Then run the drift check:

node scripts/gen-capability-registry.cjs --check

For a planning hook, verify the rendered output:

node gsd-core/bin/gsd-tools.cjs loop render-hooks plan:pre --raw

In installed workflow prose, the same resolver surface appears as gsd-tools loop render-hooks plan:pre.

The rendered JSON should include the active hook, the declared ref, and the materialised fragment.inline.

To verify a runtime-specific surface, pass the same config directory that the runtime installation uses:

node gsd-core/bin/gsd-tools.cjs loop render-hooks plan:pre --config-dir ~/.claude --raw
node gsd-core/bin/gsd-tools.cjs capability state --config-dir ~/.claude --raw

capability state is the diagnostic view for the same state that workflow dispatch consumes. For each Capability, enabled is true only when the Capability is both installed by the active profile and surfaced by the runtime surface. A hook's configured field reflects the when config key; active is true only when the Capability is enabled and the hook is configured on.

Add or change a runtime descriptor

Runtime-specific facts belong in the runtime Capability declaration, not in a parallel allowlist or runtime-name branch. When adding a first-party runtime under capabilities/<runtime>/capability.json, declare these fields in the runtime object:

• configHome: the global config root resolver, including env overrides and probes.
• configHome.skillsHome: optional separate base home for runtimes whose global skills root differs from the config root.
• artifactLayout.global and artifactLayout.local: command, agent, skill, and Kimi-agent destinations.
• hooksSurface, hookEvents, and extendedHookEvents: hook registration surface and event dialect.
• installSurface, writesSharedSettings, and permissionWriter: install-time config mutation behavior.

The runtime homes, artifact layout, and install-plan resolvers read those descriptor fields directly. Adding a descriptor-backed runtime should not require editing a second list in runtime-artifact-layout or a fallback branch in runtime-config-adapter-registry.

Own config in the Capability

Declare feature config keys in the Capability manifest:

"config": {
  "workflow.example_enabled": {
    "type": "boolean",
    "default": true,
    "description": "Enable the example planning step."
  }
}

Do not add migrated Capability keys to gsd-core/bin/shared/config-schema.manifest.json. The central schema remains for host/core keys; Capability-owned keys validate through the generated registry and are merged into loadConfig by the federated config overlay. Existing nested .planning/config.json values such as { "workflow": { "example_enabled": false } } continue to override the Capability default.

Wire the workflow through the registry

Host workflows should ask the resolver for active hooks and then dispatch from the resolved data. Do not add new direct config-get workflow.<feature> checks to a host workflow for a migrated feature.

The Phase 6 migrations follow this pattern:

• research registers a plan:pre step that invokes gsd-phase-researcher.
• ai-integration owns the AI-SPEC planning activation and config key.
• pattern-mapper registers a plan:pre step that invokes gsd-pattern-mapper.
• plan-phase.md reads the resolved PLAN_PRE_HOOKS_JSON and dispatches ref.agent or ref.skill from that data.
• code-review registers an execute:post step and review workflows dispatch from the active hook's ref.skill.
• security registers a plan:pre contribution, a verify:post step, and a blocking ship:pre gate.
• nyquist registers a verify:post step for validation coverage auditing.

This keeps "off means off" enforceable by construction: disabled Capabilities are absent from the active hook set, so the host workflow has nothing feature-specific to run.

Test the Capability

Add focused tests before changing behaviour:

• Registry tests for manifest validation, ordering, config ownership, command family dispatch, or fragment materialisation.
• Workflow text tests for host workflow cutovers when a workflow stops reading direct feature config.
• Behavioural command tests when a command family moves behind dispatchCapabilityCommand.
• Documentation tests when the feature changes developer-facing behaviour.

Run the smallest affected tests first, then the full suite before opening a ready PR:

node --test tests/capability-registry.test.cjs
node --test tests/phase6-planning-capabilities.test.cjs
node --test tests/phase6-capstone-conformance.test.cjs
npm test

Run the Phase 6 capstone test whenever a Capability adds a when key or moves activation logic in a host workflow. It checks that migrated activation keys are resolved through Capability hooks or state, and that Capability-owned config keys stay out of the central schema. If a host workflow must keep core behaviour behind a workflow.* key, document why it is core rather than Capability-owned before adding it.

Keep the docs with the slice

Every Phase 6 slice that changes capability behaviour must update the relevant docs in the same PR. Use this manual for developer-facing Capability authoring facts, use how-to guides for task flows, and use ADRs only for decisions and trade-offs.

The capability ecosystem (1.6.0)

From GSD 1.6.0, capabilities are versioned (the version field is required in capability.json) and can be installed directly from a URL, a git ref, an npm package, or a local path — without modifying the core repo.

• Tutorial — Build your first capability: scaffold and install a declarative capability end-to-end in under ten minutes.
• How-to — Publish a capability: package and distribute a capability via a URL or registry.
• How-to — Import a capability from a URL: install a third-party capability from a git URL, tarball, or npm package.
• How-to — Version and update a capability: manage version, engines.gsd, and compatVersions; use gsd capability update.
• How-to — Remove a capability: uninstall cleanly with gsd capability remove, including the --purge-data option.
• How-to — Ship a reviewer lane in your capability: declare a reviewer body (GSD 1.9.0+) so /gsd-review discovers and invokes your external review CLI or model endpoint.
• How-to — List your reviewer lane in the registry: publish a lane to the Reviewer Lane Registry (GSD 1.9.1+) so other people can find and install it.
• Reference — Capability manifest: all fields and validation rules for capability.json.
• Reference — Capability matrix: which first-party capabilities exist, their extension points, and their compatibility matrix.
• Explanation — Capability trust model: how declarative and executable capabilities are treated differently at install time.
• ADR-1244 — docs/adr/1244-capability-ecosystem.md: the architectural decision that introduced the installable capability ecosystem.