Backends¶

How agentgrep learns a working directory¶

Agents do not agree on how to write down where a session ran, so a cwd reaches a record through one of three tiers. The tier decides what you can filter on, and it is a property of the store, not of your query.

Lossless. The store wrote the path, or an encoding that inverts exactly: a cwd column (codex.state_db, pi.context_mode_db), a nested key (cursor-ide.state_vscdb), a sibling file (gemini.tmp.chats), or a %2F-escaped directory name (grok.sessions). Records carry origin.cwd and answer --cwd and cwd: with the real path.

Lossy. The store folded the path into a name that cannot be inverted on its own. Cursor CLI’s projects/<name>/ segment replaced every separator with - and escaped nothing, so foo-bar is equally consistent with /foo/bar and /foo-bar. agentgrep reconstructs the name against the filesystem and keeps the answer only when exactly one reconstruction resolves to a directory that exists. Ambiguity, a directory that has since moved, and a pathological name that exhausts the probe budget all leave origin.cwd unset: a fabricated path does not merely omit a result, it makes a repo-scoped filter silently skip your own project, so a known-unknown is the safer answer.

Digest. The store only ever knew a hash of the path (cursor-cli.chats). Records carry origin.cwd_hash and nothing else, so they answer cwd_hash: and not cwd:. A digest does not invert, so agentgrep never reverses one into a cwd — and it never runs the hash the other way either: a cwd_hash is always read from the name the store chose, never computed from a cwd recovered somewhere else. A path segment is admitted as a cwd_hash only when it has a digest’s shape, so a backup.db sitting beside a real database does not publish its own file name as a searchable project identity.

The tiers stack. A store that hashes its directory name and repeats the literal path inside — pi.context_mode_db, gemini.tmp.chats — gives a record both cwd and cwd_hash.

Only cwd_hash is a fact about where a source lives, so it is the only origin field agentgrep trusts to skip a store before opening it. A cwd learned from a sibling workspace.json or a project directory name describes the source, not a promise about every record inside it — a Cursor composer bubble can name its own worktree — so those stores are still opened and filtered record by record.

Where each agent puts them¶

Five agents create worktrees, and they disagree about where.

Cursor CLI and Cursor IDE share one root outside the repository; Claude Code and Gemini CLI create theirs inside it, on a branch named worktree-<name>. Codex, Pi, OpenCode, Antigravity, VS Code Copilot Chat and Windsurf have no worktree feature at all.

Only Grok keeps an index. ~/.grok/worktrees.db holds a worktrees table with the checkout path, source repository, git ref, head commit, status, and — the useful part — a session_id joining each worktree straight to the session that created it. Everything else is recovered from the directory tree and from git’s own .git/worktrees/ registry.

What agentgrep can tell you¶

worktree: matches records whose session ran in a git worktree. Today Cursor IDE is the only backend that populates it, because its gitWorktree block is the only place an agent writes the worktree path into the record. Cursor writes that block only for a checkout it created, never for an ordinary clone, so its presence is itself the evidence.

The other four are recoverable but not yet recorded: Claude Code and Gemini CLI repeat the working directory inside every record, so their worktree sessions stay resolvable by cwd: even after the checkout is deleted, and Grok’s database has the join but no adapter reads it.

Deleted worktrees are the sharp edge. Claude Code’s encoded project directory survives the checkout, so its transcripts keep pointing at a path that no longer exists. That is harmless for cwd:, which matches text — but Cursor CLI recovers its working directory by probing the filesystem, so a removed worktree there leaves the record with no cwd at all rather than a stale one. See How agentgrep learns a working directory for why that tier answers with a known-unknown instead of a guess.