collie/ harness
GitHub ↗
collie
A coding agent that proves its work

Watch the gate go green.

collie finds the bug, writes a reproduction that must fail, makes the smallest diff that flips it to passing — and hands you the receipt. Quietly. For pennies. With the proof kept on disk.

$ pip install collie-agentcopy › star on GitHub
$ collie -p "fix parse_duration compound units" → exit 0 = verified
collie  ~/timeparse  · gpt-5.6 · auto

        

        

          REPRO armed
          ·0 tok
          ·0.0s
          ·$0.000
          ·ctx 8%
          v verbose · u undo · esc steer
        

      

      
    

  


  
  

    

      

        
The signature

        

          
The verification gate is the hero, not a footnote.

          
Every other agent buries “I ran the test” in a tool log. collie commits to a falsifiable target first — a reproduction it wrote and executed — then refuses to call the job done until that same assertion flips from failing to passing. One object, two states, tracked by your eye.

        

      

      

        

          
✗ armed · failing

          assert parse_duration("1h30m")
   == 5400
          
Written before a single edit. Executed on the unfixed code — exit 1, got 1800. If it can’t fail here, it’s not a real gate, and collie throws it out.

        

        
→the diff
that flips it

        

          
✓ verified · passing

          total += SECONDS[unit] * int(val)
# was: total =
          
A +1 −1 change. The assertion runs again — exit 0. Only now is the run “done,” and the repro is kept on disk as evidence.

        

      

    

  


  
  

    

      

        
Seven principles

        
Lean is the brand — so the interface earns every line.

      

      

        
01 Gate first
The proof is the most prominent live thing
A persistent REPRO gate sits in the HUD, failing→passing. A run isn’t “done,” it’s “verified ✓ in 12.8s.”

        
02 Quiet by default
~1/6 the tokens should feel true
One narrative line per action, thinking collapsed, no chain-of-thought wall. A whole fix fits in ~16 lines. Verbosity is one keypress, never the default.

        
03 Show where it looked
Localization is visible
Semantic code_search prints ranked candidate sites with similarity scores — you watch it reason about location before it touches code.

        
04 Honest receipt
tokens · time · $ always on screen
Glanceable in a one-line HUD, never dumped per-turn. The “most-capable-per-token” claim is a receipt you can read, not marketing.

        
05 Smallest diff
Scope bound by the assertion
The repro defines “done,” so the edit is the smallest change that flips the gate. Diff size (+7 −2) shows as a virtue, not a wall of new code.

        
06 Git is undo
Approvals are a dial, not a modal storm
Auto-checkpoint before each edit; u reverts. Because reversal is cheap and every edit is verified, collie defaults to low-friction — the #1 documented killer, sidestepped.

        
07 Degrade cleanly
Meaning never lives in color alone — and it runs everywhere
The gate reads ✓ passing / ✗ failing in words, not just green and amber. Plain output when it’s not a TTY, NO_COLOR respected, progress to stderr and result to stdout. Lean also means uncorrupted over SSH, CI logs, and Windows Terminal.

      

    

  


  
  

    

      

        
The receipt

          
Truth-in-advertising, rendered as UI.

          
The field sells scores on benchmarks it now calls contaminated. collie sells the opposite: a live, real tally of what the run cost and the executed evidence it produced. Expand it any time with /cost.

        

      

      

        
verified · session receipt

        
repro gate✓ passing

        
tokens loc 1.1k / edit 0.9k / vfy 1.4k3,410

        
wall time12.8s

        
cost$0.006

        
diff+5 −2 · 1 file

        
evidencerepro_1h30m.py

      

    

  


  
  

    

      

        
Interaction model

        
Approvals are a dial. The repro and git are the safety net.

        
collie can afford to be quiet because every edit is reversible and executed-verified. Modes cycle live with Shift+Tab; the active one always shows in the HUD.

      

      

        

          
ModeReads · code_searchIn-repo editsShell · test runNetwork · out-of-repo

          
            
plansilentshown, not applied——

            
defaultsilentask once / patternaskask

            
autorecommendedsilentsilent · git-reversibleverify pre-allowed
other shell classifier-gated		ask

            
yolosilentsilentsilentsandbox only

          
        

      

      

        deny → ask → allow, first match wins · approvals session-sticky · a blocked action returns to the agent (“find a safer path”), escalating to a human only after 3 denials — deny-and-continue, never silent-halt.
      


      

        
/planpropose, don’t write

        
/verifyre-run repro + regression tests

        
/reproopen the reproduction collie wrote

        
/costtokens · time · $ by phase

        
/diffworking-tree diff since start

        
/undorevert to a checkpoint

        
/modelswitch model or effort

        
/quiet /verboseset default verbosity

        
/dashboardopen the control tower

      


      

        
0exit 0 — repro verified passing. collie is a CI verifier, not just an editor.

        
1exit 1 — couldn’t reach a passing assertion

        
2exit 2 — error

      

    

  


  
  

    

      

        
The control tower

        
The CLI is the cockpit. The dashboard is history and 2-D review.

        
A terminal streams one run beautifully and reviews a session badly. The web view owns what the terminal can’t — and never becomes required.

      

      

        

          
run timeline — locate │ edit │ verify · width = time · ✓ marks the gate flip

        

        

          
flask-5014✓ 9.1s

          
seaborn-3069✓ 21.4s

          
django-13121✓ 14.0s

          
sympy-21596✗ 18.2s

        

        

          locateeditverify
          retries show as repeated verify segments — seaborn re-armed once, then held.
        

      

      

        Plus the one view no competitor has: a verification ledger — every run with its kept repro, the executed assertion, and its exit code. One click to the exact script and its output. The anti-contaminated-benchmark artifact: here is the assertion, here is that it ran, here is that it passed.
      

    

  


  
  

    

      

        
What we won’t build

        
Restraint is a feature. Every surface must serve locate → edit → verify.

      

      
    
        
  • No full-screen TUI framework. Keep the append-only stream — it works over SSH, CI, and Windows. Repaint only the HUD line.
    • 
        
  • No approval-modal maze. Repro + git are why collie can be quiet. Per-edit prompting imports the exact fatigue everyone complains about.
    • 
        
  • No chain-of-thought wall. Reasoning is dimmed and opt-in. A verbose collie undercuts the 1/6-token story on sight.
    • 
        
  • No fuzzy search/replace edit engine. Exact line-anchored edits + re-read on failure. Aider’s single biggest pain — refused.
    • 
        
  • No emoji chrome, no warm closers, no persona zoo. The output is receipts, not personality.
    • 
        
  • No 40 slash-commands, no plugin marketplace, no per-turn cost spam. Nine commands. The receipt lives in the HUD.
    • 
      

    

  


  
  

    
collie shows you where it looked, arms a failing assertion, makes the smallest diff that flips it to green, and hands you the receipt — quietly, for pennies, with the proof kept on disk.