Audits every plugin HTTP route in a profile for a browser-trust fence: plugin routes win the web server's longest-prefix match ahead of the /api gateway, so they never see its trust check and must pin the Host to loopback themselves. Grades PASS/WARN/FAIL per route and fails a fence that compares Origin to Host without pinning it (bypassable by DNS rebinding). Ships as a CLI for CI and a route_fence_scan tool.
Install
# from GitHub (first run asks for allowBuilds approval — follow the hint, retry)
dsh plugin --profile web add github:Vladimir-Kryshchenko/dsh-route-fence-linter
Any plugin you install runs third-party code with your own permissions — it can read your files, use your credentials, and reach the network, and tool approvals don’t sandbox it. GitHub-sourced plugins also run build scripts at install time — pnpm blocks those until you allow them, so an install can stop with ERR_PNPM_GIT_DEP_PREPARE_NOT_ALLOWED or ERR_PNPM_IGNORED_BUILDS; dsh prints the exact key to add under allowBuilds in your profile’s pnpm-workspace.yaml, and the install works on the next run. Allowing a build is a trust decision: only install sources you trust, and pin a commit (github:owner/repo#sha).
README
Every plugin route registered on webServer is dispatched by longest-prefix
match, ahead of the host's /api gateway — so it never passes the gateway's
unified trust check. Each plugin author must ship their own browser-trust
fence, and most do not. This linter finds the ones that did not.
Usage
dsh-route-fence scan # $DSH_HOME/profiles/web
dsh-route-fence scan /path/to/profile
Exit codes: 0 clean · 1 at least one FAIL · 2 usage/IO error.
Verdicts
| Verdict | Meaning |
|---|---|
PASS |
The handler gates on a fence that pins the Host before Origin handling. |
WARN |
Either: inside the fence, Origin / sec-fetch-site is read before the Host is pinned; or the registered routes are defined outside the scanned package, so their fence could not be read. Confirm by hand. |
FAIL |
One of: no Host check anywhere; a fence that compares Origin to Host but never pins the Host to loopback (bypassable by DNS rebinding — the attacker controls both headers and they match); or a fence that exists but this handler never calls. |
Bundles whose sources cannot be read are reported as SKIP and counted in the
summary — a skipped bundle is not a clean bundle.
registerFallback is checked too, and reads worse than any route when
unfenced: the fallback seat answers every request no named route matched,
so one unfenced fallback is a wider surface than a whole prefix tree.
Scan scope
The scan covers what a package publishes (its package.json files field);
a package that declares no files is scanned whole. Anything excluded is named
in the output as a note line, never dropped silently.
Scoping by directory name was tried and reverted: skipping anything called
tests/ or examples/ would have let a plugin hide an unfenced route in a
directory with that name and be reported clean. Publication scope cannot be
gamed that way — the loader can only import what was installed.
The rebinding-bypassable shape is graded FAIL, not WARN: it was found live
in a real, widely-installed plugin and confirmed exploitable against a running
profile — a request with Host: evil.example and Origin: http://evil.example
passed the check and executed a state-changing method.
The fence this checks for
The shape the host itself uses (and dsh-better-sidebar/src/trust-fence.ts
copies): pin the Host header to loopback or a configured trusted authority
first, refuse cross-site fetch markers, and only then compare Origin.
Comparing Origin before pinning the Host is not a fence.
Routes defined in another file
A route registered as routes.map(r => webServer.register(r)), where routes
came from buildWebRoutes() in src/web/routes.ts, has no fence anywhere near
its registration — the fence is one module away. Grading such a route by the
text around the registration produced a false FAIL on real plugins, so the
linter follows the value instead.
When the registered argument is not an inline object — a bare identifier, a
factory call build(...), a spread, or the element parameter of a .map /
.forEach / .flatMap over one — the linter resolves that name to where the
handlers are actually written and grades them there. Each handler in the traced
definition is graded on its own; one unfenced handler fails the whole
registration, so a factory that fences four routes and forgets the fifth is a
FAIL, not a PASS.
The walk is fail-closed at every step. It reports PASS only when it found real
handler bodies that gate on the fence. Routes coming from a bare-specifier
import are WARN ("outside the scanned package — check by hand"), never PASS;
so is a relative import that resolves to nothing scanned. Anything else it
cannot establish — a handler: that is a bare function reference, a factory it
cannot find, a chain longer than three module hops — falls back to the previous
FAIL. Import cycles terminate on a visited-set of (file, name) pairs.
Specifier resolution covers what dsh plugins actually write: relative paths
with .ts / .mts / .tsx / .js / .mjs / .cjs or no extension, the
.js-spelled-for-a-.ts-file convention, and directory imports resolved
through index.<ext>. It is not a Node resolver: package.json exports,
imports (#alias), tsconfig paths, and workspace links are not followed —
they resolve to nothing scanned, which is WARN.
Limits
Heuristic over source text, not dataflow analysis. A fence is identified as the smallest function body that both references the Host and constrains it; a handler counts as gated when it calls that fence, receives it as a value, or performs the check inline — cross-file tracing holds a traced handler to exactly that same standard.
Still out of reach: a fence installed by wrapping (register(withFence(route))
where the wrapper is what checks), routes assembled at runtime from
configuration, handlers reached only through a bare-identifier reference, and
route sources behind a non-relative specifier. These read WARN or FAIL, never
PASS. The linter also checks that a fence is present in a handler, not that
it is textually the first statement.
Report false verdicts; they are bugs — a false PASS most of all.
Links
More in this category
toby-bridges/api-relay-audit★ 824
Runs local security audits of AI API relays and LLM proxies from DeepSeek Harness, producing Markdown reports for prompt injection, model substitution signals, tool-call rewriting, error leakage, stream integrity, and profile-gated Web3 risks.
howmp/dsh-pentest★ 374
Authorized pentest mode for DeepSeek Harness — exploration chain, assets and findings with a Web view.
SeaOf0/dsh-redteam-model★ 250
Authorized-security DSH collection: nine work modes (redteam coordinator, pentest, code audit, binary analysis, attack-defense, AV evasion, incident response, cloud security, CTF solving) and fifteen runtime plugins, managed from a settings page with one-click deploy, install, update and uninstall.
PerryLink/dsh-auto-review★ 134
Second-model auto-review on the approval answerer chain: a read-only reviewer subagent returns structured allow/deny verdicts with reasons, fail-closed by default.
PerryLink/dsh-permission-rules★ 113
Claude Code-style declarative permission rules: ordered allow/deny/ask YAML rules matching tool names, arguments, workspace paths, and agent identity on the tools/pre-execute waterfall, with full session-log audit, dry-run mode, and hot reload.
PensiveFei/dsh-secure-audit★ 87
Read-only security and compliance plugin for DeepSeek Harness: prompt-injection detection, Chinese-PII redaction, and a local configuration audit with redacted, reproducible reports.
Community comments
Comments are public GitHub Discussions. Loading them connects to GitHub and Giscus; a GitHub account is required to post.