fable-mode is a small tooling + prompt bundle, not a versioned library. Security fixes land on the main branch — please run the latest main.

Installing fable-mode wires two hooks into Claude Code that run on your own machine, with your permissions:

• fable-trigger.py (UserPromptSubmit) — reads ~/.claude/FABLE_PLAYBOOK.md and injects it into the prompt context. It does not execute project code and makes no network calls.
• test-after-edit.py (PostToolUse on Edit/Write/MultiEdit) — runs your project's own test command (npm test, pytest, cargo test, go test, make test) automatically after a code edit, to report pass/fail. Editing a file can therefore trigger execution of that project's test suite.

Neither hook sends anything over the network. They read/write only under ~/.claude and the system temp dir (debounce/marker files), and run the detected test command in the edited project's directory.

Because the test hook runs a project's test command, only enable fable-mode in repositories you trust — the same caution you would apply to running their tests yourself. The command is derived from repo content (package.json scripts, a Makefile test: target, or a .fable-test file), so a hostile repository could point it at an arbitrary command; treat it exactly like running that repo's build yourself.

Set FABLE_TEST_HOOK_ALLOW to an os.pathsep-separated list of directory prefixes (: on Unix, ; on Windows). When it is set, the hook auto-runs only for projects at or under one of those roots and stays silent everywhere else — so editing a freshly-cloned, untrusted repo triggers nothing. When it is unset the default is zero-config (runs in any project), which is why the trust caution above applies.

• Set FABLE_NO_TEST_HOOK=1 to disable it entirely.
• Set FABLE_TEST_HOOK_ALLOW=<trusted dirs> to restrict it to roots you trust.
• Tune FABLE_TEST_HOOK_DEBOUNCE / FABLE_TEST_HOOK_TIMEOUT (seconds).
• Drop a .fable-test file in a project root to pin the exact command it runs (e.g. a fast, scoped subset instead of the whole suite). Note this file is repo content, so it is governed by the same FABLE_TEST_HOOK_ALLOW gate.
• Or remove the PostToolUse entry from ~/.claude/settings.json — uninstall.py does this for you.

reference/fable-system-consumer.md is Anthropic's leaked Claude Fable 5 consumer system prompt, kept for reference/provenance only — it is never installed or injected. It is third-party content, not authored or audited here, and is removable on request. The skills under skills/ (webapp-testing, mcp-builder, skill-creator, explore-data) are vendored from upstream Apache-2.0 repos. Treat all of it as untrusted-origin text.

Please report security issues privately, not in a public issue:

• Use GitHub's "Report a vulnerability" (the repo's Security → Advisories tab), or
• if private advisories are disabled, open a minimal public issue asking for a private contact channel — do not include details there.

Include what you found, how to reproduce it, and the impact. We'll acknowledge the report and work on a fix; please allow a reasonable window before public disclosure.