What is collected
Tracemason monitors agent activity in development environments, which contain sensitive material. This page states exactly what leaves a developer's machine. It is written so that an organization can share it with its developers as-is.
Tracemason observes agents, not people. It does not record keystrokes, screens, prompts or time spent.
Never collected, at any level
- Prompts and assistant responses
- File contents, including the text written or replaced by an edit
- Command output (stdout / stderr) and tool responses
- Environment variables
- Conversation transcripts
- Error messages (only an error type is recorded), unless the level is extended
Collection levels
The level is set per organization (Settings → General) by an Owner or Admin, is visible to every member, and is printed by npx @tracemason/cli init and npx @tracemason/cli status. The server enforces it again on arrival, so a sender cannot store more than the organization's level allows.
| Field | Minimal | Metadata (default) | Extended |
|---|---|---|---|
| Session id, timestamps, agent name, model | ● | ● | ● |
| Tool name, action category, outcome, duration, exit code | ● | ● | ● |
Programs invoked by a command (git, psql) |
● | ● | ● |
| Repository slug and branch | ● | ● | ● |
| Whether a sensitive path was touched (yes/no) | ● | ● | ● |
File paths (project-relative; home directory collapsed to ~) |
● | ● | |
| Network hostnames referenced by commands (no paths, no query strings) | ● | ● | |
| Command text, after redaction | ● | ||
| Tool error message, after redaction | ● |
How commands are handled
At minimal and metadata, the command string never leaves the machine. The collector parses it locally and uploads only the classification — for example database.destructive, the program names, and the hostnames it refers to.
The classifier tokenizes the command and inspects its structure; it is not a list of regular expressions over the raw string, so echo 'rm -rf /' is not mistaken for a deletion. It is also not a shell interpreter. Commands it cannot see through — eval, command substitution, encoded payloads piped into a shell — are marked opaque, and the dashboard says their classification is best-effort.
Redaction
Two layers, applied on the sender and again on the server:
- Allowlist. Only known metadata keys are forwarded; everything else is dropped. Nested objects are never forwarded.
- Pattern scan. Remaining strings are scanned for credential shapes: private keys, AWS / GitHub / Anthropic / OpenAI-style / Stripe / Slack / Google keys, JWTs, bearer tokens, URL credentials, and
NAME=valueassignments whose name suggests a secret.
Pattern matching is a second line of defence, not the first. At the default level there is very little free text to scan.
Where data lives
- In transit: collectors and SDKs refuse any endpoint that is not HTTPS, except
localhost. - At rest: Tracemason does not encrypt rows itself. Encryption at rest is provided by your PostgreSQL deployment's storage encryption; whether it is on depends on how you deploy.
- Credentials: ingestion keys and invitation tokens are stored only as SHA-256 hashes. A key is shown once, when created.
- On the developer's machine:
~/.tracemason/holds the ingestion key (mode 0600) and the outgoing queue.
Retention and deletion
Events are deleted after the organization's retention window by an hourly worker; each run is recorded under Settings → Data & retention. Owners can delete an organization and everything in it. See Data retention.
Access
Five roles — Owner, Admin, Security Analyst, Developer, Viewer — enforced on the server. The full permission matrix is shown under Settings → Members. Sensitive changes are written to an audit log.
Third-party AI providers
Tracemason does not send event data to any AI provider. There is an organization setting recording consent for future AI-assisted trace summaries; that feature is planned and not implemented, so the setting currently has no effect.
Claims we do not make
Tracemason holds no security certifications. Hook-based enforcement is not tamper-proof (see Claude Code collector).