ThoughtKhoral is a human-led collaborative workspace where people and role-specific AI agents deliberate together while humans retain authority over room context and decisions.
The project is organized as independently versioned repositories. Its current focus is a technical MVP for governed, real-time rooms and human-controlled decisions, alongside an experimental provider-free Codex conversation POC. ThoughtKhoral is in active development and welcomes technical evaluation and feedback. The project home and repository map describe the architecture and cross-project boundaries.
For current project status, interface support, and remaining Codex milestones, see the cross-project Codex status guide, the compatibility matrix, and capability roadmap.
The core MVP demonstrates authenticated rooms, mention-aware message delivery,
a human /decisions workflow, and audited decision governance. The former
Decision: chat-prefix facilitator is retired; its proposal port is reserved
for a separately approved memory-derived implementation.
The experimental, provider-free Codex conversation POC currently supports:
- Explicit human targeting and authorized room-wide history, including relevant discussion that took place while Codex was not responding.
- Shared conversation continuation and explicit fresh-session controls; starting a fresh session preserves the room transcript.
- Model and effort controls plus context-usage status in the workspace UI.
- Gateway-mediated room access; the independent worker has no direct room database access.
- Provider-free component suites and synthetic composed checks.
Codex acceptance remains open for:
- Packaged-stack acceptance and live tool/egress-isolation checks.
- A separately authorized live multi-human check covering an intervening room fact, worker restart and continuation, and a fresh session with the authorized room baseline.
- Live settings, denied-model, and context-telemetry checks, plus Linux x86_64 native tool-policy and Rust 1.85 portability evidence.
The published conversation v1.0.0 artifact is unchanged; the additive v1.1 defaults candidate remains unreleased. These are experimental POC capabilities, not production-readiness evidence. See the cross-project status guide for exact evidence and gates.
The memory engine remains a separate, incubating room-scoped ingestion proof; Codex does not use it. A locally controlled deterministic A2A reference agent is also available. Cognee integration, open remote-agent admission, durable memory storage, and production deployment remain deferred.
If you are evaluating the system for the first time, follow the repositories in this order:
- Run it: start with the local MVP platform.
- Use it: explore the workspace UI.
- Trace it: inspect the room gateway, which owns the authenticated event boundary.
- Verify it: read the contracts, the compatibility authority for the room protocol.
The repository READMEs contain setup and verification details. To evaluate the separate Codex conversation POC, start with its cross-project status and verification guide.
- Authenticated, governed rooms.
- Real-time room events with a defined protocol boundary.
- Room-wide and mentioned-only chat delivery with governed audiences.
- Human creation, confirmation, editing, dismissal, and audited deletion of
decisions through the
/decisionsworkflow.
| Repository | Owns | Evaluation focus |
|---|---|---|
| Platform | Rootless local composition and Kubernetes-manifest validation. | Can the MVP be reproduced and verified locally? |
| Workspace UI | Browser workspace, chat stream, agent task updates, and decision controls. | Is the human-in-the-loop room experience understandable? |
| Room Gateway | Authenticated WebSocket rooms, event persistence, replay, and authorization. | Are the runtime and security boundaries clear? |
| Contracts | Versioned room and conversation schemas, protocol documentation, and compatibility fixtures. | Are the protocols durable, inspectable, and sufficient? |
These repositories have active proofs or experimental integrations, but are not production services:
- Memory Engine — a room-scoped, provenance-preserving ingestion proof; Codex memory integration, Cognee, and durable storage are deferred.
- Agent Gateway — a deterministic local A2A reference agent plus opt-in Codex mediation; open remote-agent admission remains deferred.
- Codex Agent — an independent provider-free worker for history-aware room conversations; packaged-stack and live-provider acceptance remain open.
The most useful feedback is concrete: an unclear project boundary, a missing verification step, a protocol or security concern, an unexpected MVP behavior, or a high-value next step. Open an issue in the repository that owns the behavior and include enough context to reproduce or evaluate it. See the contribution guide for the issue-first, specification-driven workflow. Accepted issues become approved specification updates before implementation and supporting documentation.