How to Find & Resume OpenAI Codex CLI Sessions on Mac

When using OpenAI Codex in your terminal, the agent executes complex multi-file refactors, runs bash tests, and generates code diffs across your projects. Finding past prompt reasoning, reviewing modified files, and resuming interrupted sessions shouldn't require manual file hunting. Here is how to find, inspect, and resume OpenAI Codex sessions on macOS.

Updated October 2026 · By the team behind L2Cache

Where Are Codex Session Logs Stored on macOS?

OpenAI Codex organizes transcripts into date-partitioned JSON Lines files in your home directory:

macOS & Linux Path:

~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl
~/.codex/memories.sqlite

Windows Path:

%USERPROFILE%\.codex\sessions\YYYY\MM\DD\rollout-*.jsonl

Each session is assigned a unique rollout ID containing the full conversation history, agent actions, tool inputs, and return results formatted in JSON Lines.

Inside the JSONL: What OpenAI Codex CLI Tracks

Every Codex CLI rollout is an append-only JSON Lines (.jsonl) stream. Each line captures a distinct phase in the agent's turn loop:

{"type":"session_meta","session_id":"rollout-2026-09-30-0814-jwt-auth","cwd":"/Users/developer/backend","timestamp":"2026-09-30T08:14:10.052Z","model":"o3-mini","aiTitle":"Refactor JWT token validator with clock skew leeway"}
{"type":"event_msg","role":"user","turn":1,"timestamp":"2026-09-30T08:14:11.120Z","content":"Update our auth middleware to allow 60 seconds clock skew leeway on JWT token expiration"}
{"type":"tool_call","turn":1,"tool_name":"file_search","arguments":{"pattern":"*auth*.go","dir":"pkg/auth"}}
{"type":"item_completed","turn":1,"tool_name":"file_search","status":"success","result":"pkg/auth/jwt_validator.go\npkg/auth/middleware.go"}
{"type":"tool_call","turn":1,"tool_name":"edit_file","arguments":{"file_path":"pkg/auth/jwt_validator.go","instructions":"Add WithLeeway(60 * time.Second) to jwt.ParseWithClaims"}}
{"type":"item_completed","turn":1,"tool_name":"edit_file","status":"success","diff":"+ claims := &CustomClaims{}\n+ token, err := jwt.ParseWithClaims(tokenString, claims, keyFunc, jwt.WithLeeway(60*time.Second))"}
{"type":"response_item","role":"assistant","turn":1,"content":"I have updated `pkg/auth/jwt_validator.go` with 60s leeway. All test suites pass."}
{"type":"turn_metrics","turn":1,"tokens":{"prompt_tokens":3120,"completion_tokens":420,"cached_tokens":18500,"total_tokens":22040},"duration_ms":1890}

Key Fields Tracked in Codex Rollout Logs:

• aiTitle: A concise, model-generated synopsis of the session's objective (e.g., "Refactor JWT token validator with clock skew leeway"). L2Cache uses this to render clean, readable session cards instead of cryptic UUID hashes.
• session_id & model: Unique identifier along with the exact model (such as o3-mini or gpt-4o) utilized during that session.
• event_msg & response_item: Full user prompt inputs paired with the assistant's final response and explanation.
• tool_call & item_completed: Tool executions (such as file searching, command execution, or file editing) including inputs, return data, and inline git-style code diffs.
• Token Analytics & Prompt Caching: Detailed breakdown of prompt_tokens, completion_tokens, and cached_tokens (showing tokens retrieved from OpenAI's prompt cache).

Why Codex Conversation History Changes How You Work

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

How to Resume OpenAI Codex Sessions

OpenAI Codex CLI provides built-in terminal commands to inspect and pick up where you left off in past coding sessions:

Limitations of Directory-Based Rollout Files:

• Date Fragmented: Because sessions are partitioned by YYYY/MM/DD, finding a prompt from 3 weeks ago requires recursively searching across dozens of nested folders.
• Nested Tool Payloads: Tool executions are stored as JSON strings inside JSON objects, making quick visual scanning difficult in standard text editors.
• Context Loss: The native CLI picker shows limited metadata without token counts or search across tool diffs.

Searching Codex Sessions via Terminal (ripgrep & jq)

If you prefer shell scripting, you can search across all Codex session files recursively using find, jq, and grep:

# Search all past Codex prompts for a keyword
find ~/.codex/sessions -name "rollout-*.jsonl" -exec jq -r 'select(.type=="event_msg") | .content' {} + | grep -i "docker"

# Find recently executed tool commands and file paths
find ~/.codex/sessions -name "rollout-*.jsonl" -exec jq -r 'select(.type=="tool_call") | .arguments' {} +

Native Codex Resume vs. L2Cache History

While the native codex resume CLI command is great for quick terminal rollbacks, L2Cache for Mac adds visual search, token analytics, and cross-project indexing: