Codex How To v0.4.0 — Safer Living Wiki

Codex How To v0.4.0

Codex How To now routes prior engineering knowledge through a review-first
Living Wiki and exposes its flagship workflows through the open agent skills
ecosystem.

Highlights

• Index-first knowledge routing: AGENTS.md directs work involving prior
  decisions, experiments, or established guidance through
  knowledge/index.md before broad repository search.
• Engineering capture: maintain-codex-wiki can preserve an explicitly
  requested lesson from a merged change, incident, review finding, or measured
  run. No material change remains a valid result.
• Explicit archive: useful query synthesis can be saved as experimental,
  cited knowledge without silently becoming verified guidance.
• Stronger provenance: source records can pin revisions, identify
  superseded evidence, and declare the wiki pages they affect.
• Deterministic integrity: wiki lint now catches invalid source
  relationships, affected pages that do not cite their source, duplicate page
  titles, malformed metadata, broken links, and missing index entries.
• Open skill installation: engineering-loop and
  maintain-codex-wiki are indexed at
  skills.sh.

Install a flagship workflow

Install the complete engineering loop into Codex:

npx skills add Phelan164/codex-howto --skill engineering-loop -g -a codex -y

Install review-first knowledge maintenance:

npx skills add Phelan164/codex-howto --skill maintain-codex-wiki -g -a codex -y

List all nine repository skills without installing:

npx skills add Phelan164/codex-howto --list

Inspect every skill before use. Skills run with the permissions available to
the agent that invokes them.

Knowledge safety model

The Living Wiki remains plain Markdown under Git review:

• queries are read-only by default;
• capture, ingest, archive, and promotion require explicit intent;
• chat prose and model output are not evidence by themselves;
• official OpenAI documentation remains authoritative for current Codex
  behavior;
• generated factual changes require semantic review; and
• scheduled maintenance may report drift or prepare a draft but cannot merge
  factual changes automatically.

The release does not add a database, vector index, MCP server, copied external
corpus, automatic transcript ingestion, or autonomous merge path.

Verification

The release gate covers:

• repository structure and every relative Markdown link;
• 26 measurement and Living Wiki utility tests;
• both dependency-free engineering playground tests;
• six seeded wiki pages with zero lint warnings;
• source revisions, supersession relationships, affected-page citations, and
  duplicate-title detection;
• official skill validation for maintain-codex-wiki;
• isolated npx skills listing and Codex installation smoke tests; and
• GitHub Actions validation on the release pull request.

Feedback wanted

Useful contributions include:

• an engineering lesson that should produce no durable wiki change;
• a source revision that should mark a dependent page stale;
• a query that reads too many files or misses relevant evidence;
• an anonymized no-skill/full-skill/lean-skill measurement; or
• a maintained stack, team, language, or translation fork.

Codex How To is an independent community project. Official OpenAI
documentation remains authoritative for current product behavior.

Codex How To v0.3.0 — Living Wiki

Codex How To v0.3.0

Codex How To now includes a review-first Living Wiki and a model-adaptive
engineering skill system.

Highlights

• Living Codex Wiki: compile repeated research into small, interlinked
  Markdown evidence pages instead of reconstructing the same context for every
  question.
• Review-first maintenance: register source authority, preserve conflicts
  and uncertainty, lint deterministic structure, and promote stable knowledge
  through reviewed changes.
• Four wiki operations: query, ingest, lint, and explicitly promote
  verified conclusions into modules, skills, repository rules, or checks.
• Model-adaptive skills: keep nine installable workflows focused on
  decisions, evidence contracts, and failure boundaries that a stronger model
  should not guess.
• Three-way measurement: compare no-skill, v0.2 full-skill, and current
  lean-skill runs while holding the model, reasoning effort, repository
  revision, tools, prompt, and done conditions constant.
• Clearer presentation: a readable engineering-loop diagram, refreshed
  social preview, and tracked community-distribution experiments.

Try the Living Wiki

Ask a read-only question before adding infrastructure:

$maintain-codex-wiki What does this repository know about orchestration
efficiency? Cite wiki pages, separate evidence from recommendation, and do not
modify files.

Run its deterministic health check:

python3 skills/maintain-codex-wiki/scripts/wiki_lint.py .

The initial implementation deliberately uses Markdown, repository search, and
Git review. It does not add a database, vector index, MCP server, or autonomous
merge path without evidence that those layers are needed.

Verification

The release gate covers:

• repository structure and local-link validation;
• all 14 measurement and wiki utility tests;
• both dependency-free engineering playground tests;
• six seeded wiki pages and their registered sources;
• wiki metadata, index coverage, and local links; and
• Python compilation for the bundled scripts.

Feedback wanted

Useful contributions include:

• a source that contradicts an existing wiki conclusion;
• a query the index routes poorly;
• an anonymized no-skill/full-skill/lean-skill measurement;
• a workflow step that adds cost without improving quality; or
• a maintained stack, team, or translation fork.

Codex How To is an independent community project. Official OpenAI
documentation remains authoritative for current product behavior.

Codex How To v0.2.0

Codex How To v0.2.0

Codex How To now provides a runnable, measurable engineering workflow instead
of only a learning catalog.

Highlights

• Engineering loop: take a task through scope, reproduction, the smallest
  coherent change, focused checks, required checks, findings-first review, and
  an evidence-based handoff.
• Nine focused skills: route work across frontend, backend, DevOps,
  testing, code review, security review, and bounded orchestration.
• Dependency-free playground: practice the complete loop against seeded
  defects without connecting Codex to a production repository.
• Measurement protocol: compare paired ad hoc and workflow-assisted runs
  using quality gates, retries, human corrections, elapsed time, and available
  token or cost data.
• Community launch package: present the project responsibly, collect
  structured workflow feedback, and support useful stack, team, or translation
  forks.

Try it

Clone the repository, follow the
five-minute engineering demo, and
record the result with the
paired-run protocol.

Feedback wanted

Useful feedback includes:

• a step that was unclear or could not be reproduced;
• a verified defect or safety gap in a skill;
• a task where the workflow added unnecessary overhead;
• anonymized paired-run results;
• a maintained stack-specific or translated fork.

Open a structured workflow feedback issue after removing private code,
credentials, and production data.

Codex How To is an independent community project. Official Codex documentation
remains the source of truth for product behavior.