AgentSkillsCN

pr-context-packer

基于 PR 关键信息（PR 描述 + Git Diff + 全部变更文件 + Scribe 提供的相关文件），构建符合预算且经过 Token 检查的上下文包（使用 o200k-base 模型）。

SKILL.md
--- frontmatter
name: pr-context-packer
description: Build PR-focused context packs (PR description + git diff + full changed files + related files via Scribe), then budget-fit and token-check with o200k-base.

PR Context Packer

Use this skill when the user wants high-signal PR context for an LLM:

  • •include GitHub PR title/body by default,
  • •include the actual diff,
  • •include full current code for changed files,
  • •optionally expand with related files discovered via Scribe,
  • •fit related files into the remaining token budget,
  • •verify final fit against a context window budget.

What it does

prepare-pr-context.sh:

  1. •Resolves PR base (--base or auto-detect)
  2. •Pulls GitHub PR title/body by default (via gh, with graceful fallback)
  3. •Captures changed files + full git diff (base...HEAD)
  4. •Includes full current content of changed files (safe filters applied)
  5. •Optionally expands related files using Scribe covering-set (dependents included by default)
  6. •Ranks related candidates and greedily fits them into remaining token budget
  7. •Writes a single final text file plus manifests (changed/related/omitted + related-omitted + selection + scribe-target audit)
  8. •Counts tokens with tokencount --encoding o200k-base and trims low-priority related files if needed

Command

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

Common invocations

bash
# Typical PR pack (tmp output by default)
$HOME/dev/pi-skills/pr-context-packer/prepare-pr-context.sh ~/dev/mobile-1 --base origin/main

# Disable Scribe expansion (diff + changed files only)
$HOME/dev/pi-skills/pr-context-packer/prepare-pr-context.sh ~/dev/mobile-1 --base origin/main --no-scribe

# Keep output in repo prompt/ instead of /tmp
$HOME/dev/pi-skills/pr-context-packer/prepare-pr-context.sh ~/dev/mobile-1 --base origin/main --in-project-output

# Tight budget + fail if over
$HOME/dev/pi-skills/pr-context-packer/prepare-pr-context.sh ~/dev/mobile-1 --base origin/main --budget 180000 --fail-over-budget

Key options

  • •--base <ref> base ref to diff against (default: auto-detect)
  • •--pr <ref> explicit PR number/url/branch for gh pr view lookup
  • •--no-pr-description skip GitHub PR title/body section
  • •--tmp-output write output to /tmp/context-packer/... (default)
  • •--in-project-output write output to <repo>/prompt/
  • •--no-scribe disable Scribe related-file expansion
  • •--no-dependents disable dependent-file expansion (dependents are included by default)
  • •--scribe-max-depth <n> and --scribe-max-files <n> control each covering-set query
  • •--scribe-target-limit <n> cap number of changed files used as Scribe targets (default: 0 = all)
  • •--max-related <n> cap ranked related candidate pool before budget fitting (0 disables related section)
  • •--budget <tokens> token budget (default: 272000)
  • •--fail-over-budget return non-zero when over budget
  • •--include-lockfiles, --include-env, --include-secrets for explicit opt-ins

Output manifests

For <output>.txt, the script writes:

  • •<output>.changed.files.txt
  • •<output>.related.files.txt
  • •<output>.omitted.files.txt (changed files omitted by safety filters)
  • •<output>.related.omitted.files.txt (related candidates omitted with reasons)
  • •<output>.related.selection.tsv (per-related candidate decision + token estimate)
  • •<output>.scribe.targets.tsv (per-target Scribe status + possible max-files cap signals)

Requirements

  • •tokencount (cargo install tokencount)
  • •Optional GitHub CLI (gh) for auto PR title/body inclusion
  • •Optional Scribe for related-file expansion:
    • •npm i -g @sibyllinesoft/scribe, or
    • •cargo install scribe-cli, or
    • •npx @sibyllinesoft/scribe available
  • •Optional clipboard tools:
    • •macOS: pbcopy
    • •Linux Wayland: wl-copy