Six items remain before the 1.0.0 release. One follows it.
End-to-end coverage of the generator, the modelling tool and the applications it produces.
Access control, authentication, secrets handling and dependencies, audited across the platform and across generated applications.
A full manual for the generator, from installation through to deployment.
The complete generation pipeline running on the DeepSeek harness.
Complete testing of the enterprise reporting application, and its integration with AppWithAI. Source: enterprise_reporting_tanstack.
Enable selected parts of each generated application's URL space to be securely exposed to the public internet through the customer's firewall, allowing authorised customers, suppliers and other external users to interact with designated application functions. External exposure will be limited to explicitly approved routes and services, with authentication, authorisation, transport security and appropriate network controls applied so that public access does not require exposing the application's internal services or infrastructure.
The repository currently reads 5.1.1. That number tracked releases rather than production readiness, so 1.0.0 is a deliberate reset.
The whole path — business description, model, review, generated application, migrations, tests, deployment — driven end to end by agents on DeepSeek Harness, with the human at the review gate rather than at the keyboard.
Item 4 puts the existing pipeline on the harness. This item makes the pipeline agentic: the harness supplies the loop, the delegation and the sandbox, and AppWithAI supplies the tools and the standards those agents are held to.
The harness is built on Cordis, where the model adapter, the tool registry, the session log and the agent loop are all plugins on a shared context — there is no privileged core to patch. AppWithAI mounts beside them as a bundle, so a harness release does not become a merge.
Domain analysis, ERD design, the Application Dictionary and generation register on
ctx.tools, and so do the published validators — checker.js and
fixer.js. An agent then meets the same diagnostics a person does, and a model
that would fail the checker cannot reach code generation.
Human-in-the-loop review of the model is step 3 of the pipeline and does not move. The
harness has ask_user_question for exactly this, plus approval gates on
operations that need explicit consent — autonomy up to the model, and a stop there.
ctx.subagents is a swappable seam: spawn or fork in process, or delegate to an
external coding agent over ACP. One subagent per entity or module, and the harness's
ralph loop — fresh agent, same task, until it holds — to iterate on a
generated application until the checks are clean.
Filesystem and subprocess are seams too, with sandboxed providers alongside the local ones. The generated application gets built, migrated and tested in that sandbox. This is the part that makes unattended delivery safe rather than merely possible.
The same composition boots as web, headless or sdk, so
a delivery can run from CI with no terminal attached. The session log is the authoritative
context and survives reloads, so a run that fails in its third hour resumes instead of
starting again.
DeepSeek Harness is in developer preview under the MIT licence and expects compatibility-breaking changes. Building the autonomous path on it is the right target and the wrong thing to make a release depend on — so 1.0.0 is the pipeline running on the harness, and this is the pipeline running itself.
Detailed scope and acceptance criteria for each item are maintained with the source: ROADMAP.md. The harness behaviour described above is taken from its own architecture, agent-lifecycle, capability-seams and tool-catalog documentation.