Run output

Author: ryan-williams

out.dvc

@@ -1,5 +1,8 @@
 outs:
-  - path: out
+- md5: 4f98f59e877ecb84ff75ef0fab45bac5
+  size: 3
+  hash: md5
+  path: out
 meta:
   computation:
     cmd: echo "v1" > out

src/dvx/cli/run_cmd.py

@@ -17,9 +17,10 @@
 @click.option("--no-provenance", is_flag=True, help="Don't include provenance in .dvc files.")
 @click.option("-p", "--push", type=click.Choice(["each", "end"]), default=None, help="Push strategy: 'each' (after each commit) or 'end' (once at finish). Also via $DVX_PUSH.")
 @click.option("-P", "--no-cache-push", is_flag=True, help="With --push, only git-push; don't push cache blobs to the remote.")
+@click.option("-D", "--no-pull-deps", is_flag=True, help="Don't auto-fetch materializable trans-deps from remote (default: try fetch before rerunning).")
 @click.option("-U", "--no-prune-fresh", is_flag=True, help="Walk full upstream chain even past fresh artifacts (default: stop at fresh).")
 @click.option("-v", "--verbose", is_flag=True, help="Show detailed output.")
-def run_cmd(targets, force, force_upstream, cached, jobs, commit, dry_run, no_provenance, push, no_cache_push, no_prune_fresh, verbose):
+def run_cmd(targets, force, force_upstream, cached, jobs, commit, dry_run, no_provenance, push, no_cache_push, no_pull_deps, no_prune_fresh, verbose):
     """Execute artifact computations from .dvc files.
 
     Run computations defined in .dvc files, respecting dependencies and
@@ -63,6 +64,7 @@ def run_cmd(targets, force, force_upstream, cached, jobs, commit, dry_run, no_pr
         push=push or "never",
         cache_push=not no_cache_push,
         prune_fresh=not no_prune_fresh,
+        pull_deps=not no_pull_deps,
     )
 
     try:

src/dvx/run/executor.py

@@ -47,6 +47,12 @@ class ExecutionConfig:
     # auto-disabled when --force-upstream patterns are set (we have to walk to
     # find pattern matches).
     prune_fresh: bool = True
+    # Try to materialize "materializable" stages (deps fresh, output missing
+    # locally) from the configured remote before running the cmd. CI runners
+    # on fresh checkouts otherwise rebuild every stage from upstream roots
+    # even when the cache already holds byte-identical outputs.
+    # See specs/done/run-auto-pull.md.
+    pull_deps: bool = True
 
 
 def _matches_patterns(path: str, patterns: list[str]) -> bool:
@@ -258,8 +264,37 @@ def _should_run(self, artifact: Artifact) -> tuple[bool, str]:
         if fresh:
             return False, reason
 
+        # Materializable trans-deps: if the output is just missing locally
+        # but the deps would otherwise let this skip, try fetching the .dvc's
+        # recorded hash from the remote cache. On success, re-evaluate and
+        # short-circuit the rerun. See specs/done/run-auto-pull.md.
+        if self.config.pull_deps and reason == "output missing":
+            if self._try_materialize_from_remote(path):
+                fresh2, reason2 = is_output_fresh(Path(path))
+                if fresh2:
+                    return False, f"fetched ({reason2})"
+
         return True, reason
 
+    def _try_materialize_from_remote(self, path: str) -> bool:
+        """Pull ``path``'s recorded hash from the remote into local cache and
+        materialize the workspace file. Returns True on success.
+
+        Non-fatal: any failure (no remote configured, blob missing remotely,
+        network error) just logs at verbose and returns False so the stage
+        falls through to its normal rerun.
+        """
+        try:
+            from dvx import Repo
+            dvc_path = f"{path}.dvc"
+            with Repo() as repo:
+                repo.pull(targets=[dvc_path])
+        except Exception as e:
+            if self.config.verbose:
+                self._log(f"    ↓ {path}: pull failed ({e})")
+            return False
+        return True
+
     def _execute_level(self, artifacts: list[Artifact]) -> list[ExecutionResult]:
         """Execute all artifacts in a level in parallel.