Conditions

A condition is a single-line expression string, evaluated against the current memory values whenever the runtime needs it. The same grammar is used everywhere a condition appears.

requires: "session.project.authenticated"

Where conditions appear

Site	Where it is written	What it controls
`requires:`	`skill.md` frontmatter, top level	Whether the orchestrator may route to the skill
`requires:`	A `tool_constraints` entry in frontmatter	Whether a tool is offered to the LLM
`if:`	First line of a prose paragraph	Whether the paragraph is included in the prompt
`complete_when:`	Frontmatter, or an `instructions:` step	Whether the step is already satisfied
`if:`	A `next:` branch list on a step	Which step runs next

Memory references

Every memory reference has three segments:

session.<namespace>.<entry_name>

<namespace> is a skill id or the literal project. <entry_name> is the bare field name as declared in memory.yml.

session.card_replace.replacement_reason    # declared in skills/card_replace/memory.yml
session.project.selected_card_id           # declared in the project-root memory.yml

The namespace is the skill’s directory name, which is also its catalog id. A skill in skills/card_replace/ with name: Card Replace is referenced as session.card_replace.<entry>.

Operators

Category	Available
Comparison	`==` `!=` `<` `<=` `>` `>=`
Boolean	`and` `or` `not`
Grouping	`( )`
Literals	strings, integers, floats, `True`, `False`, `None`

Comparisons chain, so session.a.x < session.a.y < 10 behaves as it does in Python. These are the operators the expression parser accepts. To test membership in a set, write it as a disjunction:

requires: >
  session.card_replace.replacement_reason == 'lost'
  or session.card_replace.replacement_reason == 'not_received'

Truthiness

An unset memory entry reads as None, and the whole expression is coerced with bool(). A bare reference is therefore a presence test:

requires: "session.project.selected_card_id"              # set and non-empty
requires: "not session.card_replace.replacement_reason"   # unset, empty, or false

This matters for a boolean with three meaningful states: unset, True, and False. Compare explicitly to tell them apart:

requires: "session.card_replace.wants_lock == None"    # we have not asked yet
requires: "session.card_replace.wants_lock == False"   # the user said no

0, 0.0, "", and [] are falsy. For a numeric entry that may legitimately be zero, compare against None rather than relying on truthiness.

Which entries you may reference

A condition on a skill can read:

that skill’s own public and private fields
other skills’ public fields
project fields, declared in the root memory.yml

minus anything listed in the skill’s access.deny_read. These are the same rules the runtime applies when it assembles memory for a turn, so a reference that passes validation resolves at runtime.

Writing long conditions in YAML

Conditions are often longer than a line. Use a folded scalar (>), which joins the lines with spaces into a single-line expression:

tool_constraints:
  - lock_card:
      requires: >
        not session.card_replace.card_locked_this_flow
        and (
        session.card_replace.replacement_reason == 'stolen'
        or session.card_replace.wants_lock == True
        )

Two quoting rules:

Quote or fold a condition containing : followed by a space, so YAML reads it as a string rather than a mapping.
Use single quotes for string literals inside a condition, so the whole expression can sit in double quotes: "session.card_replace.replacement_reason == 'stolen'".

Gates fail closed

Every condition guards access to something: a skill, a tool, a paragraph. When a condition does not hold, the thing it guards is withheld:

Site	When the condition is false
Skill `requires:`	The skill is not offered for routing
`tool_constraints` `requires:`	The tool is absent from the LLM’s schema, and blocked if dispatch is attempted
Prose `if:`	The paragraph is left out of the prompt
`next:` branch `if:`	Evaluation moves to the next branch

Withholding is also the outcome when a condition cannot be evaluated at all. This is deliberate: offering a tool and then refusing the call would waste a turn, so the safe direction is to keep it hidden.

What `rasa train` checks

Prose if: conditions are verified before packaging, so branching errors surface at build time rather than mid-conversation:

Code	Meaning
`calm_v2.validation.prose.empty_if_condition`	An `if:` marker with no expression
`calm_v2.validation.prose.invalid_if_expression`	An expression the parser does not accept
`calm_v2.validation.prose.unknown_memory_key`	A key the skill cannot read at runtime

The last one is the useful one: it catches a mistyped entry name or a reference to a field the active skill has no access to. The Inspector’s tools panel shows the current gate result for every declared tool, which is the quickest way to confirm a requires: expression behaves the way you intended.

Worked examples

Inline requires: patterns (same shape as tool constraints in a banking card-replace skill):

# record the reason only while it is still unset
- set_replacement_reason:
    requires: "not session.card_replace.replacement_reason"

# gate on a value another skill wrote into project memory
- check_number_of_transactions_returned:
    requires: >
      (
      session.card_replace.replacement_reason == 'lost'
      or session.card_replace.replacement_reason == 'not_received'
      )
      and session.project.selected_card_id
      and not session.card_replace.transaction_review_complete

# a tri-state boolean: offer the lock question only before it has been asked
- record_lock_choice:
    requires: >
      session.card_replace.wants_lock == None
      and not session.card_replace.card_locked_this_flow

Prose scoping:

if: session.card_replace.replacement_reason == 'stolen'
Lock the card immediately and tell the customer it was locked to protect
their account.

Branch routing inside an ordered block, where else: provides the fallback:

- id: gate_card_selected
  noop: true
  next:
    - if: "session.project.selected_card_id and session.project.selected_card_label"
      then: check_card_eligible
    - else: card_not_found

File Reference

Language

Tools

Runtime

Prompts

Where conditions appear

Memory references

Operators

Truthiness

Which entries you may reference

Writing long conditions in YAML

Gates fail closed

What `rasa train` checks

Worked examples

See also

​Where conditions appear

​Memory references

​Operators

​Truthiness

​Which entries you may reference

​Writing long conditions in YAML

​Gates fail closed

​What rasa train checks

​Worked examples

​See also

Where conditions appear

Memory references

Operators

Truthiness

Which entries you may reference

Writing long conditions in YAML

Gates fail closed

What `rasa train` checks

Worked examples

See also