Terms and product disclosure

Before you decide whether to enable capture, read this disclosure in full: One fact first: this version has four capture lanes. Three share one on/off switch (capture_enabled): Claude Code sessions from ~/.claude/projects, Codex CLI sessions from ~/.codex/sessions/**/rollout-*.jsonl, and a short note (date, project name, machine name, agent, tags, duration, working directory, up to 20 changed filenames) written to /Users/Example/ObsidianVault/Sessions/ when each Claude Code session ends. Claude Code markdown goes to /Users/Example/ObsidianVault/conversations/; Codex markdown, including this conversation, goes to /Users/Example/ObsidianVault/conversations/Codex/. Captured text and command input/output are checked against a fixed set of patterns (AWS and other vendor API-style keys, credential-style key=value pairs, email, phone, SSN, 16-digit card numbers) and matches are replaced before writing; this is pattern matching against known shapes, not a guarantee every secret is caught. Codex notes include command output; Claude Code notes omit it. Codex note metadata -- its session ID, timestamps, machine name, and the source file's absolute path -- is written as-is and is not covered by this masking; an absolute path can include your Mac username. The session-end note above is NOT masked; its fields are written as-is. Pausing delays exports. When you turn capture back on, saved conversation text from the pause can also be exported. A fourth lane, not covered by the capture_enabled switch, mirrors your per-project Claude memory files from ~/.claude/projects/*/memory/* into /Users/Example/ObsidianVault/ClaudeMemory/<this Mac's name>/<project name>/, copied byte-for-byte with no masking. It runs on its own hourly schedule regardless of that switch. When a mirrored memory file changes or is removed at the source, its previous vault copy is written to that project's _snapshots folder with a timestamped filename before the live mirror is updated or removed. It is kept at least 90 days, then pruned on a later sync. The product reads Anthropic's source memory files but never modifies them. - For automatic tagging, the installed default checks a literal loopback Ollama endpoint. When local Ollama is present, tagging is processed by Ollama on this Mac; configured Ollama and embedding URLs outside literal loopback are refused before a request and the product uses its visible local fallback instead. This loopback traffic is the capture engine's only ongoing network activity; the separately disclosed daily update check is installer-program activity. - Separately, this installer program (not the capture engine) contacts two destinations of its own. It sends your license token, a device fingerprint, this Mac's hostname, and its operating system name to our activation server -- once per attempt, so a failed and retried activation sends it again. It also may download the Auto Note Mover plugin from GitHub during the post-install steps above; that download is skipped if a valid copy is already present, and is not guaranteed to succeed. Activation registers this Mac against your license's 3-machine limit. Re-running on this same Mac reuses its fingerprint-based slot. If the limit is reached, the activation error shows the recovery page where you can free a slot. Once a day at most, this product checks https://obsidian-setup.fly.dev/version?product=obsidian-second-brain&os=mac for the latest published version. It sends only that fixed product/os query -- no license token, no machine identity. The response is a version number and a release date, nothing else: it never returns or redirects to a URL, and this product never downloads anything from it. If you're told about an update, it always points you to the same fixed page, https://dontsleeponai.com/obsidian-claude-code, never anywhere the server names. A slow, unreachable, or erroring server is treated as "no update" and never delays or affects capture. Selected vault: /Users/Example/ObsidianVault