[dittops]

Follows #18. Bud Agent Flow replayed tests/conformance/ and sent the cases that differed before recording any as its own. Three were PACT's.

What a runtime must match

• A workflow's answer is held to what it says it gives. run_workflow handed back whatever the answer stage bound; it now reads each field of answers-with: by its shape and stops (halted: stopped) when one does not fit.
• Before anything has been remembered, a repeat's remembers: entry with no starts-as: is what its combines-by: makes of nothing: an empty list for a rule that keeps several, absent for one that keeps one. It was null for every rule.
• A workflow that hands back nothing answers with an empty record.
• An agent a workflow's stage calls is an agent's run, under its own limits, approvals and reactions (docs/99 section 5). The reference workflow runtime does not run agents, and now says so.

What was wrong in what PACT ships

• examples/language/help-centre's clear-inbox said it gives text and handed back a list.
• The shared case on examples/kernel/tree-search scripted ideas with no text. Corrected; the wrong data is kept as a case of its own that expects the run to stop.
• The six routing cases ran on a workspace whose desks a reaction covers, so a runtime that refuses an uncarried guard refused the whole workflow. They run on tests/conformance/trees/routed-desk; their expect is unchanged to the letter.

Checked

scripts/test-all.sh exits 0: 4872 tests (1623 Rust + 3249 adapter). The live set (10 cases on gpt-4o-mini) passes on this commit. mkdocs build --strict builds. newest-pydantic-ai fails on master with the same five failures and is not touched here.

Owed

Nothing holds an answer stage's bind: to the workflow's answers-with: at check time.

🤖 Generated with Claude Code