By 2.6.17, NEXO already had the big parity story in place: Claude Code and Codex shared one brain, Codex could start as NEXO, and Deep Sleep could read both transcript stores. But a few practical gaps still mattered in daily use. 2.6.18 closes those gaps without pretending Codex and Claude Code expose the exact same surfaces.

Codex Now Keeps a Managed Shared-Brain Contract in Config

Before this patch, Codex wiring still leaned too much on the CLI path that registers MCP servers ad hoc. That worked, but it also meant the shared-brain contract could drift if local Codex state got out of sync.

2.6.18 now persists a managed mcp_servers.nexo entry directly in ~/.codex/config.toml, alongside the existing bootstrap, model, and reasoning profile sync. In practice, that makes Codex much harder to knock out of shared-brain mode accidentally.

Runtime Doctor Audits Real Recent Sessions

The original parity work proved NEXO could wire Codex correctly. The next question was stricter: did real recent sessions actually start like NEXO sessions?

2.6.18 teaches runtime doctor to inspect recent Codex session artifacts for startup markers, startup tool usage, and heartbeat discipline. It also verifies Claude Desktop shared-brain metadata explicitly, instead of treating Desktop as an invisible best-effort MCP attachment.

Deep Sleep Now Keeps Weekly and Monthly Memory

Version 2.6.16 gave Deep Sleep a longer 60-day horizon. That was the right first move, but it still meant the overnight system could end up rediscovering similar patterns day after day.

2.6.18 adds reusable weekly and monthly summary artifacts, so Deep Sleep can carry higher-horizon memory forward explicitly. It also adds project-priority weighting based on diaries, learnings, followups, and decisions, which means the overnight system can rank work by leverage and pressure instead of raw recency alone.

Retrieval Gets More Honest Without Getting Noisier

NEXO already auto-enabled HyDE and shallow spreading when the query shape suggested semantic or associative recall would help. The remaining problem was explainability and precision: users and agents should know why a result surfaced, and exact lookups should not be polluted by weak associative neighbors.

2.6.18 adds confidence and auto-strategy hints to retrieval explanations and trims associative expansion back to top_k after ranking. So the system becomes more transparent without regressing into noisier exact lookups.

The Reliability Hotfix That Matters in Practice

This patch also includes a targeted Deep Sleep reliability fix: the current headless model flow can write synthesis output into a nested path that the previous verification step did not always accept cleanly. That meant a run could produce the right JSON and still look failed.

2.6.18 accepts that nested output path correctly, which removes one of the most annoying false-failure modes in the overnight pipeline.

Install or Update

New install:

npx nexo-brain

Existing installation:

nexo update

Need to reapply shared-brain client wiring later?

nexo clients sync

See the full version history in the changelog, or help tighten the next edges through issues and PRs on GitHub.