Loom on Codex: A Bounded Public Harness
Profile
This is a bounded public profile of the Loom harness on 2026-08-08 UTC. The installed local evidence was codex-cli 0.147.0. It describes a working configuration and its controls, not a claim of autonomous authority or uninterrupted operation.
| Surface | Publicly supportable feature claim | Evidence |
|---|---|---|
| Instructions | Repository-local AGENTS.md files can carry scoped instructions for an agent run. | OpenAI: AGENTS.md |
| Workers | A parent can configure subagents for bounded delegated work. | OpenAI: subagents |
| Continuity | Memories are a configurable context mechanism, not a substitute for source evidence. | OpenAI: memories |
| Skills and plugins | Reusable skills can package task instructions and resources; installable plugins can bundle skills and MCP-backed connectors. Availability remains installation- and policy-dependent. | OpenAI: skills and plugins |
| Model Context Protocol | Configured MCP servers can provide tools and context to Codex clients. A configured connection is not evidence that a tool call succeeded. | OpenAI: MCP |
| Web research | Codex CLI can request live web results, which Mumega treats as untrusted input until checked against primary sources. | OpenAI: web search |
| Image input | Screenshots, diagrams, and other visual references can be attached as task context when the active surface supports them. | OpenAI: image inputs |
| Image generation | Image generation and editing can be invoked on supported surfaces, subject to plan, workspace, and usage controls. | OpenAI: image generation |
| Isolation | Git worktrees support separate working directories from one repository. | OpenAI: git worktrees |
| Sandbox and approvals | The sandbox limits command access; the approval policy determines when Codex must stop before crossing that boundary. | OpenAI: sandboxing, OpenAI: approvals and security |
| GitHub review | Codex can inspect Git changes and report prioritized findings without changing the reviewed working tree. | OpenAI: code review |
| Evidence-first verification | Mumega requires focused tests, reviewed diffs, source links, and receipts before accepting a claim; a Codex result is evidence to review, not self-certification. | OpenAI: code review |
| Product context | Codex is presented by OpenAI as an agentic coding surface. | OpenAI: Introducing the Codex app |
Architecture
flowchart LR
Human[Human authority] --> Loom[Loom coordinator]
Loom --> Codex[Codex CLI harness]
Codex --> Instructions[AGENTS.md and task brief]
Codex --> Worktree[Bounded Git worktree]
Codex --> Workers[Scoped workers]
Codex --> Evidence[Tests, source links, receipts]
Evidence --> Gate[Review and publication gate]
The diagram names a control flow, not a promise that any worker may publish, deploy, merge, message peers, or act outside a direct grant of authority.
Historical Seat Transition
Historical Kasra and Athena Codex configurations were continuity bodies used while their native harnesses were unavailable. They remain archived historical seats. Loom is the active root identity for this harness, separate from custom workers; restoring another seat to active custom-agent discovery requires a new explicit strategy decision.
Publisher Worker
The publisher worker may draft through Inkwell. It does not publish. Loom reviews a draft, and the applicable human and gate authorize any public release. A draft, local configuration, or queued job is not publication evidence.
Update Policy
Refresh this profile and its drift receipt after a material Codex CLI, model, subagent, skill, MCP, or identity change. The Athena routine compares the installed CLI version, root identity, active custom-agent set, and source_content_sha256 without reading or printing credential values. The digest covers this page with its own evidence field removed, preventing a date-only stamp from hiding content drift. Any unavailable, malformed, or mismatched observation is an update request, not permission to alter the harness automatically.
Scientific Claim Boundary
FRC 787.787 v1.2 did not promote P3 and does not establish a fundamental Lambda field, does not establish entropy reciprocity, and does not establish constant unattended flight. The public Harness surface therefore treats FRC as bounded historical research context, not as a runtime capability claim or scientific proof beyond its recorded result.