Provider-neutral tools shared by the Claude Code, Kimi and Codex agent harnesses. Each one lived inside a private harness bundle or a private deployment first, where CI cost billed minutes and in practice it was verified only on whichever machine last ran its suite by hand. Out here the matrix is free, so every tool is tested on every push against as much of its supported surface as it actually has.

These are not on PyPI. Each project publishes a wheel on its release tag, with a SHA256SUMS file beside it, and that checksum is the point: a consumer fetching "the newest release" has no other way to prove it got the artifact CI built. Verify before installing.

gh release download v1.0.0 --repo Nitjsefnie-Harness-Commons/<project>
sha256sum -c SHA256SUMS
pip install ./<wheel>.whl

Every project here follows the same shape. Where one of them does not, it is named, because a house rule with a silent exception is worse than no rule.

• A runtime dependency only where the job needs one. session-parsers, git-coauthorship and usage-query are standard library only — the parsers are imported inside hooks that run on every tool call, where a dependency would be paid for on every one of them. discord-mailbox needs discord.py and psutil; session-archivers needs boto3. daedalus and docs-hub are each a server plus a client, and each ships a client wheel that declares no dependencies at all: daedalus keeps its MCP server behind a [mcp] extra, and docs-hub keeps its server requirements out of the wheel entirely, because nobody installing a client should be made to install a web framework.
• The suites are standard library, run by python3 run_tests.py, with no pytest — in every project except docs-hub, which has a Postgres-backed server to test and runs pytest against a real database.
• The same gates on every push: tests, lint (pycodestyle + pylint), types (pyright), actionlint, audit and codeql. daedalus and docs-hub add eslint for the JavaScript they ship, and daedalus also runs speed, which times a commit against its own base and discards a warm-up round so the comparison is not measuring the runner.
• The test matrix is as wide as the thing under test. Five of these run Linux, macOS and Windows × Python 3.11, 3.12 and 3.13. daedalus adds 3.14, for twelve legs. docs-hub runs Linux and Python 3.13 against Postgres 16 and 17, because a single database version is a claim about the other one.
• A wheel probe on everything whose product is the wheel: the built artifact is installed into a venv with no checkout in reach and exercised there, so a packaging mistake fails CI rather than the first person to install it. docs-hub is the exception again — what it ships is a running service, and its client wheel rides along.
• MIT.