Codex Taskforce

A coordinated agent team for substantial work

Built for token-efficient agent orchestration, with swarm execution, persistent project context, and compact knowledge handoffs between agents.

Codex Taskforce keeps the main agent responsible for direction, architecture, scope, synthesis, integration gates, and acceptance. Specialized workers handle bounded implementation, operational checks, research, independent verification, and documentation.

What makes it useful

• Knowledge-bearing task capsules give each worker the ownership, evidence, guidance, and acceptance intent needed to work independently.
• Compact evidence-linked reports and batched decisions reduce operational noise in the main conversation.
• Six roles cover Explorer, Investigator, Executor, an optional Senior Executor, Tester, and Archivist. Use only the roles a task needs.
• Independent verification keeps production repair with its owning Executor and acceptance decisions with the main agent.
• Persistent project memory in agent_docs/ records goals, architecture, decisions, progress, and session handoffs.
• A bundled reporter produces a six-column per-role token table from available local Codex session evidence.

Included

One standalone skill, orchestration and dispatch instructions, six worker role resources, configurable model profiles, six project-context templates, a non-overwriting project initializer, and a local token reporter. Five default worker profiles use GPT-6 Luna; the optional Senior profile uses GPT-6.1 Sol. Explicit user model choices take precedence.

Requirements

Local Codex with native subagent tools, per-worker model/reasoning selection, and the filesystem/tool access required by the task. Python 3.11+ is required for the bundled helpers. Access to the configured models must be supplied by the user's Codex account/provider.

Permissions and limitations

The initializer is read-only by default and creates only missing agent_docs/ documents when setup is authorized. Existing documents, AGENTS.md, Codex settings, and Git state are preserved. Role instructions are passed in memory; no worker configuration is installed.

The reporter reads local session metadata, token counters, and deployment boundaries. The bundled helpers make no outbound requests and do not upload session content.

This is an instruction-based workflow, not a deterministic engine or an independent sandbox. It targets main-agent overhead; total worker tokens may increase, and no fixed savings percentage is guaranteed. Token reporting depends on readable local JSONL evidence and a supported schema.

Through Agensi MCP, linked references and permitted script execution must be exposed to Codex and/or its workers. Missing capabilities are reported rather than bypassed. Local packaging and synthetic filesystem/reporting checks passed; live Agensi MCP execution and a full multi-agent task replay have not yet been verified.

Use it for substantial feature work, difficult investigations, non-overlapping module changes, and recoverable project handoffs.