A run installs the repository's dependencies in a scripted step, so agents don't spend their steps hunting for them #109

New issue

Closed

opened 2026-09-29 02:36:06 -04:00 by cmoriarty · 1 comment

cmoriarty commented

2026-09-29 02:36:06 -04:00

Owner

What happens

Agents set up the repository's environment themselves, partway through a step. In the 8 runs whose logs are still on production (runs 17 to 29), about 116 of 1,474 bash calls (8%) were dependency hunting: pip install, npm install, which, listing site-packages.

ModuleNotFoundError or ImportError came back 36 times, in runs 22, 23, 25 and 26.
Run 22 (cmoriarty/scratch, the walking skeleton) couldn't find fastapi and went looking in osfd's own virtualenv at /opt/openspecflow/app/.venv.
Run 26 made a .venv in the worktree during openspec.apply[1] to run its tests, which is how #105 happened.
soundcheck's node_modules was installed during openspec.apply (#67's step-by-step measurement).

Expected

A scripted step installs the repository's dependencies from its manifests, including manifests in subdirectories such as backend/ and frontend/:
- uv.lock or pyproject.toml → uv sync
- requirements.txt → a .venv and pip install -r
- package-lock.json → npm ci
- and so on for the stacks #107 recognises
It runs before openspec.apply. It runs again before the test steps if the run added or changed a manifest, because a new repository has none until apply writes them.
The step shows what it ran and how long it took. A failed install fails the step with the installer's output, rather than an agent finding out later.
The AGENTS.md block tells agents the environment is ready and where it is, so they use it rather than making their own.

It should use #67's shared npm and uv caches once they land. #105 already keeps these environments out of commits. It's also a precondition for #104's scripted test steps, and for LSP diagnostics if runs ever turn them on (#46).

Adding nodes to the built-in pipelines runs into #102 for runs already in flight, so deploy when idle.

Related: #46, #67, #104, #105, #107.

**What happens** Agents set up the repository's environment themselves, partway through a step. In the 8 runs whose logs are still on production (runs 17 to 29), about 116 of 1,474 bash calls (8%) were dependency hunting: `pip install`, `npm install`, `which`, listing `site-packages`. - `ModuleNotFoundError` or `ImportError` came back 36 times, in runs 22, 23, 25 and 26. - Run 22 (cmoriarty/scratch, the walking skeleton) couldn't find `fastapi` and went looking in osfd's own virtualenv at `/opt/openspecflow/app/.venv`. - Run 26 made a `.venv` in the worktree during `openspec.apply[1]` to run its tests, which is how #105 happened. - soundcheck's `node_modules` was installed during `openspec.apply` (#67's step-by-step measurement). **Expected** - A scripted step installs the repository's dependencies from its manifests, including manifests in subdirectories such as `backend/` and `frontend/`: - `uv.lock` or `pyproject.toml` → `uv sync` - `requirements.txt` → a `.venv` and `pip install -r` - `package-lock.json` → `npm ci` - and so on for the stacks #107 recognises - It runs before `openspec.apply`. It runs again before the test steps if the run added or changed a manifest, because a new repository has none until `apply` writes them. - The step shows what it ran and how long it took. A failed install fails the step with the installer's output, rather than an agent finding out later. - The `AGENTS.md` block tells agents the environment is ready and where it is, so they use it rather than making their own. It should use #67's shared npm and uv caches once they land. #105 already keeps these environments out of commits. It's also a precondition for #104's scripted test steps, and for LSP diagnostics if runs ever turn them on (#46). Adding nodes to the built-in pipelines runs into #102 for runs already in flight, so deploy when idle. Related: #46, #67, #104, #105, #107.

cmoriarty referenced this issue

2026-09-29 02:36:21 -04:00

Agents serve the app through one Braid tool that starts it on the run's port, waits until it answers, shows its log and stops it #110