Relationships
#4682 Workflow guard does not see data written by an earlier step in the same run (data.latest evaluates empty)
Opened by owonwo · 11/12/2026· Shipped 11/12/2026
Workflow guard does not see data written by an earlier step in the same run
Summary
Per the docs (references/workflow/reference.md, "When Expressions Resolve"), a
guard resolves at step time and sees "Data written by earlier steps in this
run". Empirically, in a scan step -> guarded notify step workflow, the guard
evaluates false even though the scan step just wrote the resource and a plain
data query finds it with isLatest: true.
Minimal repro (same data, same engine, minutes apart):
- Unguarded step whose input reads
size(data.latest("inspace-doctor", "doctor").?attributes.?needsHuman.orValue([]))→ evaluated38. Notification delivered. - Guarded step with guard
size(data.latest("inspace-doctor", "doctor").?attributes.?needsHuman.orValue([])) > 0→ skipped: "guard evaluated false". No scan step in that probe workflow, no other data writer in between.
Same expression, same data stream (38 needsHuman items, versions=11 — also
confirmed via data.listVersions in an unguarded step). The only difference is
guard vs step-input evaluation.
Real-world impact: the pattern step A writes resource -> step B guarded on size(data.latest("A", "res").attributes.x) > 0`` never fires for findings
produced by the current sweep. In production this caused a health-check
workflow to silently never alert: with a sliding 90-minute lookback window, a
failure found by sweep N is outside sweep N+1's window, and sweep N's notify
read sweep N-1's (clean) data — so transient failures could never ship, for
days, while every run reported SUCCEEDED.
Steps to reproduce
- Behavior method writes a resource (
writeResource("doctor", "doctor", {...needsHuman: [..38 strings..]})). - Workflow: step 1 runs the method; step 2 (any task type) with
guard: ${{ size(data.latest("<behavior-name>", "doctor").attributes.needsHuman) > 0 }}, dependsOn step 1. - Run the workflow while the resource's latest version is non-empty.
- Step 2 is skipped with "guard evaluated false"; an unguarded step reading the same expression in a separate workflow sees the data fine.
Expected
Guard expressions at step time see data written by earlier steps in the same run (or at minimum see the pre-run latest version), per the documented resolution table.
Workaround
Move the notification into the behavior method itself (it already has the findings in hand), instead of a guarded follow-up step.
Environment
- swamp 20261111.234122.0-sha.cd35f58c, macOS (darwin-aarch64), local repo (no serve)
- Engine notes that may matter: the step's data write appears to commit asynchronously ~70-400ms after step completion status flips; the guard evaluated ~74ms after the scan step reported success.
Shipped
Click a lifecycle step above to view its details.
zack commented 11/12/2026, 7:57:25 PM
Thanks @owonwo for reporting this! We shipped: Show what a guard returned on the log-mode skipped-step line, so a guarded skip reads as a truthy guard instead of looking like a guard that failed to see data. Presentation-only: guard semantics, JSON output and the run record are unchanged.. The fix has been merged and a release is on its way. We appreciate your contribution to swamp.
zack commented 11/12/2026, 8:11:28 PM
The guard actually evaluated true here, and things are working as expected. I added some more to the log output so that this is more clear to humans and machines
Sign in to post a ripple.