deepx-ai/dx-compiler/dx-agent-verify

Name: dx-agent-verify
Author: deepx-ai

Claude Code and agent skill by deepx-ai/dx-compiler

Trust tier 3

community

Source check pending

Checking that the SKILL.md file is still in the repository.

Verify work before claiming done

Install

npx vskill@latest install deepx-ai/dx-compiler/dx-agent-verify

SKILL.md

/dx-agent-verify — Verify Before Claiming Done

RIGID skill. Iron Law: No completion claims without fresh evidence.
Run every verification step and confirm passing output BEFORE telling
the user compilation is complete, successful, or ready.

Trigger Words

"done", "complete", "finished", "ready", "ship it", "verified"

Type: RIGID

Evidence-before-assertions. Every claim of "it works" must be backed by command output captured in the current session. Stale evidence does not count.

Gate Function (5 Steps)

Execute ALL five steps in order. If any fails, STOP and fix before continuing.

Step 1: Input Model Integrity

bash

md5sum model.onnx
python -c "
import onnx; m = onnx.load('model.onnx')
print(f'Model: {m.graph.name or \"unnamed\"}  Opset: {m.opset_import[0].version}')
print(f'Inputs: {[i.name for i in m.graph.input]}')
print(f'Outputs: {[o.name for o in m.graph.output]}')
"

Evidence: File exists, hash captured, metadata printed.

Step 2: Config Validity

bash

python -c "
import json, onnx
cfg = json.load(open('config.json')); model = onnx.load('model.onnx')
print(f'Target: {cfg.get(\"target\", \"MISSING\")}')
onnx_name = model.graph.input[0].name
cfg_name = list(cfg['inputs'].keys())[0]
status = 'PASS' if onnx_name == cfg_name else 'FAIL'
print(f'Input name match: {status} (ONNX=\"{onnx_name}\", config=\"{cfg_name}\")')
"

Evidence: Config parses. Target set. Input names match.

Step 3: Output .dxnn File

bash

DXNN=$(ls output/*.dxnn 2>/dev/null | head -1)
if [ -z "${DXNN}" ]; then echo "FAIL: No .dxnn file"; exit 1; fi
SIZE=$(stat --format="%s" "${DXNN}")
echo "File: ${DXNN}  Size: ${SIZE} bytes ($(( SIZE / 1024 )) KB)"
[ "${SIZE}" -lt 1024 ] && echo "WARNING: suspiciously small" || echo "PASS: reasonable size"

Evidence: File path, size, PASS/WARNING status.

Step 4: Compilation Summary Report

bash

REPORT=$(ls output/*_summary.html 2>/dev/null | head -1)
if [ -z "${REPORT}" ]; then
    echo "FAIL: no *_summary.html — recompile with --export_html"
else
    echo "=== Summary Report ==="
    echo "File: ${REPORT}  Size: $(stat --format=%s "${REPORT}") bytes"
    echo "PASS: summary report generated"
fi
grep -iE "npu|cpu|partition|fallback" output/compiler.log 2>/dev/null | head -20

Evidence: Report path and size, plus the NPU/CPU partition lines from compiler.log. The report's interactive graph viewer is for human review; the compiler.log lines are the machine-checkable proof of the partition split.

Step 5: Compiler Log Review

bash

if [ -f output/compiler.log ]; then
    echo "=== Compiler Log Summary ==="
    grep -i "quantiz\|accuracy\|error\|warning\|npu\|cpu\|partition" output/compiler.log
    echo "PASS: compiler.log reviewed"
else
    echo "SKIP: No compiler.log (was --gen_log enabled?)"
fi

Evidence: Compiler log reviewed, or noted as unavailable.

Step 5.5: Cross-Validation with Precompiled Reference (if available)

If a precompiled reference DXNN exists in dx-runtime/dx_app/assets/models/, compare it against the generated model to isolate compilation vs verification issues.

bash

MODEL_NAME="<model_name>"
REF_DXNN="$SUITE_ROOT/dx-runtime/dx_app/assets/models/${MODEL_NAME}.dxnn"
if [ -f "$REF_DXNN" ]; then
    echo "=== Cross-Validation ==="
    python verify.py --dxnn "$REF_DXNN"
    echo "Reference model result: $?"
    python verify.py --dxnn output/${MODEL_NAME}.dxnn
    echo "Generated model result: $?"
fi

Evidence: Cross-validation results and diagnosis (PASS/Compilation problem/verify.py problem). See dx-dxnn-compiler.md Phase 5.7 for the full Differential Diagnosis Decision Matrix.

Step 6: Portability / Self-Contained Verification (HARD GATE — retrain/showcase)

A compile/retrain session dir MUST be relocatable — moving it must not break it. Per ultralytics-train-eval.md §6: NO absolute paths serialized into *.json/configs; prefer regenerate-if-missing over skip-if-metadata-exists. Verify by copying outside the suite:

bash

APP="<session_dir>"; TMP="$(mktemp -d /tmp/dx_portable.XXXXXX)"
rsync -a --exclude venv --exclude __pycache__ "$APP"/ "$TMP/app/"; cd "$TMP/app"
grep -rn "/dx-all-suite/\|dx-agent-dev-showcase/" --include='*.json' --include='*.py' --include='*.sh' . | grep -v '/sample/' \
  && echo "FAIL: absolute/suite path serialized" || echo "OK: no absolute suite paths"
cd - >/dev/null; rm -rf "$TMP"

FAIL if any *.json/script embeds an absolute suite path (input /sample/ media is allowed). See the suite-level Self-Contained & Portable gate and ultralytics-train-eval.md §6.

Verification Checklist

Confirm ALL before claiming completion:

[ ] Input model exists and is valid (Step 1)
[ ] Config is valid JSON, input names match (Step 2)
[ ] Output .dxnn exists with reasonable size (Step 3)
[ ] *_summary.html generated, compiler.log partition lines reviewed (Step 4)
[ ] Compiler log reviewed, no errors (Step 5)

Completion Report Template

## Compilation Completion Report
 
**Session**: <session_id>
**Timestamp**: <ISO 8601>
**Status**: PASS | FAIL
 
### Input
- Model: <model_name.onnx>
- Format: ONNX (opset <N>)
- Input shape: <shape>
- MD5: <hash>
 
### Config
- Target: dx_m1
- Input name match: PASS
- PPU enabled: <yes/no>
 
### Output
- File: <output/model.dxnn>
- Size: <N> KB
- NPU ops: <N>/<total> (<percent>%)
- CPU ops: <N>/<total> (<percent>%)
 
### Validation
- Summary report: PASS
- Compiler log errors: 0
- Compiler log warnings: <N>
 
### Output Structure
dx-agent-dev/<session_id>/
├── model.onnx
├── config.json
├── model.dxnn
├── compiler.log
└── README.md

Anti-Patterns (NEVER Do)

Saying "compilation complete" without running Steps 1-5
Relying on stale evidence from a previous session
Skipping the summary report and compiler.log review because "the file exists"
Claiming success when compiler.log contains errors
Reporting completion without the verification checklist

Security

Security scanPassed, 90 of 100Pattern checks on the skill's files.
Model reviewPassed

See the full security report

Quality evaluation

Task evaluationNot runNo evaluation yet. Run one in Skill Studio.

See evaluation results

Source

SKILL.md in the repositoryChecking.deepx/skills/dx-agent-verify/SKILL.md

Popularity

7-day trend score: 0

Not refreshed yet.

Works with

Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.

Claude Code
Cursor
GitHub Copilot
Windsurf
Codex
Gemini CLI
Cline
Amp
Roo Code
Goose
Continue
Kilo Code
Junie
Trae
OpenHands
Qwen Code
AI Agent
Tabnine Chat
Sourcegraph Cody
Amazon Q
Firebase Genkit
Sweep
Devin
Phind
Mentat
Claude Desktop
Aider
GPT Pilot
Plandex
bolt.new
v0
Replit Agent
Anse
TypingMind
BetterChatGPT
LibreChat
Warp AI
Pieces
Fig AI

Embed badge

[![vskill VERIFIED](https://verified-skill.com/api/v1/skills/deepx-ai/dx-compiler/dx-agent-verify/badge)](https://verified-skill.com/skills/deepx-ai/dx-compiler/dx-agent-verify)

Something wrong with this skill or its page? Tell us.