Agent operations · public-safe proof ladder

RDLeader

RDLeader is a local-first control plane for supervising multiple AI R&D workers: task ownership, progressive context, runtime dispatch, result collection, approval gates, and evidence-first QA.

The public repo is intentionally released in safe slices. It shows the architecture, contracts, fake-data demos, walkthroughs, and verification model without publishing private DevPlan paths, app IDs, chat IDs, QR artifacts, internal links, or raw live-integration logs.

TypeScript monorepo React + Vite manager UI Fastify control plane SQLite repositories ACP-style runtime dispatch Human approval gates Demo reset QA evidence

Copy/paste pitch: local-first control plane for AI R&D workers: task ownership, context routing, runtime evidence, approval gates, and QA.

View repository See v0.1.1 proof-ladder release Read public proof ladder Watch narrated walkthrough Use distribution kit Support RDLeader packaging

Not another chat UI The focus is the operating layer around agents: ownership, context, dispatch, evidence, blocked states, and approvals.

Public-safe demos pnpm demo:reset creates deterministic fake worker, runtime, approval, and QA state for reviewers.

Architecture signal Manager UI, API, policy, brain/context assembly, ingest, runtime dispatch, and repositories are separated.

External discovery Submitted to autonomous-agent and coding-agent awesome lists; follow-up is tracked in public.

Best support note Use RDLeader #27 for external submission follow-up or RDLeader #1 for broader public-release sanitization.

What RDLeader proves

Task ownership — a worker owns a work item instead of every agent living in one undifferentiated chat stream.
Progressive context — task envelopes can assemble identity, seed, working, episodic, knowledge, and reflection layers.
Runtime evidence — dispatch, result collection, stale-task recovery, and archived runtime events are first-class concerns.
Approval boundaries — risky external actions are represented as manager-reviewable approval requests.

Dedicated employee-agent model — one worker maps to isolated runtime home and communication config in the DevPlan architecture.
QA posture — tests, smoke checks, endurance loops, browser validation, and redaction rules are summarized publicly.
Manager UI surface — overview, execution, collaboration, growth, management, QA, and ops surfaces exist in the public slice.
Honest release posture — MVP / research prototype, not a finished enterprise HR system.

Proof ladder

Open these in order if you only have a few minutes. The point is to make the technical claim verifiable without leaking private operational state.

Step	Public surface	What it proves
1	Public demo reset	Deterministic fake worker, work item, runtime, approval, and QA state.
2	Browser walkthrough	The fake state can be inspected through the manager UI without private seed workers.
3	Runtime / approval deep dive	Task envelopes, result events, fail-closed external actions, and recovery modes are explicit.
4	Employee-agent onboarding	Worker homes, runtime homes, manager-only communication, and secret references are documented safely.
5	QA evidence	Tests, smoke checks, endurance loops, and public redaction rules are summarized.
6	Narrated walkthrough	Public-safe visuals show the control-plane shape without DevPlan screenshots.
7	Submission tracker	Distribution and follow-up are tracked instead of handled as one-off announcements.

Architecture shape

React manager UI
      │
      ▼
Fastify control plane
      │
      ├── domain model        employees, performance, emotion, approvals
      ├── brain package       task-type → context-layer assembly
      ├── policy package      risk classification and approval gates
      ├── ingest package      git/doc memory extraction and direction knowledge
      ├── runtime package     ACP-style dispatch + result collection
      ├── ops workflows       project groups, review actions, QA reports
      └── SQLite repositories work items, messages, reflections, runtime events
      │
      ▼
Local worker workspace
  WORKSPACE_MAP.*        repo and direction map
  .rdleader/tasks        inbound task envelopes
  .rdleader/results      runtime result events

Sponsorware / follow-up surface

RDLeader#27 tracks external submission review follow-up.
RDLeader#3 tracks license posture; no root license is implied until the owner decides.
RDLeader#1 tracks broader public release and DevPlan bundle sanitization.

External PR: awesome-autonomous-agents#20.
External PR: awesome-coding-agents#13.
Support route: support RDLeader packaging work.

Public-safety rule: do not publish real employee names, private workspace paths, app IDs, open IDs, chat IDs, message IDs, QR onboarding artifacts, internal document links, or raw terminal output from live integrations.

Copy/paste promo pack

Shortest line

RDLeader: local-first control plane for AI R&D workers — task ownership, context routing, runtime evidence, approval gates, and QA.

中文版本

RDLeader 是一个面向 AI 研发员工的本地优先控制台：任务归属、渐进式上下文、运行时派发、结果回收、审批闸门和 QA 证据，而不是又一个聊天窗口。

English version

RDLeader is a local-first control plane for supervising AI R&D workers. It focuses on what happens after multiple agents are running: who owns the task, what context was loaded, what runtime ran it, what result came back, what is blocked, and which actions need approval.

Best 3-link pack

Repo: https://github.com/happysnaker/RDLeader
Project page: https://happysnaker.github.io/rdleader/
Proof ladder: https://github.com/happysnaker/RDLeader/blob/main/docs/public/landing-page.md