I'll be direct: if you only use Claude Code, this skill is not for you. But if you've ever bounced between OpenAI Codex and Claude Code — or if you've ever stared at a state_*.sqlite file wondering what the hell happened in a previous session — then read-codex-history by daymade is one of the most thoughtfully engineered skills I've reviewed on the SkillsMP marketplace.
1388 stars. It's not trending (zero gains in the last seven days), which tells me it's a niche tool that found its audience and then plateaued. That's fine. Niche tools that solve real problems don't need to go viral. They need to work.
Here's what this skill actually does. It reads your local OpenAI Codex history files — the ones sitting on disk from your Codex installations — and makes them queryable from within Claude Code. Not by continuing the old task, not by resuming a session, but by reading the evidence. It extracts prompt-ledger inputs, locates exact rollouts by verified session identity, reconstructs chronological user/assistant timelines, and performs bounded keyword searches across live and archived Codex rollouts.
Why does this matter? Because the gap between tools is where context dies. You start a complex task in Codex, the Agent makes progress, you fork, you compact, and then you switch to Claude Code because you want different tooling or model capabilities. Suddenly you're starting from scratch with a vague memory of what happened. You're either retracing steps manually through log files or — worse — you pretend you remember and make wrong assumptions. This skill kills that problem. It turns your Codex history from opaque SQLite databases and JSONL files into structured, queryable evidence.
What stands out to me:
First, the architecture is honest. The SKILL.md explicitly calls out three different history surfaces — history.jsonl for what the user submitted, state_*.sqlite for inventory metadata, and sessions/**/rollout-*.jsonl for the full event stream. It refuses to let you substitute one for another. A prompt-ledger row proves what was submitted, not what the Agent answered. A state DB path is only a candidate until the rollout's session_meta.id matches. That's not pedantry; that's the difference between a useful tool and a confident liar.
Second, the script routing is dead specific. "Show my recent original inputs" gives you a chronological raw-input table — not topic mining, not sentiment analysis, not an interactive app. The skill resists the urge to be clever. When I see a tool that says "the requested output wins over the motivation," I know the author has been burned by over-engineered abstractions before.
Third, the lineage reconstruction is serious. read_codex_session.py --session <ID> --full gives you verified identity, root-to-child fork lineage, exact parent byte boundaries, compaction history, tool calls, errors, end reasons. The skill explicitly warns that if the state DB points to a rollout with the wrong identity, the reader must reject it and try the exact session_meta.id locator. Never continue from the wrong file because its title looked close. That's a principle I wish more tool builders followed.
Fourth, the bounded search is practical. analyze_sessions.py search --codex-only lets you keyword-search across live and archived rollouts with date ranges and session exclusions. The skill even calls out that the exact-ID locator is "seconds cheaper than a corpus scan" — a reminder that good engineering is also about knowing when not to scan everything.
Who should install this: Anyone who uses both Codex and Claude Code and has ever lost context between them. Also anyone doing audit work, behavior analysis, or forensic review of Codex sessions. If you maintain a complex project where Codex ran preliminary work, this skill is a cheat code for onboarding.
Who should skip it: If you don't have Codex installed locally, this skill is a dead weight. If you only use Claude Code, it adds nothing. And if you're looking for a magic "continue my Codex task in Claude" button — that's a different skill (continue-codex-work), and this one explicitly says it won't do that.
Installation is straightforward: Clone or copy the skill into your .claude/skills/ directory. The skill lives at daymade-claude-code/read-codex-history in the GitHub repo. You'll need Codex installed locally for the scripts to find anything to read.
My honest concerns: The complexity is real. The SKILL.md is dense — deliberately so, but dense nonetheless. You need to understand Codex's internal file structures to use this effectively. The session_meta.id matching requirement means you need exact IDs, not fuzzy names. And the fact that it gained zero stars in the last week tells me either the audience is already served or the tool hasn't hit a wider awareness moment. The dependency on continue-codex-work for any execution after reading also means you're buying into a two-skill workflow, not a standalone solution.
The verdict: This is a well-built, opinionated, and honest tool for a specific problem. It won't appeal to everyone, but for the developers who need it, it's genuinely valuable. The author clearly understands the data structures, the failure modes, and the temptation to be clever at the wrong time. If you're in the Codex-to-Claude Code pipeline, install it and see what your old sessions actually contain.
Links: - SkillsMP page: https://skillsmp.com/creators/daymade/claude-code-skills/daymade-claude-code-read-codex-history - GitHub: https://github.com/daymade/claude-code-skills/tree/main/daymade-claude-code/read-codex-history