Every new session starts from zero. Decisions get forgotten, ideas interrupt, context dies silently. thruline is one Rust binary that fixes the collaboration — for Claude Code, Cursor, Antigravity, Codex, and every MCP agent.
Decisions are recorded, locked, and injected into every prompt the agent ever sees — this session, next week's session, a different agent entirely. Propose against a lock and the agent gets warned before code changes.
The agent is deep in a task. You think of something. Today your options are: derail it, or forget your idea. thruline adds a third: drop the note from any terminal — the moment the agent finishes, your notes arrive as review comments. Watch both sides:
Enforced by hooks in Claude Code — the agent physically cannot stop without seeing your notes. Other agents get the queue via AGENTS.md and usually comply; we document exactly how reliably.
One command bottles the session — summary, decisions, open notes, TODOs. One command pours it back. Into tomorrow's session, or into a completely different agent.
Anything injected into an agent is an attack surface — we treat it that way. Edit the memory outside thruline (a malicious git pull, a rogue script) and it's quarantined until a human reviews it. Secrets are scrubbed before they ever touch disk. Fingerprints live outside the repo, so tampering can't bless itself.
Nine-finding public threat model. CI-built binaries with Sigstore provenance. npm installs verify pinned SHA-256. No telemetry — there is nothing to opt out of.
Agents differ in what they let outside tools do. Most products hide that. We publish it.
| agent | memory + locks | mcp tools | auto review queue | context gauge |
|---|---|---|---|---|
| claude code | full | full | enforced (hooks) | statusline |
| antigravity | full · tested live | full | best-effort | not exposed |
| cursor | full | full | best-effort | not exposed |
| command code | full | full | best-effort | not exposed |
| codex · gemini cli · copilot | full | full | best-effort | not exposed |
No agent lets outside tools draw inside their chat. So we built our own surface. The VS Code panel shows your review queue live, lets you reply to specific agent responses (like WhatsApp reply-to-message), and auto-refreshes when the agent finishes.
Works alongside Cursor, Copilot, Claude Code desktop — any agent in VS Code. File watcher auto-refreshes; no polling, no tokens.