The previous filepath.IsLocal-up-front pattern in safeJoin/safeJoin-equivalents
turned out not to satisfy CodeQL's go/path-injection rule — the post-validation
filepath.Join still constructs the joined string from tainted input, so the
analyser conservatively assumes the os.* sink that consumes it is tainted
too. Only one of 34 alerts closed on the previous attempt.
Switch to *os.Root (Go 1.24+, available on the project's 1.26.3 toolchain).
The Go runtime guarantees that operations on a Root cannot escape the
anchored directory regardless of what's in the relative path, and CodeQL has
a built-in model that recognises *os.Root.* methods as path-traversal
sanitisers. Result: every os.* sink in the datastore, marge, recorder,
mirror parity-mismatch writer, and docs handler is now reached only via a
*os.Root, which closes the rule-level alerts cleanly.
Changes per file:
* pkg/service/datastore/datastore.go — Adds a `root *os.Root` to DataStore,
lazily opened at first use (after MkdirAll-ing baseDir) and closed by a
new `(*DataStore).Close()`. Adds package-private helpers
(rootStat / rootReadFile / rootWriteFile / rootMkdirAll / rootRemove /
rootRemoveAll / rootRename / rootReadDir / rootOpen / rootExists) plus
three exported wrappers (ReadDirUnderBase, MkdirAllUnderBase,
WriteFileUnderBase) for the cross-package marge / handlers callers.
Every os.* call that previously consumed safeJoin output now goes through
these helpers. The post-join belt-and-suspenders prefix check inside
safeJoin is preserved as a defence-in-depth fallback.
* pkg/service/marge/marge.go — Replaces the five `os.ReadDir(devicesDir)`
call sites with `ds.ReadDirUnderBase(...)` so the datastore's root
enforces containment.
* pkg/service/proxy/recorder.go — Mirrors the datastore pattern with its
own `root *os.Root` anchored at Recorder.BaseDir, lazily opened. New
helpers convert the eight existing `os.*` sites that consume sessionID
/ relPath / sanitizedSegments inputs. The earlier safeJoin (filepath.IsLocal
pre-check) stays in place as the same belt-and-suspenders guard.
* pkg/service/handlers/handlers_docs.go — Opens a *os.Root at "docs" via
sync.Once and reads file content (and SUMMARY.md sidebar) through it.
Removes the prior filepath.IsLocal pre-check; the runtime now guarantees
containment.
* pkg/service/handlers/mirror_middleware.go — Routes the parity-mismatch
JSON write through `s.ds.WriteFileUnderBase` so the datastore's root
performs the path-traversal sanitiser.
Behavioural fix: *os.File.ReadDir(-1) returns directory entries in
filesystem order, but os.ReadDir is documented to sort by name and at least
one regression test
(handlers.TestMargeAccountFullExcludesEmptyAmazonSource) depends on the
sorted contract. Both rootReadDir helpers explicitly sort by name to match.
All test suites pass for the touched packages; the unrelated
TestDocsConsistency failure about untracked working-tree docs is
pre-existing. golangci-lint reports 0 issues.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
CodeQL flagged 34 go/path-injection alerts across datastore.go, marge.go,
recorder.go, handlers_docs.go and mirror_middleware.go. The existing
defences (DataStore.safeJoin's post-join prefix check, handlers_docs's
HasPrefix(filepath.Clean(...))) are functionally correct but sit
downstream of the join, so CodeQL's interprocedural taint tracking
treats every os.* sink that consumes them as still tainted.
Move the validation up-front using filepath.IsLocal, which CodeQL
recognises as a path-traversal sanitiser. IsLocal rejects absolute
paths, ".." segments, and (on Windows) reserved device names — the
same set the existing checks intended to block, just expressed in the
shape the analyser understands.
Changes:
* DataStore.safeJoin (datastore.go) — pre-validates each non-empty
element with filepath.IsLocal before joining. Existing post-join
prefix check stays as belt-and-suspenders. ~30 of the 34 alerts
flow through this helper.
* Recorder (recorder.go) — adds a new (*Recorder).safeJoin method
with the same sanitiser. getRecordingDir, DeleteSession,
GetInteractionContent and ArchiveSession route through it; their
signatures already returned error so plumbing it through is local.
* HandleDocs (handlers_docs.go) — replaces the post-join HasPrefix
check with an up-front filepath.IsLocal gate.
* Mirror parity recorder (mirror_middleware.go) — also strips
backslash separators (Windows) and gates the resulting filename
component on filepath.IsLocal, falling back to "invalid" rather
than letting malformed paths reach os.WriteFile.
No behaviour change for legitimate inputs (account IDs, device IDs,
session IDs, doc paths all satisfy IsLocal). Datastore and proxy
test suites pass; handler suite's pre-existing TestDocsConsistency
failure is unrelated.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Update `getRecordingPath` to use a timestamp format that includes the date (`20060102-150405.000`).
- Update `parseInteractionFile` and `getFullTimestamp` to handle both the new filename format and the legacy format for backward compatibility.
- Improved parsing logic to reliably extract date, time, and HTTP method from interaction filenames.
This commit addresses the data race detected in TestRecordMiddleware: - Updated Recorder.Record to clone Request and Response objects (including bodies) before background processing. - Ensures background workers can safely access data after the main request handler has finished. - Enabled synchronous recording in handler tests to ensure deterministic results and avoid race conditions.
This commit addresses the test failures in pkg/service/proxy: - Ensures synchronous recording in tests by setting RECORDER_ASYNC=false. - Adds a Close() method to the Recorder for proper cleanup. - Fixes a panic in TestRecorder_Record_Redaction caused by race conditions.
This commit introduces several key improvements: Performance Optimization (asynchronous recording), Legacy Proxy Control (Soundcork proxy toggle), X-Forwarded-For Sanitization, consistent Soundcork naming across the stack, and various code quality improvements.