Maintain a review-first Markdown knowledge base for Codex practices with source provenance, engineering capture, citation-aware queries, explicit archive and promotion, and deterministic linting.
SkillScore64Maintained
Automatic
◆Knowledge Management pack
Last verified 9/1/2026
What it changes
Passed 11 of 17 test cases with the skill, and 8 of 17 without it.
Without skill45%
With skill64%
Checks it passed that the model alone missed
Rejects symlink files and symlinked parent paths
Treats target symlinks as path validation errors
Classifies local benchmark evidence as experiment
and 4 more in the full report
6 fixed8 passed either way6 still failing2 broke
Pass-rate lift
+18 pp
Tokens
+154%
Agent turns
0%
Response time
+9%
Model gemini-3.6-flashMeasured 9/1/2026v114 of 22 cases passed with skill · 17 comparable
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 6 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 6
Prompt
We are ingesting benchmarking data from a temporary local load test run into `knowledge/sources.json`. Which standard category should be assigned in the source record?
What was checked, without and with the skill
Checked
Without skill
With skill
Classifies local benchmark evidence as experiment
Missed
Passed
case-11 · gemini-3.6-flash · 9/1/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md235 lines4 filesincludes scripts
Maintain Codex Wiki
Treat the wiki as compiled maintainer knowledge, not as an automatic source of
truth. Keep official documentation authoritative and require human review
before shared knowledge or curriculum changes land.
Choose one operation
Query: read knowledge/index.md, search candidate pages, and answer with
links. Do not write unless the user explicitly asks.
Capture: preserve a durable lesson from repository evidence such as a
merged change, incident, review finding, or measured run.
Ingest: register one source, update affected wiki pages, run lint, and
prepare a reviewable diff.
Archive: save a requested query result as an experimental, cited page.
Lint: run the bundled deterministic checker, then review semantic drift
that a script cannot prove.
Promote: move a verified conclusion into the appropriate module, skill,
or repository rule through a separate, reviewable change.
Confinement invariant
Apply this before any operation reads, searches, or changes wiki state. For
every wiki page, index, registry, log, cache target, and working-tree repository
file inspected while capturing new evidence:
require a normalized project-relative path;
reject absolute paths and .. components;
resolve symlinks;
for an existing read or update target, reject a symlink at the target or any
parent below the project root, require a regular file, and verify the
resolved target remains inside the project root; and
for a new page, require a nonexistent target under an existing,
project-contained directory, reject symlinked parents and name collisions,
then repeat the existing-file check immediately after creation.
Do not begin Query, Capture, Ingest, Archive, Lint, or Promote until every file
the operation will touch passes the applicable check. Treat an unsafe path as a
reported validation error, never as content to inspect.