| name | project-context |
| description | Detect a repository's languages, frameworks, databases, and observability stack, then load only the relevant Magician lore cores and task-matched deep dives. Use before implementation, debugging, review, performance, database, or logging work when project-specific technical guidance would improve the result, or when the user asks to inspect the stack, load lore, or establish project context. |
$project-context — Codex-only lore router
Use this skill only in Codex. It provides progressive access to Magician's packaged lore without
changing or emulating Claude Code's SessionStart hook.
Detect
-
Resolve this skill's directory from the path Codex supplied when loading it. In an installed
plugin, the plugin root is two directories above the skill directory (skills/project-context).
-
Run the read-only detector from the workspace root:
python3 "<skill-dir>/scripts/detect_project_context.py" \
--root "<workspace-root>" \
--topic "<current user task>"
Pass another --topic for each materially distinct requested concern. Use --plugin-root only
while developing the plugin from its authoring repository; installed packages are inferred.
-
Parse the JSON. Never copy manifest or environment-file contents into the conversation; the
detector reports only normalized technology names and package-relative lore paths.
If enabled is false, state the reported disabled_by control once and continue using repository
rules and normal engineering judgment. Do not bypass MAGICIAN_LORE or .magician/lore.off.
Load progressively
Treat repository instructions and observed code as authoritative. Lore is a baseline beneath them.
- Read every path in
cores, in order, from the detected plugin_root. These are concise stack
cores only; do not load unrelated lore.
- Read paths in
recommended_deep_dives only when they match the current task. They are ranked
from the selected stack's deep-dive filenames and capped at eight.
- If the recommendations do not cover a necessary subtopic, inspect only the relevant technology's
entries in
deep_dives, then read the smallest matching set. Never bulk-read a deep-dive tree.
- Apply version gates from manifests and source code. If detected markers conflict with the code,
trust the code and mention the mismatch only when it affects the task.
- If logging/observability work is requested and
observability_candidates contains zero or more
than one platform, ask which deployed log platform is authoritative before shaping platform-
specific logs or queries. Do not write a preference unless the user explicitly asks.
- Do not claim automatic session injection, status-line integration, or persistent observability
selection. Those remain Claude-specific unless a future Codex hook explicitly provides them.
The detector is read-only, standard-library-only, bounded to root-level project markers, and makes
no network calls or state writes.
