Giminger Consulting · Codex Plugin · Current benchmark

ContextGuard

37.1% lower API proxy cost across larger tasks.

0
API proxy savings
$0.150588 → $0.094695
0
Total token savings
1,594,223 → 1,250,429 tokens
0
Run capacity equivalent
same API-dollar budget
12/12
Functional arms green
3/6 strict pairs accepted

                

              

            

          

        


        

          ContextGuard keeps Codex on its normal shell workflow while injecting evidence once, compacting eligible noisy output locally,
          and preventing successful results from creating another retrieval roundtrip.
        


        

          
            
            
              
            
            
          
        

      

    


    

    

      

        
[ ARCHITECTURE ]

        
How ContextGuard works

        
Four layers between your shell and Codex, local-first with evidence preserved.

        

      

    


    

      

        
[ CURRENT MULTI TASK A/B · 2026-09-11 ]

        
Does it help on larger tasks?

        

          Three larger maintenance tasks, six counterbalanced pairs, GPT-5.6 Luna on low reasoning. The pooled sample cut total tokens by 21.6% and the API cost proxy by 37.1%; individual task medians still vary.
        

        

        

        

        

          API costs are an OpenAI Standard API proxy calculated from measured per-turn usage. Codex subscription billing is not exposed by the CLI. The rate card includes GPT-6 Astra for comparison.
        

      

    


    

      

        
[ CURRENT API COST ]

        
Measured cost, then transparent scaling

        
Six paired GPT-5.6 Luna runs at low reasoning · GPT-6 Astra Standard: $10/M input, $1/M cached, $50/M output for short context · long context: $20/M, $2/M, $75/M.

        

          

        

        

      

    


    

      

        
[ GET STARTED ]

        
Install in four steps

        

          Install from the GimingerConsulting/ContextGuard marketplace source, supported on macOS, Linux, and Windows wherever Codex runs. Start a new thread after setup.