One-way task display synchronization
The source pushes new or changed selected task groups into a dedicated remote display workspace. It does not compare remote state, pull changes, propagate source deletions or remove directories dropped by a selection rule. Matching destination directories are replaced completely; unlisted remote directories remain untouched.
Configuration
extends: default
targets:
- address: http://display:1024
token: ${DISPLAY_TOKEN}
components:
sync:
default:
backend: local
target: http://display:1024
model_names: ["experiment-*"]
rules:
research: terminal
online: live
compare: terminal
max_archives_per_flush: 4
timeout_seconds: 300
jobs:
publish_display:
backend: scheduled
expression: "* * * * *"
enable_serve: false
steps:
- backend: sync_flush_step
sync: defaultThe internal Job runs once per minute in the App timezone. It reuses the pipeline Steps, argument validation and cancellation lifecycle. Scheduled invocations are sequential and missed times are not replayed. Set enable_serve: true to also allow manual HTTP/MCP invocation. Stopping the service cancels its scheduled Jobs; research Task processes retain their existing daemon ownership.
Source selection and change detection
rules maps workflow registrations to terminal or live. Terminal publication waits for the parent and selected descendants to finish; live publication captures current evidence. model_names uses shell-style patterns, with an empty list including all models. Empty rules select nothing.
Each invocation scans source metadata. Version detection uses execution identity, file inventory, sizes, nanosecond mtimes and publication policy without reading payloads for content hashing. Successful versions are retained in sender memory. Unchanged groups are skipped, failures are retried on the next invocation, and source service restart republishes selected groups once. Remote deletion or modification does not trigger retransmission. Changing file contents while preserving size and mtime is not detected.
Model selection reuses model task_name, logical child node_name and progress name/group. Selected views retain available matching child executions and shared descendants. Parent parameters, results and progress are projected to the selected models; mixed parent logs and aggregate artifacts are omitted. Replaced child executions are excluded rather than associated with old parent calls. Sources remain unchanged.
Transfer and display
The sender creates complete bounded display directories and submits them through POST /upload?mode=replace. A workflow group and its selected descendants are packed into as many archives as needed. Descendants are sent before the root. max_archives_per_flush bounds each invocation; subsequent invocations resume partially uploaded groups and continue through other roots. Successful Task versions are retained until the group completes, so a retry does not resend completed batches. Each supplied directory is replaced atomically; a multi-archive group is not an atomic transaction. The receiver only validates and replaces files. Its existing task_index discovers supported metadata at the next scan, every five seconds by default. No receiver-side synchronization component, receipt, dedicated Job or task metadata parser is required for file transfer.
transfer_budget_bytes defaults to 512 MiB. Each Task independently receives this display budget, including reserved archive overhead; it is never divided by the number of Tasks. The same limit separately bounds both upload bytes and expanded size for each archive. Configure the source budget no larger than the receiver's. Oversized date-keyed Parquet artifacts can be sampled to a shared suffix of complete trading days; other files can be omitted explicitly. Failures and omissions are reported. Display metadata retains source identity, source/policy revisions, file coverage and the selected workflow snapshot. Display replicas cannot execute, become research inputs or be forwarded as source tasks.
Source staging under tmp/sync/ is removed after each group. Receiver staging under tmp/files/ is removed after success or a completed rollback; failed rollbacks retain original directories for recovery. The operation is file replacement, not a backup or process/environment migration. Use separate display workspaces because supplied paths overwrite matching directories.