Feature request: User-facing workflow retrospective for past Codex sessions, similar to Claude Code /insights

What variant of Codex are you using?

CLI

What feature would you like to see?

Summary

I would like Codex to provide an optional, user-facing workflow retrospective over past sessions—for example through a command such as:

/insights

or:

/retrospective

The feature would analyze selected Codex session history and produce an evidence-linked report about recurring workflows, user corrections, successful patterns, common sources of friction, and opportunities to improve AGENTS.md, skills, configuration, or the way the user delegates work to Codex.

Claude Code’s /insights command is a useful reference for this interaction pattern, but the proposed feature should be designed around Codex’s own session, memory, skill, and project-instruction systems.

Motivation

I recently started using Codex CLI after previously working with Claude Code. One feature I found both engaging and practically useful was Claude Code’s /insights.

Its value is not merely usage statistics. It helps users reflect on questions such as:

• What kinds of tasks do I repeatedly delegate to the agent?
• Where do I frequently need to correct or redirect it?
• Which prompting and collaboration patterns work well?
• Which failures, unnecessary expansions, or repeated recovery steps keep occurring?
• Which repeated instructions should become project guidance or reusable skills?

Codex already persists sessions and has memory-related infrastructure intended to extract reusable knowledge, user preferences, successful procedures, and failure patterns from prior work.

However, this learning is primarily consumed internally by future agents. Users do not currently have a clear way to inspect what Codex has inferred, verify whether those inferences are accurate, or turn them into explicit workflow improvements.

This creates an observability gap:

1. Users cannot easily audit or correct inferred preferences.
2. Users may repeat inefficient prompting or recovery patterns without noticing them.
3. Valuable recurring instructions remain scattered across sessions instead of being promoted into AGENTS.md, skills, or configuration.
4. Users cannot clearly distinguish evidence-based personalization from incorrect or outdated assumptions.

Proposed behavior

A command such as /insights or /retrospective could generate a Markdown report covering a user-selected scope.

For example:

/insights --project --since 30d

Possible scopes could include:

• the current project;
• selected sessions;
• the last 7 or 30 days;
• all locally available Codex sessions;
• optionally, Codex sessions synchronized across supported surfaces.

The report could include:

Recurring workflows

Common task families and repeated sequences of actions, such as:

• planning followed by implementation;
• debugging and repeated test cycles;
• repository review;
• issue-to-PR workflows;
• documentation generation;
• recurring environment setup or verification procedures.

Steering and correction patterns

Instructions that the user repeatedly gives after Codex starts working, such as:

• keeping the task narrowly scoped;
• not introducing unnecessary frameworks;
• requiring verification before implementation;
• preferring a specific tool or command;
• asking Codex to stop after analysis rather than modifying files.

Successful collaboration patterns

Patterns associated with efficient or successful sessions, such as:

• starting with a read-only repository audit;
• using a planning pass before implementation;
• assigning bounded tasks to subagents;
• running a specific verification command before completion.

Friction and failure patterns

Repeated causes of wasted effort, including:

• scope drift;
• unnecessary abstraction;
• repeated tool failures;
• incorrect environment assumptions;
• modifications made before sufficient investigation;
• recurring rollback or repair cycles.

Actionable recommendations

Recommendations should preferably map to Codex-native configuration surfaces, for example:

• add a stable rule to AGENTS.md;
• create or update a skill;
• change a model or reasoning-effort workflow;
• use a project-specific instruction instead of repeating a prompt;
• remove obsolete or contradictory memory;
• introduce a verification or stop rule.

Evidence and auditability

Important insights should be traceable to their supporting sessions.

For example:

Observation:
You repeatedly asked Codex not to expand focused research tasks into
general-purpose statistical frameworks.

Evidence:
Observed in 4 sessions in this project during the last 30 days.

Suggested action:
Add a scope-control rule to AGENTS.md.

Confidence:
High

Each observation could optionally include:

• project or working directory;
• date range;
• number of supporting sessions;
• short evidence summaries;
• links or identifiers that allow the user to reopen relevant sessions;
• a confidence level;
• conflicting evidence, where applicable.

This would help prevent generic or fabricated personality analysis and make the result useful as an engineering artifact.

Privacy and user control

Because this feature analyzes session history and user behavior, it should be explicitly user-controlled.

Suggested safeguards:

• run only when manually requested;
• show the intended data scope before analysis;
• support project-only and time-limited scopes;
• avoid analyzing excluded or deleted sessions;
• clearly distinguish local and remotely synchronized data;
• never modify AGENTS.md, skills, memories, or configuration automatically;
• require explicit confirmation before persisting any recommendation;
• provide a way to inspect, correct, dismiss, or delete inferred observations.

A read-only report should be the default.

Minimal viable version

The first version would not require a graphical dashboard or elaborate analytics interface.

A useful CLI-only MVP could:

1. analyze sessions from the current project over a selected period;
2. identify several recurring workflow and friction patterns;
3. produce three to five actionable recommendations;
4. attach compact evidence references;
5. render the result as Markdown in the terminal or save it to a user-selected file.

For example:

/insights --project --since 30d

This would already provide substantial value while keeping the initial scope bounded.

Why this fits Codex

This request is not primarily for entertainment, token accounting, or a clone of another product’s personality features.

It would expose a user-visible and auditable layer over capabilities Codex already appears to be developing around:

• persisted sessions;
• memories;
• reusable skills;
• project instructions;
• workflow learning;
• failure avoidance.

The main goal is to improve long-term user control and agent effectiveness by allowing users to see and act on what Codex has learned from their own work.

Additional information

Claude Code’s /insights is a useful reference because it demonstrates that cross-session reflection can be both enjoyable and operationally valuable.

Related Codex proposals appear to address adjacent infrastructure, but not this user-facing retrospective:

• #19758 — topic-based memory and a /memory command;
• #24810 — structured context preservation across long-running sessions;
• #18469 — versioned export of persisted session actions.

This request is distinct from ordinary usage analytics. It is about qualitative workflow patterns, repeated corrections, successful procedures, failure modes, and actionable improvements to future human-agent collaboration.