Imported from beril-doe/BERIL-research-observatory (.claude/skills/remote-compute/SKILL.md). Install upstream with npx skills add beril-doe/BERIL-research-observatory --skill remote-compute. Copyright stays with the author.

Remote Compute Skill

Run containerized jobs on KBase compute nodes via the CDM Task Service (CTS).

Overview

CTS runs containerized jobs on remote compute clusters. Two execution paths are available:

Always discover the live image registry first before deciding which path to take:

# JupyterHub
tscli = get_task_service_client()
tscli.get_images()        # returns list of registered images

# curl (any environment)
AUTH_TOKEN=$(grep "KBASE_AUTH_TOKEN" .env | cut -d'"' -f2)
curl -s -H "Authorization: Bearer $AUTH_TOKEN" \
  https://berdl.kbase.us/apis/cts/images | python3 -m json.tool

The registry grows over time — always check live rather than relying on any static list in this document.

When to use remote compute vs JupyterHub:

• JupyterHub: Interactive analysis, Spark SQL queries, exploratory work
• Remote compute: Batch processing, CPU/memory-intensive tools (CheckM2, Bakta, GTDB-Tk), long-running jobs, anything that would block your notebook

Preconditions

1. KBASE_AUTH_TOKEN set in environment or .env
2. MinIO client (mc) installed — see berdl-minio skill for installation
3. Proxy running (when off-cluster) — MinIO is not directly reachable from external networks. See the berdl-query skill's references/proxy-setup.md. The CTS API itself (https://berdl.kbase.us/apis/cts) does NOT require a proxy — only MinIO data staging does.
4. CTS access and your S3 paths — run whoami before staging any files. It confirms your role and tells you the exact S3 path prefixes you are allowed to use for input_files and output_dir. Every example in this skill uses cts/io/<username>/... as a placeholder — whoami tells you what to substitute.

# JupyterHub
tscli = get_task_service_client()
tscli.whoami()
# Returns, e.g.:
# {
#   "user": "mcashman",                       ← your username; use this in place of <username>
#   "roles": ["cts_user"],                    ← confirms CTS access; must include "cts_user" or "full_admin"
#   "allowed_paths": [
#     {"path": "cts/io/mcashman", "perm": "write"},   ← you can read and write here
#     {"path": "cts/io/mcashman", "perm": "read"}
#   ]
# }
# The "path" values are the S3 prefixes you use for input_files and output_dir.
# Example: input_files=["cts/io/mcashman/input/genome.fna.gz"]
#          output_dir="cts/io/mcashman/output/bakta_run"

# curl equivalent
curl -s -H "Authorization: Bearer $AUTH_TOKEN" \
  https://berdl.kbase.us/apis/cts/whoami | python3 -m json.tool

If roles is empty or does not include cts_user or full_admin, you do not have CTS access — ask a CTS admin to grant the cts_user role.

Two Interfaces

Python client — JupyterHub only

tscli = get_task_service_client()    # CTS client
minio = get_minio_client()           # MinIO client for file staging

REST API (curl) — any environment

All endpoints at https://berdl.kbase.us/apis/cts. Auth via Bearer token:

AUTH_TOKEN=$(grep "KBASE_AUTH_TOKEN" .env | cut -d'"' -f2)

Stage Input Files to MinIO

This step is required for both Path A and Path B. CTS does not read files from your local machine or from the BERDL Lakehouse directly — all input files must be pre-uploaded to MinIO before job submission. Every file must carry a CRC64NVME checksum or the submission will be rejected.

Find your MinIO path prefix

Run tscli.whoami() (see Preconditions) to find your allowed_paths. Your input and output directories should be nested under the writable path — typically cts/io/<username>/.

Upload with the MinIO CLI (mc)

# Off-cluster only: set proxy so mc can reach MinIO
export https_proxy=http://127.0.0.1:8123
export no_proxy=localhost,127.0.0.1

# Upload a single file with a CRC64NVME checksum
mc cp --checksum crc64nvme genome.fna.gz berdl-minio/cts/io/<username>/input/

# Upload a directory of files (e.g. all .fna.gz in a folder)
mc cp --checksum crc64nvme --recursive ./input_genomes/ berdl-minio/cts/io/<username>/input/

# Verify what was staged
mc ls berdl-minio/cts/io/<username>/input/

The berdl-minio alias must be configured. See the berdl-minio skill for setup instructions.

Upload from Python (JupyterHub)

minio = get_minio_client()
import os

local_file = "genome.fna.gz"
minio_path = f"io/<username>/input/{local_file}"  # path within the "cts" bucket

# Upload a file — the MinIO client handles checksums automatically on JupyterHub
minio.fput_object("cts", minio_path, local_file)
print(f"Staged: cts/{minio_path}")

# List staged files to confirm
for obj in minio.list_objects("cts", prefix="io/<username>/input/", recursive=True):
    print(obj.object_name, obj.size)

Build your input_files list

After staging, construct the input_files list for submit_job from the staged paths:

# Option 1: list them explicitly
input_files = [
    "cts/io/<username>/input/genome1.fna.gz",
    "cts/io/<username>/input/genome2.fna.gz",
]

# Option 2: discover from MinIO (useful for large batches)
minio = get_minio_client()
objs = list(minio.list_objects("cts", prefix="io/<username>/input/", recursive=True))
input_files = [f"cts/{o.object_name}" for o in objs]
print(f"{len(input_files)} files staged and ready")

Path A: Pre-built Tool Images

Pre-built images have the tool already installed and set as the container entrypoint. You do not write a bash script — you pass arguments directly to the tool. Some images include large reference databases (refdata) that CTS mounts automatically.

What is refdata?

Some tools require large reference databases (e.g., CheckM2's marker gene sets, GTDB-Tk's genome sketches, Bakta's annotation database). These are pre-staged by the CTS admins and mounted read-only into the container at a fixed path. The image registration records which refdata bundle it needs and where to mount it (default_refdata_mount_point).

You do not pass a refdata_id to submit_job — it is determined by the image. You only need to know the mount path so you can reference it in your args if needed (e.g., --db /ref_data). Most images bake the path into the tool config or environment variables so you do not need to reference it at all — see the per-tool notes below.

If you want to override the default mount point, pass refdata_mount_point="/your/path" inside the params via the Python client's submit_job call. This is rarely needed.

General pattern — pre-built image

tscli = get_task_service_client()

job = tscli.submit_job(
    "<image>:<tag>",          # full image name from the registry
    input_files,              # list of MinIO paths: ["cts/io/<user>/input/file1.fna.gz", ...]
    "cts/io/<user>/output/run_name",  # MinIO output directory (bucket/prefix)
    cluster="kbase",          # compute site — "kbase" for on-prem KBase hardware
    output_mount_point="/out", # path inside the container where the tool must write output;
                               # CTS copies everything under this path to S3 after the job finishes
    args=[...],               # arguments appended after the image entrypoint command
    num_containers=1,         # how many parallel containers to run (see per-tool guidance)
    cpus=4,                   # CPU cores per container
    memory="8GB",             # RAM per container
    runtime="PT30M",          # ISO 8601 duration: PT5M=5min, PT1H=1hour, PT2H30M=2.5hours
)

print(f"Job ID: {job.id}")    # save this — needed to check status or retrieve results later

Available pre-built images

Full per-tool examples, resource guidance, and special notes are in individual reference files under references/. Each tool has its own file named after the image pattern without the cdm_ prefix:

Claude should Read only the specific tool's reference file after identifying the user's tool from get_images().

Path B: Generic Ubuntu Script

Use this path when your tool is not in the pre-built registry, or when you need custom logic that doesn't fit a single tool invocation. You write a bash script, upload it alongside your data, and the container executes it.

The generic image (ghcr.io/kbasetest/cts_ubuntu_test:0.1.0) is a minimal Ubuntu — Python, pip, and other tools are NOT pre-installed. Your script installs what it needs at runtime.

Step 1: Write your script

#!/bin/bash
# my-analysis.sh — runs on the compute node

# Install dependencies at runtime (nothing is pre-installed)
apt-get update -qq && apt-get install -y -qq python3 python3-pip > /dev/null 2>&1
pip install biopython pandas --quiet

# The script receives arguments from the CTS job submission args list.
# A common convention: first arg is the output directory, remaining args are input files.
OUTPUT_DIR="$1"
shift   # remove $1 so "$@" now contains only the input files

# Process each input file
for FILE in "$@"; do
    BASENAME=$(basename "$FILE")
    python3 -c "
import Bio.SeqIO
records = list(Bio.SeqIO.parse('$FILE', 'fasta'))
print(f'{len(records)} sequences in $BASENAME')
" > "${OUTPUT_DIR}/${BASENAME}.result.txt"
done

Step 2: Upload script and data to MinIO

Stage your input data using the instructions in Stage Input Files to MinIO above. Your bash script must also be uploaded — it is treated as an input file by CTS and staged alongside your data.

# Upload the script (same checksum requirement as data files)
mc cp --checksum crc64nvme my-analysis.sh berdl-minio/cts/io/<username>/scripts/

# Upload input data
mc cp --checksum crc64nvme input_data/*.fna.gz berdl-minio/cts/io/<username>/input/

Step 3: Submit the job

Python client (JupyterHub)

tscli = get_task_service_client()

job = tscli.submit_job(
    "ghcr.io/kbasetest/cts_ubuntu_test:0.1.0",  # generic ubuntu image
    [
        "cts/io/<username>/scripts/my-analysis.sh",  # the script is an input file
        "cts/io/<username>/input/genome1.fna.gz",
        "cts/io/<username>/input/genome2.fna.gz",
    ],
    "cts/io/<username>/output/run_name",
    cluster="kbase",
    input_mount_point="/in",              # all input files (script + data) are staged here
    output_mount_point="/out",            # script must write all output under this path
    args=[
        "/in/my-analysis.sh",             # the script to execute; CTS stages it at this path
        "/out/",                          # first argument to the script: the output directory
        tscli.insert_files(),             # remaining input files (excluding the script) injected here
    ],
    num_containers=1,
    cpus=4,
    memory="8GB",
    runtime="PT30M",
)

print(f"Job ID: {job.id}")

REST API (curl)

curl -s -X POST \
  -H "Authorization: Bearer $AUTH_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "cluster": "kbase",
    "image": "ghcr.io/kbasetest/cts_ubuntu_test:0.1.0",
    "params": {
      "args": [
        "/in/my-analysis.sh",
        "/out/",
        {"type": "input_files", "input_files_format": "space_separated_list"}
      ],
      "input_mount_point": "/in",
      "output_mount_point": "/out"
    },
    "input_files": [
      "cts/io/<username>/scripts/my-analysis.sh",
      "cts/io/<username>/input/genome1.fna.gz",
      "cts/io/<username>/input/genome2.fna.gz"
    ],
    "output_dir": "cts/io/<username>/output/run_name",
    "num_containers": 1,
    "cpus": 4,
    "memory": "8GB",
    "runtime": "PT30M"
  }' \
  https://berdl.kbase.us/apis/cts/jobs

Job Management and Results

Step 1: Save the job ID immediately after submission

The job ID is the only handle you have on a submitted job. Save it before doing anything else.

job = tscli.submit_job(...)

print(f"Job ID: {job.id}")

# Also write it to a file so you can recover it if the notebook closes
with open("projects/<id>/data/cts_job_id.txt", "a") as f:
    f.write(f"{job.id}\n")

Step 2: Check job status

# Lightweight check — returns the current state string without loading full job details
job.get_job_status()
# Returns one of:
#   created, download_submitted, job_submitting, job_submitted,
#   upload_submitting, upload_submitted, complete,
#   recovering, canceling, canceled,
#   error_processing_submitting, error_processing_submitted, error

# curl equivalent
curl -s -H "Authorization: Bearer $AUTH_TOKEN" \
  https://berdl.kbase.us/apis/cts/jobs/{job_id}/status

To block until the job finishes (useful in a notebook thread):

import threading, json

# Run in a background thread so the notebook stays responsive
thread = threading.Thread(
    target=lambda: print(json.dumps(job.wait_for_completion(), indent=4)),
    daemon=True
)
thread.start()
# job.wait_for_completion() returns the full job dict when the job reaches a terminal state

Spark auto-import — some images (currently confirmed: CheckM2) have a registered event importer on the CTS side. When the job finishes, this pipeline automatically reads the output files from MinIO and loads them into a BERDL Iceberg table, tagged by cts_job_id. You can then query results directly with Spark instead of downloading files manually.

Pass wait_for_event_importer=True to block until both the job and the import pipeline are done:

thread = threading.Thread(
    target=lambda: print(json.dumps(
        job.wait_for_completion(wait_for_event_importer=True), indent=4
    )),
    daemon=True
)
thread.start()
# After this completes, results are in the BERDL table — no download needed:
# spark.sql(f"SELECT * FROM u_<username>__autoimport.checkm2 WHERE cts_job_id = '{job.id}'")

If the image does not have an importer, omit wait_for_event_importer=True and retrieve results from MinIO manually (see Step 4 below).

Step 3: Check exit codes on failure

When get_job_status() returns "error", check exit codes first — it tells you which containers failed without pulling full logs.

job.get_exit_codes()
# Returns e.g. {'exit_codes': [0, 2]}
# One entry per container: 0=success, non-zero=failure
# Use container index to target log reads below

curl -s -H "Authorization: Bearer $AUTH_TOKEN" \
  https://berdl.kbase.us/apis/cts/jobs/{job_id}/exit_codes

Step 4: Read logs (especially on failure)

# View stderr — most useful for diagnosing tool failures
job.print_logs(container_num=0, stderr=True)

# View stdout — tool progress output
job.print_logs(container_num=0, stderr=False)

# For multi-container jobs, increment container_num: 0, 1, 2, ...
job.print_logs(container_num=1, stderr=True)

# curl equivalents
curl -s -H "Authorization: Bearer $AUTH_TOKEN" \
  https://berdl.kbase.us/apis/cts/jobs/{job_id}/log/0/stderr

curl -s -H "Authorization: Bearer $AUTH_TOKEN" \
  https://berdl.kbase.us/apis/cts/jobs/{job_id}/log/0/stdout

Step 5: Find output files after completion

The output location is the output_dir you passed to submit_job. After the job completes, the full list of output files (with checksums) is available from the job object:

job_details = job.get_job()

# List every output file and its checksum
for output in job_details["outputs"]:
    print(output["file"])        # full S3 path, e.g. cts/io/<username>/output/run/quality_report.tsv
    print(output["crc64nvme"])   # CRC64NVME checksum for integrity verification

# Resource usage stats (available after completion)
print(f"CPU hours used:   {job_details.get('cpu_hours')}")
print(f"Max memory used:  {job_details.get('max_memory')} bytes")
print(f"Output file count: {job_details.get('output_file_count')}")

Download all outputs to a local directory:

# The output_dir you specified in submit_job is the MinIO path to download from
mc cp --recursive berdl-minio/cts/io/<username>/output/run_name/ ./results/

# Python (JupyterHub) — list and read outputs via MinIO client
minio = get_minio_client()

# List all objects under the output directory
objects = list(minio.list_objects("cts", prefix="io/<username>/output/run_name/", recursive=True))
for obj in objects:
    print(obj.object_name)

# Read a specific result file into memory without downloading
response = minio.get_object("cts", "io/<username>/output/run_name/quality_report.tsv")
print(response.read().decode("utf-8"))
response.close()
response.release_conn()

Recovering a lost job ID

If your notebook closed or you lost the job ID, use list_jobs() to find it. Results are sorted newest-first. Maximum 1000 jobs returned.

tscli = get_task_service_client()

# List your 20 most recent jobs
result = tscli.list_jobs(limit=20)
for j in result["jobs"]:
    image = f"{j['image']['name']}:{j['image']['tag']}"
    submitted = j["transition_times"][0]["time"]   # time job was created
    print(f"{j['id']}  {j['state']:<30}  {image:<55}  {submitted}")

Filter by state to find jobs that are still running or recently failed:

# Only running jobs
result = tscli.list_jobs(state="job_submitted")

# Only completed jobs
result = tscli.list_jobs(state="complete")

# Only failed jobs
result = tscli.list_jobs(state="error")

# Jobs updated after a specific date (ISO 8601)
result = tscli.list_jobs(after="2026-06-01T00:00:00Z")

# Jobs updated before a specific date
result = tscli.list_jobs(before="2026-05-01T00:00:00Z")

# Time window: completed jobs between two dates
result = tscli.list_jobs(state="complete", after="2026-05-01T00:00:00Z", before="2026-06-01T00:00:00Z")

# Combine filters: completed jobs on the kbase cluster since June 2026
result = tscli.list_jobs(cluster="kbase", state="complete", after="2026-06-01T00:00:00Z")

# curl equivalents
curl -s -H "Authorization: Bearer $AUTH_TOKEN" \
  "https://berdl.kbase.us/apis/cts/jobs?limit=20"

curl -s -H "Authorization: Bearer $AUTH_TOKEN" \
  "https://berdl.kbase.us/apis/cts/jobs?state=error&limit=20"

curl -s -H "Authorization: Bearer $AUTH_TOKEN" \
  "https://berdl.kbase.us/apis/cts/jobs?state=complete&after=2026-06-01T00:00:00Z"

Once you have the job ID, reconstruct the job object:

job = tscli.get_job_by_id("paste-job-id-here")

# Then use it as normal
job.get_job_status()
job.get_job()["outputs"]
job.print_logs(container_num=0, stderr=True)

Coordinating multiple jobs

This skill focuses on individual job submission, but common patterns when running a pipeline across tools:

# Submit jobs for independent tools in parallel, collect IDs
bakta_job   = tscli.submit_job("cdm_bakta:...",   ...)
checkm2_job = tscli.submit_job("cdm_checkm2:...", ...)

# Save both IDs
job_ids = {"bakta": bakta_job.id, "checkm2": checkm2_job.id}

# Wait for each in sequence (or use threads for true parallel wait)
bakta_result   = bakta_job.wait_for_completion()
checkm2_result = checkm2_job.wait_for_completion(wait_for_event_importer=True)

# Feed bakta outputs into a downstream job
bakta_outputs = [o["file"] for o in bakta_result["outputs"]]
downstream_job = tscli.submit_job("cdm_kofamscan:...", bakta_outputs, ...)

Track all job IDs in a file so the chain survives notebook restarts.

Canceling a job

job.cancel()
job.get_job_status()   # confirm — should transition to "canceling" then "canceled"

curl -s -X PUT -H "Authorization: Bearer $AUTH_TOKEN" \
  https://berdl.kbase.us/apis/cts/jobs/{job_id}/cancel

Key Concepts

Job lifecycle

created → download_submitted → job_submitting → job_submitted
    → upload_submitting → upload_submitted → complete
                                                OR → error
(cancel at any point → canceling → canceled)

insert_files() / input_files placeholder

In the Python client, tscli.insert_files() is a sentinel in the args list. At job runtime, CTS replaces it with the paths of the input files assigned to that container (space-separated). In the REST API, use:

{"type": "input_files", "input_files_format": "space_separated_list"}

Script is an input file (generic path only)

Your bash script is listed in input_files alongside your data. CTS stages it to the compute node at the input_mount_point, then the container executes it. Reference it in args using the staged path (e.g., /in/my-analysis.sh).

CRC64NVME checksums

Every file uploaded to MinIO for CTS must carry a CRC64NVME checksum. Always use mc cp --checksum crc64nvme when uploading. Files without checksums are rejected at job submission time.

Multi-container parallelism

Set num_containers > 1 to split input files across containers. CTS distributes files evenly. Each container receives its share via insert_files(). Use declobber=True when multiple containers could write output files with the same name (e.g., CheckM2 always writes quality_report.tsv). With declobber=True, CTS prepends the container index as a directory to every output path:

# declobber=True output layout (two containers):
cts/io/<username>/output/checkm2_run/0/quality_report.tsv
cts/io/<username>/output/checkm2_run/1/quality_report.tsv

Without declobber=True, container 1 would overwrite container 0's file.

Runtime format

ISO 8601 duration strings: "PT5M" (5 minutes), "PT1H" (1 hour), "PT2H30M" (2.5 hours), "PT4H" (4 hours). The format is case-sensitive — "PT1H" works, "pt1h" does not.

Image tags vs digest pinning

Every example in this skill uses image:tag format (e.g. cdm_bakta:0.1.3). Tags are human-readable but mutable — a tag can be re-pushed to point to a different image without warning. For reproducible science, pin to the digest instead:

ghcr.io/kbaseincubator/cdm_bakta:0.1.3@sha256:ab3d636d9381fdd57a8b21e3a2403663987f57f123027fa7a773474e66d9a827

This format uses both the tag (for readability) and the digest (for pinning). The container runtime uses the digest to pull the exact image regardless of what the tag currently points to.

Get the current digest for any image from the live registry:

images = tscli.get_images()
for img in images["data"]:
    if "bakta" in img["name"]:
        print(f"{img['name']}:{img['tag']}@{img['digest']}")
        # Use this full string as the image argument in submit_job

curl -s -H "Authorization: Bearer $AUTH_TOKEN" \
  https://berdl.kbase.us/apis/cts/images | python3 -c "
import sys, json
for img in json.load(sys.stdin)['data']:
    tag = img.get('tag', 'none')
    print(f\"{img['name']}:{tag}@{img['digest']}\")
"

Always retrieve the digest fresh from get_images() — do not rely on digests hardcoded in notebooks or this skill, as they become stale when images are updated.

Refdata and mount points

refdata_id identifies which pre-staged reference database bundle the image needs. It is registered with the image — you do not pass it to submit_job. CTS mounts the refdata read-only into the container at the default_refdata_mount_point specified in the image registration (always /ref_data for current BERDL images). To override the mount path, pass refdata_mount_point="/your/path" to submit_job. This is rarely needed since all current images use /ref_data as their default.

To verify a refdata bundle is registered and ready before submitting:

tscli.get_images()   # the refdata_id and default_refdata_mount_point are in each image's metadata

# List all registered refdata bundles
curl -s -H "Authorization: Bearer $AUTH_TOKEN" \
  https://berdl.kbase.us/apis/cts/refdata | python3 -m json.tool

GPU support

GPU compute is not currently available on the kbase cluster. Do not pass GPU-related parameters — the submit_job call will succeed but no GPU will be allocated.

API Reference

Base URL: https://berdl.kbase.us/apis/cts

Error Handling

Instructions for Claude

1. Run whoami and get_images before anything else — call tscli.whoami() to confirm CTS access and resolve the user's actual S3 path prefix (substituting it everywhere <username> appears in examples). Call tscli.get_images() to see currently registered images and retrieve their current digests. Do not rely on any static list in this skill.
2. Determine the execution path — if the user's tool is in the registry, use Path A (pre-built image). If not, use Path B (generic ubuntu script).
3. Always use digest-pinned image references — from the get_images() response, use name:tag@sha256:digest format in every submit_job call (e.g. ghcr.io/kbaseincubator/cdm_bakta:0.1.3@sha256:ab3d...). Tell the user which tag and digest you are using. Never hardcode a digest from memory or this skill — always fetch it live.
4. Ensure input files are staged before submitting — for both Path A and Path B, walk the user through uploading their input files to MinIO with mc cp --checksum crc64nvme (or the Python MinIO client on JupyterHub). Confirm files are staged before building the submit_job call. Build the input_files list from the staged MinIO paths.
5. For pre-built images — Read the tool's reference file at .claude/skills/remote-compute/references/<toolname>.md (e.g. references/gtdbtk.md, references/bakta.md). The filename matches the image name pattern without the cdm_ prefix — the lookup table is in the "Available pre-built images" section above. Use the reference file as the starting template. Fill in the user's actual MinIO paths for input_files and output_dir. Help the user determine num_containers and resource parameters based on their input count and the per-tool guidance in that file.
6. For the generic script path — help the user write a self-contained bash script, upload it with checksums alongside their data, and submit. The script should install its dependencies at the top.
7. Before submitting — confirm with the user — summarize the job: tool, image+digest, input file count, num_containers, memory, runtime, and output path. Ask the user to confirm before calling submit_job. This prevents wasted cluster time from wrong arguments.
8. After submission — display the job ID prominently and save it to a file — print it clearly and write it to projects/<id>/data/cts_job_id.txt (or equivalent). The user needs it to check status or retrieve results, and may lose it if the notebook closes.
9. Monitor when asked — use get_job_status() for lightweight polling. For blocking waits, use wait_for_completion() in a background thread. For CheckM2 (confirmed auto-import), use wait_for_completion(wait_for_event_importer=True) and then query the Spark table directly. For other images, check with the user or the CTS team before assuming an importer exists.
10. Help find previous jobs — if the user lost their job ID or wants to check past runs, use tscli.list_jobs() with appropriate filters (state, after, before, cluster). Once found, reconstruct the job object with tscli.get_job_by_id().
11. On completion — tell the user where their results are — state the output location explicitly (the output_dir passed to submit_job). List the actual output files from job.get_job()["outputs"]. Show the mc cp --recursive command to download them.
12. On error — call job.get_exit_codes() first to see which containers failed, then read stderr logs for those containers (job.print_logs(container_num=N, stderr=True)). The most common causes are wrong args, missing --force for Bakta, insufficient memory, or missing checksums.
13. Retrieve and interpret results — help the user download outputs from MinIO and explain what the tool produced.

Pitfall Detection

When you encounter errors, unexpected results, retry cycles, performance issues, or data surprises during this task, follow the pitfall-capture protocol. Read .claude/skills/pitfall-capture/SKILL.md and follow its instructions to determine whether the issue should be added to the active project's projects/<id>/memories/pitfalls.md.