Load a TaskEnv from a config file with TaskEnv.from_config (v0.1.17)

[Volv-G]

(AI-assisted)

Adds a generic TaskEnv.from_config(path) so an execution environment can be declared in a YAML/JSON file instead of hardcoded in the pipeline module. Every TaskEnv dataclass subclass inherits it, returns its own type, and is validated against its own fields.

ENV = TaskEnv.from_config("tangle/envs.yaml")  # relative to THIS file

TaskEnv.from_config is the only new public name in this PR.

Behaviour

• A relative path resolves against the calling file's directory, never the working directory, so a pipeline script selects the same config wherever the CLI is run from. No __file__ (exec/REPL) is a clear error, never a silent CWD fallback.
• The document is read with ArgsContainer._load_config_file — the loader --config already uses — so the _select directive and the accepted document shapes cannot drift from the CLI. It must resolve to exactly one object.
• Accepted keys come from inspect.signature(cls) rather than dataclasses.fields(), because that is what cls(**case) actually accepts: InitVar pseudo-fields in, ClassVar and field(init=False) out. *args/**kwargs are excluded so a typo stays fail-closed. Unknown keys are rejected with the allowed names, which keeps these files environment-only — pipeline concerns like versioning or schedules have no field to land in.
• A relative dependencies_from in the object anchors to the config file's directory rather than the module's.

Diagnostics never echo a config value

A config file is untrusted input, so nothing from it reaches a compile/CI log.

Keys go through the existing capped, control-character-scrubbing renderer, widened to accept non-string YAML keys and now applied to the _select sibling/unexpected paths that previously used bare repr().

A constructor's own TypeError/ValueError text is never quoted — it may embed a value directly, as a leaf nested inside a structure, or transformed (lower cased, sliced, re-encoded), and no content scan can undo a transformation. The rejected exception is also kept off __cause__ and __context__, because traceback.format_exception prints a chain verbatim; the error is raised after leaving the except handler, since raise ... from None would still leave the object reachable on __context__.

GpuEnv.from_config: /repo/envs.yaml case is not a valid GpuEnv (fields present:
accelerator, image). Its validation message is withheld because it can contain
config values; construct GpuEnv(...) directly to see it.

Calling GpuEnv(...) directly is unaffected and raises the ordinary ValueError in full.

Tests

tests/test_task_env_from_config.py — 29 tests: caller-relative and absolute resolution, config-file-relative dependencies_from, _select with and without a default, subclass field/type inheritance, InitVar/ClassVar/init=False acceptance, environment-only rejection, and an end-to-end compile of a generated task. Disclosure is covered for raw, nested-leaf, transformed, and single-character values, asserted against the rendered traceback rather than just str(exc). Reintroducing the exception chain fails 5 of them.

Full suite: 1540 passed. git diff --check clean.

Version

0.1.16 → 0.1.17 across the four surfaces: pyproject.toml, packages/tangle-cli/src/tangle_cli/__init__.py, tests/test_packaging.py, and the editable self-entry in uv.lock (hand-edited, 1 1 diff; the lockfile was not regenerated).