Codex 0.142 Alpha Stops Treating Skill Descriptions Like Prompt Stuffing
Codex 0.142.0-alpha.7 is the kind of release that looks too small to matter until you realize it is fixing exactly where modern coding agents tend to rot: the boundary between rich system metadata and whatever compressed catalog the model sees in-context.
The release has only two commits in the compare from rust-v0.142.0-alpha.6 to rust-v0.142.0-alpha.7, touching nine files. That is not a launch. It is a scalpel. But the change is aimed at a real architectural question for anyone building with agent skills, repo instructions, external commands, or portable agent runtimes: should a skill description be treated as source metadata, or as prompt text that needs to be chopped down before it enters the system?
OpenAI’s answer here is the right one. Codex now preserves full skill descriptions and external-agent command descriptions in metadata, then applies a 1,024 Unicode-character cap only when rendering model-visible skill lists. PR #29006 states the problem directly: “Skill descriptions are used in model-visible lists,” and “a single overlong description should not consume a disproportionate share of either list.” Translation: don’t let one verbose skill eat the prompt budget, but also don’t throw away the original record because a downstream renderer has constraints.
The cap belongs at the rendering boundary
This distinction sounds academic until you operate enough agents. A SKILL.md description has multiple audiences. Humans read it to understand what a tool does. Runtime components read it to decide what exists and how to route. Migration tools read it to preserve workflows across systems. The model sees a shortened list of available options and needs enough signal to choose correctly without burning the whole context window on documentation.
Those are different surfaces. Treating them as one blob is how teams end up with either bloated prompts or lossy registries. If the loader enforces the model-visible cap, you have permanently damaged the artifact. If migration enforces the cap, you silently break portability. If the renderer does not enforce any cap, one lovingly documented skill can crowd out the rest of the tool catalog. Codex alpha 7 lands on the sane API-gateway pattern: store the full thing, expose a constrained view.
The concrete implementation spans the loader, renderers, list-tool serialization, orchestrator provider code, tests, and external-agent migration. Codex preserves full description and metadata.short-description values when loading skills. It preserves full external-agent command descriptions during source-command-* migration. It preserves normalized orchestrator descriptions in the underlying skills catalog. Only when rendering the default available-skills context in codex-core-skills, codex-skills-extension, and serialized skills.list responses does it trim each description to 1,024 Unicode characters.
That number is less interesting than the seam. The seam is the product.
Agent skills are becoming infrastructure, not prompt garnish
The release was published at 2026-06-20T00:40:41Z with 149 assets, and the Codex repository snapshot in the research brief showed 92,263 stars, 13,634 forks, and 7,413 open issues. This is not a niche plugin deciding how to format a tooltip. Codex is one of the gravity wells in agentic coding, and it is making decisions about how skill metadata survives loading, rendering, listing, and migration.
That matters because AGENTS.md, skills, slash commands, MCP tools, and repo-local instructions are becoming the portability layer for coding agents. Teams are already trying to reuse the same operating instructions across Codex, Claude Code, Cline, OpenClaw, local runners, and internal orchestration. The first generation of this work has been vibes: write a markdown file, hope the model reads it, hope the tool understands it, hope the migration path does not eat half the useful context. The next generation needs typed-ish boundaries, lossless storage, budgeted rendering, and auditability.
Codex alpha 7 is a small step in that direction. It says a skill is not just prompt stuff. It is an artifact with source fidelity. The prompt is merely one consumer.
The migration angle is especially important. Skipping or truncating external-agent command descriptions because a model-visible list has a budget is the wrong failure mode. Migration tools should preserve user workflow unless the source is invalid, unsafe, or intentionally unsupported. If the target runtime needs to compress a description for the model, compress the view. Do not corrupt the source. Developers have been burned by this pattern in every config system ever built: a UI simplifies a field, then writes the simplified version back over the canonical config. Nobody enjoys debugging that on a Friday.
What practitioners should do now
If you maintain agent skills, the practical takeaway is not “write 1,024-character descriptions.” It is “front-load the useful part.” The full description may be preserved in metadata, but model-visible surfaces will increasingly summarize, trim, rank, or chunk it. The first sentence should identify the capability, when to use it, and any major safety constraint. Do not bury destructive-action warnings, required credentials, or non-obvious prerequisites after three paragraphs of prose.
If you are building an internal skill registry, separate stored metadata from prompt-rendered summaries now. Keep the long description, owner, version, examples, safety notes, and operational constraints in the registry. Generate one or more compact model-facing views from that source. Make truncation explicit and testable. If your system cannot tell whether the model saw the safety note because it was character 1,400, that is not a documentation problem. That is a runtime-design problem.
If you are adopting Codex skills across a team, test the list boundary deliberately. Create a skill with a long description. Confirm it loads fully. Confirm skills.list and model-visible catalogs render the truncated view. Confirm migration does not skip the command or drop the description. Then decide whether your authoring guidelines need a “short description must stand alone” rule. They probably do.
The broader industry point is that context engineering is leaving the era of prompt superstition. Everyone wants more agent capability, but capability without boundary discipline becomes context sludge. Codex preserving metadata losslessly while budgeting model-visible renderings is not glamorous, but it is exactly the kind of runtime hygiene that lets skills scale beyond a handful of hand-written markdown files.
Alpha releases usually deserve caution. This one deserves a note in your agent-platform backlog: preserve source metadata, render compact model views, and never confuse the two. The difference is small in code and large in systems design.
Sources: GitHub release — openai/codex rust-v0.142.0-alpha.7, GitHub compare rust-v0.142.0-alpha.6...rust-v0.142.0-alpha.7, PR #29006 — Preserve skill descriptions outside model context