controlstackai/truss is a Rust agent harness (agent harness) that lets AI agents resume tasks they have interrupted, and keeps a record of executions so they can be verified. The README describes this repository as a Rust agent harness for resuming, looking over, and verifying work. The October 16, 2026 commit 733d097 newly added a static landing page to this repository and also worked on the GitHub Pages deployment method. The size of the change is 37 lines: 35 lines added, 2 lines deleted.
The changes are all on the documentation and deployment side. .github/workflows/docs.yml merges the documentation built with mdbook (book/) and the new static site (docs/site/) into a single site-output and then uploads it, and the upload path was changed from book to site-output. The site-url in docs/book.toml also moved from /truss/ to /truss/guide/. The newly added docs/site/index.html contains descriptive metadata plus Open Graph, Twitter card, and JSON-LD (WebPage) schema, and sitemap.xml, llms.txt, and a search console verification file were added as signals for indexing. A link connecting to this site and the documentation was attached at the top of the README. Public deployment is set to work only on manual workflow_dispatch execution.
The newly published site also shows what this project claims. The site states that it is a development-stage project using Rust and the MIT license, that it declares a task contract (task contract) in TOML to include execution checks and review clauses, that it records executions in a SQLite journal so that already-recorded model and tool results are reused when resuming, and that Ed25519 signature items bound by a hash chain allow checking whether the record has been changed against a trusted key. In other words, the core this repository puts forward is the contract and the execution record rather than the model's answers themselves.
That said, this commit did not fix any runtime code. The added lines are static HTML, CSS, metadata, and deployment configuration, and there is no evidence in the given material that execution behavior has changed. The site itself writes that this is a runtime architecture rather than a live benchmark, and makes clear that controlled tests alone cannot prove superiority in production-scale reliability or in real-model task quality. Live provider runs and model review may incur API costs, and because the contract executes commands, unfamiliar projects should be run in an isolated environment. There are also no repository statistics or performance figures in this material. Since deployment is bound only to manual execution and the site's design description stops at the project's own account of itself, it is realistic for now to treat it as a suitable target for reading the code and checking the boundaries.