Claude Code¶

Sibling file of the Claude home, not a member of it, so no path pattern rooted at ${CLAUDE_CONFIG_DIR or ${HOME}/.claude} reaches it. A projects object keyed by absolute project root records per-project state including lastSessionId, lastSessionFirstPrompt, lastSessionModified, and per-project MCP server lists; the top level carries oauthAccount, userID, machineID, and usage counters. lastSessionFirstPrompt is the only prompt text, and it duplicates the transcript, so the row is an inventory of which projects Claude knows rather than a search surface. Older versions kept per-project history arrays here; those are now empty.

Sidecar beside each subagent transcript, carrying spawnDepth, agentType, and — when the dispatch came from the Task tool — toolUseId, description, model, and parentAgentId. description is the dispatch text, the only record of why a subagent was spawned; agentgrep reads it, else name, as the title of every record in the transcript beside it. model is only the requested alias (sonnet, inherit); the transcript’s assistant turns carry the resolved model, which is what records report. Worktree-scoped runs also carry worktreePath and spawnedWithWorktree.

Git worktrees Claude Code creates with -w/--worktree, checked out inside the repository on a branch named worktree-<name>. There is no index file: the record is git’s own .git/worktrees/ registry plus the cwd and gitBranch fields every transcript line repeats. Each worktree mints its own encoded directory under projects/, so one repository’s history splits across several store keys — see the worktree section of the backends index. Catalogued so the adapter never indexes a working tree as chat, parity with cursor-cli.worktrees.

Global prompt history JSONL. Each line carries display, pastedContents, timestamp (Unix milliseconds), project, and sessionId. Large text pastes may be content-addressed through paste-cache/<contentHash>.txt.

User-prompt audit log used by Claude Code prompt history; project transcripts remain the full conversation replay.

JSONL; stream fragments grouped by uuid, dedup across /resume, skip isCompactSummary: true. Keys: type, uuid, parentUuid, timestamp, sessionId, cwd, gitBranch, version, message.role, message.content[] (text/thinking/tool_use/tool_result), message.usage. Non-message record types also appear — queue-operation (queued prompt text in content, no role), mode, permission-mode, ai-title, last-prompt, agent-name, pr-link, attachment, file-history-snapshot, system, bridge-session, and the api-request, api-request-shape, and api-request-blob snapshots — only role-bearing message records are indexed, and api-request-blob, which copies each message sent to the API, is skipped. Each record takes the event’s top-level timestamp; the nested message carries none. custom-title (set by /rename) and ai-title name the session: agentgrep reads the last of each from the file’s final 64 KiB and titles every record with the custom name, else the AI one.

Same JSONL line format as the parent session. Each file is one sub-agent dispatch from the Task tool.

Sub-agent transcripts are conversation content; de-duplicate with the parent session by uuid.

Markdown files with YAML frontmatter; the auto-memory feature. Each file holds one fact/feedback/project/reference memory.

Persistent user and project memory Markdown loaded into Claude Code. Project-root files are discovered only from roots already referenced by local Claude project transcripts.

Claude Code task JSON with id, subject, description, status, blocks, blockedBy, and optional activeForm / metadata fields.

Persistent todo lists keyed by agent UUID.

Shell environment snapshots; rarely contains conversation text.

SQLite conversation/app state. Observed tables include base_messages, user_messages, assistant_messages, and conversation_summaries; message tables can duplicate JSONL transcript content, so this stays opt-in.

Markdown plans stored under the configured Claude Code root. Project settings may redirect plan creation, but this root is the global default observed in current installs.

Orchestration-workflow driver scripts emitted by the Workflow tool, each carrying an embedded meta = {name, description} that names the workflow intent. Machine-generated driver code; inspectable opt-in, not searched by default.

Claude Code’s own derived per-session reflection summaries. The readable natural-language fields are brief_summary, underlying_goal, and friction_detail; counts and category fields are metadata. Derived state, not transcript — inspectable opt-in, not searched by default.

Per-session memory summary Markdown plus optional session-memory template/prompt configuration. It can contain prompt context but is not part of the default transcript search surface.

User configuration, local overrides, keybindings, policy limits, and update/auth cache state. Documented for inventory, not searched as prompt history.

Authentication and account credentials. Documented but never enumerated.

Update and cleanup marker files; not prompt history.

Cached usage and statistics state; not conversation content.

Debug output and diagnostics. Catalogued separately from transcripts.

Backup files retained by Claude Code; opaque inventory only.

General cache directory; not a prompt/history source.

Content-addressed file-history backups referenced by transcript snapshots. This can be large and is not conversation history.

Shell integration snapshots; runtime state rather than prompt history.

Base64/image payload cache for pasted image references.

Context-mode counters and local app state, sometimes backed by SQLite.

Security warning and repository baseline state.

User and managed skill instructions loaded into Claude Code behavior.

Legacy custom slash-command Markdown loaded through the skill loader.

Project-local commands, agents, and skills. Roots are bounded to projects already observed in local Claude transcript metadata.

Team config JSON with member prompts and coordination metadata.

Environment state captured for active or resumed sessions.

IDE integration state and bridge metadata.

Claude-in-Chrome native-host and browser bridge state.

Native installer and local binary-management state.

Background job state and metadata, summarized without raw values.