← dashboard

file-transfer-service

model: opus-4-8 python static analysis off 0 PR(s) seen
workspaceninjarmm jira projectNJ review targetsdevelop · main · master · release/* thresholdsseverity ≥ minor · confidence ≥ medium · cap 12 max diff0.4 MB pollevery 300s

What this is

Paired file-transfer relay of the ninja-remote system: HMAC-signed upload and download legs stream a file through memory, never persisted. Python (ftsv2/aiohttp, deployed) + parallel Go implementation (fts_go).

Project primer

Durable orientation injected into every review as .ai-review/project.md.

file-transfer-service — paired file-transfer relay (review primer)

An internet-facing service in the ninja-remote system that relays a file between an
uploader and a downloader leg paired by transfer id — streamed chunk-by-chunk
through memory, never persisted. Authorization per leg = HMAC-SHA256-signed URLs
(nonce + hex timestamp + shared gateway secret from env/AWS SSM), same scheme family as
websocket-switch.

Two implementations live side by side in this repo — review each diff with the
matching language skill (per FILE, not per repo):

live implementations need the same fix.

Architecture (durable map)

download endpoint (streamed GET) keyed by transfer id + signed leg URL, plus /status.

machine (idle → ready → in-progress → completed/failed) and a bounded queue/channel
the uploader feeds and the downloader drains, chunked (~2 MB).

reaps completed/failed/stale pairs.

response headers. Deployment: Docker; secrets via AWS SSM.

What matters most when reviewing here (durable risk areas)

signed material coverage, constant-time HMAC comparison, timestamp expiry actually
rejecting, and path-normalization alternates that could bypass the check. The Python
and Go validators must stay behaviorally identical — a divergence is itself a finding.

moment: pair creation races on the same transfer id, a downloader waiting for an
uploader that never comes (bounded waits, not forever), cleanup reaping a pair the
other leg still uses, entries removed on EVERY exit path. In Go this is also channel
discipline: who closes, double-close panics, goroutines blocked on a channel nobody
drains after a disconnect. In Python: state shared across await points.

full-file buffering in memory, multipart size limits actually enforced, one slow or
dead leg must not pin memory or stall the loop (asyncio) / leak goroutines (Go).

Content-Disposition — header injection / traversal characters must be sanitized;
filesize claims must not be trusted for allocation.

codes, and error shapes — clients treat them as one service.

Security-sensitive zones

Signature validation in both languages; gateway-secret handling (never logged);
filename/header echo; anything that would persist or log transferred content (it must
remain a pass-through relay); unbounded buffering.

Compatibility surfaces (contracts that must not silently break)

Ninja backend, and agents/NinjaRemote clients on OLD versions keep using the current
shape.

(filename/size) that clients parse.

Conventions & gotchas

polling loops with capped retries — keep waits bounded, don't introduce infinite ones.

logrus for logging; config from YAML + SSM lookups.

don't spend findings on it.

at most a minor test finding.

_Orientation only — verify current specifics (names, paths, constants) in the checkout._

Ecosystem map

Shared cross-project relationships, injected as .ai-review/ecosystem.md.

NinjaRemote / NinjaOne project ecosystem (relationship map)

How the monitored repos relate — so a change in one repo is reviewed with its blast
radius in mind (a change upstream of ncpeer can break ncpeer even if it looks local).
Orientation only; verify specifics in code.

NinjaRemote product

speak the NCRP wire protocol (protocol/format changes must stay compatible across
client and server versions).

Cloud-side relays of the ninja-remote system (AWS-hosted, Python/Go)

URLs, it pairs them and forwards frames verbatim. Signed join URLs are minted by the
Ninja backend — the signature scheme is a backend⇄relay contract, and the frame
semantics are a peer⇄peer contract.

a signed upload leg streams to a signed download leg through memory (never stored).
Two parallel implementations (Python ftsv2/ + Go fts_go/) must stay behaviorally
identical; the URL/signature scheme is shared-family with websocket-switch.
NinjaRemote file-transfer features ride on these relays when peers can't go direct.

Pulled INTO ncpeer (and sometimes nmsnj) — upstream of the products

or behaviour change here affects both consumers.

build against**. A dependency bump changes what those products link (security/ABI/build).

through the ncpeer_vcpkg overlay.

Largely independent

shares nccommon and ncpeer_vcpkg with ncpeer.

part of the remote-desktop family). Runs as a child process of the NinjaRMM agent (the
parent owns install/upgrade/policy/launch); NinjaFlow owns the CLI contract, config schema,
credential precedence, DB schema/migrations, and exit-code meanings the parent consumes
those are cross-process contracts. Submits telemetry to the Ninja Backend. Shares no code with
ncpeer/nmsnj; adjacent to nmsnj only in domain (network), not in implementation.

not part of any product's build. It consumes the SBOMs the other products' pipelines emit
(Ninja-Remote/ncpeer, NMS/nmsnj, NinjaFlow all appear in its scope maps), scans them for CVEs
and files Jira issues — a downstream *security-scanning* relationship, not a code dependency.
Shares no code with the others; has its own DB/schema and a toolbox submodule.

authentication component: its pam_ninja.so gates SSH reverse-tunnel access into
Ninja infrastructure with a time-based OTP. Shares no code with the other repos; its
blast radius is operational (who can SSH-tunnel into hosts), not code-level. The client OTP
generators and the module must agree on the OTP scheme + shared secret.

Review implications

these are *upstream* of ncpeer (and nccommon/vcpkg also of nmsnj) — weigh the consumer
side: API/ABI compatibility, build impact, and cross-platform reach.

contract.

Also in every review's context

The reviewer instructions CLAUDE.md (rules, house C++ style, output schema); the PR/ticket block pr.md (title, description, linked Jira + acceptance criteria, commits, changed files); the full diff.patch.