Remote Skill Test

Orchestrate end-to-end testing of agent skills on a remote jump host. After the user updates a skill locally, this skill SSHs to the remote host, updates the installed skills, runs the target skill via opencode run in a dedicated test directory, retrieves the generated report, and compares it against the previous run to verify that changes match the skill update.

Output Language Rule

Detect the language of the user's conversation and use the same language for all output.

• Chinese input -> Chinese output
• English input -> English output

Prerequisites

Required tools:

• ssh — access to the remote jump host
• scp — for retrieving reports from the remote host
• npx — installed on the remote host (for npx skills add)
• opencode — installed and configured on the remote host

Workflow

digraph test_flow {
    "Collect SSH config\
+ skill name + prompt" [shape=box];
    "SSH: create test dir" [shape=box];
    "SSH: install skills\
(project level in test dir)" [shape=box];
    "SSH: opencode run\
(output to log)" [shape=box];
    "SCP: retrieve report + log" [shape=box];
    "Find previous report" [shape=diamond];
    "Compare reports\
vs SKILL.md changes" [shape=box];
    "Output analysis" [shape=box];

    "Collect SSH config\
+ skill name + prompt" -> "SSH: create test dir";
    "SSH: create test dir" -> "SSH: install skills\
(project level in test dir)";
    "SSH: install skills\
(project level in test dir)" -> "SSH: opencode run\
(output to log)";
    "SSH: opencode run\
(output to log)" -> "SCP: retrieve report + log";
    "SCP: retrieve report + log" -> "Find previous report";
    "Find previous report" -> "Compare reports\
vs SKILL.md changes" [label="Found"];
    "Find previous report" -> "Output analysis" [label="First run"];
    "Compare reports\
vs SKILL.md changes" -> "Output analysis";
}

Step 1: Collect Information

Gather the following from the user. Do NOT proceed until all required items are provided.

Required:

1. Target skill name — which skill to test (e.g., aws-fis-experiment-execute)
2. SSH access — how to connect to the remote jump host. Ask the user:

   How do I SSH to the remote jump host?
   Please provide: user@host, and the SSH key path (or SSH config alias).
   Example: [email protected] with key ~/.ssh/my-key.pem
   

3. Test prompt — the prompt to run on the remote host. Ask the user to provide it directly in the conversation. Example:

   What prompt should I run for the test?
   Example: "使用 aws-fis-experiment-execute skill 执行 ~/fis-experiments/2026-04-10-az-power-int-my-cluster-EXTabc123/ 目录下的实验"
   

   If the user doesn't provide a prompt, construct one from the conversation context (e.g., the skill name + any dependency paths mentioned earlier).

Append the following suffix to the user's prompt (always):

如果需要跨目录读取文件，直接操作不要确认。
所有操作自动执行，不要等待用户确认。

Store the assembled prompt as FULL_PROMPT.

Step 2: SSH — Create Test Directory

Create the timestamped test directory on the remote host first, before installing skills. All subsequent steps operate inside this directory.

TIMESTAMP=$(TZ=Asia/Shanghai date +%Y-%m-%d-%H-%M-%S)
TEST_DIR="~/skill-tests/${TIMESTAMP}-{SKILL_NAME}"

ssh -i {SSH_KEY} -o StrictHostKeyChecking=no {USER}@{HOST} \
  "mkdir -p ${TEST_DIR}"

Store TEST_DIR for subsequent steps.

Step 3: SSH — Install All Skills (Project Level)

Install all skills from the repository inside the test directory at project level (not global). Skills have inter-dependencies (e.g., aws-fis-experiment-execute loads app-service-log-analysis at runtime), so installing only the target skill would cause missing-dependency failures.

Important: The remote host may use nvm for Node.js. Use bash -i -c to load .bashrc environment variables (LLM provider URL, API keys, etc.) that the interactive guard in .bashrc would otherwise block.

ssh -t -i {SSH_KEY} -o StrictHostKeyChecking=no {USER}@{HOST} \
  "bash -i -c 'source ~/.nvm/nvm.sh 2>/dev/null; \
   cd ${TEST_DIR} && npx skills add panlm/skills -y'"

Verify the output shows "Installation complete" and lists all installed skills. If it fails, show the error and stop.

Step 4: SSH — Execute Skill via OpenCode (Output to Log)

Run opencode run on the remote host inside the test directory. Capture all output (stdout + stderr) to a log file for diagnostics.

ssh -t -i {SSH_KEY} -o StrictHostKeyChecking=no \
  -o ServerAliveInterval=60 {USER}@{HOST} \
  "bash -i -c 'source ~/.nvm/nvm.sh 2>/dev/null; \
   cd ${TEST_DIR} && \
   opencode run \
     --dangerously-skip-permissions \
     \"${FULL_PROMPT}\" \
     2>&1 | tee ${TEST_DIR}/opencode-run.log'"

Key flags:

• --dangerously-skip-permissions — auto-approve all permission prompts (cross-directory reads, file writes, etc.) so the run is fully non-interactive
• 2>&1 | tee ...log — capture all output to opencode-run.log while also displaying it in the terminal for real-time monitoring
• bash -i — loads .bashrc environment variables (LLM provider config)
• ServerAliveInterval=60 — prevents SSH timeout for long-running skills

This step may take several minutes depending on the skill being tested (e.g., FIS experiments run for minutes). Wait for the command to complete.

If the command times out or fails, the log file may still contain partial output useful for diagnostics. Proceed to Step 5 to retrieve it.

Step 5: SCP — Retrieve Reports and Log

Each test run is stored in its own timestamped subdirectory under test-results/{SKILL_NAME}/. The {TIMESTAMP} used here is the same one from Step 2 (when the remote test directory was created).

List files in the remote test directory to find generated reports and the log:

ssh -i {SSH_KEY} -o StrictHostKeyChecking=no {USER}@{HOST} \
  "ls -la ${TEST_DIR}/"

Create the local run directory and copy all report files (.md files, excluding README.md) and the execution log. Keep the original remote file names — do not rename them:

LOCAL_RUN_DIR="./test-results/{SKILL_NAME}/{TIMESTAMP}/"
mkdir -p "${LOCAL_RUN_DIR}"

scp -i {SSH_KEY} -o StrictHostKeyChecking=no \
  {USER}@{HOST}:"${TEST_DIR}/*.md" "${LOCAL_RUN_DIR}/"

# Also retrieve the execution log for diagnostics
scp -i {SSH_KEY} -o StrictHostKeyChecking=no \
  {USER}@{HOST}:"${TEST_DIR}/opencode-run.log" "${LOCAL_RUN_DIR}/"

Step 6: Find and Retrieve Previous Report

Search for the previous test run of the same skill on the remote host:

ssh -i {SSH_KEY} -o StrictHostKeyChecking=no {USER}@{HOST} \
  "ls -d ~/skill-tests/*-{SKILL_NAME} 2>/dev/null | sort | tail -2 | head -1"

This returns the second-to-last directory (the previous run). If only one directory exists (first run), skip comparison and output the current report analysis only.

If a previous directory is found, retrieve its report into the same local run directory (LOCAL_RUN_DIR) so all comparison materials are co-located. The previous report's file name naturally differs from the current one (both have different timestamps embedded), so there is no collision:

PREV_DIR="{result from above}"
scp -i {SSH_KEY} -o StrictHostKeyChecking=no \
  {USER}@{HOST}:"${PREV_DIR}/*.md" "${LOCAL_RUN_DIR}/"

After this step, the local run directory contains everything needed for comparison:

test-results/{SKILL_NAME}/{TIMESTAMP}/
├── 2026-04-14-05-51-49-demo-cluster-assessment-report.md   # current report
├── 2026-04-14-05-29-47-amazon-eks-best-practice-checklist.md  # previous report
├── opencode-run.log                                        # execution log
└── test-analysis.md                                        # (generated in Step 8)

Read both report files from LOCAL_RUN_DIR for comparison in Step 7. The current report is the one whose timestamp is closest to {TIMESTAMP}; the other .md file(s) are from the previous run.

Step 7: Analyze and Compare Reports

Perform the following analysis:

7a. Report Structure Compliance

Read the target skill's SKILL.md from the skills repository (this repo). The path is ./{SKILL_NAME}/SKILL.md relative to the repo root (the current working directory). If the skill is not found at the repo root, check whether it is under others/{SKILL_NAME}/SKILL.md.

Extract the report template (look for markdown code blocks defining the report structure — headings, tables, required fields). If the target skill's SKILL.md does not define a report template (e.g., the skill generates CFN templates or directories instead of assessment reports), skip structure compliance and proceed directly to 7b.

Check the new report against each required element:

7b. Diff Against Previous Report (if available)

Compare the structural differences between the new and previous reports:

• Added sections — new H2/H3 headings not in previous report
• Removed sections — H2/H3 headings in previous but not in new
• Changed fields — fields present in both but with different structure

Do NOT compare data values (timestamps, resource IDs, metrics) — only structure and format.

7c. Correlate with SKILL.md Changes

Read the recent git changes to the target skill's SKILL.md from the repo:

git log --oneline -5 -- ./{SKILL_NAME}/SKILL.md
git diff HEAD~1 -- ./{SKILL_NAME}/SKILL.md

For each structural change in the report (from 7b), check whether it corresponds to a SKILL.md update. Flag:

• Expected changes — report differences that match SKILL.md updates
• Unexpected changes — report differences with no corresponding SKILL.md change
• Missing changes — SKILL.md updates that should have affected the report but didn't

Step 8: Output Results

Present the analysis to the user:

## Remote Skill Test Results

**Skill:** {SKILL_NAME}
**Test directory:** {TEST_DIR}
**Report file:** {REPORT_FILENAME}

### Structure Compliance
| Required Section | Present | Notes |
|---|---|---|
| {section} | Yes/No | {details} |

### Changes vs Previous Run
(Skip if first run)
| Change Type | Section | Details |
|---|---|---|
| Added | {section} | {description} |
| Removed | {section} | {description} |

### Correlation with SKILL.md Updates
| SKILL.md Change | Report Impact | Status |
|---|---|---|
| {change description} | {expected report change} | Match / Missing / Unexpected |

### Verdict
{Overall assessment: PASS / PARTIAL / FAIL with explanation}

Save this analysis to ./test-results/{SKILL_NAME}/{TIMESTAMP}/test-analysis.md.

Error Handling

Safety Rules

1. Never store SSH credentials in files. Always ask the user at runtime.
2. Never expose IP addresses, hostnames, or usernames in committed files.
3. Never modify files on the remote host beyond creating the test directory and running opencode run.
4. Never delete remote directories — previous test results are kept for comparison.