A case expects only what the specification fixes

CHANGELOG.md

@@ -3,6 +3,22 @@
 What changed in PACT, newest first. Each entry says what an author can now write, or what now
 holds that did not; the commits are listed under it for whoever wants the detail.
 
+## 2026-11 — a case expects only what the specification fixes (12 November)
+
+Bud Agent Flow's second replay: 21 of 24 agreed, and the three that did not were each a question
+the cases had answered by accident.
+
+### What a runtime must match
+
+- **A run that stopped because its answer does not fit hands back an empty record.** PACT handed
+  back the unfitting answer beside the word `stopped`, which a caller could still read fields
+  off. Why it stopped is in `said` (`workflow_runtime.run_workflow`).
+- **A field the run did not set is absent from its answer**, not `null`: `gives reply: text,
+  optional` and a run that never set it answers `{}`.
+- **The order agents are asked in across the items of an `each`** is the runtime's to choose
+  unless the stage says `at-once-at-most: 1`, as the schema's help has always said. The shared
+  case that expected an order now runs on a tree that writes it.
+
 ## 2026-11 — the shared cases, replayed by a second runtime: three things PACT had wrong (12 November)
 
 Bud Agent Flow replayed `tests/conformance/` and sent the cases that differed before recording

adapters/python/src/pact_adapters/workflow_runtime.py

@@ -745,6 +745,11 @@ async def run_workflow(
     if wrong:
         result.halted = "stopped"
         result.said = f"the answer does not fit `answers-with:` — {wrong}"
+        # And it is not handed back. A value that fails the workflow's own
+        # contract, returned beside the word `stopped`, is a value a caller can
+        # still read fields off — fields the workflow said it does not give.
+        # Why it stopped is in `said`; what the stages answered is in `steps`.
+        result.answer = {}
         return result
     result.halted = "final"
     return result

tests/conformance/trees/routed-desk/desk.pact

@@ -69,6 +69,10 @@ workflow clear-inbox
     identified-by: [id]
     limits:
       items-at-most: 50
+      # One at a time, in order. Left out, how many run at once is the
+      # runtime's to choose, and the order agents are asked in with it — which
+      # a case could then not expect.
+      at-once-at-most: 1
       when-it-runs-out: stop-and-say-so
     one = call route with message: item.text, channel: item.channel
   done-all = answer with replies: every.reply

tests/conformance/workflows.json

@@ -1,5 +1,5 @@
 {
- "about": "Workflows: stages in the order they run, what each called agent and each person was given, and the answer. Replayed as conformance_kit.replay_workflow does: `agents` is what each agent answers when called (a list is one answer for each time it is asked, the last repeating); `tools` is what each `<tool>/<action>` hands back; `people` is what a person answers to each question by name (absent: nobody is there to ask, and the run parks). In `expect`: `trace` is the stages that ran, nested ones as `outer/inner`; `given` is each agent called and its inputs, in order; `shown` is each question put to a person and what they were shown; `called` is each tool call. An agent a stage calls is an agent's run, under its own limits, approvals and reactions; here each is answered by the case, so what is compared is the workflow.",
+ "about": "Workflows: stages in the order they run, what each called agent and each person was given, and the answer. Replayed as conformance_kit.replay_workflow does: `agents` is what each agent answers when called (a list is one answer for each time it is asked, the last repeating); `tools` is what each `<tool>/<action>` hands back; `people` is what a person answers to each question by name (absent: nobody is there to ask, and the run parks). In `expect`: `trace` is the stages that ran, nested ones as `outer/inner`; `given` is each agent called and its inputs, in order; `shown` is each question put to a person and what they were shown; `called` is each tool call. An agent a stage calls is an agent's run, under its own limits, approvals and reactions; here each is answered by the case, so what is compared is the workflow. A case expects only what the specification fixes: where the order things happen in is the runtime's to choose (an `each` with no `at-once-at-most:`), the tree a case runs on says `at-once-at-most: 1`. In `answer`, a field the run did not set is absent, not null; a run that stopped hands back an empty record, and why is in `said`.",
  "cases": [
   {
    "name": "a-written-rule-decides-and-nobody-is-asked",
@@ -644,14 +644,7 @@
     ]
    },
    "expect": {
-    "answer": {
-     "ideas": [
-      {
-       "idea": "Lumen",
-       "score": 0.95
-      }
-     ]
-    },
+    "answer": {},
     "called": [],
     "given": [
      {