talont-org/setup-environment

Environment Setup

Pre-flight environment gate for the research recipe. Decouples environment decisions (Docker vs host-micromamba vs none) from implement-experiment's inline Docker build. Reads the experiment plan to determine the required environment type, probes Docker availability, builds a Docker image when possible, falls back to a host micromamba environment for pure conda-forge Python deps, and emits an env_mode verdict consumed by downstream steps. PASS and WARN proceed to decompose_phases; FAIL escalates immediately.

Arguments

/autoskillit:setup-environment <worktree_path> <experiment_plan>

• worktree_path — Absolute path to the shared research worktree (positional).
• experiment_plan — Absolute path to the experiment plan YAML (positional).

When to Use

• Called by the research recipe's setup_environment step between stage_data and decompose_phases
• Whenever a pre-flight environment check is needed before experiment implementation

Critical Constraints

NEVER:

• Modify the experiment plan
• Write files outside {{AUTOSKILLIT_TEMP}}/setup-environment/
• Skip the Docker availability probe before attempting a build
• Run subagents in the background (run_in_background: true is prohibited)

ALWAYS:

• Parse environment.type from the experiment plan first
• Try Docker before falling back to micromamba
• Write the environment setup report before emitting the verdict token
• Use model: "sonnet" for all subagents

Workflow

Step 0 — Parse the Experiment Plan

Read the experiment plan at experiment_plan. Extract:

• environment.type — "standard" or "custom"
• The name: field from the environment.yml embedded section (the conda environment slug, used as the Docker image tag and micromamba env name)

If environment.type is absent or not one of "standard" / "custom", or if the name: field is missing from the environment.yml section, set env_mode = unavailable, verdict = FAIL, and proceed directly to Step 7 (write report) then Step 8 (emit tokens). Do not proceed to Steps 2–6 with an undefined slug.

Step 1 — Short-Circuit for Standard Environments

If environment.type == "standard", no custom environment is needed:

• Set env_mode = none, verdict = PASS
• Skip Steps 2–6 and proceed directly to Step 7 (write report) then Step 8 (emit tokens)

Step 2 — Probe Docker Availability

Run:

timeout 5 docker info --format '{{.OSType}}' 2>/dev/null

If the command exits 0, Docker is available.

Step 3 — Build Docker Image (Docker Available)

Before staging, verify environment.yml exists in the worktree root. If the file is absent, set env_mode = unavailable, verdict = FAIL, and proceed directly to Step 7.

Stage the build context under {{AUTOSKILLIT_TEMP}}/setup-environment/build/:

mkdir -p {{AUTOSKILLIT_TEMP}}/setup-environment/build
cp src/autoskillit/assets/research/Dockerfile.template \
   {{AUTOSKILLIT_TEMP}}/setup-environment/build/Dockerfile
cp environment.yml \
   {{AUTOSKILLIT_TEMP}}/setup-environment/build/{slug}.yaml

The template is at src/autoskillit/assets/research/Dockerfile.template and uses base image mambaorg/micromamba:1.0-bullseye-slim. Build:

docker build \
  --build-arg MAMBA_ENV={slug} \
  -t research-{slug} \
  {{AUTOSKILLIT_TEMP}}/setup-environment/build

Step 4 — Docker Outcome Routing

• Build success: Set env_mode = docker, verdict = PASS. Proceed to Step 7 (write report).
• Build failure OR Docker unavailable: Continue to Step 5 (fallback viability assessment).

Step 5 — Assess Micromamba Fallback Viability

Parse environment.yml for:

Channels: Only conda-forge and defaults are considered viable. Any other channel (e.g., bioconda, pytorch) marks the environment as not viable.

Dependency markers (not viable):

• CUDA deps: cudatoolkit, cuda-*, nvidia-*
• Bioconductor deps: bioconductor-*
• System packages: entries that resolve to OS-level packages (detected via dpkg -l <package> 2>/dev/null returning exit 0, or which <package> returning a hit in /usr/bin, /usr/sbin, /bin, /sbin)

If channels or deps are not viable, proceed to Step 7 with env_mode = unavailable, verdict = FAIL.

Step 6 — Create Host Micromamba Environment

If viable, run:

micromamba create -f environment.yml -n experiment-{slug}

• Success: Set env_mode = micromamba-host, verdict = WARN
• Failure: Set env_mode = unavailable, verdict = FAIL

Step 7 — Write Environment Setup Report

Always write the report before emitting tokens, regardless of verdict:

{{AUTOSKILLIT_TEMP}}/setup-environment/env_setup_report_{YYYY-MM-DD_HHMMSS}.md

Report structure:

## Environment Setup Report
**Date:** {timestamp}
**Environment Type:** {standard|custom}
**Verdict:** {PASS|WARN|FAIL}
**env_mode:** {none|docker|micromamba-host|unavailable}

### Docker Probe Result
{available|unavailable} — {reason}

### Build Result (if attempted)
{success|failure} — {output or error}

### Micromamba Viability Assessment (if evaluated)
- Channels: {list}
- Problematic deps: {list or "none"}
- Viability: {viable|not viable} — {reason}

### Micromamba Install Result (if attempted)
{success|failure} — {output or error}

### Rationale
{explanation of final verdict}

Step 8 — Emit Structured Output Tokens

Emit structured output tokens as LITERAL PLAIN TEXT with NO markdown formatting on the token names. Do not wrap token names in **bold**, *italic*, or any other markdown. The adjudicator performs a regex match on the exact token name — decorators cause match failure.

env_mode = {none|docker|micromamba-host|unavailable}
env_report = /absolute/path/to/env_setup_report_{YYYY-MM-DD_HHMMSS}.md
verdict = {PASS|WARN|FAIL}

Output

env_mode = none|docker|micromamba-host|unavailable
env_report = /absolute/path/to/{{AUTOSKILLIT_TEMP}}/setup-environment/env_setup_report_{YYYY-MM-DD_HHMMSS}.md
verdict = PASS|WARN|FAIL