phistory
Phistory automatically archives versioned system prompt snapshots from agent CLIs like Claude Code, Codex, OpenClaw, and Hermes.
Phistory archives how coding agent system prompts change across versions
A tracker that captures and compares versioned system prompt snapshots from agent CLIs such as Claude Code, Codex, and OpenClaw.
What Phistory tracks
Phistory follows how system prompts change across popular coding-agent CLIs. The README says it tracks agents such as Claude Code, Codex, DeepSeek Harness, Antigravity, Grok Build, MiniMax Code, Kimi Code, MiMo Code, OpenClaw, Hermes, Kimi CLI, opencode, Pi, and Oh My Pi. The goal is to let a reader open a web viewer and compare prompt snapshots across versions, seeing how agent design shifts through prompts, tools, policies, and runtime instructions. The hosted viewer lives at phistory.cc, and the README notes the archive checks for new releases hourly, with the last update shown as 2026-08-24 04:23 UTC. Use cases listed include following how Anthropic, OpenAI, and other agent builders iterate on system prompts over time, spotting when new tools, permission checks, model defaults, or user-confirmation rules appear, and comparing how different CLIs structure agent behavior. Because the snapshots are stable and dated, they can be cited in posts, research notes, audits, or debugging reports without re-deriving them from a live install. The README also offers a Chinese translation, signaling that the audience spans both English and Chinese reading developers who follow the agent tooling space. A note in the README says the archive is regenerated hourly, so a newly published agent release tends to appear in the viewer within about a day of the package going live upstream.
How captures are produced
The README explains the capture method in detail. For each supported release, Phistory installs the exact CLI package and runs each configured snapshot through a tool called claude-tap, which captures the prompt-bearing HTTP request without calling the real model provider. The result is stored under captures/<agent>/<version>/variants/<variant>/ with three files: prompt.md, trace.jsonl, and meta.json. A default snapshot acts as the baseline, and selected models or modes are stored as additional variants. For recent Claude Code releases, Phistory also extracts static prompt-like strings from the installed package and stores them under captures/<agent>/<version>/static/, keeping the raw extraction input so matching rules can be improved later without reinstalling every historical package. GitHub Actions check automatically tracked CLI releases every hour and commit new snapshots when they appear. Local commands let a developer capture the latest release, capture only selected agents or variants, backfill a historical version range, extract static prompts, and regenerate the README and web viewer from the captured data. The capture configuration approach means a single default snapshot anchors every agent's history, with model or mode differences recorded as separate variants rather than separate trees. The capture workflow installs the exact CLI package and runs each snapshot through claude-tap, capturing the prompt request without calling the real model provider at all.
Coverage and capture status
The capture status table in the README shows the breadth of coverage. Claude Code leads with 395 versions captured and 395 snapshots, last captured on 2026-08-23. Codex CLI shows 77 versions and 89 snapshots, with its latest at 0.149.1 dated 2026-08-24. Other agents include DeepSeek Harness with 8 versions, Antigravity CLI with 33, Grok Build with 130, MiniMax Code with 31, Kimi Code with 66, MiMo Code with 13, OpenClaw with 69, Hermes Agent with 27, Kimi CLI with 21, opencode with 106, Pi with 41, and Oh My Pi with 69 versions. Each row lists the latest version, the version count, the snapshot count, and the last captured time, so a reader can see how fresh each agent's archive is. The repository metadata records 552 stars and a primary language of HTML, with a last update on 2026-08-24. The project does not declare a license in the metadata, and it presents itself as a research and citation resource rather than a runtime tool, with the viewer doing the comparison work in the browser. The hourly check cadence means a fresh agent release typically appears in the archive within a day of its publication. Each agent's row lists the latest version, the version count, the snapshot count, and the last captured time, so a reader can see at a glance how fresh the archive is.
Editorial conclusion
The repository records 552 stars, lists 14 supported agents including Claude Code with 395 captured versions, and the metadata shows a last update of 2026-08-24.
Community notes