Skip to main content

AXON EBNF — abridged grammar reference

This page reproduces the production rules an agent most often needs: program structure, declarations, the flow body, type expressions, and the session-types fragment.

The full canonical EBNF lives in the language paper papers/paper_axon_language.md; the parser implementation lives in axon-frontend/src/parser.rs. This abridged version is curated for agent consumption: tokens that matter, productions that compile, omissions clearly marked.

Lexical surface​

identifier        = letter , { letter | digit | "_" } ;
integer           = digit , { digit } ;
number            = integer , [ "." , digit , { digit } ] ;
duration          = integer , ( "ms" | "s" | "m" | "h" ) ;
string_literal    = '"' , { utf8_char } , '"' ;
bcp47             = '"' , language_subtag , { "-" , subtag } , '"' ;

Comments are // line comments. Whitespace is non-significant except where the dotted-slug grammar (legal:HIPAA.164_502) requires token adjacency.

Program​

program           = { declaration | import | run } ;
import            = "import" , identifier , "{" , identifier_list , "}" ;
declaration       = persona_def
                  | context_def
                  | flow_def
                  | anchor_def
                  | tool_def
                  | type_def
                  | session_def
                  | socket_def
                  | axonendpoint_def
                  | axonstore_def
                  | shield_def
                  | mandate_def
                  | compute_def
                  | lambda_def
                  | daemon_def
                  | (* …65+ total — see axon://grammar/top_level *) ;
run               = "run" , identifier , "(" , arg_list? , ")"
                  , [ "as" , identifier ]
                  , [ "within" , identifier ]
                  , [ "constrained_by" , "[" , identifier_list , "]" ]
                  , [ "on_failure" , ":" , failure_policy ]
                  , [ "output_to" , ":" , string_literal ]
                  , [ "effort" , ":" , effort_level ] ;

Persona​

persona_def       = "persona" , identifier , "{" , persona_field+ , "}" ;
persona_field     = "domain"               , ":" , string_list
                  | "tone"                 , ":" , tone_ident
                  | "confidence_threshold" , ":" , number
                  | "cite_sources"         , ":" , bool
                  | "refuse_if"            , ":" , bracketed_identifiers
                  | "language"             , ":" , bcp47
                  | "description"          , ":" , string_literal ;

tone_ident        = "analytical" | "assertive" | "casual" | "diplomatic"
                  | "empathetic" | "formal"    | "friendly" | "precise" ;

Flow + step​

flow_def          = "flow" , identifier , "(" , param_list? , ")"
                  , [ "->" , type_expr ]
                  , "{" , { flow_step } , "}" ;
param_list        = parameter , { "," , parameter } ;
parameter         = identifier , ":" , type_expr ;

flow_step         = step_node
                  | if_stmt | for_stmt | let_stmt | return_stmt
                  | break_stmt | continue_stmt
                  | reason_step | probe_step | validate_step | refine_step
                  | weave_step | use_step | par_block | hibernate_step
                  | listen_step | retrieve_step | persist_step | mutate_step
                  | deliberate_step | navigate_step | drill_step | trail_step
                  | (* …complete list in parser.rs::parse_flow_step *) ;

deliberate_step   = "deliberate" , "{" , { flow_step } , "}" ;
(* v2.12.0/63 — cognitive retrieval over a `pix` tree OR a `corpus`
   graph (a corpus WITH `relations:` is an MDN graph). Dispatch is by
   the referenced declaration, not by keyword. *)
navigate_step     = "navigate" , identifier
                  , "{" , { navigate_field } , "}" ;
navigate_field    = "query"  , ":" , expression
                  | "from"   , ":" , identifier        (* seed document *)
                  | "budget" , ":" , number            (* max documents  *)
                  | "trail"  , ":" , boolean
                  | "output" , ":" , identifier ;
drill_step        = "drill"  , identifier , "{" , { navigate_field } , "}" ;
trail_step        = "trail"  , identifier , "{" , { navigate_field } , "}" ;

step_node         = "step" , identifier , [ "use" , identifier ]
                  , "{" , step_field+ , "}" ;
step_field        = "given"            , ":" , expression
                  | "ask"              , ":" , string_literal
                  | "output"           , ":" , type_expr
                  | "confidence_floor" , ":" , number
                  | "navigate"         , ":" , dotted_identifier
                  | "apply"            , ":" , identifier
                  | sub_construct ;
sub_construct     = ( "use" | "probe" | "reason" | "weave" | "stream" ) , … ;

reason_step       = "reason" , [ identifier ] , [ "{" , (* skipped *) , "}" ] ;

Type expressions​

type_expr         = identifier , [ "<" , type_expr , ">" ] , [ "?" ] ;
type_def          = "type" , identifier
                  , [ "(" , number , ".." , number , ")" ]
                  , [ "where" , expression ]
                  , [ "compliance" , bracketed_identifiers ]
                  , [ "{" , type_field_list , "}" ] ;
type_field_list   = type_field , { "," , type_field } ;
type_field        = identifier , ":" , type_expr ;

Recursive generics like FlowEnvelope<List<TenantRecord>> are accepted since v2.0.0.

Anchor​

anchor_def        = "anchor" , identifier , "{" , anchor_field+ , "}" ;
anchor_field      = "require"          , ":" , identifier
                  | "reject"           , ":" , bracketed_identifiers
                  | "enforce"          , ":" , identifier
                  | "description"      , ":" , string_literal
                  | "confidence_floor" , ":" , number
                  | "unknown_response" , ":" , string_literal
                  | "on_violation"     , ":" , violation_policy ;
violation_policy  = ( "raise" | "fallback" ) , identifier
                  | identifier ;

Tool​

tool_def          = "tool" , identifier , "{" , tool_field+ , "}" ;
tool_field        = "provider"    , ":" , identifier
                  | "max_results" , ":" , integer
                  | "filter"      , ":" , filter_expr
                  | "timeout"     , ":" , duration
                  | "runtime"     , ":" , identifier
                  | "sandbox"     , ":" , bool
                  | "effects"     , ":" , effect_row ;
effect_row        = "<" , effect_term , { "," , effect_term } , ">" ;
effect_term       = effect_name , [ ":" , qualifier_value ] ;
effect_name       = "io" | "network" | "pure" | "random" | "storage"
                  | "stream" | "trust" | "sensitive" | "legal" | "ots" ;
qualifier_value   = dotted_slug ;

Session types (v2.3.0)​

session_def       = "session" , identifier , "{"
                  , "client" , ":" , session_type
                  , "server" , ":" , session_type
                  , "}" ;
session_type      = "[" , session_action , { "," , session_action } , "]" ;
session_action    = "send" , type_expr
                  | "receive" , type_expr
                  | "select" , "{" , session_branch , { "," , session_branch } , "}"
                  | "branch" , "{" , session_branch , { "," , session_branch } , "}"
                  | "loop"
                  | "end" ;
session_branch    = identifier , ":" , session_type ;

socket_def        = "socket" , identifier , "{" , socket_field+ , "}" ;
socket_field      = "protocol"      , ":" , identifier
                  | "backpressure"  , ":" , "credit" , "(" , integer , ")"
                  | "reconnect"     , ":" , "cognitive_state"
                  | "legal_basis"   , ":" , identifier ;

The duality + credit-refined backpressure constraints (peer ≡ self⊥, Presburger discharge) live in the type checker, not the grammar — they are typing rules, enforced after the parser has produced the AST.

What this page does NOT include​

This is the abridged grammar reference. Productions deliberately omitted (covered in dedicated resources or in the paper):

  • The full flow_step enumeration (~30+ kinds — see parser.rs::parse_flow_step for the canonical list).
  • The full axonendpoint, axonstore, daemon, shield, psyche, ots, agent, lambda, compute, resource, fabric, manifest, observe, reconcile, lease, ensemble, pix, mandate, mcp, taint, logic, corpus, dataspace declarations.
  • Compliance-annotation dotted-slug grammar (covered under each axon://compliance/<framework> resource).
  • The Honda-Yoshida-Carbone multiparty projection rules (covered under axon://logic/session_duality and the v2.3.0 paper).

For those productions, consult the linked sources or the per-primitive resources via axon.primitives + axon.primitive_doc.