AGENT SKILL

context-check

Audits always-on context load for bloat, redundancy, and quality. Use when context feels heavy or periodically to keep the always-on budget lean.

License 许可
MIT
最近更新
2026-11-15
Stars
★ 9
Forks
1

01 前言 Front Matter

name
context-check
tags
[ "context", "tokens", "efficiency", "audit" ]
category
agent-ops
description
Audits always-on context load for bloat, redundancy, and quality. Use when context feels heavy or periodically to keep the always-on budget lean.

02 资源文件 2

03 SKILL.md 正文

Context Check

Audit what Claude loads every session. Writes context-check-report.md at the project root after explicit confirmation.

Steps

  1. Read context-findings.md if it exists; skip logged issues.

  2. Global artifacts — check ~/.claude/CLAUDE.md. If it is absent, say so and continue. If present, read it. For each @ import, count lines and produce a size table:

    File Lines Issues
  3. Check .claude/skills/ in the current working directory.

    If the directory doesn't exist, note "No .claude/skills/ directory found" and continue.

    If it exists:

    • Run ./tools/skills.sh status if that file exists, otherwise skills.sh status if on PATH. This gives the full registered-skill list.
    • Run find .claude/skills -maxdepth 1 -mindepth 1 -type l 2>/dev/null to identify whole-dir symlinks (canon-managed, classic model).
    • Run find .claude/skills -maxdepth 1 -mindepth 1 \! -type l -type d 2>/dev/null to identify real subdirectories. For each, run readlink .claude/skills/<name>/SKILL.md 2>/dev/null — if the target contains /canon/skills/ or /.canon/skills/, it is canon-managed (per-skill model); otherwise it is project-local.

    Report two sub-lists:

    • Canon-managed (whole-dir symlinks, or real dirs whose SKILL.md symlinks to a canon path): name and symlink target only. Do not audit content — these are installed from an external source and not part of the project's own context footprint.
    • Project-local (real dirs with no canon SKILL.md symlink): name and file size. Include in the content audit at step 8.
  4. Check ~/.claude/settings.json. If it is absent, say so and continue. If present, read it and list hooks, matchers, and scripts.

  5. Find the current repo's Claude project memory first, if present, then summarize other ~/.claude/projects/*/memory/ directories. If no memory paths exist, say so and continue. Report file count and total size; use line counts or bytes as a proxy, not exact tokens.

  6. Project artifacts — check the following in the current working directory. For each: report lines if present, or "not present" if absent.

    • .claude/settings.json
    • .claude/settings.local.json
    • CLAUDE.md (project root)

    Produce a project-level size table (omit rows for absent files):

    File Lines Issues

    If none exist, state "No project-level .claude/ artifacts found."

  7. Flag size issues where line count > 30 and less than half is usually relevant, or a section is one-time/rarely needed. Include the evidence: file, line count, and a short reason the content is usually irrelevant or rarely needed.

  8. Read each imported file plus repo AGENTS.md, and each present project artifact from Step 6. Skip canon-managed skills identified in Step 3 (symlinked skill dirs or @-imports that resolve to a canon install path such as ~/.canon/, ~/.claude/skills/, or any path containing /canon/skills/). Flag only high-confidence issues in project-owned files:

    • Cross-file redundancy — the same rule or constraint appears in two or more files, verbatim or near-verbatim. Quote both occurrences.
    • Obvious statements — rules a capable model already follows. Quote the statement and explain why it adds little control.
    • Vague non-actionable rules — instructions with no specific compliance path. Quote the rule and explain what is not enforceable.
    • Dead references — paths, tools, or commands that no longer exist. Verify before flagging and name the failed check.
  9. Report in two labeled sections:

    ### Global — size table from Steps 2–5, content findings for global files.

    ### Project — size table from Step 6, content findings for project-level files.

    Note below each size table: "Line counts are a proxy for context weight, not exact token counts."

    In each size table, set the Issues column to Y if any finding was flagged for that file, — if none. For each file assessed in Steps 7–8, explicitly state either the issue found or "no relevance concern" — do not silently skip files that passed. If one section has no findings, say so and continue to the next section. If no findings exist anywhere, stop before the write prompt.

  10. Ask: Write context-check-report.md report? (y to confirm). Do not write without y. On confirmation:

  • Compute finding_count = the number of rows, across both the Global and Project sections (from Step 9), whose Status will be issues found in the report table below, and file_count = the total number of rows across both sections.
  • Read .reports/context-check-history.json at the project root (create the .reports/ directory if absent). Treat a missing or unparseable file as {}. Look up the entry keyed by this project's absolute root path (an array of {timestamp, finding_count, file_count} objects, oldest first).
  • If that array is non-empty, take its last entry's finding_count as <prev> and its timestamp as <date>.
  • Append {timestamp: <ISO 8601 now>, finding_count, file_count} to the array; if the array now has more than 20 entries, drop from the front until it has exactly 20. Write the file back (pretty-printed JSON, one key per absolute project path — do not touch other projects' entries).
  • Write context-check-report.md at the project root as a markdown table:
| File | Status | Details |
|------|--------|---------|

One row per audited file. Status must be exactly one of: clean, issues found, not present, skipped. Use skipped for canon-managed files that are not audited. Details is a one-line summary of the finding or reason, or — if none. Overwrite if the file already exists — this is a point-in-time snapshot, not a log. If a prior history entry existed (the <prev>/<date> computed above), insert one line at the very top of the file, before the first table: **Trend:** findings <prev> → <finding_count> (last run <date>). Omit this line entirely on a project's first-ever run (no prior entry).

04 同仓库更多 Skills