Codex Context Budget Audit

Codex Context Budget Audit

Use this skill when the task is about measuring or comparing Codex startup/context overhead from local configuration sources.

Workflow

1. Identify the execution location.
   • Use the user's requested path as --workdir.
   • Default Codex home is $env:USERPROFILE\.codex unless the user provides another path.
2. Run the bundled CLI first:

python "$env:USERPROFILE\.codex\skills\codex-context-budget-audit\scripts\context_budget_audit.py" --workdir <windows-project-root>

3. If comparing runs, save a baseline JSON and compare a later run:

python "$env:USERPROFILE\.codex\skills\codex-context-budget-audit\scripts\context_budget_audit.py" --workdir <windows-project-root> --json-out <temp>\codex-context-before.json
python "$env:USERPROFILE\.codex\skills\codex-context-budget-audit\scripts\context_budget_audit.py" --workdir <windows-project-root> --baseline <temp>\codex-context-before.json

4. Report the result by phase:
   • Config state: memory, plugin, MCP enabled counts.
   • Memory summary estimate.
   • AGENTS/instructions estimate.
   • Local skill manifest estimate.
   • Plugin skill manifest estimate.
   • MCP enabled surface.
   • Optional triggered skill body upper bound.
5. State limits clearly:
   • Platform system/developer/tool-schema tokens are not fully observable from local files.
   • Existing sessions do not shrink after config changes; a new Codex session is required.
   • Token counts are estimates based on character ratios, not tokenizer telemetry.

ASCII Prompt Ratio Sketch

Use the CLI dashboard as the main visualization. The visual style should stay pure ASCII and terminal-safe while borrowing these interface cues:

• Claude Code style: compact statusline-like metrics for subtotal, memory, MCP, and platform-base state.
• Gemini CLI style: terminal-first source rows that expose tools, MCP, and local workspace surfaces clearly.
• Codex audit requirement: keep unobservable platform prompt/tool-schema tokens separate from measured local estimates.
• Color is allowed when the terminal supports ANSI. Use --color auto by default, --color always for screenshots/log demos, and --color never for plain logs.
• For Codex-visible colored output, prefer --render codex; it emits Markdown/HTML inline colors instead of ANSI escape codes.
• Keep long vertical separators in source rows so the chart stays readable even without color.

If summarizing manually, use this shape:

+--------------------------------------------------------------------------------------------------+
| CODEX CONTEXT BUDGET                                            local estimate / startup surface |
| STYLE RAIL  | dense statusline | clear tool surface | measured-local estimate                    |
+==================================================================================================+
| STATUS      | LOCAL SUBTOTAL  10.9k-14.6k tok                           MEMORY OFF | MCP 0 enabled |
| RUNTIME     | PLATFORM BASE   unknown                      system + developer + native tool schema |
+--------------------------------------------------------------------------------------------------+
| ID  | SOURCE                           | BAR                      | TOKENS       |   % | STATE     |
+--------------------------------------------------------------------------------------------------+
| SYS | platform system/developer/tools   | [????????????????????????] |      unknown | n/a | runtime-owned |
| MEM | memory summary                   | [------------------------] |      0-0 tok |   0% | config-gated |
| DOC | AGENTS/workspace instructions    | [==========--------------] | 2.7k-3.6k tok|  42% | local files  |
| SKL | local skill manifests            | [#############-----------] | 3.5k-4.7k tok|  55% | local roots  |
| PLG | plugin skill manifests           | [+-----------------------] |  180-241 tok |   3% | enabled plugins |
| MCP | enabled MCP tool surface         | [------------------------] |      0-0 tok |   0% | tool schema hidden |
+--------------------------------------------------------------------------------------------------+

Never present the sketch as exact telemetry. Label it as an estimate or measured-local subset. Keep SYS with question marks unless runtime telemetry exists.

CLI

The script is intentionally self-contained and uses only Python standard library modules.

Useful options:

• --workdir PATH: target execution location.
• --codex-home PATH: Codex home directory.
• --config PATH: explicit config file.
• --json-out PATH: save report for future comparison.
• --markdown-out PATH: save the Codex-friendly colored Markdown/HTML dashboard.
• --baseline PATH: compare against a saved report.
• --color auto|always|never: control ANSI colors.
• --render terminal|codex|both: choose the ASCII/ANSI terminal dashboard, the Codex Markdown/HTML dashboard, or both.
• --include-skill-bodies: estimate full SKILL.md body size as an upper-bound, not startup cost.

Source: pingqLIN/UniText — distributed by TomeVault.