Agent Overview

This page summarizes the current OrangeBot working model: what exists today, what is still a proposed operating role, and how work should move from planning to implementation.

Current status

Main obot / Orchestrator: active. Runs in Telegram/OpenClaw, handles intake, status, routing, safe bounded fixes, and recovery when worker sessions fail.
PM/Design cycle: operating role. Uses GitHub issues as durable memory, owns Project / Requirement / Task structure, Human ToDos, roadmap/Hub tracking, and development handoffs.
Development cycle: operating role. Owns implementation through PRs, evidence, checks, demos, and issue/PR updates.
Testing cycle: proposed/specialized role. Should own browser, GHA, Playwright, visual/regression, and evidence-hardening tasks when needed.

Today these roles are mostly prompt-scoped work sessions, not fully isolated permanent OpenClaw agents. The framework work below is intended to make them more reliable and auditable.

Responsibilities

Main obot

Intake Jeremy's requests and decide whether PM, Development, Testing, or direct bounded action is appropriate.
Keep Telegram replies concise and plain-text safe.
Verify work-session liveness before saying a cycle is active.
Recover stalled work sessions and report failures plainly.

PM / Design

Run the portfolio PM audit before selecting development work.
Maintain Project, Requirement, and Task issue structure.
Keep Hub homepage, roadmap, reports, and Human ToDos aligned with GitHub issues.
Apply the Human ToDo delegation gate before asking Jeremy for anything.
Produce exact development handoffs with scope, evidence, checks, and non-goals.

Development

Implement through branches and PRs under obot-claw.
Use clean worktrees or explicit branch strategy; do not overwrite dirty unrelated work.
Run feasible checks and capture evidence.
Update linked issues/PRs with concise status and blockers.
Do not merge unless Jeremy explicitly approves.

Testing

Own browser and GHA evidence when feature work needs stronger validation.
Document local Playwright/Chromium blockers separately from product readiness.
Compare demos, console output, accessibility, and visual behavior when UI output is involved.
Escalate missing test infrastructure as Tasks, not vague Human ToDos.

Human ToDo delegation gate

Before creating a 🙋 Human ToDo, PM must ask:

Can PM decide or document this?
Can Development implement or verify this?
Can Testing produce the missing evidence?
Is a new specialized agent/task needed?

Only if the answer is no should the Hub show a Jeremy ToDo. When Jeremy is needed, the linked issue or PR must contain @jwildfire followed by explicit questions/instructions, decision criteria, and the relevant evidence.

Portfolio framework tasks

#21 Agent Overview page — this page.
#22 Portfolio audit helper and PM report workflow.
#23 Hub sync gate for portfolio tracking.
#24 PM / Development / Testing launch contracts.
#25 Durable work-session supervision — implemented with scripts/work_session.py and the supervision contract.