feat(vscode): add a minimum low-token baseline config for agent sessions

[JustAGhosT]

Problem

In large workspaces, prompt/context cost is dominated by always-on context expansion rather than actual task content.

Goal

Establish a minimum baseline profile that starts lean, then opt in to heavier features only when needed.

Motivation

Referenced instructions, nested agent files, skill-adherence prompt, session projection, custom hooks, memory, and broad MCP discovery can compound into large prompt overhead.

Proposed baseline config

{
  "chat.includeReferencedInstructions": false,
  "chat.experimental.useSkillAdherencePrompt": false,
  "chat.agentSessionProjection.enabled": false,
  "chat.useNestedAgentsMdFiles": false,
  "chat.useCustomAgentHooks": false,
  "github.copilot.chat.copilotMemory.enabled": false,
  "chat.mcp.discovery.enabled": {
    "claude-desktop": false,
    "windsurf": false,
    "cursor-global": false,
    "cursor-workspace": false
  },
  "github.copilot.chat.switchAgent.enabled": false
}

Follow-up

Add a balanced profile and document which toggles reintroduce the most prompt weight.

Acceptance criteria

1. Document the minimum baseline profile in the repo.
2. Explain the likely token cost drivers in descending order.
3. Provide guidance for upgrading from lean to balanced/full-context.
4. Cross-link this issue with Context budget: referenced instruction expansion is the largest token source #449, Context budget: large workspace metadata snapshot injected per session #450, Context budget: tool and policy payload grows unbounded with plugins and MCP servers #451, Context budget: conversation history replay is expensive in long sessions #452, and Context budget: audit and document VS Code settings that inflate context #453 rather than duplicating that analysis.

Related: #449 #450 #451 #452 #453

[blocksorg]

Mention Blocks like a regular teammate with your question or request:

@blocks can you add technical details to this issue?
@blocks make the following changes ...
@blocks create an issue from what was mentioned in the following comment ...
@blocks explain the following code ...
@blocks are there any security or performance concerns?

Run @blocks /help for more information.

Workspace settings | Disable this message

[JustAGhosT]

Balanced profile (recommended next step)

If teams need more capability than the lean baseline, enable features in this order while watching token/cost impact:

{
  "chat.includeReferencedInstructions": false,
  "chat.experimental.useSkillAdherencePrompt": false,
  "chat.agentSessionProjection.enabled": false,
  "chat.useNestedAgentsMdFiles": false,
  "chat.useCustomAgentHooks": false,
  "github.copilot.chat.copilotMemory.enabled": true,
  "github.copilot.chat.switchAgent.enabled": true,
  "chat.mcp.discovery.enabled": {
    "claude-desktop": true,
    "windsurf": false,
    "cursor-global": false,
    "cursor-workspace": false
  }
}

Upgrade path from lean → balanced → full

1. Turn on github.copilot.chat.copilotMemory.enabled first.
2. Turn on github.copilot.chat.switchAgent.enabled if workflow needs multiple agents.
3. Enable one MCP discovery source only (prefer claude-desktop) and measure.
4. Keep chat.includeReferencedInstructions off unless a task explicitly needs full instruction expansion.
5. Keep nested agent files / skill adherence / projection off by default in large repos.

Suggested measurement note

Capture before/after for: median tokens per turn, p95 latency, and qualitative answer quality. Roll back the last toggle if cost increase is disproportionate.

[JustAGhosT]

Measurement template (copy/paste)

Use this table for each profile change (lean → balanced → full-context) so we can compare cost vs quality consistently.

Run date	Profile	Toggles changed	Median tokens/turn	p95 tokens/turn	Median latency (s)	p95 latency (s)	Quality score (1-5)	Notes
YYYY-MM-DD	lean	baseline
YYYY-MM-DD	balanced	+memory,+switchAgent,+1 MCP source
YYYY-MM-DD	full-context	+referencedInstructions,+nestedAgents,+projection

Suggested quality rubric (1-5)

• 5: Correct, concise, no missing requirements.
• 4: Correct with minor omissions.
• 3: Usable but needs follow-up prompts.
• 2: Frequent misses or bloated answers.
• 1: Incorrect/unusable.

Decision rule

• If p95 tokens or p95 latency increase >30% without at least +1 quality point, roll back the last toggle.
• Prefer the lowest-cost profile that meets quality requirements.

Tracking note

Link each run to sample tasks used for evaluation (e.g., incident triage, code edit + build, repo search, docs question) so comparisons are apples-to-apples.