Project Intelligence
Automatically learns conventions, tool hints, and working notes from every AI interaction.
Overview
The Project Intelligence layer is a three-part learning system that captures institutional knowledge as a side effect of normal work. Unlike knowledge cards (which are explicit), intelligence accumulates automatically from all chat participants.
Prerequisite for full auto-capture: Register the VS Code agent hook once (Dashboard → Settings → 🪝 Agent Hooks → Install Hooks). Without it, only interactions during your active chat sessions are captured. With it, every Copilot Chat,
@workspace, and background agent response feeds the intelligence pipeline.
Three Stores of Intelligence
1. Conventions
Structured codebase conventions with confidence tracking:
| Field | Description |
|---|---|
| Subject | What the convention covers (e.g., “error handling”) |
| Category | architecture, naming, patterns, testing, tooling, pitfalls |
| Content | The convention itself |
| Confidence | confirmed ✅, observed ⏳, or inferred 🔮 |
Conventions start as observed when the LLM first extracts them. You can confirm or discard them from the Intelligence tab.
Example:
Naming → confirmed: All React components use PascalCase with a
.tsxextension. Utility functions use camelCase in.tsfiles.
2. Tool Hints
Learned search patterns that help the AI navigate your specific codebase:
✅ Pattern: Search "TabStripController" not "tab strip controller"
❌ Anti-pattern: "tab strip" returns noise from CSS files
Tool hints track use count - frequently used hints are prioritized in injection.
3. Working Notes
Free-form agent exploration memory for relationships and insights:
| Field | Description |
|---|---|
| Title | What the note covers |
| Content | Markdown content |
| Related Files | Files associated with this note |
| Related Symbols | Code symbols referenced |
| Staleness | fresh, possibly-stale, stale (file-based tracking) |
Working notes now have file-based staleness detection, displayed as 🟢 fresh → ⚠️ possibly-stale → 🔴 stale. Staleness is detected by comparing relatedFiles modification times against the note’s last update timestamp. Checks run on dashboard open and on file save (debounced 2 s). Controlled by the enableStalenessTracking setting.
Working notes are matched by file path and keyword when queried via #ctx - only task-relevant notes are returned.
Intelligence Delivery
Intelligence is delivered through two complementary mechanisms — no per-prompt injection budgets to tune.
Managed Block in copilot-instructions.md
ContextManager maintains an auto-synced managed block at the bottom of your .github/copilot-instructions.md file. This block contains:
#ctxtool usage instructions — so every agent knows how to invoke the tool- Pinned knowledge card titles — cards flagged Pinned appear as summary lines with their IDs, prompting agents to call
#getCardwhen relevant
The managed block is regenerated whenever pinned cards change. You never edit it by hand.
#ctx Tool (On-Demand)
All intelligence is queryable via the #ctx tool. Agents invoke it when they need conventions, notes, hints, or queued card candidates. Available modes:
| Mode | Purpose |
|---|---|
#ctx query:"error handling" | Full-text search across cards, conventions, notes |
#ctx mode:"list" type:"conventions" | List all items of a type (conventions, workingNotes, toolHints, cards, queue) |
#ctx mode:"learn" learnType:"convention" ... | Store a new convention, tool hint, or working note |
#ctx mode:"getCard" id:"<cardId>" | Read the full content of a knowledge card |
#ctx mode:"getQueueItem" id:"<candidateId>" | Read a queued card candidate in full |
#ctx mode:"approveQueueItem" id:"<candidateId>" | Approve a queued item into a saved knowledge card |
#ctx mode:"distillQueue" | Synthesize queued items into proposed knowledge cards |
#ctx mode:"fetch" observationIds:["<id>"] | Fetch full observation details by IDs |
#ctx mode:"retrospect" | End-of-task reflection and capture |
Because delivery is on-demand, there are no token budgets to configure. Agents pull exactly the intelligence they need, when they need it.
Auto-Capture
The Auto-Capture service records every AI response as a lightweight observation:
- Source tracking - which participant generated the response (Copilot, @workspace, hook, etc.)
- Content-hash deduplication - identical interactions are never recorded twice (30-second window)
- Privacy tags -
<private>content</private>tags are stripped before storage - Type badges - observations are classified internally (bugfix, feature, discovery, etc.) and shown as summary counts in the observations header
Auto-Learn
When enabled, a lightweight LLM extraction runs on chat interactions to learn:
- Conventions - coding patterns the AI observes in your codebase
- Working notes - relationships and insights discovered during exploration
- Tool hints - which search terms work and which don’t
- Knowledge cards - high-confidence cards extracted from conversation context
This means even regular Copilot Chat conversations contribute to your project intelligence.
Auto-Learn fires one LLM call per captured response (a small number of tokens per call). This is inexpensive but not free. Disable via intelligence.autoLearn in Settings or the Intelligence tab toggle — Auto-Capture (observation recording) continues without it.
Knowledge cards and architecture notes detected by Auto-Capture / Auto-Learn are staged in the Card Queue for user review — they are never silently persisted. You remain in full control of what enters your knowledge base.
Dashboard: Intelligence Tab
The Intelligence tab provides a unified view:
- Auto-Capture / Auto-Learn toggles with live stats
- Observation feed with per-source filter pills (Copilot, @workspace, hook)
- Convention list with confidence badges and confirm/edit/discard actions
- Tool hints with anti-patterns and delete actions
- Working notes with staleness badges, related files, and promote-to-card actions
- AI Distill - run LLM extraction on selected observations