fix(ai-assistant): use MCP server's manifest description in approval dialog instead of hardcoded "Inspektor Gadget" fallback

[Saksham842]

What does this PR do?

Fixes the MCP tool approval dialog showing "Inspektor Gadget debugging tool: <name>" for every non-Gadget MCP server (Flux, Prometheus, Vault, custom tools, etc.).

Closes #1096

---

Problem

getToolDescription() in packages/ai-common/src/tools/catalog/getToolDescription.ts used keyword heuristics tuned to Inspektor Gadget naming conventions, with a catch-all fallback that hardcoded the "Inspektor Gadget" brand name:

// Before
return `Inspektor Gadget debugging tool: ${toolName}`;

The MCP server's own tool description was already stored in RuntimeToolInfo.description at the getMCPTools() call sites — but it was never threaded through to getToolDescription. Every non-Gadget MCP tool therefore showed completely wrong text in the human approval dialog.

---

Solution

1. getToolDescription.ts — added an optional catalogDescription?: string third parameter. When present for an MCP tool it is returned immediately, before any keyword heuristic runs. The "Inspektor Gadget" catch-all is replaced with the generic "MCP tool: <name>":

// After
export function getToolDescription(
  toolName: string,
  isMCPTool: boolean,
  catalogDescription?: string   // ← new optional param
): string {
  if (isMCPTool) {
    if (catalogDescription) return catalogDescription;  // server's own description wins
    // keyword heuristics as fallback …
    return `MCP tool: ${toolName}`;                    // generic, not Gadget-specific
  }
  …
}

2. LangChainAssistantSession.ts — both toolCallsForApproval construction sites changed from .some() → .find() to capture the full RuntimeToolInfo object, then pass mcpToolEntry?.description as the third argument. The private getToolDescription wrapper is updated to forward the new parameter.

3. getToolDescription.test.ts — updated the stale 'Inspektor Gadget' fallback assertion; added two new tests:

   • Catalog description wins over keyword heuristics (simulates a Flux tool containing run in its name)
   • Empty-string catalog description falls through to keyword heuristics

---

Before / After

Scenario	Before	After
flux__reconcile_kustomization (manifest desc: "Reconcile a Flux Kustomization object")	"Inspektor Gadget debugging tool: flux__reconcile_kustomization"	"Reconcile a Flux Kustomization object"
Unknown tool, no manifest description	"Inspektor Gadget debugging tool: some__tool"	"MCP tool: some__tool"
Gadget gadget__trace_open, no manifest description	"Traces system calls and processes for debugging"	unchanged

---

Checklist

• Bug fix (non-breaking — new parameter is optional; all existing callers remain valid)
• Tests updated / added (getToolDescription.test.ts — 11/11 )
• No changes to public API surface
• LangChainAssistantSession.ts — both approval construction sites updated consistently

---

Files changed

File	Change
packages/ai-common/src/tools/catalog/getToolDescription.ts	Added catalogDescription param; replaced Gadget fallback
packages/ai-common/src/assistant/LangChainAssistantSession.ts	Pass mcpToolEntry?.description at both call sites
packages/ai-common/src/tools/catalog/getToolDescription.test.ts	Updated stale test + 2 new tests

[illume]

Thanks for these changes.

Can you please have a look at the git commits to see if they meet the contribution guidelines? We use a Linux kernel style of git commits. See the contributing guide for general context, and please see previous git commits with git log for examples.

Commits that need attention

• fix(ai-assistant): thread MCP manifest description into approval dialog — Missing area: description prefix — e.g. frontend: HomeButton: Fix so it navigates to home or backend: config: Add enable-dynamic-clusters flag.

Commit guidelines

• Use atomic commits focused on a single change.
• Use the title format <area>: <Description of changes> — description must start with a capital letter.
• Keep the title under 72 characters (soft requirement).
• Explain the intention and why the change is needed.
• Make commit titles meaningful and describe what changed.
• Do not add code that a later commit rewrites; squash or reorder commits instead.
• Do not include Fixes #NN in commit messages.

Good examples:

• frontend: HomeButton: Fix so it navigates to home
• backend: config: Add enable-dynamic-clusters flag

[Copilot]

Copilot review overview

🟡 Changes recommended

The multi-tool orchestration approval path still uses an LLM-generated description instead of authoritative MCP metadata.

Review effort: Balanced
Findings: 1

Open (1)

• Use MCP manifest descriptions for multi-tool approval dialogs · New

What changed in this PR

Updates MCP approval dialogs to prefer server-provided tool descriptions over Gadget-specific heuristics.

Changes:

• Adds optional catalog descriptions and a generic MCP fallback.
• Threads descriptions through direct and streaming approval paths.
• Adds unit tests for precedence and fallback behavior.

File	Description
getToolDescription.ts	Adds manifest-description support and generic fallback.
getToolDescription.test.ts	Tests description precedence and fallbacks.
LangChainAssistantSession.ts	Passes MCP metadata into approval descriptions.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Saksham842]

Addressed in the latest rebased commit. In handleMultipleToolExecution(), we now resolve the matching RuntimeToolInfo from this.toolManager.getMCPTools() and route it through getToolDescription(toolName, isMCPTool, mcpToolEntry?.description). This ensures manifest descriptions take precedence over planner LLM output during multi-tool execution approvals. Added a unit test in LangChainAssistantSession.test.ts covering this path.

[illume]

Thanks for the contribution.

There are some open Copilot review comments — could you take a look at them? Please mark each one as resolved once you've addressed it.

[Saksham842]

Thanks for the review @illume!

1. Commit Guidelines & Rebase:

   • Rebased the branch onto the latest main.
   • Squashed into a single atomic commit adhering to the Linux kernel style:
     • Title: ai-assistant: Use MCP manifest description in approval dialog (< 72 characters, capitalized description).
     • Removed Fixes #... from the commit message and included the intention, rationale, and DCO sign-off.

2. Copilot Review:

   • Addressed the comment in handleMultipleToolExecution() so orchestrated multi-tool approvals now also source their descriptions from the MCP server manifest rather than the planner's LLM output.
   • Added a corresponding unit test in LangChainAssistantSession.test.ts to verify description precedence and fallback behavior for multi-tool orchestration. Marked the Copilot conversation as resolved.