Terragrunt, the Go tool that wraps OpenTofu and Terraform for layered infrastructure layouts, saw 43 commits land on main in the last week. The diffstat reads 1244 files changed with 50602 insertions, and most of that volume sits in docs and generated reference pages. The interesting parts for operators are a fix for dependencies whose config_path points at a config file instead of a directory, a series of parse performance changes, and an early outputs block experiment.
Discovery now handles file valued config paths ¶
The most operationally relevant commit of the window is the fix for reading a dependency whose config_path names a config file. Until the dependency-config-path-file strict control becomes the default, a dependency block can set config_path to an HCL file rather than a directory. Discovery treated such a unit’s path as the file itself, which broke relative path resolution for anything that dependency referenced.
The fix adds a configDir field on internal/component.Unit, with ConfigDir() and SetConfigDir() accessors. During the parse phase in internal/discovery/phase_parse.go, when a component path is a file, the code records filepath.Dir(componentPath) as the config dir and uses the file base name as the config filename. Path extraction in internal/discovery/helpers.go then resolves relative dependency paths against that directory instead of the file path. The commit includes a new test, relationship_config_file_test.go, that builds a fixture where two apps depend on ../../shared/another-name.hcl and asserts the discovered relationship is correct.
The Go doc comments on ConfigDir() are explicit that both accessors are temporary scaffolding. Once the file valued config_path behavior is removed, every unit path is a directory and this goes away. If you maintain tooling that walks Terragrunt component graphs, note the deprecation path before building on these methods.
Related discovery fixes in the same week: rejecting a null or missing exclude if during discovery, and returning matches for non canonical fileset() patterns. The exclude fix closes a footgun where a malformed exclude block silently mis behaved during graph construction.
Parse store work targets slower config loads ¶
Four commits this week attack the cost of reading Terragrunt configs. The parse store, which caches parsed HCL to avoid repeated work across a run, is now threaded through more code paths: config rewrite parse store, plumbing the store down, reading run configs through it, and a partial parse consistency fix that keeps cached results coherent.
None of these commits publish benchmark numbers, so treat the win as unproven until you measure it on your own stack. Large monorepos with hundreds of units, or stacks that reuse OCI and HTTP archive sources, are the shapes most likely to benefit.
Two micro optimizations also landed. Allocating the Terragrunt function table at its final size avoids repeated map growth when registering builtin HCL functions. Removing marks while converting cty values to Go maps strips ctype marks earlier in the conversion, which saves allocation on every dependency output decode. Both are cheap and safe; the winners are anyone running terragrunt run over wide graphs, the losers are nobody in particular.
The outputs block experiment ¶
The outputs block on units behind the outputs-block experiment is the only new feature of the week. It lets a unit declare its outputs in HCL instead of relying on dependency blocks to fish values out of remote state on every read. That is how Terragrunt’s own docs describe the direction: explicit, cheap to read, and independent of state backend latency.
Caveats worth naming. It is gated behind an experiment flag, so do not build production pipelines on it yet. The project also added missing discussion threads for experiments and set up a weekly RFC feedback request, which signals the maintainers want input before stabilization. The docs work this week touched the HCL blocks reference page eight times, mostly documenting these new surfaces.
Strict controls and stack fixes ¶
A new strict control cas-source-inside-stack landed, part of the ongoing migration of behavioral defaults behind opt in strictness. Separately, a fix for rewriting update_source_with_cas in OCI and HTTP archive stacks matters if you vendored stack sources from archives; the rewrite step previously mishandled those layouts. Plumbing --filter through render is a small but useful CLI improvement: filtered rendering means you can inspect a subset of a large stack without paying to render everything.
Housekeeping also landed: the toolchain moved to Go 1.27.4, an in memory venv abstraction (pkg/venv with venv.MemVenv) replaced the older test helper for faster tests, and a vfs overlay filesystem was added for test fixtures. Those changes explain a good chunk of the 50602 insertions and are purely internal.
What to watch ¶
- Track the
dependency-config-path-filestrict control. When it flips to default,ConfigDir()disappears and any external tooling that assumed unit paths are directories of HCL files must be updated. - Watch the
outputs-blockexperiment discussions. If it stabilizes, dependency heavy stacks can drop a large class of remote state reads. - Measure the parse store changes on your own repos before crediting them with anything. No benchmarks were published this week.
All of this is on the terragrunt repository, default branch main.