CodeClone

Deterministic Structural Change Controller for AI-assisted Python development

[CodeClone @codeclone_app]

https://t.co/XXIhHPfxpP

[CodeClone @codeclone_app]

@thsottiaux I switched because GPT-5.6 Sol is genuinely excellent at coding.

And the broader difference matters: winning adoption with temporary unlimited Codex access feels very different from pushing users through the threat of model retirement.

[CodeClone @codeclone_app]

Everyone seems to be building a CodexBar.

So… we built a LimenBar. 😄

We already know where the money goes.

Watching the balance disappear doesn't make it hurt less.

What we actually wanted to see:
- all our limits in one glance — Codex, Cursor, Claude, Grok;
- how fresh each reading is — live vs stale, no guessing;
- burn rate and time-to-limit ("~1.7h left" beats a surprise 429);
- session window vs weekly quota, side by side
local models next

Still early. We're building it because we use it every day.

[CodeClone @codeclone_app]

The further we take CodeClone, the clearer one thing becomes: powerful models have already outgrown chat.

Most agent interfaces today are still natural-language CLI. They can execute commands, but they do not provide a persistent state model, causal structure, visible intent, bounded actions, or verifiable transitions between “before” and “after”.

CodeClone has made us think about what the same engineering principles could mean beyond software: an interface where AI is not the conversation layer, but part of the operating medium itself.

Not another agent that “does everything”. Something different.

Once CodeClone is stable and the team has the capacity, this may be the next product worth exploring.

[CodeClone @codeclone_app]

Released ckdn 1.1.0…

…and 1.1.1 immediately afterwards. 🙈

Sometimes the fastest bug finder is a swarm of coding agents.

New in 1.1.x:

✅ pre-commit parser
✅ worktree-aware cwd for CLI & MCP
✅ workspace command policy
✅ deterministic config locking for CI
✅ shared MCP guidance for clients
✅ lots of worktree and artifact fixes
The biggest lesson:
If you’re building tools for AI agents, cwd is part of your API.
It sounds trivial until one agent runs from a worktree, another from /tmp, and a third through MCP.
Those edge cases are now covered. On to the next release. 🚀

[CodeClone @codeclone_app]

Some of the problems we’re fixing in CodeClone v2.1.0a2 have existed in the detection core from the beginning.

They were visible, but our golden tests were doing exactly what they were designed to do: preserving compatibility and preventing silent semantic drift. I would not allow the detector contract to be changed casually.

That restraint was right. But it also meant that correcting the foundation required a deliberate compatibility epoch—not another patch.

Now the time has come.

We are rebuilding the detection core around a Python-version-stable canonical fingerprint format, explicit semantics, stronger domain separation, and fewer redundant AST traversals. No shims, no quiet reinterpretation of old results, and no pretending the previous contract was more universal than it really was.

Strictness and honesty must begin with CodeClone’s own code. We cannot leave it outside the scrutiny we expect CodeClone to provide for everyone else.

You can build an impressive product quickly. Building an honest and deterministic one takes distance, evidence, trial and error—and the willingness to revisit your earliest assumptions.

That is what v2.1.0a2 is about.

[CodeClone @codeclone_app]

We said CodeClone v2.1.0a2 would add no new features.

Then the a1 audit exposed a harder truth: a codebase can pass every local check while the same architectural responsibility is independently implemented across several subsystems—and slowly drifts.

Why couldn’t CodeClone detect this in itself?

That question led us to Semantic Authority Analysis: a new dimension of CodeClone’s existing multidimensional analysis, designed to detect split ownership, shadow producers, and competing projections of the same fact, contract, or effect.

Because strictness and honesty must begin with CodeClone’s own code. We cannot leave it outside the scrutiny we expect it to provide for others.

A feature born from pain.

[CodeClone @codeclone_app]

v2.1.0a2 is less about adding another visible feature and more about making CodeClone dependable at its core.

We are working on it intensely, with source audits, executable proofs, failure testing, and benchmarks.

Determinism and explicit trust remain non-negotiable.

[CodeClone @codeclone_app]

CodeClone v2.1.0a1 did what a good alpha should do: it exposed problems while we still have the freedom to fix them properly.

For v2.1.0a2, we are not hiding those issues behind patches. We are redesigning several foundations of the product.

[CodeClone @codeclone_app]

Some of these changes are deep and intentionally breaking.

That is exactly why we are doing them during the alpha stage.

We would rather establish strict, durable contracts now than preserve convenient mistakes that become impossible to remove later.

[CodeClone @codeclone_app]

In the article on https://t.co/Hxm1rRuHRj me, I talked in detail about ckdn

https://t.co/eEwGd1pdDi

[CodeClone @codeclone_app]

Your coding agent shouldn’t read 10,000 lines of test output and guess whether it passed.

In football, a checkdown is the short, safe pass when the long play is too risky.

ckdn gives coding agents that same safe option.

Try it on your repo. Break it. Star it.

https://t.co/3zEmtYkP3X

[CodeClone @codeclone_app]

ckdn is live.

Try it on a real project. Let your coding agent use it instead of reading raw test and coverage logs—and tell me where it helps, where it gets confused, and what still feels rough.

Early feedback matters a lot. Issues, ideas, integrations, and real-world edge cases will shape the next releases.

And if the idea resonates, a GitHub star is a simple way to support the project.

https://t.co/3zEmtYkP3X

[CodeClone @codeclone_app]

Your coding agent doesn’t need another 10,000-line test log.

It needs a safe answer it can trust.

In American football, a checkdown is the short, safe pass when the long play is too risky.

That’s exactly what ckdn is for AI coding agents.

[CodeClone @codeclone_app]

ckdn runs your project checks and gives the agent a small, clear result instead of a wall of terminal noise.

No guessing. No wasted context. No “looks green.”

The full evidence stays on disk when you need it.

Open source. MIT.

uv tool install 'ckdn[mcp]'

[CodeClone @codeclone_app]

We promised to ship rundown — and tomorrow, we will.

One small twist: the name changed. The product got better.

It was born while stabilizing CodeClone, a project with nearly 5,000 tests. Raw test, coverage, lint, and type-check logs became a real pain for agentic workflows.

So we built a tool that sits between your coding agent and pytest, Ruff, coverage, mypy, ty, and others. It owns the real exit code, preserves full evidence, and gives the agent one tiny deterministic digest instead of thousands of lines of noise.

Less context waste. No “looks green.” No silent parser failures.

New name and v1.0.0 tomorrow.

[CodeClone @codeclone_app]

One more thing that deserves recognition.

Thank you to the OpenAI team for what appears to be a very thoughtful user experience around usage limits.

During a long multi-agent coding session, I hit the 5-hour usage limit—but the work that was already in progress kept running. Even more interesting, the same seemed to apply to the sub-agents that had already been launched.

Eventually you'll still run into a weekly quota or context limits, but allowing in-flight work to finish makes a huge difference. Losing a complex orchestration halfway through would be incredibly frustrating.

It's a small design decision, but one that shows an understanding of how people actually use agentic coding tools.

For me, this preserves one of the best aspects of ChatGPT Codex: once the work has started, the system tries to let it finish instead of stopping everything the moment a quota is reached.

👏

[CodeClone @codeclone_app]

Spent some time pushing GPT-5.6 Sol on a real-world workload, and I'm impressed.

Instead of a single coding task, I gave it a pool of GitHub issues that had to be distributed, coordinated, and executed in parallel. That's where it really stood out—the orchestration was consistently solid.

Another pleasant surprise: CodeClone v2.1.0a1 behaved very well with sub-agents. The workspace coordination model held up under parallel execution, keeping tasks scoped and reducing the usual agent drift.

It's still early, but this feels like a meaningful step toward reliable multi-agent development workflows.

The downside? 😄 Token limits disappear at an alarming speed. Complex orchestration is powerful... but definitely not cheap...

[CodeClone @codeclone_app]

Every code-review tool starts after the patch exists. That's too late for agent work — by then, every decision has already been made.

CodeClone 2.1 moves control before the edit:

declared intent → blast radius → explicit boundary → patch → deterministic verification → receipt

Your reviewer stops reverse-engineering intent from a diff and starts reviewing a contract: what the agent said it would do, what it did, and evidence they match.

[CodeClone @codeclone_app]

Your AI agent just "fixed pagination." It also refactored a shared helper nobody asked about — and in the diff, that looks like diligence, not scope creep.

CodeClone 2.1.0a1 is out: a deterministic Structural Change Controller for AI-assisted Python development.

Agents declare intent before the edit. CodeClone maps the blast radius, sets an explicit boundary, verifies the real patch against the declared scope, and issues an auditable receipt.

Let agents move fast. Keep structural change explicit, bounded, and verifiable.

https://t.co/OUMaPYvo3P