Context Harness v0.7.0 is about making the project feel less like a promising CLI and more like local-first infrastructure you can trust: faster semantic retrieval when you have embeddings, cleaner storage boundaries, predictable config locations, organized project docs, and no telemetry.

The release pulls together the work from v0.6.0 through today: the workspace split and documentation foundation, the vector-index design work, XDG config compliance, the zvec sidecar implementation, and the telemetry removal.

Release notes

The latest release is v0.7.0, published on June 8, 2026.

What’s changed since v0.6.0:

Download:

# Apple Silicon macOS
curl -L https://github.com/parallax-labs/context-harness/releases/latest/download/ctx-macos-aarch64.tar.gz | tar xz
sudo mv ctx /usr/local/bin/

# Linux x86_64
curl -L https://github.com/parallax-labs/context-harness/releases/latest/download/ctx-linux-x86_64.tar.gz | tar xz
sudo mv ctx /usr/local/bin/

Windows users can download ctx-windows-x86_64.zip from the v0.7.0 release.

Faster semantic search with zvec

Context Harness still uses SQLite as the canonical local store: documents, chunks, checkpoints, FTS5 keyword search, embedding metadata, stats, and export all live there. v0.7.0 adds a separate optional acceleration layer for semantic search: a rebuildable zvec vector-index sidecar.

That means the storage model is now explicit:

.ctx/data/
  ctx.sqlite             # canonical store
  vector-index/zvec/     # derived semantic-search sidecar

The sidecar is derived state. If it is missing, stale, or unavailable, Context Harness can rebuild it from SQLite or fall back to the exact SQLite vector scan. Keyword search continues to use SQLite FTS5.

The benchmark that motivated the sidecar is stark: on a real corpus shape with 4,917 chunks, zvec HNSW candidate retrieval measured around 0.8 ms median compared with roughly 112 ms for the exact SQLite scan, with high top-k overlap against the SQLite baseline. The full write-up is in Vector Search Bake-Off: SQLite Brute Force vs zvec.

The important design point is not “replace SQLite.” It is: SQLite owns truth; zvec owns speed.

Storage boundaries are now part of the contract

v0.7.0 introduces the app-level AppStore and optional VectorIndex boundary. This is plumbing, but it matters: the reusable core search flow can keep working against the Store trait, while the native application owns operational storage concerns like initialization, checkpoints, embedding maintenance, stats, and export.

The new boundary keeps the responsibilities sharp:

This gives Context Harness room to improve retrieval performance without turning the canonical local database into a grab bag of special cases.

XDG config and cleaner workspace state

Context Harness now has a clearer directory model for CLI users. New workspaces prefer .ctx/config.toml, .ctx/data/, and .ctx/cache/; global defaults can live under the usual XDG paths such as ~/.config/ctx/config.toml.

The config resolution order is designed to be predictable and backward-compatible:

  1. --config <path>
  2. CTX_CONFIG
  3. ./.ctx/config.toml
  4. ./config/ctx.toml
  5. $XDG_CONFIG_HOME/ctx/config.toml
  6. built-in defaults

Existing deployments that pass --config or use config/ctx.toml keep working. The change gives new projects a less noisy default and gives multi-workspace users a place for global defaults.

Telemetry removed

v0.7.0 removes product telemetry entirely. No analytics client, no product event stream, no local analytics identity, and no PostHog dependency.

Context Harness often indexes source code, project notes, runbooks, planning docs, and other private context. Even anonymous analytics creates a networked reporting path that weakens the local-first trust boundary. The rule is simpler now:

Context Harness will stay telemetry-free forever.

Network access still happens when you explicitly configure networked behavior, such as Git, S3, OpenAI embeddings, Ollama, model downloads, or registry installation. Product telemetry is different, and it is gone. The longer explanation is in Telemetry-free forever.

Documentation grew up

The docs are now organized around how the project is actually built:

That structure is already paying off. The vector-index work has an explicit storage contract, the XDG work has a config-resolution spec, and release/debug workflows have places to live besides issue comments and memory.

Upgrading

Use the latest release binary for your platform, or install from the repo:

# From a release asset
curl -L https://github.com/parallax-labs/context-harness/releases/latest/download/ctx-macos-aarch64.tar.gz | tar xz
sudo mv ctx /usr/local/bin/

# From source
cargo install --path crates/context-harness --force

No config migration is required. Existing --config and config/ctx.toml workflows remain supported. New workspaces can move toward .ctx/config.toml when convenient.

What this release means

v0.7.0 is not just a feature drop. It tightens the shape of Context Harness:

That is the direction from here: local-first context infrastructure that stays boring where it should be boring, fast where speed matters, and honest about where your context goes.