iterion-dsl

name	iterion-dsl
description	Use when asked to write .iter workflow files, create Iterion DSL workflows, define multi-agent orchestration pipelines, or debug .iter syntax errors. Triggers on: "write an .iter file", "iterion workflow", "iter DSL", "multi-agent pipeline", "workflow orchestration", "agent judge router".
version	0.1.0

Iterion DSL Reference

Iterion is a workflow orchestration engine for multi-agent AI pipelines. Workflows are defined in .iter files using an indentation-sensitive DSL (2 spaces per level). One workflow per file.

File Structure

Blocks must appear in this order:

vars:             ## optional — global variables
mcp_server X:     ## optional — MCP server declarations
prompt X:         ## prompt templates (referenced by name)
schema X:         ## typed data contracts (referenced by name)
agent X:          ## LLM nodes with tools
judge X:          ## LLM nodes producing verdicts (typically no tools)
router X:         ## deterministic or LLM-based routing
human X:          ## human-in-the-loop interaction points
tool X:           ## direct command execution (no LLM)
compute X:        ## deterministic expression node (no LLM, no shell)
workflow X:       ## entry point, edges, budget

Comments start with ## and extend to end of line.

Node Types

Type	Purpose	Key Properties
agent	LLM node with tools	model/backend, input, output, system, user, tools, session
judge	LLM verdict node	Same as agent (semantically: produces verdicts, rarely uses tools)
router	Fan-out or conditional routing	mode (fan_out_all/condition/round_robin/llm)
human	Human-in-the-loop pause	interaction (human/llm/llm_or_human), instructions
tool	Direct shell command	command, [input], output
compute	Deterministic expression node (no LLM, no shell)	input, output, expr (key→expression mapping)
done	Terminal: success	Reserved identifier, not declared — used as edge target
fail	Terminal: failure	Reserved identifier, not declared — used as edge target

Agent / Judge Properties

Agents and judges share identical syntax. All properties are optional except model or backend (one is required).

agent my_agent:
  model: "claude-sonnet-4-20250514"       ## LLM model identifier (supports ${ENV_VAR})
  backend: "claude_code"                   ## execution backend: claw | claude_code | codex
  input: my_input_schema                   ## schema reference for input data
  output: my_output_schema                 ## schema reference for structured output
  publish: my_artifact                     ## persist output as named artifact
  system: my_system_prompt                 ## prompt reference for system message
  user: my_user_prompt                     ## prompt reference for user message
  session: fresh                           ## fresh | inherit | fork | artifacts_only
  tools: [Read, Edit, Write, Bash]         ## tool capability names
  tool_policy: [Bash]                         ## tool policy refs (same syntax as tools)
  tool_max_steps: 15                       ## max tool-use iterations (0 = unlimited)
  max_tokens: 8000                         ## per-call output cap (0 = backend default)
  reasoning_effort: high                   ## low | medium | high | xhigh | max
  readonly: true                           ## not considered mutating for workspace safety
  interaction: llm_or_human                ## none | human | llm | llm_or_human
  interaction_prompt: my_interaction_prompt ## prompt ref for LLM interaction decisions
  interaction_model: "claude-haiku-4-5-20251001"  ## model for interaction (fallback to model)
  await: wait_all                          ## wait_all | best_effort (convergence strategy)
  mcp:                                     ## MCP configuration block
    inherit: true
    servers: [my_mcp_server]
    disable: [unwanted_server]
  compaction:                              ## per-node compaction override (optional)
    threshold: 0.85                         ## fraction of context window in (0, 1]
    preserve_recent: 4                      ## minimum recent turns kept verbatim (>= 1)

Backends:

• claw — default, in-process LLM call. Recommended for read-only LLM nodes (judges, reviewers, planners).
• claude_code — recommended for nodes that need real tool/shell access (implementers, fixers).
• codex — supported but discouraged (raises C030); the integration is less ergonomic and tool gating is limited.

Session modes:

• fresh — new context (default)
• inherit — inherit parent session (forbidden on convergence points)
• fork — non-consuming fork from parent (forbidden on convergence points)
• artifacts_only — only persistent artifacts in context

Interaction modes:

• none — no interaction capability (default for agent/judge)
• human — always pause for human input (default for human nodes)
• llm — LLM auto-answers interaction questions
• llm_or_human — LLM decides whether to answer or escalate to human

Router Properties

router my_router:
  mode: fan_out_all          ## fan_out_all | condition | round_robin | llm

Modes:

• fan_out_all — send to ALL outgoing edges in parallel
• condition — route based on when clauses on edges
• round_robin — cycle through targets one at a time
• llm — LLM decides which target(s) to route to (requires additional properties)

LLM router additional properties (only valid with mode: llm):

router my_llm_router:
  mode: llm
  model: "claude-sonnet-4-20250514"   ## required for llm mode
  system: router_system_prompt        ## prompt ref (optional)
  user: router_user_prompt            ## prompt ref (optional)
  multi: true                         ## allow multi-target selection (optional)

Routers do NOT support await — they are fan-out sources, not convergence targets.

Human Properties

human my_gate:
  input: gate_input_schema
  output: gate_output_schema
  publish: gate_result
  instructions: gate_instructions_prompt   ## prompt ref shown to human
  interaction: human                       ## human | llm | llm_or_human
  interaction_prompt: decision_prompt      ## prompt ref for LLM decisions
  interaction_model: "claude-haiku-4-5-20251001"
  min_answers: 1                           ## minimum human answers required
  model: "claude-sonnet-4-20250514"        ## required for llm/llm_or_human modes
  system: human_system_prompt              ## prompt ref for LLM system prompt
  await: wait_all                          ## convergence strategy

Tool Properties

Direct command execution without LLM. Supports ${ENV_VAR} in command (resolved at compile time) and {{input.field}} for runtime values.

tool run_tests:
  command: "make test"
  input: test_input_schema       ## optional — lets {{input.X}} render in command
  output: test_result_schema
  await: wait_all

A string[] field rendered via {{input.list_field}} expands as space-joined items, so a tool can take an arbitrary list of arguments without manual concatenation.

Compute Properties

Deterministic node — evaluates a list of expressions over vars / input / outputs / artifacts / loop / run and emits a structured output. No LLM, no shell. Useful for streak detection, boolean combinations, counters, simple aggregations that shouldn't burn tokens.

schema streak_state:
  consecutive_passes: int
  ready: bool

compute streak:
  input: streak_input               ## optional input schema
  output: streak_state              ## required output schema
  expr:
    consecutive_passes: "loop.refine.iter"
    ready: "outputs.review.passed && loop.refine.iter >= 2"
  await: wait_all

Built-ins available inside expressions: length(x), concat(a, b, …), unique(list), contains(list, item). A compute node with no expr entries raises C039; an unparseable expression raises C040.

Schema Syntax

Typed data contracts referenced by nodes. Field types: string, bool, int, float, json, string[].

schema review_result:
  summary: string
  blockers: string[]
  approved: bool
  confidence: string [enum: "low", "medium", "high"]
  details: json
  score: float
  retry_count: int

Prompt Syntax

Free-form text templates with {{...}} interpolation. Referenced by name from node properties.

prompt review_system:
  You are a senior code reviewer.
  Analyze the code changes and produce a structured review.

prompt review_user:
  Review this PR.

  Title: {{input.pr_title}}
  Diff: {{input.diff}}
  Previous feedback: {{outputs.previous_review.summary}}
  Saved context: {{artifacts.pr_context}}
  Config: {{vars.review_rules}}

Template references:

• {{vars.X}} — workflow variable
• {{input.X}} — current node's input field
• {{outputs.node_id}} — full output of a node
• {{outputs.node_id.field}} — specific field from a node's output
• {{outputs.node_id.history}} — array of all outputs across loop iterations (only valid for nodes in a loop)
• {{artifacts.X}} — persistent artifact by name
• {{loop.<name>.iter}} — current 0-based iteration of a declared loop
• {{loop.<name>.previous.<field>}} — previous iteration's field value on the loop's controlling node
• {{run.id}}, {{run.store_dir}} — run-scoped metadata
• ${ENV_VAR} — environment variable (resolved at compile time)

Workflow Block

workflow my_workflow:
  vars:
    pr_title: string
    max_retries: int = 3
    verbose: bool = false

  entry: first_node

  default_backend: "claude_code"   ## fallback backend for nodes that don't set their own
  tool_policy: [Bash]                 ## workflow-level tool policy refs
  worktree: auto                    ## auto | none — see "Worktree isolation" below

  budget:
    max_parallel_branches: 4
    max_duration: "30m"
    max_cost_usd: 10.0
    max_tokens: 500000
    max_iterations: 50

  compaction:                       ## workflow-level compaction defaults
    threshold: 0.85
    preserve_recent: 4

  interaction: llm_or_human     ## workflow-level default for all nodes

  mcp:
    autoload_project: true
    servers: [my_server]
    disable: [unwanted]

  ## Edges
  first_node -> second_node
  second_node -> done

Worktree isolation

worktree: auto runs the workflow inside a per-run git worktree at <store-dir>/worktrees/<run-id>/ so the user's main working tree stays untouched and WIP edits stay invisible to the workflow. Clean exits automatically remove that worktree. Failed, cancelled, or error exits preserve it and log the path so the operator can inspect it and, when desired, clean it up manually with git worktree remove --force <path>. Omit the field (or set none) to run in place. See examples/vibe_feature_dev.iter for a workflow that opts in.

Edge Syntax

src -> dst [when [not] field] [as loop_name(max_iter)] [with { mappings }]

Components:

• when field — route only if field is true in source output (field must be bool)
• when not field — route only if field is false
• when "<expression>" — route on a quoted compound expression (same namespaces and built-ins as compute: vars / input / outputs / artifacts / loop / run, plus length / concat / unique / contains)
• as loop_name(N) — declare a bounded loop (max N iterations)
• with { ... } — data mappings for the target node's input

Examples:

## Simple edge
agent_a -> agent_b

## Conditional edges (must cover all cases or have a fallback)
judge -> done when approved
judge -> fixer when not approved

## Bounded loop
judge -> agent as refine_loop(5) when not approved with {
  feedback: "{{outputs.judge.summary}}",
  history: "{{outputs.agent.history}}"
}

## Data mapping
builder -> reviewer with {
  pr_context: "{{outputs.builder}}",
  rules: "{{vars.review_rules}}"
}

Edge rules:

• Non-router nodes can have at most ONE unconditional edge
• Conditional edges must either be exhaustive (when X + when not X) or have an unconditional fallback
• fan_out_all / round_robin / llm routers can have multiple unconditional edges
• llm router edges must NOT have when conditions (the LLM decides routing)
• round_robin and llm routers need at least 2 outgoing edges
• Cycles MUST have a loop declaration (as name(N)) — undeclared cycles produce diagnostic C019
• done and fail are reserved targets — do not declare them as nodes

MCP Server Declarations

Top-level declarations for reusable MCP servers.

mcp_server code_tools:
  transport: stdio
  command: "npx"
  args: ["-y", "@anthropic-ai/claude-code-mcp"]

mcp_server api_server:
  transport: http
  url: "http://localhost:3000/mcp"

Transport types: stdio (requires command), http (requires url). SSE is not supported in V1.

Convergence (Await)

There is NO join keyword. Convergence is handled by the await property on the receiving node.

When a node receives edges from multiple parallel branches (e.g., after a fan_out_all router), set await on the target:

router fan_out:
  mode: fan_out_all

agent worker_a:
  ## ...

agent worker_b:
  ## ...

judge merge:
  await: wait_all       ## waits for BOTH worker_a and worker_b
  ## ...

• wait_all — wait for all incoming branches before executing
• best_effort — proceed when possible, tolerate branch failures

Important: session: inherit and session: fork are forbidden on convergence points (diagnostic C009). Use session: fresh or session: artifacts_only.

Variables

Top-level or workflow-level. Types: string, bool, int, float, json, string[].

vars:
  workspace_dir: string = "${PROJECT_DIR}"
  max_retries: int = 3
  verbose: bool = false

Workflow-level vars override top-level vars with the same name. Pass values at runtime: iterion run file.iter --var key=value.

Common Mistakes

1. Missing model or backend — Every agent/judge needs one. Set model: "${MODEL}" for env-based config or backend: "claude_code" for delegation. (Earlier drafts used delegate:; that keyword has been removed — use backend: everywhere.)

2. when field not boolean — Condition fields must be bool in the source output schema. Use approved: bool not status: string.

3. Undeclared cycles — Every back-edge must have as loop_name(N). Without it you get C019 (infinite loop risk).

4. session: inherit on convergence point — Nodes with await: or multiple incoming sources cannot use inherit or fork. Use fresh or artifacts_only.

5. Non-exhaustive conditions without fallback — If you have when approved you must also have when not approved OR an unconditional fallback edge.

6. Schema not declared — Every schema referenced by input: or output: must be declared as a top-level schema block.

7. when on LLM router edges — LLM routers select targets directly. Don't add conditions to their edges.

8. Single edge on round_robin/llm router — These routers need at least 2 outgoing edges.

Minimal Example

prompt review_system:
  You are a code reviewer. Evaluate the code and decide if it passes review.

prompt review_user:
  Review this code:
  {{input.code}}

schema review_input:
  code: string

schema review_output:
  approved: bool
  summary: string

agent reviewer:
  model: "${MODEL}"
  input: review_input
  output: review_output
  system: review_system
  user: review_user

workflow code_review:
  entry: reviewer
  reviewer -> done when approved
  reviewer -> fail when not approved

CLI Quick Reference

iterion validate file.iter              # parse + compile + validate
iterion run file.iter --var key=value   # execute workflow
iterion diagram file.iter --view full   # generate Mermaid diagram
iterion inspect --run-id <id>           # view run state
iterion resume --run-id <id> --file f --answers-file a.json  # resume paused run

See Also

• references/dsl-grammar.md — formal grammar specification
• references/patterns.md — common workflow patterns with examples
• references/diagnostics.md — all validation diagnostic codes
• examples/skill/ — minimal self-contained .iter examples
• examples/ — production-grade workflow examples