Within 24 hours of DeepSeek Harness going open source, developer Elie Bakouch pulled a clearly visible statistical curve from the GitHub work tree: of 984 merged pull requests, 209 came from codex/* branches — 21.2%. Those 209 PRs brought in 2,807 commits, accounting for 22.8% of the project's 12,293 total commits. In the same dataset, claude/* branches contributed only 3 PRs and 71 commits — 0.3% and 0.6% respectively. The commit history is also scattered with code review traces like "Codex review round".
Vaibhav (VB) Srivastav, a member of OpenAI's developer experience and community team, responded publicly: "Friends at DeepSeek, if you have any feedback on Codex, feel free to reach out — hope you enjoy Codex as much as I do."
The picture is somewhat absurd: DeepSeek is building an Agent Harness, and one of the tools helping it build that Harness is Codex, made by its competitor OpenAI.
Codex's tactical play
Even more interesting is Codex's tactical coordination. When DeepSeek released V4 Pro on August 13, it publicly stated that "on Code Agent projects, V4-Pro-0813 uses DeepSeek Harness minimal mode." The same week, they made their own API natively support the OpenAI Responses API format and wrote a dedicated adaptation script for Codex — users can plug V4 Pro or V4 Flash into Codex with one click.
In other words: the model layer and the Harness layer are DeepSeek's own work; but for high-contamination engineering actions like code generation, PR submission, and code review, they are willing to open up to a rival's tool.
The engineering math
There is an engineering calculation behind this. Claude Code's approach is that the model and Harness share the same origin (Anthropic's own Claude), so code generation, Harness, tool calls, and UI are all co-optimized — this is why Claude has long ranked near the top of AI coding benchmarks.
DeepSeek chose the opposite: split it apart. Model by V4 Pro, Harness by its own minimal design, and the "dirty work" of code generation and PRs by Codex — a temporary use of "rival ammunition" to fill an engineering slot before Anthropic gives away its Code Agent mode.
Short-term pragmatism, long-term liability
In the short term this is pragmatic; over the long run it is debt. A Codex share of 22.8% is already a structural dependency. If OpenAI tightens API commercial terms in some release, or if Codex falls behind on some niche task, DeepSeek Harness's iteration cadence gets dragged along with it.
"The Harness is pluggable" — easy to say in a keynote, another thing entirely to see in Git commit history. The real test of pluggability is whether core actions can run without a single vendor. DeepSeek's move made something visible to everyone: the Harness is still in its construction phase, and the engineers building it are using Codex to write the Harness.
Claude's 0.3% and what comes next
Worth a second look is Claude's position — 0.3%. While DeepSeek's Harness team (the "Black Whale" team) chose a pluggable design, they barely used Claude at all. It is not that Anthropic is unwilling — Claude Code simply did not offer the DeepSeek team an engineering entry point equivalent to Codex's.
That "~20%" line will become a hidden thread in AI coding engineering factions over the next six months: who dares to publicly admit how many commits in their own Harness come from a rival's tool, and who can keep the promise of true pluggability. With Cursor's "engineering delivery" acquired by SpaceX and Claude Code's closed-source moat still standing, DeepSeek's transparency about using Codex puts the market's real level of dependency on the table.