ThreadBack

Leave the problem. Keep the thread. A private, local-first Mac app that restores your debugging focus with NVIDIA Nemotron.

Comment

💡 Inspiration

Studies show it takes a software developer an average of 23 minutes and 15 seconds to regain deep focus after a single interruption.

When you get pulled away mid-investigation by an urgent meeting or on-call incident, you return to a messy terminal scrollback, half-baked diffs, and the agonizing dilemma: "Where was I, what hypotheses did I already disprove, and what should I run next?"

Most AI coding assistants try to autonomously rewrite code or scrape background history. We built ThreadBack to give developers their mental context back without sacrificing privacy: a local-first macOS assistant that records explicit evidence, saves checkpoints offline in 99 milliseconds, and uses NVIDIA Nemotron on Nebius Token Factory to synthesize source-linked, auditable Resume Capsules.

🚀 What It Does

ThreadBack lives quietly in the macOS menu bar, bound to approved Git project roots:

  1. Explicit, Opt-In Capture: Captures clean Git snapshots, opt-in wrapped terminal commands (tb run -- <cmd>), and developer notes (Fact, Hypothesis, Question). It never captures ambient keystrokes or background terminal history.
  2. Sub-100ms Offline Checkpoints: Hitting Save My Place writes your complete state to local SQLite WAL storage in under 100ms with zero network requests—allowing you to close your laptop immediately.
  3. Cryptographic Privacy Gate: Dual-stage regex scanner redacts live secrets (sk_live_..., tokens, PEM keys). Entries with sensitive findings are hard-blocked from transmission.
  4. Exact Payload Authorization: Displays the exact sanitized JSON payload and SHA-256 fingerprint before a single byte leaves your machine.
  5. NVIDIA Nemotron Resume Capsule: Nemotron synthesizes a structured capsule: separating Verified Facts (with clickable [evt-1] citations to local traces) from Hypotheses, recording Failed Attempts (even if code was reverted), and suggesting the next action.
  6. Repository Drift Sentinel: Upon return, ThreadBack compares the repository's HEAD, branch, and dirty working tree against the checkpoint snapshot before showing resume actions. ThreadBack never executes commands automatically; it keeps the human firmly in control.

🛠️ How We Built It

  • Core Engine (threadback-core in Rust): Air-gapped domain engine enforcing strict canonical root boundaries (std::fs::canonicalize) to block path traversals and external symlinks. Uses SQLite in WAL mode with commit-before-acknowledge durability.
  • Native Desktop Shell (src-tauri): Tauri 2 macOS menu bar popover using Cocoa AppKit screen-tracking for seamless multi-monitor positioning.
  • Frontend: React 19 + TypeScript with end-to-end type safety generated directly from Rust types via ts-rs.
  • Nebius Token Factory & NVIDIA Nemotron:
    • Connects to https://api.tokenfactory.nebius.com/v1 running nvidia/nemotron-3-super-120b-a12b.
    • API keys are stored exclusively in the macOS Keychain (com.threadback.desktop.tokenfactory) and never written to SQLite, logs, or UI renderers.
    • Features an anti-prompt-injection untrusted evidence envelope and a deterministic single-pass repair engine if schema or provenance validation fails.

📊 Evaluation & Benchmarks

ThreadBack includes an automated 10-scenario evaluation harness (npm run eval) and 5 persona workflows (W01–W05):

  • Fact Evidence Precision: 100.0% (10/10)
  • Unsupported Verified Claims: 0 (Zero Hallucination)
  • Seeded Secret Leakage: 0 bytes
  • Offline Save Latency: 99 ms max
  • Nemotron Inference Latency on Nebius: p50 = 415 ms, p95 = 490 ms
  • Resume Effort Reduction: 75% fewer inspection steps

🥊 Challenges We Ran Into

  1. Eliminating Hallucinations in Terminal Logs: Raw compiler traces and stack traces are messy. We isolated evidence into strict untrusted envelopes and built a Rust-level provenance checker requiring every verified fact to cite a valid local event ID.
  2. Multi-Monitor Display Positioning on macOS: Coordinating AppKit screen geometries across external monitors with differing display scale factors required native AppKit cursor tracking.
  3. Zero-Friction Privacy: Balancing security with developer velocity by providing deterministic two-stage redaction and an audited payload preview modal.

🏆 Accomplishments We're Proud Of

  • Achieving 100% provenance precision with zero unsupported claims.
  • Lightning-fast sub-100ms local offline saves.
  • Full auditability: preserving the immutable raw model draft alongside the user's approved capsule and edit diffs.

🔮 What's Next for ThreadBack

  • Signed Team Capsules: Exporting cryptographically verified Markdown capsules for shift handovers and PR descriptions.
  • Editor Extensions: Lightweight VS Code and Neovim plugins to jump directly to cited source lines from the capsule's next action.

Built With

Share this project:

Updates

Submission history