Team Mode: capture design ideas for privacy-safe team coaching signals

[aymenfurter]

Thank you @vanditmakwana55 for the thoughtful work in #56! The idea of giving teams visibility into aggregate coaching signals — while keeping raw prompts, code, and file names strictly local — is a compelling direction worth exploring further.

This issue captures the design ideas surfaced in that PR so we can reference them in future planning.

Design ideas worth preserving

1. Privacy-safe snapshot model — only aggregate numeric stats leave the developer's machine: practice-group scores, token totals, anti-pattern counts by severity, and week-over-week deltas. No raw prompts, source code, file names, or session transcripts.

2. Opt-in by default — Team Mode is off unless the user explicitly enables it and provides a server URL + developer ID.

3. Local snapshot assembler — converts existing analyzer outputs into a compact numeric payload after each analysis run, keeping the sync layer decoupled from the analysis pipeline.

4. One-way sync client with retry — snapshots are posted asynchronously and non-blocking; failed uploads are retried quietly without affecting the local experience.

5. Team Dashboard UI — a webview page showing team-wide scores, trends, and per-member rows with filtering by category and date range. Includes PDF export.

6. Practice-group scoring — five categories (prompt quality, session hygiene, code review, tool mastery, context management) each scored 0–100.

Current project direction

We do not plan to introduce a server-side component (backend API, hosted database, or cloud infrastructure) as part of this project. The AI Engineering Coach is designed to run entirely within the VS Code extension, and we want to keep it that way for simplicity, trust, and ease of adoption.

That said, the client-side aspects of this design (local snapshot assembly, privacy boundaries, dashboard visualization) contain ideas we may draw from in the future — for example, enabling local-only team reports or file-based sharing workflows that don't require a running server.

Next steps

• Keep this issue open as a reference for future team-oriented features
• Any future work in this space should preserve the privacy-first, no-server constraint

[vanditmakwana55]

@microsoft-github-policy-service agree

[aymenfurter]

Additional ideas to explore in the future

Building on the design ideas already captured here, a few more directions worth considering when/if we revisit team-oriented features:

1. File-based snapshot exchange — instead of a server, developers could export/import JSON snapshot files via shared drives, Git repos, or messaging. This keeps the no-server constraint while enabling lightweight team visibility.

2. Configurable aggregation granularity — let teams choose how coarse the shared signals are (e.g., daily vs. weekly rollups, category-level vs. overall scores only) to match their comfort level with data sharing.

3. Differential privacy noise — for teams that want stronger guarantees, add optional ε-differential privacy noise to numeric scores before they leave the machine.

4. Local-only team dashboard — a read-only view that renders imported peer snapshots alongside your own data, with no live connection required. Think "git log" but for coaching signals.

5. Goal-setting and nudges — team leads could define target score ranges per practice group; individual developers see gentle nudges when their rolling average dips below the team goal, all computed locally.

6. Anonymous benchmarking — show percentile bands (p25/p50/p75) derived from imported snapshots so individuals can see where they stand without revealing who is who.

These all preserve the core privacy-first, no-server philosophy while opening up useful team dynamics.

[aymenfurter]

Please do not implement any PR for this issue, it is still in design phase.