Skip to content

v0.3.0

Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 16 Oct 15:39
v0.3.0
4202d96
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

The release that makes Sensei connect itself to every AI tool it finds, handle
real coding-agent traffic, and measure itself honestly, including against
Headroom.

Switch it on once

  • sensei connect / sensei disconnect, plus the same two buttons in the
    tray and on the Connections page. Sensei wires every tool it detects and
    explains every tool it can't, with the reason. New tools are picked up while
    it runs. Every edit is backed up, hash-checked and reversible.
  • 25 tools. New: VS Code Copilot, Copilot CLI, Antigravity, Qwen Code,
    Crush, Factory, Junie, Kimi, Amp, Kiro, Warp, Augment, LM Studio and Goose.
    Fixed: opencode (shape and path), Kilo (kilo.jsonc), Devin (macOS path,
    windsurf-next), Cline (CLI and every editor), the Claude Desktop MSIX
    install, Continue (YAML), and Codex. Codex's config.toml is now edited
    table by table, so the Codex app rewriting the file no longer drops Sensei's
    settings.
  • Claude Code uses a hook by default, not a base URL. A PostToolUse HTTP
    hook compresses Bash, Grep, Glob, WebFetch and MCP output before it enters
    the conversation. Read, Edit and Write are never touched. If Sensei is not
    running, the hook does nothing and Claude Code carries on. Remote Control
    and the desktop app's Code tab keep working. Routing the whole conversation
    through the gateway is still available with --route, with a session-start
    hook that clears the base URL if Sensei has gone away. Copilot CLI,
    opencode and Kilo get the same hook approach.
  • Installer fixes. SENSEI_SETUP_TOOLS=no is now honoured; before, the
    running server wired everything 30 seconds later anyway. install.ps1 no
    longer hangs on a real desktop.
  • Remote MCP for clients that only accept HTTPS (ChatGPT, claude.ai
    connectors): bearer token, Host and Origin checks, and a tunnel recipe in
    docs/remote-mcp.md.

Compression for what agents actually read

  • New compressors for grep/rg results, diffs, git output, test and build
    output, directory listings, configs, HTML, minified files and base64 blobs.
  • Lossless folds first. Each fold is verified by a round trip before
    anything lossy runs.
  • Lossy cuts can be undone. Every lossy cut leaves a
    [sensei: …; retrieve ccr=<id>] marker. The id is the content hash, so the
    same input always gives the same bytes.
  • Nothing can grow. A result that comes out larger than the input is
    replaced by the original.
  • File reads stay byte-exact, including code that only looks like prose.
  • An independent review found nine ways the first version could hide
    information, such as diff lines folded away and distinct compiler errors
    merged into one. All nine are fixed and each has a regression test.

A gateway that respects the prompt cache

  • Live zone. Each turn replays what was sent last turn byte for byte and
    compresses only the new messages. That keeps the provider's cache hitting
    and keeps thinking-block signatures valid. Cache misses are attributed
    (cold start, TTL, prefix change). Preparing a request takes 15 ms of CPU,
    against 84 ms in 0.2.
  • sensei_retrieve, answered by the gateway. When the model asks for an
    original, the gateway fetches it and continues the upstream call, so the
    client never sees the extra round trip.
  • Tool definitions are compacted safely; Anthropic tool search is opt-in.
  • Repeated tool output across turns becomes a one-line reference.
  • Opt-in extras: a model router, an exact response cache, and per-tool
    budgets.

Honest numbers

  • The README's 79% is gone; it came from six hand-written samples. On 430 real
    Claude Code and Codex sessions, Sensei removes 3.9% of tool-output
    tokens. That is 0.68% of the bill once prompt caching is priced in
    (Headroom 0.41 on the same data: 3.6% and 0.56%).
  • On a seeded corpus of realistic agent tool output, Sensei removes 37.9%
    (Headroom: 18.1%) and keeps 74 of 74 facts.
  • cd backend && python -m benchmarks.refresh regenerates every number,
    graphic and README table. Nightly CI fails on regressions or a stale README.

Install

One command, no Python needed — downloads the binary for your machine
and checks it against SHA256SUMS before installing:

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/SenseiIssei/Sensei/main/install.sh | sh
# Windows
irm https://raw.githubusercontent.com/SenseiIssei/Sensei/main/install.ps1 | iex

Or take a single file yourself:

Platform File
Windows 10/11 x64 sensei-windows-x64.exe (or .zip)
macOS (Apple silicon) sensei-macos-arm64.dmg (or .tar.gz)
macOS (Intel) sensei-macos-x86_64.dmg (or .tar.gz)
Linux x86_64 sensei-linux-x86_64.AppImage, .deb, .rpm, .tar.gz
Docker ghcr.io/senseiissei/sensei-backend:latest
VS Code sensei.vsix

Then:

sensei up            # start the server, open the web UI
sensei setup-tools   # point the AI tools on this machine at it

These builds are not code-signed

Sensei has no Apple Developer account or Windows code-signing
certificate yet, so:

  • macOS shows "cannot be opened because the developer cannot be
    verified"    . Right-click the app and choose Open, or run
    xattr -d com.apple.quarantine /path/to/sensei.
  • Windows shows a SmartScreen warning. Click More info ->
    Run anyway    .

Verify what you downloaded instead of trusting the warning dialog:

sha256sum -c SHA256SUMS --ignore-missing

Every artifact is also signed with Sigstore. Verify one with:

pip install sigstore
sigstore verify identity \
  --cert-identity-regexp 'https://github.com/SenseiIssei/Sensei/.*' \
  --cert-oidc-issuer https://token.actions.githubusercontent.com \
  sensei-linux-x86_64.AppImage

What's Changed

  • Integrations: switch on once, connect to everything (Claude Code hook mode, one switch, installer fixes) by @SenseiIssei in #161
  • Cache-aware gateway: live zone, server-side sensei_retrieve, tool compaction, dedup by @SenseiIssei in #162
  • Compression for coding-agent traffic: search, diff, git, test/build, listings, lossless tables, retrievable markers by @SenseiIssei in #164
  • bench: real, reproducible numbers — agent corpus, session replay, Headroom head-to-head by @SenseiIssei in #163
  • Release 0.3.0 by @SenseiIssei in #165

Full Changelog: v0.2.0...v0.3.0

Contributors

SenseiIssei
Assets 50
Loading