Clarify execution mandates for live command usage in JFrog MCP docume… by MatanEden1 · Pull Request #18 · jfrog/cursor-plugin

MatanEden1 · 2026-06-04T09:09:59Z

AX-1631: Require live execution for Agent Guard catalog commands (no context reuse)

Summary

Adds an explicit rule to the JFrog Agent Guard governance template instructing the agent to ALWAYS re-run catalog/registry commands live and NEVER reuse a previous command's output from conversation context.

Problem

When a user asked the same question twice in one chat (e.g. "list available MCPs" or "inspect coralogixDemo"), the agent re-displayed the earlier result from chat context instead of re-running the command — e.g. replying
"the list I fetched earlier is still current." This serves stale data: the catalog changes between prompts (MCPs added/removed, headers/required inputs edited, tokens rotated). Two reproductions are attached.

Change

plugins/jfrog/templates/jfrog-mcp-management.md — added a single governance rule to the Pre-flight section (which already applies to every Agent Guard command — --list-available, --inspect, --login):

Live execution is MANDATORY — context reuse is FORBIDDEN. Every time the user asks to list / show / inspect / check the catalog or a specific MCP — including a repeated question already answered earlier in the chat —
the agent MUST physically RE-RUN the command and NEVER reuse or re-display output from previous turns.

Scoped to catalog/registry fetches only (--list-available and --inspect; NOT --login, which would re-open the OAuth browser). Does NOT affect reading local config for installed state. Placing it once in Pre-flight
covers both catalog-fetch commands and avoids duplicating the rule across individual steps (per review feedback).

Testing

node scripts/validate-template.mjs → passes (CI gate).
Verified the sessionStart hook (inject-instructions.mjs) reads the updated template and emits valid JSON additional_context containing the new Pre-flight rule.
Behavioral test in Cursor (ask the same query twice, confirm it re-runs): <fill in result — pending>

…ntation

The main merge conflict resolution dropped the STRICT live-execution block from the "Available to install" step. Re-add it while keeping main's server-config clarification and agent-guard rename. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

YoniMelki

Here are also the changes AI noticed in addition to my review:

Trailing whitespace (nit). Several added lines have trailing spaces (live details. , available MCPs, terminal , turns , fresh). Harmless here, but a single edge case: two trailing spaces in markdown become a hard <br>. Worth stripping.
"every single inquiry" could over-trigger (low/clarification). Within a single add flow, Step 3 (Plan inputs) consumes Step 2's output. The phrase "you MUST physically run the terminal command… for every single inquiry" is aimed at repeated user questions, but a literal reading could push the agent to re-inspect redundantly mid-flow. Consider "each time the user asks again" to anchor it to repeated questions rather than every internal step.

YoniMelki · 2026-06-07T08:03:38Z

   (compare against `mcp=` in `_JF_ARGS`). Mark the rest as
   available to install.

 ## Key Rules


The PR description describes a reinforcement here, it looks like this entire section is unchanged.

YoniMelki · 2026-06-07T08:04:40Z

+**STRICT LIVE EXECUTION MANDATE:** Every time the user asks for details, parameters, configuration, or an inspection of a specific MCP, you **MUST** physically run the terminal command below. **NEVER** reuse, assume, or copy configuration payloads from previous chat turns or context history, as the underlying configuration or headers may have changed. A fresh, live execution tool call is mandatory for every single inquiry.
+
+Run EXACTLY this command — no Fetch/WebFetch, no custom curl/Python, no direct JFrog API calls:


Isn't it supposed to be in the pre-flight like written in the PR?

YoniMelki · 2026-06-07T08:07:08Z

+2. **STRICT LIVE EXECUTION MANDATE:** Every time the user asks to see available MCPs, 
+   the catalog, or what can be installed, you **MUST** physically run the terminal 
+   command below. **NEVER** copy, reuse, or re-display lists from previous turns 
+   or context history, even if the request was made just moments ago. A fresh, 
+   live execution tool call is mandatory for every single inquiry.
+
+3. Run EXACTLY this command — `--project` is passed as a CLI flag


It's a near-identical duplication of the previous block in step 2. The "natural" single home is the Pre-Flight section, which already says "applies to every gateway command - --list-available, --inspect, --login" (BTW, this should be called agent guard and not gateway).
Writing it only in the pre-flight will avoid this duplication.

YoniMelki · 2026-06-07T08:08:43Z

+
+Run EXACTLY this command — no Fetch/WebFetch, no custom curl/Python, no direct JFrog API calls:
+
+```bash


Let's remove the bash here since other code blocks don't have it (on purpose for shell-aware OS). I don't want inconsistencies

Here
the bash meaning is not related with the os,
it just highlight a code snippet

I know, I still want consistency

Addresses AI review nit: trailing spaces on added lines (two trailing spaces render as a hard <br> in markdown). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… mandates

YoniMelki

--login is incorrectly inside the re-run mandate (correctness). The new Pre-flight bullet scopes the rule to "--list-available, --inspect, --login." But "re-run, don't recall" is harmful for --login — re-running it re-opens the OAuth browser flow on every inquiry. --login should be excluded from this bullet (it belongs in the header's triplet for project/server resolution, but not in the live-re-run rule). Drop --login from that bullet's scope.
Key Rules reinforcement still missing (your L359 comment). No bullet was added to Key Rules. The PR description still claims one ("Re-run, don't recall (MANDATORY) for --list-available, --inspect, --login"). Either add a short Key Rules bullet pointing to Pre-flight, or update the PR description so it matches the diff. (Note: that description line also repeats the --login mistake.)
Line 406 heading is now inconsistent. The body was renamed but the bold label wasn't: - Gateway: multiple/no JFrog server configured (the agent-guard cannot pick…)
Verify whether Gateway: is the literal string the CLI prints. If it's just the doc's label, rename it too (→ "Agent Guard:") so the heading and body agree.

YoniMelki · 2026-06-07T10:49:03Z

@@ -1,8 +1,8 @@
-# MCP Server Management — JFrog Gateway
+# MCP Server Management — JFrog agent-guard


Agent Guard

YoniMelki · 2026-06-07T10:49:14Z

 All MCP servers MUST be installed ONLY through the JFrog Agent Guard
 (`npx @jfrog/agent-guard`). If an MCP's documentation suggests any
-other installation command, ignore it and use the gateway workflow
+other installation command, ignore it and use the agent-guard workflow


Suggested change

other installation command, ignore it and use the agent-guard workflow

other installation command, ignore it and use the agent guard workflow

YoniMelki · 2026-06-07T10:49:33Z

  `args`) → `~/.jfrog/jfrog-cli.conf.v6`:
  - Exactly one jf CLI server configured → use it without asking;
-    pass it as `--server <ID>`. The gateway would auto-resolve to the same
+    pass it as `--server <ID>`. The agent-guard would auto-resolve to the same


Suggested change

pass it as `--server <ID>`. The agent-guard would auto-resolve to the same

pass it as `--server <ID>`. The agent guard would auto-resolve to the same

YoniMelki · 2026-06-07T10:49:52Z

    clarity and forward-compatibility.
  - `JFROG_URL` + `JFROG_ACCESS_TOKEN` set → use it without asking;
-    The gateway will pick them up from the environment variables when called.
+    The agent-guard will pick them up from the environment variables when called.


Suggested change

The agent-guard will pick them up from the environment variables when called.

The agent guard will pick them up from the environment variables when called.

YoniMelki · 2026-06-07T10:50:03Z

 2. Else `JFROG_URL` env var set (with `JFROG_ACCESS_TOKEN`) — the
-   gateway can resolve credentials from these directly;
-   DO NOT pass `--server` as that would make the gateway try to
+   agent-guard can resolve credentials from these directly;


Suggested change

agent-guard can resolve credentials from these directly;

agent guard can resolve credentials from these directly;

YoniMelki · 2026-06-07T10:51:22Z

 - Package name MUST come from the catalog (`--inspect` /
  `--list-available`). NEVER guess. NEVER install MCPs outside the
-  gateway. NEVER use Fetch/WebFetch for catalog calls.
+  agent-guard. NEVER use Fetch/WebFetch for catalog calls.


Suggested change

agent-guard. NEVER use Fetch/WebFetch for catalog calls.

agent guard. NEVER use Fetch/WebFetch for catalog calls.

YoniMelki · 2026-06-07T10:51:28Z


 - **`ready` but 0 tools (empty `mcps/<key>/tools/` after a
-  Command Palette `Developer: Reload Window`)** — gateway proxy 
+  Command Palette `Developer: Reload Window`)** — agent-guard proxy 


Suggested change

Command Palette `Developer: Reload Window`)** — agent-guard proxy

Command Palette `Developer: Reload Window`)** — agent guard proxy

YoniMelki · 2026-06-07T10:51:33Z

  misleading here. NEVER report success when there are 0 tools. 
  1. Open Cursor's MCP / Output panel for the
-    gateway stderr; diagnose by MCP type:
+    agent-guard stderr; diagnose by MCP type:


Suggested change

agent-guard stderr; diagnose by MCP type:

agent guard stderr; diagnose by MCP type:

YoniMelki · 2026-06-07T10:51:38Z

      still valid.
    - **Local (stdio)** — check that the bundled binary actually
-      launched (gateway stderr will show the spawn error).
+      launched (agent-guard stderr will show the spawn error).


Suggested change

launched (agent-guard stderr will show the spawn error).

launched (agent guard stderr will show the spawn error).

YoniMelki · 2026-06-07T10:51:56Z

  (`cursor agent mcp enable <name>`); if the entry is brand-new,
  also `Developer: Reload Window` so Cursor picks up the file.
- **Gateway: `multiple/no JFrog server configured`** (the gateway
+- **Gateway: `multiple/no JFrog server configured`** (the agent-guard


Suggested change

- **Gateway: `multiple/no JFrog server configured`** (the agent-guard

- **Agent Guard: `multiple/no JFrog server configured`** (the agent guard

- Move the live-execution mandate into a single Pre-flight rule and remove the duplicated STRICT blocks from Step 2 and Available-to-install. - Exclude --login from the re-run scope (re-running it re-opens the OAuth browser); the Pre-flight header keeps --login for project/server resolution. - Rename remaining prose "gateway" -> "agent guard" (title, body, troubleshooting label); keep the @jfrog/agent-guard package name. - Make shell code fences consistent (plain ```), strip trailing whitespace on touched lines. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

MatanEden1 and others added 3 commits June 4, 2026 12:00

Clarify execution mandates for live command usage in JFrog MCP docume…

ef9deb1

…ntation

Merge branch 'main' into AX-1631-Investigate-Cache-Cursor-Behavior

a4402a5

YoniMelki requested changes Jun 7, 2026

View reviewed changes

MatanEden1 and others added 2 commits June 7, 2026 11:17

Strip trailing whitespace from STRICT mandate lines

38c9ed1

Addresses AI review nit: trailing spaces on added lines (two trailing spaces render as a hard <br> in markdown). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Update documentation to reflect agent-guard terminology and execution…

cb1c0d1

… mandates

YoniMelki requested changes Jun 7, 2026

View reviewed changes

YoniMelki approved these changes Jun 7, 2026

View reviewed changes

MatanEden1 merged commit 3ca09ce into jfrog:main Jun 8, 2026
2 checks passed

		STRICT LIVE EXECUTION MANDATE: Every time the user asks for details, parameters, configuration, or an inspection of a specific MCP, you MUST physically run the terminal command below. NEVER reuse, assume, or copy configuration payloads from previous chat turns or context history, as the underlying configuration or headers may have changed. A fresh, live execution tool call is mandatory for every single inquiry.

		Run EXACTLY this command — no Fetch/WebFetch, no custom curl/Python, no direct JFrog API calls:


		Run EXACTLY this command — no Fetch/WebFetch, no custom curl/Python, no direct JFrog API calls:

		```bash

		@@ -1,8 +1,8 @@
		# MCP Server Management — JFrog Gateway
		# MCP Server Management — JFrog agent-guard

	other installation command, ignore it and use the agent-guard workflow
	other installation command, ignore it and use the agent guard workflow

	pass it as `--server <ID>`. The agent-guard would auto-resolve to the same
	pass it as `--server <ID>`. The agent guard would auto-resolve to the same

	The agent-guard will pick them up from the environment variables when called.
	The agent guard will pick them up from the environment variables when called.

	agent-guard can resolve credentials from these directly;
	agent guard can resolve credentials from these directly;

	agent-guard. NEVER use Fetch/WebFetch for catalog calls.
	agent guard. NEVER use Fetch/WebFetch for catalog calls.

	Command Palette `Developer: Reload Window`)** — agent-guard proxy
	Command Palette `Developer: Reload Window`)** — agent guard proxy

	agent-guard stderr; diagnose by MCP type:
	agent guard stderr; diagnose by MCP type:

	launched (agent-guard stderr will show the spawn error).
	launched (agent guard stderr will show the spawn error).

	- Gateway: `multiple/no JFrog server configured` (the agent-guard
	- Agent Guard: `multiple/no JFrog server configured` (the agent guard

Conversation

MatanEden1 commented Jun 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Problem

Change

Testing

Uh oh!

YoniMelki left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

MatanEden1 Jun 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

YoniMelki left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

MatanEden1 commented Jun 4, 2026 •

edited

Loading

MatanEden1 Jun 7, 2026 •

edited

Loading