registry

package
v0.1.0-rc.1 Latest Latest
Warning

This package is not in the latest version of its module.

Go to latest
Published: Aug 10, 2026 License: Apache-2.0 Imports: 16 Imported by: 0

Documentation

Overview

Package registry — acceptance criteria runner.

A skill's `acceptance_criteria` frontmatter is a list of Given/When/Then prose strings. The runner evaluates each criterion against the skill's body and frontmatter, returning a TestReport that callers can render.

The DeterministicEvaluator ships in-tree: it looks for "PASS:" or "FAIL:" markers inside each criterion so fixture skills can encode expected outcomes without standing up an LLM provider. Production evaluation against a live model was previously provided by the LLMEvaluator; that path moved out when the agent runtime was retired.

The runner is read-only — it never mutates the skill or its files.

Index

Constants

View Source
const (
	ResolutionAuthor   = "author"
	ResolutionDefault  = "default"
	ResolutionOverride = "override"
)

Resolution provenance values, shared with the REST surface.

View Source
const (
	ModelSourceTopLevel      = "model"
	ModelSourceMetaPreferred = "metadata.preferred-model"
	ModelSourceMetaModel     = "metadata.model"
)

Model preference source keys, in recognition precedence order.

Variables

View Source
var ErrCriterionOutOfRange = errors.New("criterion index out of range")

ErrCriterionOutOfRange is returned when RunOptions.CriterionIndex names a criterion that does not exist on the skill. The CLI maps this to exit code 2.

View Source
var ErrNoCriteria = errors.New("skill has no acceptance_criteria")

ErrNoCriteria signals that the skill has no acceptance_criteria frontmatter. The runner returns this distinct error so callers can emit a clear "nothing to test" message and exit cleanly rather than reporting zero failures (which is ambiguous: did the skill pass, or was there nothing to check?).

View Source
var ErrNotFound = errors.New("not found")

ErrNotFound is returned when a skill does not exist in the store.

Functions

func AgentHonorMatrix

func AgentHonorMatrix() map[string]HonorStatus

AgentHonorMatrix returns a copy of the full agent-target matrix, keyed by target slug.

func IsKnownModelValue

func IsKnownModelValue(v string) bool

IsKnownModelValue reports whether a value is a documented alias, a suffixed alias form, or shaped like a full model ID. Callers treat a false as advisory-warn material, never an error.

func KnownModelAliases

func KnownModelAliases() []string

KnownModelAliases returns the documented alias vocabulary, for lint messages and docs.

func NormalizeModelValue

func NormalizeModelValue(v string) string

NormalizeModelValue canonicalizes a model value for comparison: trimmed and lowercased. Two declarations that normalize equal are the same preference; a rewrite is never forced over case or whitespace.

func RenderSkillMD

func RenderSkillMD(skill *AgentSkill) ([]byte, error)

RenderSkillMD serializes an AgentSkill back to SKILL.md format.

func RenderWithModelPreference

func RenderWithModelPreference(sk *AgentSkill, value string) ([]byte, error)

RenderWithModelPreference renders SKILL.md bytes with the resolved model preference applied to a copy of the skill. It backs the projection-time policy rewrite: the registry canonical is NEVER mutated; callers write the returned bytes into projected copies only.

The honored key is always set: Claude Code reads top-level `model:` on projected skill files, so the rewrite guarantees it even for authors who declared only a metadata key. Any metadata declaration the author wrote is updated in place as well, so a projected file never carries two disagreeing declarations. All other frontmatter rides through the existing parse/render round trip untouched.

func RewriteTopLevelModelLine

func RewriteTopLevelModelLine(raw []byte, value string) ([]byte, bool)

RewriteTopLevelModelLine returns frontmatter-bearing bytes with the top-level `model:` key set to value (or removed when value is empty), preserving every other byte verbatim. It is deliberate line surgery, not a parse/re-render, for files whose bytes are the contract (identity-projected agents, and adopt's reversal of a projected rewrite). ok is false when the file has no recognizable frontmatter block or an existing model value is not a single-line scalar (never valid per the client docs, and not safely replaceable line-wise); callers fall back to pass-through. Callers that can parse the result should verify it, since text surgery cannot see every YAML form.

func SkillHonorMatrix

func SkillHonorMatrix() map[string]HonorStatus

SkillHonorMatrix returns a copy of the full skill-target matrix, keyed by target slug (REST and lint consumers).

func ValidateSkill

func ValidateSkill(s *AgentSkill) error

ValidateSkill validates an AgentSkill and returns just the error (convenience wrapper).

func ValidateSkillName

func ValidateSkillName(name string) error

ValidateSkillName validates a skill name against the agentskills.io spec.

Types

type AgentSkill

type AgentSkill struct {
	// --- Frontmatter fields (from YAML between --- delimiters) ---
	Name          string        `yaml:"name" json:"name"`
	Description   string        `yaml:"description" json:"description"`
	License       string        `yaml:"license,omitempty" json:"license,omitempty"`
	Compatibility string        `yaml:"compatibility,omitempty" json:"compatibility,omitempty"`
	Metadata      SkillMetadata `yaml:"metadata,omitempty" json:"metadata,omitempty"`
	AllowedTools  string        `yaml:"allowed-tools,omitempty" json:"allowedTools,omitempty"`
	// AcceptanceCriteria documents expected skill behavior as human-readable
	// Given/When/Then scenarios. Gridctl extension; not part of agentskills.io spec.
	// See https://agentskills.io/specification
	AcceptanceCriteria []string `yaml:"acceptance_criteria,omitempty" json:"acceptanceCriteria,omitempty"`

	// --- Gridctl extensions (not in agentskills.io spec) ---
	State ItemState `yaml:"state,omitempty" json:"state"`

	// Extra holds frontmatter keys this struct does not model, preserved
	// through parse, render, and the JSON API so imports and edits never
	// strip them (clients read keys like argument-hint and
	// disable-model-invocation from projected skills). Values are the
	// decoded YAML; gridctl never interprets them. The yaml tag is unused:
	// UnmarshalYAML populates it and RenderSkillMD emits it explicitly.
	Extra map[string]any `yaml:"-" json:"extra,omitempty"`

	// --- Parsed from file content (not in frontmatter YAML) ---
	Body string `yaml:"-" json:"body"` // Markdown content after frontmatter

	// --- Computed fields (not serialized to YAML) ---
	FileCount int    `yaml:"-" json:"fileCount"`     // Number of supporting files (scripts/, references/, assets/)
	Dir       string `yaml:"-" json:"dir,omitempty"` // Relative path from skills/ root (e.g., "git-workflow/branch-fork")
}

AgentSkill represents an Agent Skills standard SKILL.md file. See https://agentskills.io/specification for the full spec.

func ParseSkillMD

func ParseSkillMD(data []byte) (*AgentSkill, error)

ParseSkillMD parses a SKILL.md file into an AgentSkill. The file format is YAML frontmatter between --- delimiters followed by a markdown body.

func (*AgentSkill) UnmarshalYAML

func (s *AgentSkill) UnmarshalYAML(value *yaml.Node) error

UnmarshalYAML decodes frontmatter into the typed fields and captures every unmodeled top-level key into Extra, so unknown keys survive the registry's parse/render round trip instead of being silently dropped. The mapping node is walked by hand for the same reason SkillMetadata's unmarshaler does it: yaml.v3's map decoding would reject a duplicate key outright, and a plain struct decode has no way to see unmatched keys at all. Known fields keep struct-decode strictness (a type mismatch fails the parse, as before); unknown values that fail to decode are skipped rather than fatal.

func (*AgentSkill) Validate

func (s *AgentSkill) Validate() error

Validate checks the skill against the agentskills.io specification.

type DeterministicEvaluator

type DeterministicEvaluator struct{}

DeterministicEvaluator is the zero-LLM evaluator used by CI and unit tests. It inspects each criterion for the case-sensitive prefix markers "PASS:" or "FAIL:" and produces the corresponding verdict; criteria without a marker return TestSeverityError so test authors know the fixture needs to be explicit.

This adapter exists so the acceptance contract can be exercised end to end (CLI → API → evaluator) without standing up an LLM provider.

func (DeterministicEvaluator) Evaluate

func (DeterministicEvaluator) Evaluate(_ context.Context, _ *AgentSkill, idx int, criterion string) TestResult

Evaluate parses the marker convention. Markers are checked case-sensitively to avoid false positives in natural-language criteria like "should pass validation" — only explicit "PASS:" or "FAIL:" at the start of the trimmed criterion fires.

func (DeterministicEvaluator) Name

Name returns "deterministic".

type Evaluator

type Evaluator interface {
	// Name identifies the evaluator backend for the TestReport.
	Name() string

	// Evaluate produces a verdict for criterion idx against skill.
	// Implementations MUST NOT return both a TestResult and an error;
	// transient infrastructure failures should be returned via a
	// TestResult with Severity = TestSeverityError so the report
	// remains complete for the user.
	Evaluate(ctx context.Context, skill *AgentSkill, idx int, criterion string) TestResult
}

Evaluator runs a single acceptance criterion against a skill and returns the verdict. Implementations MUST honor ctx cancellation.

type HonorStatus

type HonorStatus string

HonorStatus is what one projection target does with a declared model preference. The matrix is static, maintained truth about client behavior; it will go stale on client-doc timescales, so every entry cites its source in the tables below and the sync packages carry exhaustiveness tests that fail when a projection target gains no row.

const (
	// HonorHonored: the client reads the key and uses it as a model
	// resolution input (below its own env var and per-invocation
	// overrides).
	HonorHonored HonorStatus = "honored"
	// HonorIgnored: the client documents no model key for this surface.
	HonorIgnored HonorStatus = "ignored"
	// HonorUnknown: consumer-dependent or unverified; surfaced as such
	// rather than guessed.
	HonorUnknown HonorStatus = "unknown"
	// HonorDropped: the projection render deliberately drops the key
	// (client model vocabularies are not Claude's); reported in
	// Rendered.Dropped at sync time too.
	HonorDropped HonorStatus = "dropped-on-render"
)

func AgentHonor

func AgentHonor(slug string) HonorStatus

AgentHonor returns the honor status for one agent projection target slug; unknown slugs report HonorUnknown.

func SkillHonor

func SkillHonor(slug string) HonorStatus

SkillHonor returns the honor status for one skill projection target slug; unknown slugs report HonorUnknown rather than guessing.

type ItemState

type ItemState string

ItemState represents the lifecycle state of a skill. Note: state is a gridctl extension, not part of the agentskills.io spec.

const (
	StateDraft    ItemState = "draft"
	StateActive   ItemState = "active"
	StateDisabled ItemState = "disabled"
)

type ModelPolicy

type ModelPolicy struct {
	// Rewrite opts the scope into projection rewrite. False means
	// surfacing-only: Resolve still answers (for status and REST), but
	// nothing on disk changes.
	Rewrite bool
	// Default applies where the author declared nothing.
	Default string
	// Overrides maps exact registry names to the preference applied
	// regardless of the author's declaration, in either direction.
	Overrides map[string]string
}

ModelPolicy is one scope's compiled model preference policy (the stack.yaml `model_preferences:` block for either the skills or the agents scope), consumed by the projection engines. The controller compiles it from the loaded stack; CLI call sites without stack context run with a nil policy, which is pure pass-through and, for projections a policy previously rewrote, preserve-not-revert; the sync packages own that rule.

func (*ModelPolicy) NeedsRewrite

func (p *ModelPolicy) NeedsRewrite(name, declared string) (string, bool)

NeedsRewrite reports whether projection must write a resolved value that differs from the author's declaration. Equality is judged after normalization, so case or whitespace never forces a rewrite (and never forces a skill off its symlink channel).

func (*ModelPolicy) Resolve

func (p *ModelPolicy) Resolve(name, declared string) (value, resolution string)

Resolve answers the effective preference for one name: override beats the author's declaration beats the default. The returned resolution names which source won ("override", "author", "default"), or "" when nothing applies.

type ModelPreference

type ModelPreference struct {
	// Values holds the declared preference(s), most preferred first.
	Values []string `json:"values"`
	// SourceKey names where the author declared it: "model",
	// "metadata.preferred-model", or "metadata.model".
	SourceKey string `json:"sourceKey"`
}

ModelPreference is the typed read-side view of a skill or agent author's declared model preference. It is a preference, never an enforcement point: clients resolve their own model (env vars and per-invocation parameters outrank projected frontmatter), and gridctl only surfaces, defaults, and overrides what the projected files declare.

Values is a list so a future agentskills.io ordered-preference shape is a consumption change, not a migration; today it always holds one element. The author's raw key is untouched wherever it was written (top-level Extra or metadata); this view is derived, never stored.

func ExtractModelPreference

func ExtractModelPreference(sk *AgentSkill) *ModelPreference

ExtractModelPreference derives the typed preference from a skill's frontmatter. Recognition precedence: top-level `model:` beats `metadata.preferred-model` beats `metadata.model`. Non-string or empty declarations yield nil. Extraction never mutates the skill.

func ModelPreferenceFromKeys

func ModelPreferenceFromKeys(topLevel string, metadata map[string]string) *ModelPreference

ModelPreferenceFromKeys derives the typed preference from raw key values, for callers whose documents are not AgentSkill-shaped (agent definitions keep frontmatter as raw YAML nodes). topLevel is the top-level `model:` scalar; metadata the `metadata:` map (nil allowed).

func (*ModelPreference) Value

func (p *ModelPreference) Value() string

Value returns the single effective preference (the first value), or "" for a nil preference.

type RegistryStatus

type RegistryStatus struct {
	TotalSkills  int `json:"totalSkills"`
	ActiveSkills int `json:"activeSkills"`
}

RegistryStatus contains summary statistics.

type RunOptions

type RunOptions struct {
	// CriterionIndex, when non-negative, scopes the run to a single
	// criterion. -1 (the zero value via NewRunOptions) means "all".
	CriterionIndex int

	// DryRun lists criteria without evaluating them. Severity on each
	// result is empty.
	DryRun bool

	// Now overrides the wall-clock used for EvaluatedAt and
	// GeneratedAt. Tests inject a fixed time; production callers
	// leave this zero so the runner defaults to time.Now().
	Now time.Time
}

RunOptions tunes a single Run pass.

func NewRunOptions

func NewRunOptions() RunOptions

NewRunOptions returns RunOptions with CriterionIndex set to -1, the "run every criterion" sentinel. Use this instead of a literal zero value so future-added fields stay backward-compatible.

type Server

type Server struct {
	// contains filtered or unexported fields
}

Server is an in-process MCP server that serves Agent Skills as prompts. It implements mcp.AgentClient so it can be registered with the gateway router, and mcp.PromptProvider so the gateway can serve skills via MCP prompts and resources.

func New

func New(store *Store) *Server

New creates a registry server that serves skills as MCP prompts.

func (*Server) CallTool

func (s *Server) CallTool(_ context.Context, _ string, _ map[string]any) (*mcp.ToolCallResult, error)

CallTool always returns an error. See Tools().

func (*Server) GetPromptData

func (s *Server) GetPromptData(name string) (*mcp.PromptData, error)

GetPromptData returns a specific active skill's content as MCP PromptData.

func (*Server) HasContent

func (s *Server) HasContent() bool

HasContent returns true if the registry has any skills.

func (*Server) Initialize

func (s *Server) Initialize(ctx context.Context) error

Initialize loads the store.

func (*Server) IsInitialized

func (s *Server) IsInitialized() bool

IsInitialized returns whether the server has been initialized.

func (*Server) ListPromptData

func (s *Server) ListPromptData() []mcp.PromptData

ListPromptData returns active Agent Skills as MCP PromptData. Each skill gets a single optional "context" argument for clients to pass additional context when requesting the skill via prompts/get.

func (*Server) Name

func (s *Server) Name() string

Name returns "registry".

func (*Server) RefreshTools

func (s *Server) RefreshTools(ctx context.Context) error

RefreshTools reloads the store from disk.

func (*Server) ServerInfo

func (s *Server) ServerInfo() mcp.ServerInfo

ServerInfo returns server information.

func (*Server) Store

func (s *Server) Store() *Store

Store returns the underlying store for REST API access.

func (*Server) Tools

func (s *Server) Tools() []mcp.Tool

Tools always returns an empty slice. Skills are served as MCP prompts via the PromptProvider surface, not as executable tools. The method survives to satisfy mcp.AgentClient so the gateway router can keep the registry in its client set without a special case.

type SkillFile

type SkillFile struct {
	Path  string `json:"path"`  // Relative path within the skill dir (e.g., "scripts/lint.sh")
	Size  int64  `json:"size"`  // File size in bytes
	IsDir bool   `json:"isDir"` // True for directories
}

SkillFile represents a file within a skill directory.

type SkillMetadata

type SkillMetadata map[string]string

SkillMetadata holds the frontmatter metadata mapping. The agentskills.io spec defines it as string-to-string, but ecosystems like openclaw/ClawHub publish nested values there, so decoding is lenient: non-string values are coerced to strings (see UnmarshalYAML in frontmatter.go).

func (*SkillMetadata) UnmarshalYAML

func (m *SkillMetadata) UnmarshalYAML(value *yaml.Node) error

UnmarshalYAML decodes metadata leniently. Values that are not strings (nested mappings, sequences, booleans, numbers) are coerced to their string form instead of failing the whole SKILL.md parse. Non-mapping metadata (e.g. a bare string) is ignored rather than treated as an error. The mapping node is walked by hand so one bad entry (or a duplicate key, which yaml.v3's map decoding rejects) cannot discard the valid keys.

type Store

type Store struct {
	// contains filtered or unexported fields
}

Store manages skill directories on disk. Each skill is a directory containing a required SKILL.md and optional supporting files (scripts/, references/, assets/).

func NewStore

func NewStore(baseDir string) *Store

NewStore creates a store rooted at the given directory.

func (*Store) ActiveSkills

func (s *Store) ActiveSkills() []*AgentSkill

ActiveSkills returns only skills with State == "active". Returned pointers are copies.

func (*Store) DeleteFile

func (s *Store) DeleteFile(skillName, filePath string) error

DeleteFile removes a file from a skill directory.

func (*Store) DeleteSkill

func (s *Store) DeleteSkill(name string) error

DeleteSkill removes a skill directory and cache entry.

func (*Store) Dir

func (s *Store) Dir() string

Dir returns the store's base directory path.

func (*Store) GetSkill

func (s *Store) GetSkill(name string) (*AgentSkill, error)

GetSkill returns a skill by name.

func (*Store) HasContent

func (s *Store) HasContent() bool

HasContent returns true if there is at least one skill.

func (*Store) ListFiles

func (s *Store) ListFiles(skillName string) ([]SkillFile, error)

ListFiles returns all files in a skill directory (excluding SKILL.md).

func (*Store) ListSkills

func (s *Store) ListSkills() []*AgentSkill

ListSkills returns all skills (all states). Returned pointers are copies.

func (*Store) Load

func (s *Store) Load() error

Load scans the skills/ subdirectory for SKILL.md files and checks for legacy YAML registry files.

func (*Store) ReadFile

func (s *Store) ReadFile(skillName, filePath string) ([]byte, error)

ReadFile reads a specific file from a skill directory.

func (*Store) RefreshFileCount

func (s *Store) RefreshFileCount(name string) error

RefreshFileCount recomputes a skill's supporting-file count from disk. Callers that write files into a skill directory after SaveSkill (the git importer installs scripts/ and references/ once the skill validates) use this so the cached count matches what actually landed.

func (*Store) RenameSkill

func (s *Store) RenameSkill(oldName, newName string) error

RenameSkill renames a skill directory and updates its frontmatter.

func (*Store) SaveSkill

func (s *Store) SaveSkill(sk *AgentSkill) error

SaveSkill creates or updates a skill (validates, writes SKILL.md, updates cache).

func (*Store) Status

func (s *Store) Status() RegistryStatus

Status returns registry summary counts.

func (*Store) WriteFile

func (s *Store) WriteFile(skillName, filePath string, data []byte) error

WriteFile writes a file to a skill directory, creating parent directories as needed.

type TestReport

type TestReport struct {
	// SkillName echoes the skill the runner ran against.
	SkillName string `json:"skill_name"`

	// Results is the per-criterion verdicts in criterion order.
	Results []TestResult `json:"results"`

	// PassCount, FailCount, ErrorCount summarize Results so callers
	// don't re-tally.
	PassCount  int `json:"pass_count"`
	FailCount  int `json:"fail_count"`
	ErrorCount int `json:"error_count"`

	// Evaluator names the backend that ran the criteria (e.g. "llm",
	// "deterministic"). Surfaced in JSON output so CI logs make the
	// provenance explicit.
	Evaluator string `json:"evaluator"`

	// DryRun is true when Results was populated by listing criteria
	// without executing them. Each TestResult in that case carries
	// Severity = "" so renderers can show "—" instead of a verdict.
	DryRun bool `json:"dry_run,omitempty"`

	// GeneratedAt is the wall-clock time the report was assembled.
	GeneratedAt time.Time `json:"generated_at"`
}

TestReport is the runner's full output for one skill.

func RunAcceptance

func RunAcceptance(ctx context.Context, skill *AgentSkill, ev Evaluator, opts RunOptions) (TestReport, error)

RunAcceptance evaluates the skill's acceptance_criteria using the supplied evaluator and returns a populated TestReport.

Returns:

  • ErrNoCriteria when the skill has no acceptance_criteria.
  • ErrCriterionOutOfRange when opts.CriterionIndex is set and points outside the skill's criteria slice.

Per-criterion infrastructure failures are reported as TestResults with Severity = TestSeverityError; only the two errors above stop the whole run.

func (TestReport) HasErrors

func (r TestReport) HasErrors() bool

HasErrors reports whether any criterion errored during evaluation. The CLI uses this to map to exit code 2.

func (TestReport) HasFailures

func (r TestReport) HasFailures() bool

HasFailures reports whether any criterion failed. The CLI uses this to map to exit code 1. Error-severity results are separate and map to exit code 2 in the CLI; here we only flag fail.

type TestResult

type TestResult struct {
	// Index is the criterion's position in the skill's
	// AcceptanceCriteria slice (zero-based). Stable across runs.
	Index int `json:"index"`

	// Criterion is the verbatim Given/When/Then string from the
	// skill's frontmatter.
	Criterion string `json:"criterion"`

	// Severity classifies the outcome.
	Severity TestSeverity `json:"severity"`

	// Message is a short human-readable rationale. Required for fail
	// and error; optional for pass.
	Message string `json:"message,omitempty"`

	// EvaluatedAt is the wall-clock time the verdict was rendered.
	EvaluatedAt time.Time `json:"evaluated_at"`
}

TestResult is the verdict for one criterion. Mirrors optimize.Finding's shape so the JSON envelope feels familiar.

type TestSeverity

type TestSeverity string

TestSeverity classifies a single criterion's outcome. The vocabulary mirrors pkg/optimize.Severity so renderers can share badge styling.

const (
	// TestSeverityPass means the evaluator verified the criterion is
	// satisfied by the skill as written.
	TestSeverityPass TestSeverity = "pass"

	// TestSeverityFail means the evaluator believes the criterion is
	// NOT satisfied. This is the case the runner reports back to CI.
	TestSeverityFail TestSeverity = "fail"

	// TestSeverityError means evaluation could not complete (LLM
	// unavailable, malformed criterion, etc.). Mapped to exit code 2
	// by the CLI so infrastructure failures never look like a real
	// fail verdict.
	TestSeverityError TestSeverity = "error"
)

type ValidationResult

type ValidationResult struct {
	Errors   []string // Fatal validation failures
	Warnings []string // Non-fatal advisories (e.g., body too long)
}

ValidationResult contains errors and warnings from skill validation.

func ValidateSkillFull

func ValidateSkillFull(s *AgentSkill) *ValidationResult

ValidateSkillFull validates an AgentSkill and returns both errors and warnings.

func (*ValidationResult) Error

func (v *ValidationResult) Error() error

Error returns the first error as a Go error, or nil if valid.

func (*ValidationResult) Valid

func (v *ValidationResult) Valid() bool

Valid returns true if there are no errors (warnings are OK).

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL