How to Find & Resume Claude Code Sessions on Mac

Software engineering workflows have undergone a massive shift. Developers no longer rely solely on basic autocomplete; today, we pair with autonomous CLI coding agents like Claude Code and Codex to execute multi-turn refactors, write unit tests, and resolve complex edge-case bugs across full repositories.

Figure 1: L2Cache floating panel surfacing Claude Code & Codex session summaries, token consumption, prompt cache hit ratios, and 1-click terminal resumption.

Where Are Claude Code & Codex Session Logs Stored on macOS?

When running agentic coding assistants in your terminal, every prompt, tool response, and diff is recorded locally on disk:

• Claude Code stores conversation logs inside your user home directory at ~/.claude/projects/<project-folder>/<session-uuid>.jsonl. Project directories contain raw JSONL session records with timestamps, tokens, tool arguments, and bash tool output.
• OpenAI Codex CLI maintains session memories and tool execution logs under ~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl as well as persistent memory in SQLite.

Inside the JSONL: What Claude Code Actually Tracks

Every Claude Code session is a continuous JSON Lines (.jsonl) stream. Each line is an independent, valid JSON object recording a discrete event in the lifecycle of the session:

{"type":"session_meta","sessionId":"a85fd1e8-3b4a-4f12-89cd-7e62a104f982","cwd":"/Users/developer/apps/l2cache","timestamp":"2026-09-30T07:15:22.104Z","model":"claude-3-7-sonnet-20250219","aiTitle":"Fix memory leak in panel animation timer"}
{"type":"USER_INPUT","turn":1,"timestamp":"2026-09-30T07:15:23.002Z","content":"Fix the memory leak in panel animation where NSTimer is retained across window close"}
{"type":"PLANNER_RESPONSE","turn":1,"content":"Let's search for timer allocations in PanelAnimation.swift.","tool_calls":[{"name":"grep_search","args":{"Query":"Timer.scheduledTimer","SearchPath":"/Users/developer/apps/l2cache/src"}}]}
{"type":"TOOL_RESULT","turn":1,"tool_name":"grep_search","status":"DONE","output":"PanelAnimation.swift:42: self.animTimer = Timer.scheduledTimer(timeInterval: 0.016, target: self, ...)"}
{"type":"PLANNER_RESPONSE","turn":1,"content":"I will replace the strong timer reference with a weak target and invalidate on deinit.","tool_calls":[{"name":"replace_file_content","args":{"TargetFile":"/Users/developer/apps/l2cache/src/PanelAnimation.swift","StartLine":42,"EndLine":46,"ReplacementContent":"self.animTimer = Timer.scheduledTimer(withTimeInterval: 0.016, repeats: true) { [weak self] _ in\n    self?.updateFrame()\n}"}}]}
{"type":"TOOL_RESULT","turn":1,"tool_name":"replace_file_content","status":"DONE","output":"Successfully replaced lines 42-46 in PanelAnimation.swift"}
{"type":"turn_metrics","turn":1,"metrics":{"input_tokens":1280,"output_tokens":310,"cache_creation_input_tokens":6400,"cache_read_input_tokens":48920,"duration_ms":1140}}

Key Fields Tracked in Every Agent Session:

• aiTitle: Claude Code generates a descriptive, high-level summary of the session's goal (e.g., "Fix memory leak in panel animation timer"). This title is crucial for finding past discussions quickly without reading raw terminal traces.
• USER_INPUT vs. PLANNER_RESPONSE: Full conversation flow between your requirements and the agent's internal chain-of-thought planning.
• tool_calls & Arguments: The exact tools invoked (grep_search, run_command, replace_file_content), target file paths, line numbers, and exact code diffs.
• TOOL_RESULT: Real-time stdout, stderr, compiler errors, unit test output, and return codes captured during execution.
• Prompt Cache Analytics: cache_read_input_tokens tracks how many tokens were served directly from Anthropic's prompt cache (saving 90% in cost and reducing latency by 80%), while cache_creation_input_tokens records new prompt state written to cache.

Why Agent Conversation History Changes How You Work

Treating your Claude Code and Codex transcripts as a searchable repository memory—rather than a pile of disposable terminal sessions—fundamentally transforms day-to-day engineering velocity:

How to Search and Resume Past Claude Code & Codex Sessions

Both Claude Code and Codex CLI offer built-in terminal resumption commands:

Native CLI Resume vs. L2Cache History Dashboard

While native CLI commands are helpful for local directories, L2Cache for Mac bridges cross-project indexing, visual search, and token analytics:

Tracking Claude Code Token Usage and Prompt Cache Hit Rates

In L2Cache 1.4, we introduced native AI Agent History & Analytics—bringing full visibility to your CLI agent sessions directly from your Mac's menu bar or global hotkey.

Preventing Agent Memory Loss: Promoting Gotchas to AGENTS.md

When an agent spends 20 turns debugging a subtle AppKit threading bug, a SQLite concurrency issue, or a custom build flag constraint, it learns vital facts about your codebase. But when that session finishes, those insights often stay buried in local log archives.

L2Cache's Repository Knowledge Graph bridge converts transient agent discoveries into permanent repository documentation:

# Example AGENTS.md rule promoted from agent session #a85fd1e8
## AppKit & SwiftUI Concurrency Rules
- guard `view.window != nil` before displaying popovers in async tasks to prevent EXC_BAD_ACCESS.
- Never perform DB reads/writes on the main thread; use `Task` with `await`.

1. Extract Discovered Gotchas: L2Cache parses agent execution logs to surface fragile components, recurring error tracebacks, and technical debt notes.
2. 1-Click Promotion to AGENTS.md: Review extracted constraints and click "Promote to AGENTS.md" to write the rule directly into your project's root AGENTS.md or CLAUDE.md file.
3. Cross-Agent Alignment: Ensure all coding agents (Claude, Codex, Antigravity, or Cursor) share the exact same rules, preventing agents from repeating known bugs in future sessions.

100% Local, Zero-Cloud Privacy

Just like L2Cache's core clipboard management, all Agent History & Analytics parsing occurs strictly on-device on your Mac. L2Cache reads local session logs directly from your home directory, stores analytics in local SQLite databases using GRDB, and never sends your prompts, code snippets, or token statistics to external servers.

Looking to find a conversation you already had? Browse the Claude Code session history viewer or Codex CLI history viewer for L2Cache’s Mac workflow.