The agent coding workbench gives Codex safe hands.
Commands, file reads, patches, verification, tags, Beads, and artifact paging are routed through caller-bound tools instead of loose ad hoc access.
System shape
One execution core, five surrounding capabilities.
Codex is not left alone in a terminal. It operates with controlled coding tools, fetched evidence, durable memory, repeatable playbooks, and outside review with strict authority boundaries.
Grounded Evidence turns claims into fetched proof.
PDF/document evidence stays citation-grade and separate from memory. Search finds candidates; fetch supplies the proof chunk.
Long-Term Project Memory preserves operating context.
Durable Markdown memory keeps decisions, runbooks, project orientation, and reusable findings available for the next session.
Operating Playbooks load the right expert behavior.
Review, bug hunting, SoT updates, contracts, wiki curation, and helper routing become repeatable procedures rather than improvised prompts.
Safe Review Gateway adds a constrained outside lens.
ChatGPT can inspect approved knowledge and propose work while shell, git, build, service control, and destructive authority remain outside its lane.
The split makes the system legible.
Evidence proves claims from documents. Memory carries durable context. Review stays useful because execution authority remains local.
Operating path
From first question to verified work.
Start with the full harness, then follow each surface into the capability it delivers: workbench control, grounded evidence, durable memory, operating playbooks, and safe review.
Shows the full Codex-centered operating system at a glance.
Keeps commands, edits, diffs, and verification inside a bounded workflow.
Turns PDF and document search into fetched, citation-ready proof.
Preserves decisions, runbooks, orientation, and curation history.
Loads repeatable expert workflows at the moment they are needed.
Adds outside analysis without handing over shell, git, build, or service control.
Capability surfaces
Choose the capability you want to inspect.
Each surface keeps its own visual identity while staying connected to the same controlled AI engineering harness.
Workbench
Agent Coding Workbench
Caller-bound tool execution, verification gates, structured evidence, and operator-safe workflows.
Open workbench EvidenceGrounded Evidence Engine
PDF/document evidence, search/fetch flow, and citation-ready proof boundaries.
Open evidence engine MemoryLong-Term Project Memory
Durable memory, curation, proposal/diff/write, and trash-inspection boundaries.
Open memory surface PlaybooksOperating Playbooks
Skill packages as repeatable operating doctrine for review, bug hunting, documentation, contracts, and handoff work.
Open playbooks ReviewSafe External Review Gateway
The safe review gateway where ChatGPT can inspect, summarize, and propose without holding destructive controls.
Open review gatewayThe AI engineering harness turns agent work into a controlled system.
Start with the operating map, then inspect the workbench, evidence, memory, playbooks, and safe review surfaces that make the system reliable.