Context Enrichment

Context enrichment is the process of assembling everything the planner needs to generate a good command plan. Without context, the planner is a talented stranger — it knows shell syntax but nothing about your project.

The buildPlannerContext function

buildPlannerContext() is a pure function that takes raw inputs and produces a structured PlannerContext. It lives in apps/desktop/src/lib/buildPlannerContext.ts.

Inputs

Input	Source	Purpose
sessionId	Active session	Links plan to session
cwd	Session state	Current working directory
shell	Session state	Shell type (bash, pwsh, etc.)
os	Runtime detection	Operating system
memoryItems	MemoryStore	All accepted memory items
workflows	WorkflowStore	All saved workflows
lastRunByWorkflowId	WorkflowRunStore	Last run status per workflow
recentHistory	HistoryStore (filtered)	Current session’s history

Processing

1. Effective memory resolution: calls resolveEffectiveMemory(items, cwd) which merges project-scoped items (for the current directory) with non-shadowed global items. Project-scoped items take precedence over globals with the same key.

2. Recent commands: extracts the last 5 executed commands from session history. Gives the planner awareness of what you just did.

3. Relevant workflows: filters to workflows matching the current directory (or global workflows), takes the top 5. Each workflow becomes a projectFact with:

   • kind: "workflow"
   • label: workflow name
   • value: step commands joined with “ → “, plus last-run status if available

Output

{
  sessionId, cwd, projectRoot, os, shell,
  recentCommands: string[],
  memoryItems: { kind, key, value, confidence }[],
  projectFacts: { kind, label, value }[]
}

How the backend uses context

Ollama prompt

The build_planner_prompt() function in Rust formats the context into the LLM prompt:

• System section: cwd, OS, shell
• Memory section: ## Known context with each memory item
• Workflow section: ## Known workflows with each workflow’s commands and last-run status
• Recent commands: listed for recency awareness
• Rules: 10 rules including “prefer known workflow commands over inventing new ones”

Mock planner

The mock fallback uses context for:

• Setting cwd on the response
• Populating retrievedContext in the plan review
• Workflow matching: fuzzy-matching intent against workflow labels to return the workflow’s commands instead of a generic stub

Plan review

retrievedContext in the PlanReviewPayload lists what context was assembled:

• cwd: ~/projects
• projectRoot: ~/projects
• workflow:deploy
• workflow:test-suite

This is displayed in the PlanPanel footer so you can see what influenced the plan.

The enrichment loop

Context enrichment creates a virtuous cycle:

Use CommandUI → History accumulates
    ↓
Detectors find patterns → Suggestions appear
    ↓
Accept suggestions → Memory items created
    ↓
Save workflows → Workflow library grows
    ↓
Submit semantic request → Rich context assembled
    ↓
Planner generates better plan → User approves
    ↓
More history → Better detection → Richer context

The system gets better the more you use it, without any explicit training step.