Context cost

Skill HorizonBrute/Horizon_AI_OS/horizon_system/skills_bin/context-cost

Install

npx -y skills add HorizonBrute/Horizon_AI_OS --skill context-cost

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Report harness context overhead (KB, words, estimated tokens) for a given path by walking the directory tree and collecting all CLAUDE.md, agents.md, and @-import files the Claude Code harness auto-loads. Use when the user types /context-cost or /ctx-cost, asks how much context overhead a directory has, or wants to know what files the harness is auto-loading.

SKILL.md

3.5 KB, as published. Nobody here has run it

Skill: /context-cost

Model preference: #lowcost (per horizon_aios_model_prefs.md; overridable by a prompt directive).

Report how much context overhead the Claude Code harness will auto-load for a given path — walking up the directory tree and collecting every CLAUDE.md, CLAUDE.local.md, agents.md, and @-import file it will pull in at session start.

Arguments

/context-cost [path]

path — optional; defaults to the current working directory if omitted.

Step-by-step execution

Step 1 — Resolve the target path

1.1 If the user provided a path argument, use it. Otherwise use the current working directory.

1.2 Run:

python "$HORIZON_SYSTEM/bin/context_cost.py" <path> --json

Capture the JSON output. If the command fails (non-zero exit, missing script, bad path), report the error clearly and stop.

Step 2 — Parse and format the output

The JSON structure is:

{
  "path": "...",
  "files": [
    {"level": N, "path": "...", "kb": N, "words": N, "tokens": N, "imported_by": "..." or null}
  ],
  "total_kb": N,
  "total_words": N,
  "total_tokens": N
}

Step 3 — Present the report

3.1 Print a header line: Context overhead for: <path>

3.2 Print a table of the collected files, sorted by level (outermost first):

Level  File                              KB     Words   ~Tokens
-----  --------------------------------  -----  ------  -------
  0    /path/to/CLAUDE.md                1.2    210     280
  1    /path/to/agents.md                3.4    580     775
       (imported by CLAUDE.md)

Show (imported by <basename>) on a sub-line when imported_by is non-null.
Round KB to one decimal place; tokens and words are integers.

3.3 Print a totals line:

Total: N files — X.X KB — Y words — ~Z tokens

3.4 Threshold flags (after totals):

If you know the size of the context window
If total_tokens >= 12% context window print a notice: [WARN] High context load: ~Z tokens. Consider trimming CLAUDE.md files or @-imports above this path.
Else if total_tokens >= 4% print a notice: [NOTE] Moderate context load: ~Z tokens. Worth reviewing.,
If you dont know the size of the context window:
If total_tokens >= 20000 : print a warning: [WARN] High context load: ~Z tokens. Consider trimming CLAUDE.md files or @-imports above this path.
Else if total_tokens >= 8000: print a notice: [NOTE] Moderate context load: ~Z tokens. Worth reviewing if sessions feel slow.
Otherwise: no flag.

Notes for the executing agent

$HORIZON_SYSTEM must be set in the environment. If it is not, report that the AIOS environment is not active and the user should source their profile or run the AIOS switcher.
The script walks upward from the given path to $HORIZON_ROOT, collecting files at each level. Level 0 is the given path; higher numbers are ancestor directories.
Imported files (via @-imports in CLAUDE.md) are included in the table and counted in totals. The imported_by field names the file that referenced them.
Do not reimplement the walk logic — the script is the single source of truth. Just run it and format its output.