Cora Code v0.21.1: Install Docs That Agents Can Execute
Cora Code v0.21.1 ships install instructions agents can execute: release archive, checksum verify, no source builds. Tested live: 3,025 symbols indexed in 4.3s and a planted bug caught CRITICAL in 93 seconds.
Most install documentation for CLI tools ends at "download the binary." That leaves the questions agents ask in practice unanswered. Which asset matches this machine? Did the download survive intact? Does semantic search work before the first API call? Cora Code v0.21.1 answers those questions in its own docs, and the release is the first one where our skills instruct agents to install from the release binary instead of building from source.
This is a patch release (v0.21.1, November 17, 2026), but the change lands where agents read. The cora-code and cora-review skills now carry a references/install.md: install with install.sh or the release archive, verify the checksum, never build from source to use cora. A source build lacks the bundled embedding model, which turns meaning-based search off when you work offline. Running cora install afterwards refreshes the cora-written skill copies. The skills also state that reviewing a branch with cora before opening a pull request is required, not recommended.
Why does the wording matter? Because agents follow instructions literally. A skill that says "building from source is recommended" gets a working binary that behaves worse than the release build, and nobody notices until semantic search returns nothing on a plane. Instructions that say "verify the checksum, use the release archive, run cora index" get exactly that.
We ran the install path ourselves
Before writing this post, we installed v0.21.1 from the GitHub release on an ARM64 Linux server and used it on a real repository (our uteke codebase, 325 files, about 121k lines). The steps below are the same ones the new install doc prescribes.
Pick the asset that matches uname -m. On our host that was cora-aarch64-unknown-linux-gnu-v0.21.1.tar.gz. The x86_64 binary fails silently on ARM through Linux binfmt, so checking the architecture first saves a confusing dead end. Compare the tarball's SHA-256 against the release's checksums-sha256.txt before extracting anything. Ours matched on the first byte: bcc2511e7d32....
Then the smoke test, before replacing anything: ./cora --version should print cora 0.21.1. Only after that did we copy the binary over the old install and remove a stale copy that was shadowing it on PATH. The archived archives carry the code embedding model, so no extra download happens on first use. Four checks, four passes:
cora indexon uteke: 3,025 symbols from 324 files in 4.3 seconds.cora brainhybrid search (FTS5 + vector + graph): 20 relevant results in 0.7 seconds, vector signals present. Semantic search worked offline, out of the archive.cora scan --no-llm: 325 files, 121k lines, 50 findings, 7 seconds.cora review --stagedthrough the LLM path: we planted a deliberate bug (add()returninga - b) in a scratch repo, and the review flagged it CRITICAL with the right fix suggestion, verdict BLOCKED, in 93 seconds.
That last number is the honest one. A review that catches a planted logic bug in about a minute and a half is what the pre-PR requirement asks agents to run before every pull request.
Why release binary over source builds
The release archive ships the code embedding model inside the tarball. A source build skips it, and cora then falls back to keyword-only search unless it downloads the 33.5 MB model later. The install doc also covers the failure mode we hit for real: old copies of the binary in ~/.cargo/bin can shadow the fresh install on PATH, and running MCP child processes keep the previous version until the next gateway restart. Kill those children mid-session and you trade a stale binary for a broken tool connection.
The practical sequence is short: install from the archive or install.sh, verify the checksum, run cora install to refresh the skill copies, then cora init and cora index in your project. Everything after that is the ordinary review loop, now with a required cora pass before the PR opens.
Upgrade today: grab the v0.21.1 archive from the GitHub release, verify the checksum, and let the skills do the rest. If you have been building from source, the release build will feel familiar and semantic search will start working offline.