kotobuki09 Instructree
Use Instructree to establish which instruction files exist, which may apply to a target, and whether their metadata or links are malformed.
Run the audit
Work from the repository root. Prefer an already installed instructree command or the checked-out package's local binary. If neither is available, ask before downloading executable packages, then use the pinned release:
npx github:kotobuki09/instructree#6e6694c0cac60c06bb2d33162771eb9375ed13ad scan .
Do not add --yes unless the user has authorized non-interactive package downloads.
Choose the narrowest command that answers the request:
instructree scan . --jsoninventories supported files and emits stable diagnostics.instructree explain <file> --root .shows instructions that may apply to one target.instructree explain <file> --root . --client codexselects the Codex repository project-instruction chain.- Add repeatable
--fallback <filename>values and--max-bytes <n>when the Codex project configuration uses non-default fallbacks or byte limits. instructree explain <file> --root . --effectiveincludes recursive Copilot CLI imports.instructree imports . --jsonaudits the recursive@pathgraph.instructree skills . --client codexconcisely audits the shared~/.agents/skillsroot, Codex's deprecated default~/.codex/skillscompatibility root, repository-local candidates, and supported user~/.codex/config.tomlskill settings without installing, writing, or uploading anything. It reports disabled candidates, unmatched rules, unsupported relevant syntax, configured catalog pressure, and up to five largest enabled contributors under the fallback-character estimate. Add--allfor every human-readable candidate or--jsonfor the complete structured object. The--client codexflag may be omitted becauseskillsis Codex-specific; use optional--home <home>only to inspect a different user scope.instructree starter .checks a focused six-skill Codex engineering stack against that local inventory, classifies companions as ready, disabled, invalid, or missing, and prints install commands only for missing skills. It is read-only and does not call the directory or install anything. Review each linkedSKILL.mdbefore installation; add--jsonfor deterministic output.instructree doctor .produces one share-safe Codex pre-session report covering supported user project settings, user instructions, the current project chain, and summarized skill state. Add--jsonfor deterministic structured output.instructree scan . --sarifemits SARIF 2.1.0 for code-scanning integrations.- Add
--strictonly when warnings should fail the check.
Interpret the result
Report file paths, line numbers, diagnostic codes, and the command's exit status. Separate schema or path errors from warnings. Describe always/never conflicts as possible conflicts requiring human review, not proof of agent behavior.
Instructree is static analysis. It does not upload repository content or call a model. Treat plain explain as a cross-client map of what may apply; use --client codex only for the documented repository project-instruction chain.
Use doctor when the user asks why Codex instructions or skills are missing, which project root or instruction files will be selected, or for a compact report suitable for an issue. It reads only a supported user-config subset and fails closed on unsupported relevant project settings. Treat its output as a pre-session preview, not live or resumed-session state; repeat its limitations when they matter.
The skills inventory follows Codex's documented shared ~/.agents/skills root, its deprecated default ~/.codex/skills compatibility root, and repository .agents/skills candidate scopes. It recursively scans depth-bounded roots, follows symlinked folders, deduplicates canonical targets, skips hidden descendants, and reports logical paths without exposing the home directory. It flags possible duplicate names with source lines, malformed metadata, scan failures, and an approximate initial-list character estimate including name, description, and logical path. The configured estimate ranks up to five largest enabled contributors and applies current Codex's 1,024-character per-description cap. It applies the supported skill subset of user ~/.codex/config.toml with current Codex name/path and later-rule precedence, failing closed on unsupported relevant syntax. The 8,000-character comparison is only the unknown-context-window fallback; live Codex can use a model-context token budget and path aliases, so the ranking is a review aid rather than a prompt dump. The audit excludes custom CODEX_HOME roots, admin/system and plugin skills, session overrides, project-level skill rules, product restrictions, and configured project-root markers, so it does not claim an exact loaded list. Read the official Codex skills documentation, the skill-overload guide, Instructree's pinned user-root research note, and the skill-config research note when interpreting the estimate.
Do not edit instruction files unless the user asked for changes. After an authorized fix, rerun the same command and report the before-and-after diagnostics.