What our run establishes
Our run on October 9, 2026, at commit f8ea3cd recorded a successful install step followed by unsuccessful build and test commands. These observations belong to our sandbox, rather than to a production application or a new execution for this correction.
The harness stores each step's success flag, exit code, elapsed time, timeout flag, and log tail. Install was recorded as successful in 90 seconds; build and test commands exited unsuccessfully in 4 and 2 seconds. Neither command timed out. The build and test tails both show Yarn rejecting the workspace lockfile state before producing compiler diagnostics or test results. A failed test command here does not establish failed assertions.
The MrKeyoor testing methodology explains how recorded command outcomes differ from runtime and security validation. That page supplies provenance context; its current procedure does not establish which methodology version governed this older run.
Installation and audit evidence disagree
The install step's recorded success does not erase its log warnings. The tail includes an ERESOLVE dependency conflict involving Vite and an aliased tsx package, then later package-install output. Without a complete-log reconciliation, we cannot identify the full sequence or claim that an ordinary install completed without intervention.
The attached stored parser summaries report 432 dependencies, 145 MB of dependency footprint, and an npm audit total of 0. The raw install tail reports different package and advisory totals. We preserve the stored values and flag the disagreement instead of substituting the tail's figures. The audit total is a recorded advisory result with unknown scan coverage, not proof that the dependency tree is secure.
Test metrics are null. There is no supported aggregate of passed, failed, or skipped tests, and no basis for describing the parser summaries as exhaustive unique coverage. We did not validate a running plugin host, hot reload, teardown order, workload performance, or deployment.
Why evaluate Cordis despite that boundary?
The linked DeepSeek Harness primer describes Cordis as its vendored plugin framework. It explains a context as a repository of services addressed by stable keys. Plugins declare required services through injection and wait until those services exist. This gives a plugin host a way to express dependencies without manually arranging every activation.
The primer describes registrations as reversible effects that reload and teardown unwind. Its examples include prompt sections, tool schemas, adapters, providers, and listeners. This is documented behavior, not a result demonstrated by our run. The primer also advises giving every registration a disposer and keeping related work in one effect when disposal order matters.
That lifecycle model is the reason to prototype Cordis. Start with a small dependency relationship, remove the providing plugin, and check what happens to the consumer and its registrations. Then exercise failed activation and repeated reload. These are proposed acceptance checks, not tests we completed.
API stability and documentation
Cordis warns that its API is not yet stable and may change without notice. Its repository also says official documentation is under construction and points readers to the primer.
The primer provides concrete explanations of contexts, injection, dispatch modes, and disposal, but it addresses authors working inside DeepSeek Harness. General adopters must distinguish those host-specific examples from the framework behavior they need. Current documentation also cannot prove that every described behavior matches the measured commit.
When a server framework fits better
NestJS is a TypeScript framework for Node.js server-side applications. Consider it when your main decision concerns server application architecture. The supplied extract does not establish an equivalent to Cordis's runtime disposal model.
Fastify is a web framework with hooks, plugins, and decorators. Consider it when HTTP routes and server extensions define the application. Its plugin architecture alone does not establish interchangeable lifecycle behavior with Cordis.
These are alternatives for different application goals, rather than demonstrated replacements for Cordis's teardown behavior. The supplied sources do not establish an integration between them and Cordis.
How to read the scores
Assessed scores use a 1 to 5 editorial rubric: 1 means a demonstrated blocker for the dimension; 2 means substantial unresolved constraints; 3 means useful evidence with material gaps; 4 means strong evidence with limited gaps; 5 means comprehensive evidence within the stated scope. These are judgments about the supplied evidence, not performance measurements. Unassessed means the evidence cannot support a rating.
Setup receives 2 because installation was recorded as successful but subsequent commands stopped at workspace resolution. Documentation receives 3 because the primer explains the model while official documentation remains under construction. Maturity receives 2 because the project explicitly warns of API instability. Community is unassessed: the supplied primary extracts do not establish contributor breadth, support responsiveness, or adoption.

