AgentSkillsCN

context-packer

构建适用于大语言模型的代码转储，并可选附带相关文档；随后使用 tokencount（o200k-base）工具，根据上下文预算（例如 GPT-5 的 27.2 万 Token 预算）对代码进行 Token 计数。

SKILL.md
--- frontmatter
name: context-packer
description: Build LLM-ready code dumps with optional docs, then count tokens with tokencount (o200k-base) against a context budget (for example 272k for GPT-5).

Context Packer

Use this skill when the user wants to:

  • •copy a project into an LLM-friendly single text file,
  • •include/exclude docs, tests, lockfiles, etc.,
  • •verify token count against a model context window.

What it does

prepare-context.sh:

  1. •Selects relevant project files (tracked files by default)
  2. •Excludes common junk (generated files, prior dumps, lockfiles/env/sensitive files unless requested)
  3. •Builds a fenced text dump (path + ``` + contents) for all selected files
  4. •Writes dump to <project>/prompt/<output>.txt (or /tmp/context-packer/... with --tmp-output)
  5. •Writes a manifest of included files next to the dump
  6. •Counts tokens with tokencount --encoding o200k-base
  7. •Reports whether it fits the provided budget

Command

bash
$HOME/dev/pi-skills/context-packer/prepare-context.sh <project_dir> [options]

Common invocations

bash
# Default budget (272000), code + docs + tests pack
$HOME/dev/pi-skills/context-packer/prepare-context.sh ~/dev/pui

# Code-focused pack (exclude docs and tests)
$HOME/dev/pi-skills/context-packer/prepare-context.sh ~/dev/pui --no-docs --no-tests

# Include env/sensitive files explicitly (off by default for safety)
$HOME/dev/pi-skills/context-packer/prepare-context.sh ~/dev/pui --include-env --include-secrets

# Custom budget and output name
$HOME/dev/pi-skills/context-packer/prepare-context.sh ~/dev/pui \
    --budget 272000 \
    --output pui-gpt5.txt

# Fail with exit code 2 if over budget
$HOME/dev/pi-skills/context-packer/prepare-context.sh ~/dev/pui --fail-over-budget

Options

  • •--output <name> output filename (default: context-dump.txt)
  • •--tmp-output write output to /tmp/context-packer/... instead of <project>/prompt/
  • •--budget <tokens> token budget (default: 272000)
  • •--with-docs include docs/ (default: on)
  • •--with-tests include test files (__tests__, tests/, test/, *.test.*, *.spec.*, etc.) (default: on)
  • •--no-docs exclude docs/
  • •--no-tests exclude test files (__tests__, tests/, test/, *.test.*, *.spec.*, etc.)
  • •--include-lockfiles include lockfiles (pnpm-lock.yaml, Cargo.lock, etc.)
  • •--include-env include env files (.env, .env.*, .envrc) (default: off)
  • •--include-secrets include potentially sensitive files (.npmrc, .netrc, keys/certs, cloud credential files, etc.) (default: off)
  • •--no-clipboard do not refresh clipboard from output file
  • •--fail-over-budget return non-zero if budget exceeded
  • •--install-tools install missing tokencount via cargo

Requirements

  • •tokencount installed (cargo install tokencount)
  • •Optional helper utility: copy_files (not required)
  • •Optional clipboard tools:
    • •macOS: pbcopy
    • •Linux Wayland: wl-copy

If no clipboard tool is available, the script still writes output files; it just skips clipboard copy.