Skip to main content
← Back to list
01Issue
BugShippedSwamp CLIPublic
Assigneesskunk-ape

Relationships

#4765 validate: spurious ambiguous-exit warnings for evidence absent in this stage/cycle, and for complementary cel comparisons

Opened by rcrowley · 11/15/2026· Shipped 11/16/2026

Summary

validate gives ambiguous-exit warnings for sibling transitions that are exclusive. The exclusive() check in _lib/engine/graph.ts does not recognize two common gate forms:

  1. A CEL test that the evidence is absent in this stage and cycle, compared with an evidence-recorded gate on the same evidence.
  2. Complementary CEL comparisons, for example x >= 2 and x < 2.

Case 1: evidence absent in this stage and cycle

exclusive() sees that an evidence-recorded gate and a cel gate are exclusive only when absentEvidence() finds !("x" in evidence) or !has(evidence.x) at the top level of the expression.

But evidence-recorded passes only on evidence that was recorded in this stage and cycle. The evidence map in CEL holds the latest record of each name in the era, from any cycle. Thus !("x" in evidence) is not the complement of evidence-recorded. It is stricter: when one round records the evidence, the gate also fails in all of the rounds after it. The correct complement is:

- type: cel
  config:
    expr: >-
      !("review-feedback" in evidence &&
      evidence["review-feedback"].stage == stage.id &&
      evidence["review-feedback"].cycle == stage.cycle)

absentEvidence() does not recognize this form, because the negation wraps a conjunction. Thus the checker accepts only the incorrect gate, and it gives a warning for the correct gate.

Reproduction

Stage review has these transitions. The cel gate is on each automatic exit, and revise gates on evidence-recorded:

- id: review
  transitions:
    - name: accept
      to: deliver
      gates:
        - type: cel
          config:
            expr: >-
              !("review-feedback" in evidence &&
              evidence["review-feedback"].stage == stage.id &&
              evidence["review-feedback"].cycle == stage.cycle)
        - type: findings-clear
          config: { artifact: review, blocking: [critical, high] }
        - type: cel
          config:
            expr: artifacts["review"].version >= 2
    - name: again
      to: review
      gates:
        - type: cel
          config:
            expr: >-
              !("review-feedback" in evidence &&
              evidence["review-feedback"].stage == stage.id &&
              evidence["review-feedback"].cycle == stage.cycle)
        - type: findings-clear
          config: { artifact: review, blocking: [critical, high] }
        - type: cel
          config:
            expr: artifacts["review"].version < 2
    - name: revise
      to: research
      gates:
        - type: evidence-recorded
          config: { name: review-feedback }

swamp behavior run <factory> validate logs:

stages.2.transitions.0 (from stage 'review'): transition 'accept' (to 'deliver') and transition 'revise' (to 'research') from stage 'review' can both pass with no person choosing between them; their cel gates could not be compared; make one manual, give it a human-approval gate without when, or make their gates exclusive
stages.2.transitions.1 (from stage 'review'): transition 'again' (to 'review') and transition 'revise' (to 'research') from stage 'review' can both pass with no person choosing between them; their cel gates could not be compared; make one manual, give it a human-approval gate without when, or make their gates exclusive

Expected

There are no warnings. When evidence-recorded on x passes, x is in evidence with stage == stage.id and cycle == stage.cycle, so the cel gate is false.

Suggested fix

In absentEvidence(), also accept !(A && ...) where one term is "x" in evidence or has(evidence.x), and all of the other terms are evidence.x.stage == stage.id or evidence.x.cycle == stage.cycle, in either operand order. Also accept the De Morgan form: !("x" in evidence) || evidence.x.stage != stage.id || evidence.x.cycle != stage.cycle.

Case 2: complementary CEL comparisons

accept and again above gate on artifacts["review"].version >= 2 and artifacts["review"].version < 2. These can never both pass, but validate logs:

stages.2.transitions.0 (from stage 'review'): transition 'accept' (to 'deliver') and transition 'again' (to 'review') from stage 'review' can both pass with no person choosing between them; their cel gates could not be compared; make one manual, give it a human-approval gate without when, or make their gates exclusive

Expected

There is no warning when two sibling cel gates compare the same operand with the same constant using complementary operators (< with >=, <= with >, == with !=), or when one expression is exactly !(<other>).

Workaround

None that keeps the same behavior. The max-cycles pair for case 2 counts stage entries, not artifact versions. Case 1 requires a manual transition or a human-approval gate, which changes the factory.

Environment

  • swamp 20261114.210101.0-sha.537e03d1
  • @swamp/stagecraft 2026.11.12.1

Related: #2655 (another false positive in the ambiguous-exit exclusivity check).

Upstream repository: https://github.com/swamp-club/swamp-extensions

Environment

  • Extension: @swamp/stagecraft@2026.11.13.1
  • swamp: 20261114.210101.0-sha.537e03d1
  • OS: linux (aarch64)
  • Deno: 2.9.8
  • Shell: /usr/bin/zsh
02Bog Flow
✓OPEN✓TRIAGED✓IN PROGRESS✓SHIPPED+ 1 MOREASSIGNED+ 2 MOREREVIEW+ 11 MOREPR_MERGED+ 2 MORESESSION_SUMMARIZED

Shipped

11/16/2026, 12:04:43 AM

Click a lifecycle step above to view its details.

03Sludge Pulse
skunk-ape assigned skunk-ape11/15/2026, 5:59:06 PM
Editable. Press Enter to edit.

rcrowley commented 11/15/2026, 6:04:23 AM

Case 3: equality with different constants

Sibling cel gates that compare the same operand with == to different constants are also exclusive, but validate gives a warning for them. This case is not the same as the complementary comparisons in case 2 (< with >=, == with !=).

Reproduction

- id: intake
  transitions:
    - name: from-ticket
      to: plan
      gates:
        - type: cel
          config:
            expr: artifacts["intake"].payload.origin == "ticket"
    - name: from-request
      to: plan
      gates:
        - type: cel
          config:
            expr: >-
              artifacts["intake"].payload.origin == "request" &&
              "request" in artifacts["intake"].payload
    - name: from-patch
      to: develop
      gates:
        - type: cel
          config:
            expr: artifacts["intake"].payload.origin == "patch"

validate logs:

stages.0.transitions.0 (from stage 'intake'): transition 'from-ticket' (to 'plan') and transition 'from-patch' (to 'develop') from stage 'intake' can both pass with no person choosing between them; their cel gates could not be compared; make one manual, give it a human-approval gate without when, or make their gates exclusive
stages.0.transitions.1 (from stage 'intake'): transition 'from-request' (to 'plan') and transition 'from-patch' (to 'develop') from stage 'intake' can both pass with no person choosing between them; their cel gates could not be compared; make one manual, give it a human-approval gate without when, or make their gates exclusive

(from-ticket and from-request go to the same stage, so the check skips that pair.)

Expected

There is no warning when two sibling cel gates each have a term X == c1 and X == c2, with the same operand X and different literal constants. Each term must be the whole expression or a term of its top-level &&.

Workaround

Today the checker compares field values only in evidence-recorded gates, through requireField and match. To remove these warnings, the factory must record origin a second time, as evidence, only for the gates. No gate compares the fields of an artifact.

Sign in to post a ripple.