Loom: A repo-native state graph for local coding agents

I have been working on Loom, a Markdown-native project-state protocol for AI agents which has greatly enhanced the experience with Codex (both the App and CLI). The motivation is pretty simple: I wanted fewer agent surfaces. A plan (like the lovely ExecPlan) file is useful until it starts carrying too many jobs. A spec tool, issue tool, memory file, review prompt, and scratchpad can each help, but they also give the agent more places to scatter state. Loom puts the agent-facing work record into one project graph inside the repo. Research holds discovery and rejected paths. Specs hold intended behavior. Plans hold sequencing. Tickets track live execution. Evidence records observations. Critique holds review findings. Wiki holds accepted explanation. Packets hold bounded worker contracts. Memory helps recall without becoming the ledger. This is useful for local agents because there is no hosted service or runtime database. The graph is Markdown. Agents can search it with rg, follow links, edit records, and use Git for implementation history. The important part is packet compilation. A packet is a bounded snapshot of the upstream graph: project constraints, initiative, research, spec, plan, ticket, evidence, critique, source context, write scope, stop conditions, and output contract. The worker gets less context by volume, but better context by shape. Then retrospective pushes accepted learning back up into the graph. Repo: z3z1ma/agent-loom How are people here handling project state across agent sessions today? Memory files, plans, local issue trackers, worktrees, scratch docs, custom prompts?