Skip to content

v0.2.7-alpha

Choose a tag to compare

@immane immane released this 10 Nov 07:55
· 37 commits to main since this release
8b44b58

Nexus Agent 0.2.7-alpha

Alpha release rebuilt from main commit 62ff8eb9368b4a7703dcd32fdae30f6df3d3d4c7 (merge of PR #35) with rustc and cargo 1.99.0.

Highlights since the first 0.2.7-alpha publication

  • Adds a default-inherit task-thinking configuration at global, provider, and model scopes. Resolution order is selected model → provider → global → adapter default. Existing configurations and provider requests remain unchanged unless an explicit override is configured.
  • Adds Web Settings → Thinking for the global default and provider/model overrides. On (low effort) and Off are explicit requests, not guarantees: low effort may merely enable thinking on boolean-only models.
  • Supports explicit task-thinking overrides only on detected Ollama Direct / Chat Completions endpoints. An effective unsupported override fails explicitly instead of being silently ignored.
  • Keeps reasoning visibility, historical reasoning replay, and the summary low-effort hint independent of the task-thinking setting.
  • Adds metadata-only context-budget diagnostics for pre-reduction, protected-floor, summary-candidate, retained-view, and final-dispatch stages. They report payload categories without logging transcript content.
  • Adds opt-in benchmark --read-page-bytes guidance for bounded file paging, narrow searches, and selected command output. Default prompts and global tool limits are unchanged.
  • Documents the 16K small-context pressure experiments, including the final exposed SymPy regression that completed two summary reductions and was officially resolved. It is one regression result, not a general success-rate claim.

Downloads

Platform Archive SHA-256
macOS arm64 nexus-agent-0.2.7-alpha-aarch64-apple-darwin.tar.gz 4ac5e6a5ee3f83634ec1841db712be6569af529694d2e4b4c88831e43a32d324
Linux x86_64 nexus-agent-0.2.7-alpha-x86_64-unknown-linux-gnu.tar.gz 98c3a63295386af4cfce1a0ce23d2540f441c4eac293cba010a8c0c8e4665dbb

Each archive has a matching .sha256 sidecar. Verify it before extraction:

# macOS
shasum -a 256 -c nexus-agent-0.2.7-alpha-aarch64-apple-darwin.tar.gz.sha256

# Linux
sha256sum -c nexus-agent-0.2.7-alpha-x86_64-unknown-linux-gnu.tar.gz.sha256

Install

tar -xzf nexus-agent-0.2.7-alpha-<target>.tar.gz
cd nexus-agent-0.2.7-alpha-<target>
python3 install.py "$HOME/.local/bin"

The archive contains nexus-tui, nexus-server, nexus-headless, nexus-config, the installer, LICENSE, and a per-file checksum/provenance manifest. System OpenSSL and the platform sandbox backend are runtime dependencies; see the Quickstart for details.

Verification

  • Both archives were built from a clean main checkout at 62ff8eb9368b4a7703dcd32fdae30f6df3d3d4c7 (source_dirty: false, rustc 1.99.0).
  • The macOS arm64 archive was built natively on macOS. The Linux x86_64 archive was built in an amd64 Debian bookworm container with the pinned Rust 1.99.0 toolchain.
  • Extracted installations passed command smoke checks for nexus-tui, nexus-server, nexus-headless, and nexus-config on macOS arm64 and Linux x86_64.
  • PR #35 passed fmt, Clippy, docs, performance, CodeQL, and the Ubuntu/macOS test jobs before merge.
  • Archive checksums are recorded above and in the uploaded sidecars.

Scope and limitations

  • This is Alpha, not a Beta, stable, production-readiness, or security-certification claim.
  • The Web client is loopback-only. It can remember the bearer token in browser storage; treat that browser profile as credential-bearing. Loopback HTTP has no TLS and does not isolate mutually untrusted same-user processes.
  • Saved-session listing, history, restore, and deletion require server --session-store. Active runs are not reattached after reload.
  • The macOS build is native arm64. The Linux build is x86_64 and smoke-tested in Debian bookworm; it is not a general Linux distribution/filesystem validation.
  • Binaries are not signed or notarized. Checksums detect corruption, not authenticity.
  • Deliberately detached descendants and sandbox escapes are outside the documented cleanup guarantee; their effects remain uncertain.
  • Linux hardening beyond the recorded reference profile remains unverified. Container execution is a prototype, not an approved isolation boundary.
  • Task-thinking support is currently limited to detected Ollama Direct / Chat Completions endpoints. A request is not proof of model compliance, and no strict reasoning-token or task-quality guarantee is made.
  • Headless remains a scripted fake runner and must not be presented as a live-provider execution frontend.

Full changelog: v0.2.6-alpha...v0.2.7-alpha