* security(skills): drop bare Bash and the portal CLI glob from skill grants
A skill's allowed-tools pre-approves the listed tools for the turn that
invokes it, even in an untrusted folder. job-application-assistant granted
bare Bash, Edit and Write; job-scraper granted bun run on any
.agents/skills/*/cli/src/cli.ts. Every Bash command the skills document is
already allowlisted per command in .claude/settings.json, so the bare and
glob grants bought nothing but exposure. The per-portal grants stay in
settings.json, which is where a portal added by /add-portal registers too
(Step 5 now says so).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* security(settings): deny exfil-shaped shell and credential reads, always ask on permission-defining files
Deny rules block in every permission mode including bypassPermissions and
override any allow rule or skill grant; ask rules force a prompt even when
a skill's allowed-tools would pre-approve the call. settings.json now
ships both: deny for shell that uploads or opens a remote shell (curl
data/form/upload flags, wget post, scp, sftp, ssh, nc) and for reads of
.env and ~/.ssh, ~/.aws, ~/.gnupg; ask for the files whose content is the
permission surface (settings, skill frontmatter, agents, CI, this guard)
and for git push. security_guards.py requires each rule to stay present,
the same way it requires the personal-data gitignore rules. Prefix rules
are tripwires, not walls; SECURITY.md will say so.
Also allowlists tools/robots_check.py and the upcoming tools/fetch_page.py
so the 403 retry no longer needs a curl grant.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* security(guards): hold skill and command allowed-tools to a reviewed allowlist
The guard allowlisted settings.json permissions but never read skill
frontmatter, so the narrow settings allowlist sat next to a bare Bash
grant it could not see. Check 4 parses every .claude/skills/*/SKILL.md,
.agents/skills/*/SKILL.md and .claude/commands/*.md frontmatter with a
stdlib parser (scalar, folded, and list forms) and holds each
allowed-tools entry to ALLOWED_SKILL_TOOLS; a portal skill may
additionally grant exactly its own cli/src/cli.ts, matched by path so no
portal name is hardcoded. Bare Bash and mid-pattern wildcards get their
own message. On master's frontmatter this check reports two failures.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* feat(tools): add GET-only fetch_page.py for the browser-header retry
09-web-research.md documented the 403 retry as a curl command with browser
headers plus an inline python3 -c tag-stripper, which meant the workflow
wanted a Bash(curl:*) grant - and curl with a free tail is an upload
channel. fetch_page.py is the same retry as one command with a fixed GET
argv (URL after --, bounded to 45 s / 5 redirects / 10 MB), http(s) only,
robots_check.gate() enforced with no bypass flag, --out refused inside
the repo, and --text for the stripped page. urllib fallback when curl is
absent. The settings.json grant names the script.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* docs(security): describe the permission surface as a table; route the 403 retry through fetch_page.py
SECURITY.md now lists every shipped grant with its location, lifetime
and guard, and says plainly what the permission layer does not do: no
sandbox on native Windows, prefix rules are tripwires, WebFetch is
unrestricted by domain, subagents run with the parent's tools (next
change), and Edit/Write stay granted because tailoring a CV is editing
files. 09-web-research.md's retry section is the one fetch_page.py
command (framework_version 1.1.3 -> 1.1.4); apply, rank and scrape point
at it instead of curl. SETUP.md and the README tools tree follow.
CHANGELOG gains a Security section under [Unreleased].
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* security(review): ask on command frontmatter, portal skills grant only their CLI, bound the urllib fallback
Findings from a fresh-context review of the branch:
- A command's frontmatter is a grant too (check 4 already parses it), so
Edit(.claude/commands/*.md) joins the required ask rules; without it a
steered session could write `allowed-tools: Bash` into apply.md with no
prompt and the next /apply would be pre-approved for bare Bash.
- SECURITY.md promised a portal skill may grant "its own CLI and nothing
else"; the guard now enforces exactly that instead of own-CLI plus the
core allowlist. All six shipped portal skills already comply.
- fetch_page.py's urllib fallback now matches curl's 5-redirect cap and
applies the 10 MB cap after decompression, so the docstring's bounds
hold on both paths.
- Frontmatter parser: a BOM no longer reads as "missing frontmatter",
YAML flow lists split correctly, and a non-UTF-8 file fails with a
message instead of a traceback. Module docstring lists check 4.
- CHANGELOG no longer claims "no new prompt": compile and aux cleanup,
which the removed bare Bash covered in a skill-invoking turn, now
prompt, deliberately. "Once per session" wording corrected to per call.
- Three upload-shaped deny siblings added (curl -XPOST, curl --request
POST, wget --method=POST); /add-portal recon points at fetch_page.py.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
.gitignore's `node_modules/` rule matched directories only, so a symlink
named node_modules pointing at a contributor's absolute path was committed
(#623), broke `bun install` in CI with a dangling link and checked out as a
text file on Windows. The rule loses its slash, and security_guards gains a
check that no tracked path has a node_modules component (skipped quietly
outside a git checkout). Also qualifies the one bare `seen_jobs.json`
mention left in README.md after #617.
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
The two-line comment above `documents/interview/**` says interview prep and
experience records live there. Nothing has ever written to that directory:
/interview saves its pack to
documents/applications/<company>_<role>/interview_prep_<stage>.md, covered by
the documents/applications/** rule. `git grep documents/interview` returns
only the two declarations of the rule itself (.gitignore and
REQUIRED_IGNORE_RULES), `git log --all -- 'documents/interview*'` is empty,
and documents/README.md documents the applications path outright.
Nothing leaks - the comment is the defect, and it is the misleading kind. It
is the one dedicated, well-argued line about interview material in the
personal-data block, so an auditor checking that the framework's most
sensitive artifact is covered reads it and stops, at the only path in the
block with no writer.
The comment's description of what needs protecting was always right; only its
location was wrong. It now sits above documents/applications/**, the rule that
actually provides that protection, so a reader auditing the block finds the
reasoning attached to the rule doing the work. documents/interview/** stays -
REQUIRED_IGNORE_RULES pins it, so dropping it from .gitignore alone turns CI
red, and it is harmless defence in depth - relabelled in both files as
belt-and-braces rather than the primary guard.
The new check-ignore case in GitignorePatternBehaviorTests derives the
prep-pack path from /interview's own spec instead of hardcoding it. That
distinction is the whole value of the test: a hardcoded path pins only that
documents/applications/** still matches that shape, which security_guards.py
already catches first, and stays green if /interview moves its output -
leaving the corrected comment stale exactly the way this issue found it.
Since #329 the spec states the location in two pieces - Step 1 derives the
archive folder, Step 3 names interview_prep_<stage>.md - so the test pins both
fragments separately and composes the concrete path from them. Mutation-
verified on each half: repointing the folder at documents/prep_packs/, and
renaming the file, both fail this test while `python3 tools/security_guards.py`
still reports OK.
The class's temp-repo setup moved to setUp for the second case.
ayobamiseun reviewed the pre-rebase branch and called all three rebase hazards
in advance: the split literal, the released CHANGELOG context, and the setUp
re-merge. Reached independently here during the rebase; the review was posted
first.
The upskill/*.md rule is rooted, but /upskill is a skill and skills
resolve bare relative paths against their own directory - the same
observed behavior the **/job_scraper rules exist for. A report at
.claude/skills/upskill/upskill/report-*.md was not ignored, and an
upskill report records the candidate's skill gaps against named
employers. **/upskill/*.md would also ignore the skill's own SKILL.md
(the directory shares the name), so the new rule pins the report-file
prefix: **/upskill/report-*.md. Added to .gitignore and
REQUIRED_IGNORE_RULES, with a check-ignore-based test pinning both
properties.
check_permissions() read permissions.allow and nothing else, so a `hooks`
block in the same file passed the guard silently.
A hook is strictly more dangerous than a pre-approved permission. A
permission pre-approves something Claude may choose to do; a hook runs
unconditionally when its event fires, with no prompt and no model decision
in between. Cloning the repo and opening it is enough.
This is the vector the Shai-Hulud worm used in its August 2026 wave: a
SessionStart hook in .claude/settings.json chaining to .claude/math_init.js,
executing on session start.
https://research.jfrog.com/post/shai-hulud-is-back-august/
For a template that thousands of people are explicitly invited to fork,
that is the riskiest key in the file this guard already parses.
Follows the established pattern exactly - ALLOWED_HOOKS ships empty, since
the template has no hooks, so any addition must be allowlisted in the same
PR and is therefore explicit and reviewable.
Two details worth reviewing closely:
- The hook check runs *before* the permissions shape guards. Those guards
return early, so a file pairing a malformed permissions block with a live
hook would otherwise skip the hook check entirely - a fail-open. Pinned by
test_hook_is_caught_even_when_permissions_block_is_malformed.
- _hook_commands() fails closed. Any hook layout it does not recognise
yields a marker that cannot be in the allowlist, so an unfamiliar shape is
rejected rather than silently skipped, rather than trusting that the
Claude Code schema will not change.
Verified:
- 8 new HookGuardTests cases; 14 of the suite's 26 tests fail against the
unpatched guard, all 26 pass with it
- injecting the real worm shape into this repo's own settings.json makes
the guard exit 1 naming 'SessionStart:node .claude/math_init.js';
removing it returns OK
- lint_skills, check_framework_version, security_guards all OK;
python3 -m unittest discover -s tests 219 passed
check_gitignore() verified each required personal-data rule was present via set membership, but .gitignore is order-sensitive: a later !pattern re-includes a file an earlier rule excluded, so the required line stays physically present while the file is no longer ignored - the guard failed open on exactly the weakening its docstring claims to catch. Keeps the required-rules-present check and additionally rejects any negation line outside a small reviewed ALLOWED_IGNORE_NEGATIONS allowlist (same explicit-widening pattern as ALLOWED_PERMISSIONS). Fixes#194.
By @thejesh23. Verified: allowlist matches the four negations currently in .gitignore; guard test suite passes locally (17 tests) and in CI.
Closes#194
* ci: run the Python test suite - CONTRIBUTING.md asks for tests CI never executes
CONTRIBUTING.md tells contributors to put Python tool tests in tests/
and run the relevant suites, and tests/ now holds real ones
(test_salary_lookup.py, test_convert_salary_excel.py from #75) - but no
CI job executes them. A suite that never runs in CI can't gate a PR and
silently rots. New python-tests job: unittest discover over tests/,
stdlib only, no new dependencies. Future test files run without any
workflow change.
Also lands tests/test_security_guards.py, which missed #84's merge
window (pushed to the branch as #84 was being merged; the merge took
2a6cb8c, the tests were 260c37a). 13 unittest cases in the existing
tests/ style: each copies the guard script into a synthetic repo tree
and runs it as a subprocess - the same way CI invokes it - asserting
real exit codes and messages. Every forbidden state fails (Bash(*) and
Bash(curl:*) additions, each personal-data gitignore rule removed one
at a time, each forbidden lifecycle script, trustedDependencies,
invalid settings JSON, zero manifests); every non-event passes (dropped
shipped permission, extra ignore rules, benign scripts, hostile
manifest inside node_modules); and the real repo passes its own guards.
22 tests total, all passing locally via the exact command the job runs.
* test: use benign lifecycle-script values in fixtures - AV heuristics flag attack-shaped strings
Review found the curl-pipe-to-sh fixture value matches a real Defender
signature (Trojan:Script/Stealer.HAX!MTB): Windows quarantines the temp
package.json mid-test, making the suite flaky for any Windows
contributor who runs it - while proving nothing extra, since the guard
flags the script KEY and never inspects the value.
Fixture values are now 'echo test' (also in the node_modules-ignored
test, same class of string), with a comment on the key-only test
explaining why the value must stay benign so a future 'make the fixture
realistic' cleanup doesn't reintroduce the quarantine flake. Coverage
is unchanged: same keys, same assertions, 13 tests passing.
