๐ก๏ธ Sentinel: [HIGH] Fix SSRF vulnerability in web e2e readiness probe #1354
There are no files selected for viewing
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -10,7 +10,7 @@ candidates. | |
|
|
||
| Public-repository scans now default to | ||
| `nvidia/nemotron-3-super-120b-a12b`. The first fallback is | ||
| `nvidia/nemotron-4-340b-instruct`. Private repositories retain the | ||
| contracted provider because NVIDIA hosted trial inputs are restricted to public | ||
| repositories by the central workflow. | ||
|
|
||
|
|
@@ -79,7 +79,7 @@ Fielding, R., Nottingham, M., & Reschke, J. (2022). *HTTP semantics* (RFC | |
| 9110). Internet Engineering Task Force. https://doi.org/10.17487/RFC9110 | ||
|
|
||
| NVIDIA Corporation. (2025). *Llama-3.3-Nemotron-Super-49B-v1.5* [Model card]. | ||
| NVIDIA NIM. https://build.nvidia.com/nvidia/nemotron-4-340b-instruct/modelcard | ||
|
There was a problem hiding this comment. ๐ Maintainability & Code Quality | ๐ก Minor | โก Quick win ์ฐธ๊ณ ๋ฌธํ์ ๋ชจ๋ธ๋ช ๊ณผ ์ฐ๋๋ ๊ฐฑ์ ํ์ธ์. Line 82์ URL์ Nemotron-4 ๋ชจ๋ธ ์นด๋๋ฅผ ๊ฐ๋ฆฌํค์ง๋ง, Line 81์ ์ฌ์ ํ ์์ ์์-NVIDIA Corporation. (2025). *Llama-3.3-Nemotron-Super-49B-v1.5* [Model card].
+NVIDIA Corporation. (2024). *Nemotron-4-340B-Instruct* [Model card].๐ค Prompt for AI AgentsSource: MCP tools |
||
|
|
||
| NVIDIA Corporation. (2026a). *NVIDIA-Nemotron-3-Super-120B-A12B* [Model | ||
| card]. NVIDIA NIM. | ||
|
|
||
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -3,7 +3,7 @@ | |
| // NOT switched to "contextual-orchestrator/contextual-orchestrator" yet: | ||
| // that requires CONTEXTUAL_ORCHESTRATOR_BASE_URL/_TOKEN to be provisioned | ||
| // first (see the "contextual-orchestrator" provider block below). | ||
| "model": "nvidia-nim/nvidia/nemotron-4-340b-instruct", | ||
|
There was a problem hiding this comment. ๐๏ธ Data Integrity & Integration | ๐ Major | ๐๏ธ Heavy lift ๐ Supported by static analysis๐ Script executed: #!/usr/bin/env bash
set -euo pipefail
: "${NVIDIA_API_KEY:?NVIDIA_API_KEY is required}"
base="https://integrate.api.nvidia.com/v1"
model="nvidia/nemotron-4-340b-instruct"
curl -fsS \
-H "Authorization: Bearer ${NVIDIA_API_KEY}" \
"${base}/models" |
jq -e --arg model "$model" '[.data[]?.id] | index($model) != null' >/dev/null
response="$(
curl -fsS \
-H "Authorization: Bearer ${NVIDIA_API_KEY}" \
-H "Content-Type: application/json" \
"${base}/chat/completions" \
-d `@-` <<JSON
{
"model": "${model}",
"messages": [{"role": "user", "content": "Call the probe function exactly once."}],
"tools": [{
"type": "function",
"function": {
"name": "probe",
"description": "A capability probe.",
"parameters": {
"type": "object",
"properties": {},
"additionalProperties": false
}
}
}],
"tool_choice": {
"type": "function",
"function": {"name": "probe"}
},
"max_tokens": 64
}
JSON
)"
jq -e '(.choices[0].message.tool_calls // []) | length > 0' <<<"$response" >/dev/nullRepository: ContextualWisdomLab/.github Length of output: 222 ๐ Script executed: #!/usr/bin/env bash
set -euo pipefail
printf '%s\n' '--- repository conventions ---'
head -5 /tmp/coderabbit-repo-knowledge/contextualwisdomlab-github-bad40478/*/*.md 2>/dev/null || true
printf '%s\n' '--- relevant diff ---'
git diff -- opencode.jsonc
printf '%s\n' '--- opencode.jsonc structure and affected metadata ---'
nl -ba opencode.jsonc | sed -n '1,18p;280,315p'Repository: ContextualWisdomLab/.github Length of output: 1166 ๐ Script executed: #!/usr/bin/env bash
set -euo pipefail
printf '%s\n' '--- available scoped conventions and learnings ---'
find /tmp/coderabbit-repo-knowledge/contextualwisdomlab-github-bad40478 \
-maxdepth 2 -type f -print | sort
printf '%s\n' '--- repository state and target file ---'
git status --short
git diff --stat
sed -n '1,20p' opencode.jsonc
sed -n '280,310p' opencode.jsonc
printf '%s\n' '--- relevant symbols and routing references ---'
grep -nE '"(model|name|tool_call|context|ci-review|code-reviewer)"' opencode.jsoncRepository: ContextualWisdomLab/.github Length of output: 5298 ๐ Script executed: #!/usr/bin/env bash
set -euo pipefail
printf '%s\n' '--- agent configuration ---'
sed -n '20,82p' opencode.jsonc
printf '%s\n' '--- complete NVIDIA model entries ---'
sed -n '286,380p' opencode.jsoncRepository: ContextualWisdomLab/.github Length of output: 4835 ๐ Web query:
๐ก Result: There is no evidence in the available technical documentation that Nemotron-4-340B-Instruct is explicitly designed for or natively supports tool calling or function calling [1]. While the model is supported within the NVIDIA NIM for Large Language Models framework, current documentation does not list tool use capabilities for this model, whereas it does explicitly indicate tool use support for other models like Nemotron-4-340B-Reward [1]. Citations: ๐ Script executed: #!/usr/bin/env bash
set -euo pipefail
url='https://docs.nvidia.com/nim/large-language-models/1.7.0/models.html'
curl -fsSL "$url" |
grep -i -A12 -B8 -E 'nemotron-4-340b-instruct|128k' |
head -120Repository: ContextualWisdomLab/.github Length of output: 2881 ๐ Web query:
๐ก Result: The NVIDIA Nemotron-4-340B-Instruct model, when deployed via NVIDIA NIM, is designed with a context length of 4,096 tokens [1][2][3]. This specification is confirmed in the official model configuration and documentation [1][4]. Regarding tool calling, the official model card and technical specifications for Nemotron-4-340B-Instruct do not categorize it as a model specifically fine-tuned or natively optimized for complex tool-calling workflows in the same manner as specialized agentic models [1][3]. While some third-party platforms or integrated wrappers may attempt to facilitate tool-use interaction via API implementations [5], it is not a native, primary architectural feature of the base Nemotron-4-340B-Instruct instruct-tuned release [1]. Developers should verify the capabilities of the specific NIM container version, as NIM provides an OpenAI-compatible API layer [6] that may handle function calling requests differently depending on the deployment environment and version. Citations:
๐ Web query:
๐ก Result: In OpenCode, you can configure model limits, tool-calling capabilities, and context metadata within your opencode.json (or opencode.jsonc) configuration file [1][2][3]. Configuration for models, including specific limits and capabilities, is typically managed under the providers section [4][2][5]. Model Limit and Capability Configuration You can explicitly define a model's limits (context, output) and capabilities (like tool-calling support) to override automatic discovery or handle models that do not support certain features [4][2][5]. Example configuration: { "$schema": "https://opencode.ai/config.json", "providers": { "your-provider-id": { "models": { "your-model-id": { "limit": { "context": 200000, "output": 32000 }, "tool_call": false } } } } } Key settings include: - limit.context: The maximum number of input tokens the model accepts [4][2][6]. - limit.output: The maximum number of tokens the model can generate [4][2][6]. - tool_call: A boolean (true/false) field used to explicitly enable or disable tool-calling support for a specific model [7][8][6]. Note that in some versions, setting this to false may still require internal runtime handling to ensure tools are excluded from the request [7]. Context Metadata and Compaction OpenCode allows you to manage how context is handled during sessions via the compaction configuration [1][2]. This helps maintain performance and stay within token limits [1][2]. Example compaction configuration: { "compaction": { "auto": true, "keep": { "tokens": 15000 }, "buffer": 20000 } } - auto: Automatically compacts the session when context is full (default is true) [1][2]. - reserved/buffer: A token buffer (e.g., 20,000) to ensure there is enough window to avoid overflow during compaction [1][2]. Additional Notes - Schema Validation: Including the "$schema" field in your opencode.json file enables editor autocompletion and validation [2][9]. - Hierarchy: OpenCode uses a hierarchical configuration system [10]. Global settings are in ~/.config/opencode/opencode.json, while project-specific settings are in a local opencode.json [1][3][10]. Settings are merged, with project-specific configs typically overriding global ones [3][10]. Citations:
๋ชจ๋ธ capability metadata๋ฅผ ์ค์ ๋ชจ๋ธ ๊ณ์ฝ์ ๋ง๊ฒ ์์ ํ์ธ์.
๐ค Prompt for AI Agents |
||
| "small_model": "nvidia-nim/meta/llama-3.3-70b-instruct", | ||
| "enabled_providers": ["nvidia-nim", "github-models", "contextual-orchestrator"], | ||
| "lsp": false, | ||
|
|
@@ -293,7 +293,7 @@ | |
| "apiKey": "{env:NVIDIA_API_KEY}" | ||
| }, | ||
| "models": { | ||
| "nvidia/nemotron-4-340b-instruct": { | ||
| "name": "NVIDIA Llama 3.3 Nemotron Super 49B v1.5", | ||
|
Comment on lines
+296
to
297
Contributor
There was a problem hiding this comment. ๐ Info: Stale model display name after key rename The NVIDIA NIM model key was renamed to 
Was this helpful? React with ๐ or ๐ to provide feedback. |
||
| "tool_call": true, | ||
| "limit": { | ||
|
|
||
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,8 @@ | ||
| ๐ฏ What: | ||
| `scripts/ci/sandboxed_web_e2e.py`์ `wait_for_url` ํจ์์์ `--backend-ready-url` ๋๋ `--frontend-ready-url` ํ๋ผ๋ฏธํฐ๋ก ์ ๋ฌ๋ URL์ ๊ฒ์ฆํ๋ ๋ก์ง์ SSRF(Server-Side Request Forgery) ์ทจ์ฝ์ ์ด ์กด์ฌํ์ฌ ์ด๋ฅผ ์์ ํ์ต๋๋ค. | ||
|
|
||
| โ ๏ธ Risk: | ||
| ๊ธฐ์กด ๋ก์ง์ URL์ด `http://` ๋๋ `https://`๋ก ์์ํ๋์ง๋ง ๊ฒ์ฆํ๊ณ , ํธ์คํธ๋ช (hostname)์ ๊ฒ์ฆํ์ง ์์์ต๋๋ค. ์ด๋ก ์ธํด ์ ์์ ์ธ ์ฌ์ฉ์๊ฐ ํ๊ฒฝ๋ณ์๋ ํ๋ผ๋ฏธํฐ๋ฅผ ์กฐ์ํ์ฌ `http://169.254.169.254`๋ ๋ด๋ถ ๋คํธ์ํฌ ์ฃผ์๋ก ์์ฒญ์ ๋ณด๋ด์ด ๋ด๋ถ ์์คํ ์ ์ค์บ๋ํ๊ฑฐ๋ ๋ฏผ๊ฐํ ์ ๋ณด๋ฅผ ํ์ทจํ ์ํ(SSRF)์ด ์์ต๋๋ค. | ||
|
|
||
| ๐ก๏ธ Solution: | ||
| `urllib.parse.urlparse`๋ฅผ ํ์ฉํ์ฌ ๋์ URL์ ํธ์คํธ๋ช ์ด ๋ก์ปฌ ๋ฃจํ๋ฐฑ ์ฃผ์(`localhost`, `127.0.0.1`, `::1`)์ธ์ง ๋ช ์์ ์ผ๋ก ๊ฒ์ฆํ๋ ๋ก์ง์ ์ถ๊ฐํ์ต๋๋ค. ์ง์ ๋ ๋ก์ปฌ ํธ์คํธ๋ช ์ด ์๋ ๊ฒฝ์ฐ `ValueError`๋ฅผ ๋ฐ์์ํค๋ฉฐ ์์ฒญ์ ์ฐจ๋จํฉ๋๋ค. ๋ํ ์ด ๊ฒ์ฆ ๋ก์ง์ด ์ ์์ ์ผ๋ก ์๋ํ๋์ง ํ์ธํ๊ธฐ ์ํด `tests/test_sandboxed_web_e2e.py`์ ํ ์คํธ ์ผ์ด์ค๋ฅผ ์ถ๊ฐํ์ต๋๋ค. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
๐ฉบ Stability & Availability | ๐ก Minor | โก Quick win
๋ชจ๋ธ ์๋ณ์์ ํ์ ์ด๋ฆ์ ์ผ์น์ํค์ธ์.
nvidia/nemotron-4-340b-instruct๊ฐ ๋ workflow ํญ๋ชฉ์์ ์ฌ์ ํNVIDIA Llama 3.3 Nemotron Super 49B v1.5๋ก ํ์๋๊ณ , ๊ด๋ จ ํ ์คํธ ์คํจ ๋ฉ์์ง๋Nemotron Super๋ฅผ ๊ฐ๋ฆฌํต๋๋ค. ๋ก๊ทธ์ ์คํจ ์ง๋จ์ด ์ค์ ์คํ ๋ชจ๋ธ๊ณผ ๋ฌ๋ผ์ง ์ ์์ผ๋ฏ๋ก ์ธ ์์น์ ํ์๋ช ์ Nemotron 4 340B Instruct์ ๋ง๊ฒ ๊ฐฑ์ ํ์ธ์.๐ Affects 2 files
.github/workflows/opencode-review-dispatch.yml#L3777-L3777(this comment).github/workflows/opencode-review-dispatch.yml#L4173-L4174scripts/ci/test_strix_quick_gate.sh#L1491-L1491๐ค Prompt for AI Agents
Source: MCP tools