Relationships
#4887 serve refuses shell steps that splice vault.get into run on protocol 5 workers (shell_data_placement_refused)
Opened by zack · 11/18/2026· Shipped 11/18/2026
Description
During the protocol 5 deprecation window serve still enrolls protocol 5 workers (SUPPORTED_REMOTE_PROTOCOL_VERSIONS = [5,6,7]), and logs the deprecation warning when one connects. But it refuses every command/shell step that splices a vault.get into run and is placed on a protocol 5 worker. The refusal is shell_data_placement_refused, and it happens before dispatch.
The same step succeeds locally and on protocol 6 and 7 workers. A typed (extension) method that takes the same vault.get as a method argument also works on protocol 5, so the problem is specific to shell placement on the legacy protocol 5 frame.
Steps to reproduce
- Start
swamp servewith no auth. - Create a
local_encryptionvaultv1withSEC=s3cr3t-value. - Create a command/shell model
shwhoserunism="${{ vault.get('v1','SEC') }}"; printf %s "$m" | sha256sum. - Create two one-step workflows that pin
shto workers labelledprotocol=p5andprotocol=p7. - Connect a protocol 5 worker (release
20261028.201701.0-sha.c98e3ab2) with--label protocol=p5, and a current worker with--label protocol=p7. - Run each workflow with
--server.
Expected
Both steps run on their worker and receive the exact secret through the environment. This is the intended contract from swamp-club#4065: protocol 5 workers get the plaintext frame during the deprecation window.
Actual
- The p7 workflow succeeds, and its stdout is the exact SHA-256 of the secret.
- The p5 workflow fails in about 31 ms with no dispatch:
Refused to run execute: argument run holds text from data, and its text changed after evaluation, so swamp cannot tell which part is data. Pass the value in the env argument and reference the variable instead of splicing it into run (shell_data_placement_refused) - The p5 worker gets 0 dispatches.
Root cause (from source)
- The protocol-below-6 branch of
DispatchService.#frameFor(src/serve/dispatch_service.ts:801-812) first places the request on its sentinel form, and that passes. - It then calls
request.placeArgumentsagain on the legacyglobalArgs/methodArgs, where the plaintext secret has replaced the sentinel. dataSplicePlacer(src/domain/models/method_execution_service.ts:489-535, check at lines 520-531) compares that text with the unresolved text. The plaintext always differs, somerged[key] !== expected[key]is always true and it throws.
This was introduced by 1a360f11b (swamp-club#3868, data-origin inputs through the environment), layered on 9f89b8acf (swamp-club#4065, the protocol 5 legacy frame).
Fix: build the protocol 5 frame from the placement already computed on the sentinel form, substituting values afterwards, rather than running the data placer again on the resolved text.
Environment
- swamp built from main
8b8e14c6, Linux. - Found by the swamp-uat test
tests/cli/worker/dispatch/mixed_fleet_test.ts, in its two shell-step vault journeys across protocol 5, 6 and 7 workers (swamp-uat#830).
Shipped
Click a lifecycle step above to view its details.
Sign in to post a ripple.