Files
archie/internal/domain/eda/expr/expr.go
T
sam d26ba54e77 feat(eda): wire action playbooks through loader and module schemas
Pass 2 of archie-core-nx30. Pass 1 typed each prior action's result in
the CEL env; this pass wires that typing through the loader, module
schemas, and daemon, and removes the now-redundant id classifier.

Loader shape (D2): one playbook is exactly one workflow action
(unchanged) OR one or more module actions; the shapes never mix. Mixed,
unsupported, or zero actions fail the load naming the playbook.
playbook.Load now takes a KindSchemas source (a small interface this
package defines, satisfied by *module.ModuleRegistry) so playbook grows
no module import. Callers: bootstrap.loadEDAPlaybooks passes b.modules;
daemon and playbook tests pass module.New().

Per-playbook env (D3): each action's when and args values compile against
an env built from the PRIOR actions' ids typed by their kinds' Result
structs (expr.NewEnv(DeclaredResult{ID,Type})). event stays dyn. An
undeclared id, a forward reference, or a field the kind's Result lacks is
a compile-time undefined-field error.

Args schema (D4): an args key the kind's Args struct does not define
fails the load. An args VALUE whose compiled type mismatches the field
type is left as a named gap (not cheap without leaking cel-go type kinds
into playbook and handling dyn reads); a wrong-typed value still surfaces
at module invoke, not load.

D1: Store.Dispatch routes workflow playbooks only, so an action playbook
cannot hijack the definition pin. bootstrap.loadEDAPlaybooks logs a
visible Warn (not debug) for each loaded action playbook naming it and
stating it does not execute yet.

D5: ModuleRegistry.Invoke is deliberately schema-agnostic (the stored
invoker is type-erased and returns map[string]any). The map-to-Result
marshalling belongs in the coordinator named by the doc, which is the
t2db.31 run path and out of scope here.

Classifier enumeration (E) and verdict, all proven for a typed action
playbook via CEL (expr.NewEnv with a declared log.Result id):

  unknown id                        CEL compile (undefined field)       -> delete unknownActionReference
  literal map index actions["x"]    CEL compile (no _[_] overload)      -> covered
  computed index actions[key]       CEL compile (undeclared 'key')      -> covered
  dynamic index actions[event.name] CEL compile (no _[_] overload)      -> covered
  "x" in actions                    CEL compile (no @in overload)       -> covered
  size(actions)                     CEL compile (no size overload)      -> covered
  actions.all(...)                  CEL compile (not a range)           -> covered
  actions == {}                     CEL compile (no _==_ overload)      -> covered
  bare actions                      compiles; KEPT classifier           -> retained resolvable check

unknownActionReference is deleted; actionReferences/ActionReferences are
kept solely for the bare-actions spelling CEL does not reject.
2026-09-22 09:02:33 +10:00

355 lines
13 KiB
Go

// Package expr is the CEL expression environment for the EDA playbook engine:
// one mechanism for both an action's `when` condition and its `args` values,
// per eda-playbook-engine.md's resolved open question 1 (CEL decision).
//
// The environment declares two context roots -- `event` (the triggering
// event's decoded payload, kept dyn by J4) and `actions` (prior actions'
// results keyed by the action's declared id) -- and applies a cost limit to
// every program. `actions` is not a global map: its shape depends on the ids
// one playbook declares and the kind each id runs, so NewEnv takes the
// declared ids and their Go Result struct types and builds a per-playbook
// object type (multi-action-playbooks.md, D3). The same compile path serves
// the playbook loader (reject-at-load) and the lint tool, so author-time
// diagnostics and runtime evaluation cannot disagree.
//
// CEL is non-Turing-complete, side-effect-free, and panic-free by design (no
// recover() wrapper is needed; verified in the t2db.14 acceptance tests
// against hostile data). Anything the schema does not accept is a returned
// error here -- the reject-at-load philosophy of the parent design doc.
package expr
import (
"reflect"
"sort"
"strings"
"cel.dev/cel-go/cel"
celast "cel.dev/cel-go/common/ast"
"cel.dev/cel-go/common/types"
"cel.dev/cel-go/common/types/ref"
)
// DefaultCostLimit bounds evaluation of every playbook expression (J5 in the
// resolved doc: tunable, start here; linter and daemon share the same value).
const DefaultCostLimit = 100_000
// DeclaredResult is one prior action's result the environment exposes to a
// later expression. ID is the action's declared id, read as
// `actions.<ID>.result.<field>`; Type is the reflect.Type of that action
// kind's Go Result struct (the value the run marshals into the per-id
// `{result: ...}` map before a later expression reads it).
type DeclaredResult struct {
ID string
Type reflect.Type
}
// Env is the CEL environment for playbook expressions: the declared context
// roots and the default cost limit applied to every compiled program.
type Env struct {
celEnv *cel.Env
costLimit uint64
}
// Context is what an expression may read at dispatch time. Both roots are
// read-only from the expression's perspective; CEL enforces this.
type Context struct {
// Event is the triggering event's decoded payload (webhook body, forge
// issue, schedule tick). Declared dyn because its shape is unknown by
// design (schema-by-example); field-level typing is follow-up work once
// multi-action playbooks exist (per-kind generated Result structs
// declared as CEL types).
Event map[string]any
// Actions holds prior actions' results keyed by the action's id as
// declared in the playbook. Each id maps to the per-id wrapper
// `{"result": <KindResult struct>}` the run produces, so a later
// expression reads `actions.<id>.result.<field>`.
Actions map[string]map[string]any
}
// NewEnv builds a CEL environment whose `actions` root is an object type with
// one field per declared result. An empty declared set is valid: `actions` is
// then an object with no fields, and any `actions.<id>` read fails at compile
// time. `event` stays dyn (J4).
func NewEnv(declared ...DeclaredResult) *Env {
reg, err := newResultRegistry(declared)
if err != nil {
// NewEnv with an unsupported Result type (not a struct) is a
// programming error, not a playbook error. Panic is appropriate
// (package init-style invariant).
panic(err)
}
provider := newActionResultProvider(declared, reg)
celEnv, err := cel.NewEnv(
cel.CustomTypeAdapter(reg),
cel.CustomTypeProvider(provider),
cel.Variable("event", cel.MapType(cel.StringType, cel.DynType)),
cel.Variable("actions", provider.actionsType()),
)
if err != nil {
// cel.NewEnv with static variable declarations cannot fail in
// practice; a failure here is a programming error, not a playbook
// error. Panic is appropriate.
panic(err)
}
return &Env{celEnv: celEnv, costLimit: DefaultCostLimit}
}
// newResultRegistry registers every declared Result reflect.Type with CEL's
// native type system, using lower-cased Go field names so `Written` is read
// as `written`. The registry is both the provider (for the Result struct
// fields) and the adapter (so a Go Result struct value inside the eval map
// adapts to those same fields).
func newResultRegistry(declared []DeclaredResult) (*types.Registry, error) {
items := make([]any, 0, len(declared)+1)
items = append(items, types.ParseStructField(lowerFieldName))
for _, d := range declared {
items = append(items, d.Type)
}
return types.NewRegistry(items...)
}
// lowerFieldName maps a Go struct field to its CEL name: the lower-cased
// field name (Written -> written, Level -> level).
func lowerFieldName(f reflect.StructField) string {
return strings.ToLower(f.Name)
}
// nativeResultTypeName is the CEL type name cel-go's native type system
// assigns to a reflect.Type (simple package alias + struct name), e.g.
// `log.Result`. It must match the name the native registry registers, because
// the per-id wrapper's `result` field is typed by that name and its fields
// are resolved through the native registry.
func nativeResultTypeName(t reflect.Type) string {
if t.Kind() == reflect.Pointer {
t = t.Elem()
}
pkg := t.PkgPath()
if i := strings.LastIndexByte(pkg, '/'); i >= 0 {
pkg = pkg[i+1:]
}
return pkg + "." + t.Name()
}
// The `actions` root object type name and the per-id wrapper type namespace.
// Per-id names must live in a reserved namespace (eda.action.<id>): naming a
// wrapper `actions.<id>` makes the checker parse `actions.<id>` as a type
// reference rather than a field selection, which then fails with "type ...
// does not support field selection".
const (
actionsTypeName = "actions"
actionTypePrefix = "eda.action."
resultFieldName = "result"
)
// actionResultProvider is the data-driven types.Provider for the
// per-playbook `actions` object. It owns the `actions` root and each
// `eda.action.<id>` wrapper; every other type name (including the native
// Result structs) is delegated to the native registry.
type actionResultProvider struct {
ids []string
resultOf map[string]string
base *types.Registry
}
// newActionResultProvider builds the provider from the declared results,
// de-duplicating and sorting ids so field-name reporting is deterministic.
func newActionResultProvider(declared []DeclaredResult, base *types.Registry) *actionResultProvider {
p := &actionResultProvider{
ids: make([]string, 0, len(declared)),
resultOf: make(map[string]string, len(declared)),
base: base,
}
for _, d := range declared {
if _, seen := p.resultOf[d.ID]; seen {
continue
}
p.resultOf[d.ID] = nativeResultTypeName(d.Type)
p.ids = append(p.ids, d.ID)
}
sort.Strings(p.ids)
return p
}
func (p *actionResultProvider) actionsType() *types.Type {
return types.NewObjectType(actionsTypeName)
}
func (p *actionResultProvider) wrapperType(id string) string {
return actionTypePrefix + id
}
// EnumValue implements types.Provider.
func (p *actionResultProvider) EnumValue(enumName string) ref.Val {
return p.base.EnumValue(enumName)
}
// FindIdent implements types.Provider.
func (p *actionResultProvider) FindIdent(identName string) (ref.Val, bool) {
return p.base.FindIdent(identName)
}
// FindStructType implements types.Provider.
func (p *actionResultProvider) FindStructType(structType string) (*types.Type, bool) {
if structType == actionsTypeName {
return types.NewTypeTypeWithParam(types.NewObjectType(structType)), true
}
for _, id := range p.ids {
if structType == p.wrapperType(id) {
return types.NewTypeTypeWithParam(types.NewObjectType(structType)), true
}
}
return p.base.FindStructType(structType)
}
// FindStructFieldNames implements types.Provider.
func (p *actionResultProvider) FindStructFieldNames(structType string) ([]string, bool) {
if structType == actionsTypeName {
return append([]string(nil), p.ids...), true
}
for _, id := range p.ids {
if structType == p.wrapperType(id) {
return []string{resultFieldName}, true
}
}
return p.base.FindStructFieldNames(structType)
}
// FindStructFieldType implements types.Provider.
func (p *actionResultProvider) FindStructFieldType(structType, fieldName string) (*types.FieldType, bool) {
if structType == actionsTypeName {
for _, id := range p.ids {
if fieldName == id {
return &types.FieldType{Type: types.NewObjectType(p.wrapperType(id))}, true
}
}
return nil, false
}
for _, id := range p.ids {
if structType == p.wrapperType(id) && fieldName == resultFieldName {
return &types.FieldType{Type: types.NewObjectType(p.resultOf[id])}, true
}
}
return p.base.FindStructFieldType(structType, fieldName)
}
// NewValue implements types.Provider.
func (p *actionResultProvider) NewValue(structType string, fields map[string]ref.Val) ref.Val {
return p.base.NewValue(structType, fields)
}
// Compile parses and type-checks a playbook expression string against the
// declared context. A syntax error, an unknown root (anything other than
// event/actions), a dynamic `actions` read, or a field typo on a declared
// result is a returned error -- never a panic. The returned Program is safe
// to evaluate concurrently (CEL programs are stateless once compiled).
func (e *Env) Compile(src string) (*Program, error) {
ast, issues := e.celEnv.Compile(src)
if issues != nil && issues.Err() != nil {
return nil, issues.Err()
}
prg, err := e.celEnv.Program(ast, cel.CostLimit(e.costLimit))
if err != nil {
return nil, err
}
ids, resolvable := actionReferences(ast)
return &Program{prg: prg, actionIDs: ids, resolvable: resolvable}, nil
}
// actionReferences walks the compiled AST once and classifies every read of
// the `actions` context root as either a statically-resolvable action id or
// not. The invariant is exhaustive by construction: each read of `actions`
// consumes exactly one `actions` identifier node, so counting identifier
// nodes and counting the reads that match the static access shape (a field
// selection on the `actions` ident) yields
//
// resolvable = (identCount == staticCount)
//
// Any other spelling that mentions `actions` contributes an identifier node
// without a matching static access, so it reports unresolvable rather than
// slipping through as a runtime miss. The returned ids are the sorted,
// de-duplicated ids of the static accesses.
//
// With the per-playbook object type the CEL checker rejects every non-static
// `actions` read at compile time -- a dynamic index, `in`, `size`, or a
// method/field selection that is not a declared id never reaches this walk.
// The one spelling the checker does NOT reject is a bare `actions` value
// read, which compiles as an object value but cannot be pinned to an action
// id at load. This walk is retained solely to reject that spelling; the ids
// return is now informational (the unknown-id check moved to the checker's
// per-playbook object type).
func actionReferences(ast *cel.Ast) ([]string, bool) {
if ast == nil || ast.NativeRep() == nil {
return nil, true
}
var identCount, staticCount int
seen := map[string]struct{}{}
visitor := celast.NewExprVisitor(func(e celast.Expr) {
switch e.Kind() {
case celast.IdentKind:
if e.AsIdent() == "actions" {
identCount++
}
case celast.SelectKind:
sel := e.AsSelect()
if !isActionsIdent(sel.Operand()) {
return
}
staticCount++
seen[sel.FieldName()] = struct{}{}
}
})
celast.PreOrderVisit(ast.NativeRep().Expr(), visitor)
ids := make([]string, 0, len(seen))
for id := range seen {
ids = append(ids, id)
}
sort.Strings(ids)
return ids, identCount == staticCount
}
// isActionsIdent reports whether e is the `actions` context-root identifier.
func isActionsIdent(e celast.Expr) bool {
return e.Kind() == celast.IdentKind && e.AsIdent() == "actions"
}
// Program is a compiled, cost-limited playbook expression.
type Program struct {
prg cel.Program
actionIDs []string
resolvable bool
}
// ActionReferences reports the action ids the expression reads from the
// `actions` context root (sorted, de-duplicated) and whether every `actions`
// read could be statically resolved to one of those ids. ids is empty when
// the expression reads no prior-action result; resolvable is false when the
// expression reads `actions` in any form that cannot be pinned to a prior
// action id at load. Under the per-playbook object type the checker already
// rejects every dynamic `actions` read, so the only form that still reaches
// here unresolved is a bare `actions` value read; the playbook loader rejects
// that rather than evaluating it as a runtime miss.
func (p *Program) ActionReferences() (ids []string, resolvable bool) {
if p == nil {
return nil, true
}
return p.actionIDs, p.resolvable
}
// Eval evaluates the program against a dispatch-time context. A missing
// field, a wrong type against a dyn value, or an exceeded cost limit is a
// returned error. CEL's own evaluation recovers panics internally (verified
// in cel-go's program.go: Eval wraps evaluation in recover), and its
// returned values implement ref.Val -- unwrapped here to the native Go value
// the caller expects.
func (e *Env) Eval(prg *Program, ctx Context) (any, error) {
data := map[string]any{
"event": ctx.Event,
"actions": ctx.Actions,
}
v, _, err := prg.prg.Eval(data)
if err != nil {
return nil, err
}
return v.Value(), nil
}