WEIFENG2333/phistory

Phistory automatically archives versioned system prompt snapshots from agent CLIs like Claude Code, Codex, OpenClaw, and Hermes.

Phistory – Prompt History Tracker for Coding‑Agent CLIs

What it does

  • Installs a list of popular AI‑coding‑assistant command‑line tools (Claude Code, Codex, DeepSeek Harness, etc.).
  • Runs each tool with a set of predefined “snapshots” and captures the exact HTTP request that carries the system prompt, without sending it to the model provider (using the helper claude‑tap).
  • Stores the captured prompt, a trace log and metadata in a version‑ed directory structure.
  • A static web site (phistory.cc) visualises the prompts side‑by‑side, highlighting what changed between releases – new tools, permission checks, model defaults, user‑confirmation rules, etc.
  • Supports Chinese translations of the prompt diffs and provides scripts to back‑fill historic releases, translate captured prose, and rebuild the site.

Why it matters

  • System prompts are the hidden “brain” of coding agents; they dictate how the model behaves, which tools it can call, and what safety policies apply.
  • By archiving every prompt version, Phistory lets researchers, auditors, and developers see exactly how agents evolve, cite stable snapshots, and spot regressions or policy shifts.
  • The hourly GitHub‑Actions workflow keeps the archive up‑to‑date automatically.

How to use it

  1. Explore – Visit the hosted viewer at https://phistory.cc to compare prompts across agents and versions.
  2. Run locally – Install the locked Python environment with uv sync --all-groups.
  3. Capture new releases: uv run phistory capture --latest --agents <list>.
  4. Back‑fill older versions: uv run phistory backfill <agent> --from <ver> --to latest.
  5. Translate captured prompts (requires external credentials): uv run phistory translate --all-captured.
  6. Re‑generate the index and static site with uv run phistory render-index and uv run phistory build-site.

Supported agents (as of Oct 2026) include Claude Code, Codex CLI, DeepSeek Harness, Antigravity CLI, Grok Build, MiniMax Code, Kimi Code, MiMo Code, OpenClaw, Hermes Agent, Kimi CLI, opencode, Pi, and Oh My Pi.

Current coverage – The README lists the latest captured version for each agent, the number of releases tracked, and the total snapshots stored (e.g., Claude Code 2.1.296 with 758 snapshots).

Tech stack – Python project managed with uv, uses a custom claude‑tap tool to intercept prompts, stores data as markdown, JSONL, and JSON, and builds a static site for browsing.

All details are taken directly from the repository’s README.

Related

  • Project
    marckrenn/claude-code-changelog

    A community-maintained archive that tracks the evolution of system prompts and feature flags for Claude Code across different releases.

  • Project
    Piebald-AI/claude-code-system-prompts

    A repository that extracts, version‑tracks, and publishes every system‑prompt string used by Anthropic’s Claude Code code‑assistant, providing raw prompt files, token counts, and a changelog for over 500 fragments.

  • Project
    weiesky/cc-viewer

    CC‑Viewer is a Node‑based wrapper that launches Claude Code with a local web UI, offering real‑time chat, code diffs, search, mobile sharing, password‑protected remote access, and detailed logging.

  • Project
    YoanWai/agent-manager

    agent-manager is a Go‑based tmux UI that lets developers run, monitor and interact with multiple AI coding agents (Claude, Gemini, Grok, etc.) from a single, mouse‑friendly list, offering quick prompts, diff review, session lifecycle controls and extensive customization.