Codex Session Analysis

The Hidden Gap In 903 Sessions

On one side, you adopt new Codex models almost immediately. On the other side, the workflow features that cut the most friction stay unused. This story explains what happened, why it matters, and exactly how to close that gap.

903 Sessions audited
847 Missed-feature moments
318 Days covered
Same starting point Model adoption: 75.8% Workflow adoption: often 0%
The fork appears in February 2026: fast pickup of gpt-5.3-codex, almost no pickup of new orchestration and coordination features.

How The Audit Worked

We treated each session as evidence, not anecdote. Every tool call, command, timestamp, and prompt was parsed, then matched against release windows so we only count a “missed feature” after it was actually available.

1. Read Logs 903 JSON/JSONL files from sessions/ 2. Extract Signals tools, commands, durations, prompts 3. Align Dates release-aware checks post-release only 4. Score Gaps 847 opportunities ranked by impact
Analyzer: analyze_codex_sessions.py, Output: session_analysis.json
  1. Parsed every available session file (legacy rollout JSON + newer JSONL records) from your historical Codex logs.
  2. Detected concrete events: command patterns, tool calls, approval friction, long-running tasks.
  3. Linked feature usage to release dates from official Codex changelog RSS.
  4. Generated an opportunity catalog in opportunities.csv.

This matters because it prevents false alarms. A feature is only counted as “unused” in sessions that happened after that feature shipped.

The Core Tension: Speed In One Layer, Stall In Another

You clearly change behavior when the upside is obvious. But the upside is currently concentrated in model choice, while collaboration and orchestration features stay mostly dormant.

Fast where it is visible

75.8% of eligible sessions used gpt-5.3-codex after the February 5, 2026 rollout window.

Ready, but not redirected

You already use manual parallel shell patterns in 68.4% of eligible sessions, but parallel tool usage is still 0%.

Coordination features are the gap

spawn_agents_on_csv and request_user_input both remain at 0% in post-release windows.

When New Features Arrived, What Actually Changed?

This timeline is the important forensic view. Dots near the top indicate high post-release adoption. Dots on the floor show “available but untouched.”

Release-aware data from recent_feature_coverage.csv

Non-obvious takeaway 1

This is not a “you avoid change” story. You changed quickly for models. The stall is specifically around interaction patterns (coordination, approvals, and thread controls).

Non-obvious takeaway 2

The top three opportunity types account for 726 of 847 missed moments (85.7%). You do not need ten fixes. You need three defaults.

Where The Friction Actually Lives

Opportunity counts reveal where time leaks most often. The tallest bars are where behavior changes will pay back fastest.

Parallel reads are the biggest tax

457 sessions showed sequential, independent reads that could have been run with parallel.

Approval friction is measurable

138 sessions had repeated permission friction where request_user_input could have unblocked faster.

Concrete examples from your sessions

  1. Parallel opportunity (March 1, 2026): sequential independent reads were executed one after another: ls -la | cat .../code/SKILL.md | cat .../llm/SKILL.md This is exactly the pattern where parallel would cut waiting time.
  2. Approval-friction opportunity (February 25, 2026): one thread recorded 7 permission-related errors before completion. A concise request_user_input choice flow would likely have resolved this earlier.
  3. Batch fan-out opportunity (February 28, 2026): a batch-like task reached 103 tool calls without spawn_agents_on_csv, indicating heavy manual orchestration.

These excerpts are taken from structured evidence in opportunities.csv.

Impact Simulator: What If You Close Just Part Of The Gap?

This simulator uses one explicit assumption: each resolved missed-feature moment saves about two minutes of execution or coordination time. Adjust the slider to see rough upside across the analyzed period.

Closing 30% of missed moments

Assumption is intentionally conservative and transparent. Change it in code if your own benchmark differs.

Estimated time saved over this dataset window
8.5h
~0.8 hours per month equivalent

The Practical Playbook

You do not need a new workflow religion. You need a small default prompt layer that matches how you already work.

  1. Run multiple independent reads in parallel.
  2. For 20+ tool calls, maintain update_plan throughout.
  3. For long-running commands/tests, delegate via sub-agents and report checkpoints.
  4. If blocked by permissions, ask me concise choices.
  5. If sandbox/config gets in the way, use /permissions and /debug-config early.
Explore the underlying files