self.md radar — 2026-06-29
models are learning to write their own harnesses, repos are freezing old decisions into files agents have to obey, and the outside world is starting to arrive with signatures attached.
that is the useful shape today: less chat theater, more scaffolding. one drop trains coding models to generate the workbench around the solution. two repos turn AGENTS.md into a decision and command contract. one geo-memory project and one planning paper both say the same quiet thing: if an agent claims the world changed, make it bring a receipt.
1. the harness moved into training
sources:
what happened: DeepReinforce published Ornith-1.0, an open-source family for agentic coding with 9B Dense, 31B Dense, 35B MoE, and 397B MoE variants. the claim is not just another coding score. Ornith trains the model to generate both solution rollouts and task-specific harnesses, then optimizes the scaffold and the answer together. the flagship 397B model reports 77.5 on Terminal-Bench 2.1 and 82.4 on SWE-Bench Verified; the smaller 35B model reports 64.2 and 75.6.
why this matters: the harness is no longer a neutral wrapper around model quality. if the model learns to build its own test rig, every coding-agent benchmark needs to say what counted as model, what counted as scaffold, and who was allowed to move the rails.
2. decisions became repo-local law
sources:
what happened:
Lore says the blunt part out loud: give the coding agent the decisions the team already made, so it stops re-litigating them. its RAC engine keeps requirements, decisions, designs, roadmaps, and prompts as typed Markdown, serves them read-only over MCP, and avoids embeddings or LLM calls for relevance. the exported AGENTS.md block lists settled ADRs and tells the agent not to reopen or contradict them. mustflow comes at the same bruise from the command side: root AGENTS.md first, .mustflow/ for workflows and context, and commands only when the repo-local contract says they are allowed.
why this matters: agent memory is too squishy for decisions that have already cost humans time. the personal AI OS needs a place where “we decided this” is not a vibe in the transcript, but a file, a gate, and a thing the next agent can be blamed for ignoring.
3. world claims want signatures and gates
sources:
what happened:
emem describes a verifiable memory substrate for agents where every read returns an ed25519-signed receipt that can be checked offline. its earth layer addresses ground as cell64 patches, about 9.55 meters at the equator, and the live surface lists 93 /v1/* paths, 81 MCP tools, 160 algorithms, and 46 source schemes. the GILP paper attacks the same problem in planning: an LLM drafts an action and imagined state change, a small parameterized world model supplies valid actions, deltas, risk, and value, and a consistency gate asks for revision when they disagree. on GPT-4o-mini calls, the reported hallucinated-state rate falls from 0.176 to 0.035; simulator success rises from 0.668 to 0.838 for about 22% more LLM calls.
why this matters: long-running agents cannot run on “seems plausible” forever. once the system starts touching maps, sensors, plans, and downstream tools, the useful question is not whether the answer sounds grounded. it is whether the state change has a signature, a model check, or both.
supporting links
- Jon Udell on unreviewable agent PRs — a clean human-in-the-loop warning: if the agent opens thousands of lines, adding reviewer agents may just move the headache one room over.
- OpenAI Codex issue #2847
— the still-open request for repo and global ignore files around
.env, keys, AWS, SSH, and other local secrets. - MemoryGuard
— a tiny local context-sync repo that says it updates
AGENTS.md,CLAUDE.md,MEMORY.md, and Cursor rules without network calls or an external LLM. - entwurf — an ACP bridge experiment for moving Claude Code, Codex, and Gemini sessions through one garden-addressed surface.
left on the table
- Codex ignore files stayed supporting because yesterday already put agent setup and secret exposure on the main stage.
- Jon Udell’s PR warning is sharp, but the code-review choke point was already a main signal this month.
- Understanding Rollout Error in Graph World Models was too close to GILP: same authors, same date, same graph-world-model lane.
- Vibe-Trading had repo-gift energy, but not enough operator-control substance for a main slot.
- DESIGN.md stayed out as a recent repeat, not because the idea got less useful overnight.
Related self.md routes
- Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
- AI coding assistants — compare coding workbenches by review surface, permissions, cost, logs, and escape hatches
- Best Claude Code plugins — choose the Claude-specific extensions worth installing, and the ones to skip