Harness Adapters: How Neuron Wires Into Your Coding Agent

What a harness adapter is

A harness adapter is the interface between neuron and a coding agent’s harness (detect/capability/install/uninstall/verify), letting neuron init wire a real hook into that harness rather than relying on an instruction the agent may or may not follow. Capability is recorded as a map from lifecycle point (session start, pre-prompt, context reset, pre-command) to a support record — whether that point injects, its payload cap, its failure posture, its timeout — not a single label. The deterministic/best-effort/instruction-only fidelity shown in neuron init’s output and in the table below is derived from that map for display, never stored separately, so there’s one source of truth for what an adapter actually does.

Recall fidelity by harness

Harness	How recall lands
Claude Code	Every turn, automatically — hooked into SessionStart, UserPromptSubmit, and PreCompact
Codex CLI	Every turn, automatically — same three hook points
GitHub Copilot CLI	Once per session, automatically — the harness only exposes a session-start hook
Cursor	Once per session, automatically — same session-start-only constraint
Anything else	Instruction-based fallback via AGENTS.md, prompting the agent to query the store itself

Claude Code and Codex CLI additionally get a pre-command hook, firing on every shell tool call to surface a relevant memory before the command runs — see each harness’s own page for what’s wired and what isn’t.

Limitations

Recall coverage varies by harness, and the difference is structural, not a temporary gap: Claude Code and Codex CLI expose a per-turn hook point, while GitHub Copilot CLI and Cursor only expose a session-start hook — a memory recorded mid-session on those two won’t surface again until the next session starts. neuron init reports exactly what got wired, per harness, so this is never silently assumed.

Source: README.md “Recall your agent can’t skip”.