Metadata-Version: 2.4
Name: prompt-history
Version: 1.2.0
Summary: Read back every prompt you typed into Claude Code and Codex, from your own machine.
Author-email: Owais <owaizkhan96@gmail.com>
License-Expression: MIT
Project-URL: Homepage, https://github.com/alshell7/prompt-history
Project-URL: Source, https://github.com/alshell7/prompt-history
Project-URL: Issues, https://github.com/alshell7/prompt-history/issues
Project-URL: Changelog, https://github.com/alshell7/prompt-history/blob/main/CHANGELOG.md
Keywords: claude,claude-code,codex,openai-codex,anthropic,prompts,prompt-history,prompt-engineering,llm,ai,transcript,export,cli,developer-tools
Classifier: Development Status :: 5 - Production/Stable
Classifier: Environment :: Console
Classifier: Environment :: Web Environment
Classifier: Intended Audience :: Developers
Classifier: Natural Language :: English
Classifier: Operating System :: MacOS
Classifier: Operating System :: Microsoft :: Windows
Classifier: Operating System :: POSIX :: Linux
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.9
Classifier: Programming Language :: Python :: 3.10
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Programming Language :: Python :: 3.13
Classifier: Programming Language :: Python :: 3 :: Only
Classifier: Topic :: Software Development :: Libraries :: Python Modules
Classifier: Topic :: Utilities
Classifier: Typing :: Typed
Requires-Python: >=3.9
Description-Content-Type: text/markdown
License-File: LICENSE
License-File: THIRD_PARTY_NOTICES.md
Provides-Extra: dev
Requires-Dist: pytest>=7; extra == "dev"
Requires-Dist: build>=1.0; extra == "dev"
Requires-Dist: twine>=5; extra == "dev"
Dynamic: license-file

# Prompt History

**Every prompt you have ever typed into Claude Code and Codex is already sitting on your disk. This reads it back.**

[![PyPI](https://img.shields.io/pypi/v/prompt-history.svg)](https://pypi.org/project/prompt-history/)
[![Python](https://img.shields.io/pypi/pyversions/prompt-history.svg)](https://pypi.org/project/prompt-history/)
[![License](https://img.shields.io/pypi/l/prompt-history.svg)](LICENSE)

```bash
pipx install prompt-history && prompt-history
```

Your browser opens. There is everything you have asked an AI to build, searchable, grouped, one click from your clipboard.

No dependencies. No account. No network. It never writes to your session files, it only reads them.

---

## Why you want this

You spent six months writing prompts. Some of them were really good. You have no idea where any of them are.

They are in `~/.claude/projects` and `~/.codex/sessions`, buried in JSONL transcripts next to a few hundred megabytes of tool output and model replies. Technically readable. Practically gone.

This pulls out your half of the conversation and nothing else.

- The prompt you wrote at 2am that finally fixed the build
- That one perfect refactor instruction you want to reuse
- Six months of how you actually talk to these tools, as a file you own

---

## Get it

```bash
pipx install prompt-history     # recommended
uv tool install prompt-history  # or this
pip install prompt-history      # or the classic
```

Python 3.9 or newer. That is the entire dependency list.

For a regular pip installation, both commands are installed into your Python environment's scripts directory and work from any folder:

```bash
python -m pip install --upgrade prompt-history
prompt-history --port 8080
phist --port 8080
```

If your scripts directory is not on `PATH`, use `python -m prompthistory --port 8080`. For a user-wide isolated command, `pipx install prompt-history` is also supported.

Set the port with `--port 8080`, `PROMPT_HISTORY_PORT=8080`, or `port = 8080` under `[server]` in your config. CLI flags take priority. The default is 7777; if busy, the app tries up to 19 following ports and prints the actual URL. Use `--port 0` to let the OS select an available port, and `--no-browser` to start without opening a tab.

---

## Use it

### The browser UI

```bash
prompt-history
```

Search across everything, browse by tool then project then session, copy any prompt with one click, download the lot as a zip.

It binds to `127.0.0.1` and serves a single HTML file. Nothing is uploaded, because there is nothing to upload to.

### The terminal, if you live there

```bash
prompt-history list                         # everything, newest first
prompt-history search postgres migration    # both words, anywhere
prompt-history stats                        # what you have, at a glance
prompt-history sessions                     # every session, newest first
prompt-history sources                      # where it looked and what it found
```

Some real ones:

```bash
# Your long prompts from this quarter, as Markdown
prompt-history export -f md --since 2026-07-01 --min-words 60 -o good-prompts.md

# Everything you ever said to Codex about auth
prompt-history search auth --tool codex --full

# One project, one archive
prompt-history export -f zip --project checkout-api -o checkout-prompts.zip

# Pipe it anywhere
prompt-history list --json | jq -r '.[] | select(.words > 100) | .text'
```

`--tool`, `--since`, `--until`, `--min-words`, `--max-words`, `--project`, `--exclude-project`, `--search` and `--no-slash-commands` work on every command, including `serve`. Filter once, and the UI, the exports and the zip all agree.

### The zip, organised

```bash
prompt-history export -f zip
```

```
prompt-history-2026-09-21/
├── README.txt                  what is in here, and the counts
├── index.json                  the full machine readable snapshot
├── prompts.csv                 one row per prompt, opens in any spreadsheet
├── prompts.md                  everything, one document
├── prompts.txt                 everything, no markup
├── by-session/
│   └── claude/checkout-api/
│       └── 2026-09-14-fix-the-stripe-webhook.md
└── by-date/
    └── 2026-09-14.md
```

`by-session` is the folder you actually read. Each file is one conversation, your prompts in the order you typed them, with the project, model and branch at the top.

### Sync to a folder

Keep a plain folder of Markdown on disk, updated as you work. Obsidian vault, notes repo, wherever.

```bash
prompt-history sync ~/prompts --save
```

```
~/prompts/
├── claude/
│   ├── checkout-api/prompts.md
│   └── prompt-history/prompts.md
└── codex/
    └── infra/prompts.md
```

One file per project, every session inside it in the order you typed them. `--save` means it re-syncs automatically every time you scan, so the folder stays current without you thinking about it.

Change the shape however you like:

```bash
prompt-history sync ~/prompts --layout session    # a file per conversation
prompt-history sync ~/prompts --layout single     # everything in one file
prompt-history sync ~/notes --no-tool             # drop the tool folder
prompt-history sync ~/data --format json          # or txt, or csv
prompt-history sync ~/prompts --dry-run           # show me first
```

In the UI there is a **Folder sync** panel: pick a folder with a real system dialog, choose the structure, and watch the example path update as you change it.

Two things make it safe to point at a folder you care about:

- A file is only rewritten when its content actually changed, and generated files carry no "synced at" stamp. Nothing churns, so it is fine to keep in git.
- Sync keeps a manifest of exactly what it wrote. When cleaning up stale files, it will only delete things on that list. Your own notes in the same folder are never touched, and a folder with no manifest is never cleaned at all.

### As a library

```python
from prompthistory import PromptHistory

ph = PromptHistory(since="2026-01-01", tool="claude")

print(len(ph.prompts()))

for hit in ph.search("rate limit"):
    print(hit.timestamp, hit.project, hit.text[:80])

ph.write("q3.zip")                 # format inferred from the suffix
markdown = ph.render("md")         # or get it as a string

report = ph.sync("~/prompts")      # mirror to a folder
print(report.summary())            # "Wrote 3 new, 1 updated."
```

Every CLI flag is a keyword argument, and a typo raises instead of silently returning everything.

`Prompt` carries `text`, `timestamp`, `project`, `session_id`, `model`, `git_branch`, `turn`, `words` and `kind`. `Session` carries the prompts plus where they came from.

---

## Configure it

```bash
prompt-history config --init     # writes a commented template
prompt-history config --show     # prints the values actually in effect
```

Lives at `~/.config/prompt-history/config.toml`, or `config.json` if your Python is older than 3.11.

```toml
[sources]
claude_code = true
codex = true
include_subagents = false   # sub-agent prompts are written by the model, not you

[filter]
min_words = 5               # stop "yes", "continue" and "go on" from filling the UI
skip_slash_commands = true

[server]
port = 7777
open_browser = true

[sync]
sync_enabled = true         # mirror on every scan
sync_folder = "~/prompts"
sync_layout = "project"     # project | session | single
sync_include_tool = true    # tool/project/prompts.md, or project/prompts.md
sync_format = "md"
```

The UI writes its own sync choices to `sync.json` beside your config, so it never rewrites a file you hand edited.

Config file, then `PROMPT_HISTORY_*` environment variables, then CLI flags. Last one wins.

---

## Where the prompts come from

| Source | Path | What it gives |
|---|---|---|
| Claude Code transcripts | `~/.claude/projects/<project>/<session>.jsonl` | Full text, timestamps, project, model, branch |
| Claude Code recent history | `~/.claude.json` | Prompts whose transcript is gone |
| Codex rollouts | `~/.codex/sessions/**/rollout-*.jsonl` | Full text and metadata |
| Codex composer history | `~/.codex/history.jsonl` | Everything typed at the composer |
| Codex thread index | `~/.codex/state*.sqlite` | Opening prompt of threads with no rollout file |

`CLAUDE_CONFIG_DIR` and `CODEX_HOME` are respected, lists included. Windows `%APPDATA%` too.

Three Codex rollout formats are supported, current and legacy, because the on-disk shape has changed a few times.

Codex thread titles, projects, and sidebar sections are read from the local thread index when available. Internal approval-review and sub-agent sessions are excluded by default. A prompt you deliberately submit again is kept; duplicate transcript/event representations of the same submission are folded together.

Goal prompts stored as a `goal-objective.md` attachment are expanded back into their objective text. If only saved goal context survives, its objective is recovered once and labeled. Missing or unreadable files are reported alongside the original reference; they are never silently replaced with guessed text. Recovery reads local files only.

The browser renders Markdown links to files, tools, apps, and skills inline, and keeps attachment references below the prompt. Copy preserves the underlying Markdown. Local file and app links depend on your browser and installed application handlers; their targets remain visible on hover. No linked content is fetched automatically.

**Thrown away:** tool results, tool calls, assistant turns, compaction summaries, `[Request interrupted]` markers, sub-agent prompts, and injected context like `<system-reminder>`, `<environment_context>` and `<user_instructions>`. What is left is what you typed.

When a history log repeats a prompt a transcript already has, you see it once. `stats` tells you how many were folded together.

---

## What it cannot do

Claude Desktop and ChatGPT Desktop chats are not files on your disk. They are in encrypted app storage or only on a server.

Codex makes this visible: if you use the desktop app, the thread titles are cached locally but the prompt text is not. Prompt History counts those threads and tells you, rather than quietly pretending they are not there.

If somebody tells you they can export your Claude Desktop history from local files, check what they are actually reading.

---

## Privacy

Your prompts contain your API keys, your client names and your worst code. So:

- Session files are opened read only and never modified
- The server binds to loopback, and there is no outbound request anywhere in the codebase
- No telemetry, no analytics, no update check
- Zero third-party packages, so there is no supply chain to trust but this one

It is roughly 2,300 lines of stdlib Python plus one HTML file. Read it yourself, it will not take long.

---

## Contributing

Adding a source is one file.

```
prompthistory/
├── api.py             PromptHistory: scan, filter, snapshot
├── config.py          file, env and flags, layered
├── model.py           Prompt, Session, text cleaning, timestamps
├── export.py          md, json, csv, txt, zip
├── server.py          loopback HTTP
├── cli.py             the commands
├── sources/
│   ├── claude_code.py
│   └── codex.py       <- copy this one
└── web/index.html     the whole frontend
```

Write a module in `sources/` that returns `list[Session]`, call it from `PromptHistory.refresh()`, add a fixture to `tests/conftest.py`. Done.

```bash
git clone https://github.com/alshell7/prompt-history
cd prompt-history
pip install -e ".[dev]"
pytest
node --test tests/web.test.cjs   # reference rendering and link safety
```

An optional browser smoke test lives in `tests/browser_smoke.cjs`. With Playwright and Chrome installed, run `node tests/browser_smoke.cjs` to check desktop/mobile rendering, copying, search, project isolation, and offline assets against synthetic data.

Cursor, Aider, Gemini CLI and Zed all keep local transcripts. Pull requests very welcome.

---

MIT. Interface icons from [Tabler](https://tabler.io/icons); Codex and Claude icons from [Lobe Icons](https://github.com/lobehub/lobe-icons), bundled offline under MIT. See [third-party notices](THIRD_PARTY_NOTICES.md).
