trailmem 0.1.8 — hosts/ modularisation, doctor host checks, required file params

What's new

Hosts modularised

integrate.py (one 588-line file, all 9 hosts' knowledge tangled + hand-synced uninstall lists) split into trailmem/hosts/ — one module per host (claude, codex, kiro, kilo, opencode, antigravity, zed, cursor, windsurf). Every Artifact pairs install() with remove(), and run/uninstall iterate the same registry — install/uninstall drift is structurally impossible. Adding a host = one file + one registry entry. Shared plumbing in hosts/_util.py.

Narrow-write / wide-detect policy

integrate now auto-writes config only for verified hosts (Claude via claude mcp add, Codex TOML, Kiro, Kilo). OpenCode/Antigravity/Zed/Cursor/Windsurf stay detected but get the exact entry printed for manual add (flip write=True once verified). Detection stays broad and grows — stores are always attributed to the right agent_type.

trailmem doctor — host capability + drift + stale servers

• Per-host artifact status (registered / not installed / stale launcher / missing agent pin).
• Flags running stale trailmem.mcp_server processes — the recurring "old server kept writing" incident is now visible.
• Caught and fixed a real gap live: Kiro SessionStart hook was missing.

code_files / doc_files now REQUIRED on store

Empty was indistinguishable from "no files touched". Missing/empty → ValidationError; literal 'none' → NULL. Enforced in store.store() so all paths (MCP/CLI/dashboard) apply it.

Other

• Canonical identity via session_context (versioned object, authoritative; legacy session_id/agent_type params deprecated through 1.0).
• Threat-model docs added.
• .gitignore hardened (root scratch scripts caught by sdist leak-grep in the first 0.1.8 build).

Upgrade

uv tool install trailmem@latest --force
trailmem doctor   # check host artifacts + stale servers

Then restart every agent so running MCP servers pick up the new code.

Full diff: v0.1.7...v0.1.8