Does the demo upload my Claude transcripts?
No. The hosted demo cannot reach your filesystem. It renders committed synthetic utilization rows baked into the static build.
Telemetry and privacy
The Living Vault measures how reusable agentic assets perform across AI coding harnesses by reading structural metadata from each harness's own local session logs. It never reads prompt text, responses, or file contents. The public demo uses synthetic data only; real usage stays on your machine.
Try the demo Public repo coming soon
What each harness contributes
| Harness | Local source (typical) | Typical honesty grade |
|---|---|---|
| Claude Code | Project transcript JSONL under your Claude config directory | full per-asset invocations, success, errors, latency |
| Codex | Session rollout JSONL files | partial session-level signals; gaps marked |
| Grok | Session folders and unified log (metadata paths only) | partial summaries at log fidelity |
| Copilot | Session event JSONL streams | occupancy or derived; cost not faked |
Metadata only, never content
Collectors read: event types, tool and skill names, invocation and error counts, durations, token totals, ids, and timestamps.
Never read: prompt text, model responses, reasoning, command text, tool arguments, tool results, file contents, file paths, summaries, titles, or auth tokens. Parser unit tests enforce this with sentinel content in fixtures.
Notable exclusions by harness
- Grok: chat history, prompt context, system prompts, terminal output, and worktree checkouts (can contain secrets).
- Copilot: message content, tool arguments and results, session-store database text columns. The collector does not open the SQLite store.
- Codex: function output strings and message payloads.
- Claude Code: transcript message bodies; only structural invocation metadata is parsed.
Where data lives
Public demo build: ships committed synthetic utilization.json and harness-usage.json only. No access to your machine.
Live Fabric (owner push): curl living-vault-collect.mjs from the deployed app, run node living-vault-collect.mjs telemetry, then upload ./utilization.upload.json in Settings. Fabric never reads your home directory.
Local dev loop: repo collectors write gitignored *.local.json snapshots; the dev server prefers them. Demo builds never bundle real usage.
Honesty over completeness: the UI labels each metric with a fidelity grade (full | partial | occupancy | derived | none). Unobserved assets show "no telemetry", not a measured zero.
How it reaches the Monitor
- Harness writes session logs locally (already happening today).
- A metadata-only collector aggregates Claude Code usage into upload rows.
- You upload
utilization.upload.jsonin Settings (live Fabric). - Rows reconcile to governed catalog asset ids (type + slug).
- Monitor, Effectiveness, and Cost surfaces render the signal with per-harness honesty grades.
Public demo vs live Fabric push
| Question | Public demo (SWA) | Live Fabric (owner push) |
|---|---|---|
| Reads your machine? | No | Only when you run the collector locally |
| Telemetry data | Synthetic seed data in the demo build | ./utilization.upload.json you upload in Settings |
| Repo checkout required? | No | No, curl living-vault-collect.mjs from the deployed app |
| Honesty grades | Illustrative on demo assets | Reflect Claude Code transcript fidelity today |
Collector entrypoints
Deployed app (no repo):
curl -fsSL …/living-vault-collect.mjs -o living-vault-collect.mjsnode living-vault-collect.mjs telemetry→./utilization.upload.jsonnode living-vault-collect.mjs assets→./living-vault-assets.json
App repo (local dev):
node scripts/vault-telemetry.mjs→ gitignoredutilization.local.jsonnode scripts/vault-harness-usage.mjs→ gitignoredharness-usage.local.json
Parsers read metadata only. Unit tests assert sentinel prompt content never appears in output.
Honesty grade quick reference
| Grade | Meaning |
|---|---|
full |
Directly instrumented end to end (Claude Code today) |
partial |
Measured where logs allow; gaps marked (Codex, Grok) |
occupancy |
Presence and activity only (Copilot) |
derived |
Inferred from adjacent signals, labeled as inference |
none |
Not measurable; shown as a gap, never a fake zero |
Frequently asked questions
Why not read everything the harness logs?
Session logs contain prompts, responses, and file paths. The product measures asset effectiveness, not conversation content. Collectors parse structural fields only and tests enforce exclusions.
Why show Copilot cost as "not tracked"?
Copilot logs do not expose reliable per-request dollar amounts. The UI labels cost as untracked rather than inventing a number.
What blocks org-wide rollups today?
A trusted owner can already collect locally and upload into Fabric. Team-wide reach still needs every contributor to run collectors (or scheduled sync), per-user attribution on utilization rows, and RBAC for who may push whose signal. Those org-scale pieces remain on the roadmap.