data.md

description	What Pair keeps and where: logins and keys by the config, the rest in its data folder, an index in your repository if asked, and what leaves the machine.

Data and privacy

• Logins and keys live next to the config, readable by you only (models.md): auth.json (logins and stored keys; $PAIR_AUTH moves it), service.env (the variables your entries name in apiKeyEnv, stored by pair setup for the service) and device-id (the id "Sign in with ChatGPT" is given; it sits next to auth.json and moves with it).
• Everything else lives in the data directory: ~/.local/state/pairing-buddy by default, or under $XDG_STATE_HOME, or dataDir in the config.
• Pair's panel in the Herdr window, where that window is installed and "herdrPanel" is not false: pair setup puts the mod in its mods (~/.config/herdr-app/mods/pair) and adds a key and a + menu entry to its keys.json and programs.json, keeping each file first as <file>.pair-bak-<time> (herdr.md).

File	What it holds
partner.sqlite	The Partner's conversations, one per project (pi-durable)
sessions.sqlite	Sessions, goals, chat history (with the remark, place or suggestion each message is about), session notes and debriefs (with where the next session picks up), active time (in all and per project), the remarks with your verdicts, what the decision model answered about each moment it judged (its probabilities, the Watcher's score, what the rules would have done, the remark it led to, and the state it was told), and the ledger of what Pair did (what an entry opens to, a command's output or an edit's text before and after, for a week)
chat-history.json	What you sent in the chat pane, the last 200 lines of each project, for Up to bring back
runs/	The output of the latest tests on save run of each session and of the hand-over's check, and in runs/trees/ the scratch copy of each project that check runs in (safe to delete)
gen/	The folders of /gen try samples, gen/<project>-<hash>/try/<n>/: what the model wrote and what its runs left there; and the copies of the project /gen scaffold builds in, gen/<project>-<hash>/<n>/ (with its manifest <n>.json), or with --branch a git worktree there. Each try or copy goes once it has not changed, run or been used for gen.keepTries days (7), or when you delete it (x); a worktree only when you delete it (git worktree remove works too). Safe to delete, apart from a worktree with commits of yours
panes/	Pair's own pane: its shell's output log and each command's exit code, one folder per pane, removed when the pane closes (herdr.md)
voice/	A push-to-talk recording, only while whisper.cpp reads it (LFM3-Omni's and OpenRouter's are kept in memory only); and once you turn it on (:Pair voice on), voice/lfm3-omni/: LFM3-Omni's files (4.0 GB, each with its checksum note) and in cache-<revision>/ their weights packed for your CPU (4.2 GB). :Pair voice off deletes them
off	After :Pair find off with an embedding model: the note that keeps search by meaning off until :Pair find on (semantic-grep's, in Pair's data folder)
index/	Search by meaning's index, semantic-grep's, a SQLite file per repository (<repository>-<hash>.sqlite), shared by its worktrees: each piece of code's file, lines, hash, kind and name, the FTS5 index of its words (lower case and stemmed, with their places, but not the text itself), and a vector per piece (256 numbers, about 1 KB, for EmbeddingGemma); no copy of the code. :Pair find off deletes them, :Pair index rebuild makes one again
meaning/	Search by meaning's EmbeddingGemma 2, once you turn it on: its model (embeddinggemma-2-text-270m.litertlm, 165 MB, with its checksum note), and venv/, the venv with its runtime LiteRT-LM that :Pair find on makes (with a venv-made-by-pair.json note next to it) unless you made it there yourself. :Pair find off deletes the model and the venv Pair made
daemon.log	The daemon's log, when pair started it (under systemd it goes to the journal, under launchd to ~/Library/Logs/pairing-buddy.log)
herdr-panel.json	While the daemon runs, the port and token of its loopback port for Pair's panel in the Herdr window, readable by you only; only when the socket's folder is not yours alone, else it is <socket>.panel beside the socket
daemon.lock	Makes sure one daemon owns the files above

The run log (each build's and run's output and the lines you typed before it) is kept in the daemon's memory only, for the session.

Nothing is written to your repository, unless you put search by meaning's index there ("meaning": {"index": "project"}: .pair/index.sqlite, and .pair/ in .git/info/exclude). The work board (Shepherdr) is read, and written to only when you pick it: /gen board adds one task, with its brief shown to you first, as the pick you chose says (add only, add and delegate, delegate with a plan first, or delegate for research only), and a task's review tells its agent what you wrote, asks it to merge (a tell), or archives the task, each on the pick that names it, and /brief sync tells the branch's task what changed in the brief, on its pick too. What the board's working tasks change is read with git, never changed. Nothing leaves your machine except the requests to the model endpoints you configured (OpenRouter's voice too, when you set it), the one-time downloads you say yes to (search by meaning's EmbeddingGemma 2 from Hugging Face and LiteRT-LM's wheel from PyPI, meaning.md; voice's LFM3-Omni from Hugging Face and its Python packages from PyPI on :Pair voice on, voice.md), gh pr view for the context card when gh is installed and, for a login, to the provider's sign-in service. Those requests carry your code: the Watcher gets each moment's text (your edits, a command's output, a test's failure), and the Partner the files and lines it reads, the diffs it suggests and your questions; an embedding model of your own gets the text of each piece of code it indexes, and your searches; a rerank model gets a search's words and the text of its first pieces. With a cloud provider, that code goes to the provider: check that your employer allows it before you use Pair on work code; a local server keeps it on your machine. A decision model, when you set one, gets each moment's kind and text, your goal, the paths you edited, your last commands and test counts, and your verdicts on earlier remarks; never the contents of a file. With voice through OpenRouter, your spoken question's recording goes to OpenRouter's transcription model, and the answer's first sentences to its speech model.