Key points
Within 24 hours of DeepSeek Harness going open source, developer Elie Bakouch surfaced a visible statistical curve from the GitHub working tree:
- Of 984 merged PRs, **209 came from
codex/*branches — 21.2% - These 209 PRs brought in 2,807 commits — 22.8% of the project's 12,293 total commits**
- In the same data,
claude/*branches contributed only 3 PRs and 71 commits, or 0.3% and 0.6% respectively - Commit logs also show code review traces such as "Codex review round"
- In the short term this is pragmatic; in the long run it is a liability.
- Codex's share at 22.8% is already a structural dependency. If OpenAI tightens API commercial terms in some release, or Codex falls behind on certain tasks, DeepSeek Harness's iteration cadence gets dragged along with it.
- Saying "the harness is pluggable" is easy in a keynote; seeing it in git commits is another matter. The real test of plug-in architecture is whether core actions can run without a single vendor.
- DeepSeek's move showed everyone: the harness is still in its construction phase — and the engineers building it are using Codex to write the harness.
An OpenAI developer experience and community team member, Vaibhav (VB) Srivastav, publicly responded: "Friends at DeepSeek, if you have any feedback on Codex, feel free to reach out — hope you're enjoying Codex as much as I do."
The picture is somewhat absurd: DeepSeek is building an Agent Harness, and one of the tools helping it do so is rival OpenAI's Codex.
Codex's tactical配合 — DeepSeek's deliberate openness
The more interesting part is the tactical coordination. When DeepSeek released V4 Pro on August 13, it publicly stated that "on the Code Agent project, V4-Pro-0813 uses DeepSeek Harness in minimal mode." The same week, it also made its own API natively support the OpenAI Responses API format and published a dedicated adaptation script for Codex — users can one-click wire V4 Pro or V4 Flash into Codex.
In other words: the model layer and harness layer are DeepSeek's own work; but high-noise engineering actions like code generation, PR submission, and code review are willingly opened up to a competitor's tool.
The engineering calculus
Behind this is an engineering account. Claude Code's approach is model and harness from the same source (Anthropic's own Claude), so code generation, harness, tool calling, and UI are all co-optimized — this is why Claude has long ranked near the top of AI coding benchmarks.
DeepSeek chose the reverse: decouple. Model = V4 Pro; harness = its own minimal build; code generation and PR "dirty work" = Codex. Until Anthropic gives away its Code Agent mode for free, DeepSeek is temporarily filling an engineering slot with "the rival's ammunition."
Short-term pragmatism, long-term liability
Claude's position: 0.3%
Also notable is Claude's spot at 0.3%. DeepSeek Harness's team (the "Black Whale" team) went pluggable while barely using Claude at all. Not because Anthropic is unwilling — but because Claude Code never offered DeepSeek an engineering entry point equivalent to Codex's.
The "about 20%" line will become a hidden thread in AI coding engineering circles over the next six months: who dares to publicly admit how many of their harness's commits come from rival tools, and who can hold up the promise of pluggability. With Cursor's engineering delivery acquired by SpaceX and Claude Code's closed-source moat still standing, DeepSeek's transparency about its Codex usage puts the market's true degree of dependency on the table.