quietclash

---

The problem nobody is catching

You run several AI coding agents in parallel — each in its own git branch or worktree. One writes the pricing logic, another touches a caller, a third refactors a helper. Then you merge.

Git checks one thing: did two changes touch the same lines of text? If not, it reports a clean merge and waves you through. But git has no idea whether two changes that don't overlap textually are compatible in meaning.

They often aren't. A measured 5–10% of parallel-agent merges are textually clean and test-passing yet behaviorally broken (CodeCRDT, arXiv:2510.18893).

This is exactly why teams that adopt AI heavily ship +98% more PRs but spend +91% longer in review — with no net throughput gain (Faros AI, 2026). Generating code is solved. Trusting the merge is not.

- git:        "No conflicts. Safe to merge." ✓
+ quietclash: "formatTotal() now returns $4200.00 where the author meant $42.00." ✗

Most tools in this space read the diff or ask an LLM to guess. quietclash instead executes the code and compares behavior across each branch — so a flagged conflict comes with a concrete input that reproduces it, not a hunch.

	git merge	CI tests	Static analysis / LLM review	quietclash
Catches textual conflicts	✅	—	partial	defers to git
Catches a behavior that matches neither agent's intent	❌	only if you wrote the exact test	guesses	✅ runs it
Gives a reproducing input	❌	✅	❌	✅
Needs you to have written a test for it	—	✅	❌	❌

---

See it

$ quietclash check --base main --branches agent-a,agent-b

Two agents worked in parallel. Their branches merge with NO git conflict:

  $ git merge agent-a agent-b   →   clean merge, no conflicts ✓

But quietclash runs the code and sees what git can't:

✖ 1 silent behavioral conflict found — these branches merge cleanly and pass
  tests, but behave in ways no agent intended.

  formatTotal  (price.mjs)
    what: one branch changed a function this caller relies on; the caller now
          behaves differently than intended
    evidence: diverged on 5/24 probe inputs
      e.g. input ["42"] → base=(n/a) | A=$42.00 | B=$42.00 | merged=$4200.00
    why:  Agent A changed parsePrice to return cents; Agent B's new formatTotal
          still assumes dollars. They merge cleanly and break on every price.

Reproduce it yourself: bash examples/demo.sh

---

How it works

quietclash doesn't read your code — it runs it, in four parallel "worlds," and compares what each function actually does.

flowchart LR
    A["base + N agent<br/>branches"] --> B{"merge cleanly<br/>textually?"}
    B -- "no" --> X["git already warns you —<br/>quietclash stays quiet"]
    B -- "yes" --> C["find the functions<br/>both agents touched"]
    C --> D["probe behavior in<br/>4 worlds"]
    D --> E["behavioral diff"]
    E --> F["false-positive<br/>filter"]
    F --> G["explain + report"]

    style X fill:#fde68a,stroke:#b45309,color:#000
    style G fill:#bbf7d0,stroke:#15803d,color:#000
    style D fill:#ddd6fe,stroke:#6d28d9,color:#000

The four worlds are the heart of it. For every symbol that two agents touched, quietclash asks: what did this function return before, what did each agent intend, and what does the merged result actually do?