[gabe4coding]

Summary

This change implements the 13 findings of a hardening audit. Each fix moves a class of mistake from code review to a check that fails in CI.

• Types: statuses, levels, verdicts and actions are Literal types. pyright (standard mode, exhaustive match) now runs in make check.
• Exit codes: one ExitCode enum. A usage error exits with 1.
• Findings: Finding is frozen, and blocking is computed from the corpus and the verification. Only validate_findings and read_findings make a Finding (new findings.py).
• Processes: git.py and claude.py are the only modules that start processes.
• Module layers: import-linter checks that deterministic modules never import a model module, and that only shall.jev imports the TypeSafe SDK.
• Config: unknown keys and bad hooks.actions rows in the local shall.yaml stop shall at load. In the extended org config, shall drops a row that it cannot read and writes a note to stderr, so the hooks do not fail open.
• Models: the agents and the review model are pinned to claude-sonnet-5, and a test checks that they agree.
• Test fakes: tests/fakes.py has a FakeJev that pyright checks against JevLike.
• RFCs: each JTOOL RFC lists the test or check for each statement.
• CI: a new workflow runs make test and make check, and it pins the Claude Code CLI.

The plugin version is now 0.4.1, because the agents changed.

Code-review fixes

• shall list and shall show printed "None" for an invalid status. They now print "invalid" (Rfc.status_label).
• shall eval-verify ignored review.verify_model. Now Config.verify_model is the one rule, and verify and triage use it too.
• load_json crashed on a cache file that is not a JSON object.
• One unreadable org hooks.actions row made the whole hook fail open.
• CI installed an unpinned Claude Code CLI.
• pyright downloaded Node at run time. It now uses pyright[nodejs].

The self-review also found that run_claude retried claude -p at once (JRFC-0006.2). It now waits with exponential backoff and jitter.

Verification

• make test: 200 passed.
• make check: corpus lint, pyright 0 errors, import-linter 2 contracts kept, index current.
• make self-check: OK.
• make test-plugin: 7 passed.
• make self-review: 0 findings.
• /verify, through headless Claude Code with the plugin:
  • a force push was denied by JRFC-0012.2, with and without hookd;
  • a write of an AWS example key was denied by the scanner (JRFC-0004.1);
  • a file write, then Stop, gave no action and a fail-open Stop without a Jev key;
  • CLI probes showed that a bad config row stops shall or is dropped with a note, and that list and show show "invalid".

Not in this change

The self-review gave two advisory findings, JTOOL-0002.6 and JTOOL-0004.2. They are about code that existed before this change and only moved to findings.py. A separate task is fixing them.

Known friction: shall writes the note about dropped org rows each time it loads the config. A hook loads the config on each tool call, so the note repeats.

🤖 Generated with Claude Code