Updates

Every release, and what changed.

What each version of ranwhat added and fixed, newest first. Get the next one by email or RSS. The commit history has the reasoning behind every change, including what the tool got wrong.

01 / Follow

Get the next release.

One short email when a version ships, saying what changed. Nothing else, and every email has a link to stop. The emails are sent by Resend (privacy).

Or follow the RSS feed

https://ranwhat.com/rss.xml

Paste it into any feed reader, or open the feed. GitHub has a feed of its own for every commit.

02 / Releases

Releases, newest first.

Run the newest with uvx ranwhat@latest check, or bring an install up to date with pipx upgrade ranwhat.

0.10.0

clean now searches the files Claude Code keeps beside its transcripts

  • New: ranwhat clean now also searches the large tool outputs Claude Code stores in <session>/tool-results/ (a cat .env too big for the transcript lands there whole) and ~/.claude/history.jsonl, the prompts you typed. Both were left out before, so a key that only appeared there was never reported. They are only read, never rewritten: a finding says to rotate the key and then delete the file or the line.
  • Fixed: on Windows, two ledger record runs at the same time could end one with "cannot lock". A lock file that is being deleted is now a busy lock, and the run waits. The chain was never harmed.

0.9.0

A ledger of what watch flagged, with verify

  • New: ranwhat ledger keeps a local, append-only record of what watch flagged, in ~/.ranwhat/ledger.jsonl. It is opt in and free. ledger record adds one line per flagged action, holding only the facts push sends (the action id, the agent, the time, and each rule and its severity), never a command, a path or a secret. Each line carries the SHA-256 of the line before it, so the lines form a chain.
  • ledger verify names the first line that no longer fits and why: changed, taken out, put in, moved, cut short or not a ledger line, and exits with 3. A chain alone cannot show that lines were cut off the end or that the whole file was rewritten, so save the head that ledger record prints somewhere else and pass it with --head.
  • ledger record refuses to add to a ledger that does not verify, one writer at a time holds the chain, and a ledger that is a symbolic link is refused. Actions in your ignore file are left out. It sends nothing over the network.

0.8.0

Thirty coding agents, ignore, push and sixteen watch rules

  • check, watch and clean read Antigravity, GitHub Copilot in VS Code, Kilo Code, Zed, Warp, Amazon Q, CodeBuddy, Oh My Pi, Mistral Vibe and DeepSeek Harness too, thirty coding agents in all. The ten were written from each agent’s own code and documentation, not from a long-lived install, and none of them publishes its history format, so a new version of one can change what ranwhat reads. clean finds secrets in all ten, but masks them only in files it can safely rewrite: not in Warp’s, Zed’s or Amazon Q’s histories, Kilo Code’s database, Mistral Vibe’s unified sessions or DeepSeek Harness’s compressed ones. Zed’s compressed threads and DeepSeek Harness’s compressed sessions need Python 3.14 or the zstd command to be read. A task Kilo Code imported from Roo Code is listed under both.
  • watch has three new rules, for sixteen in all, up from thirteen. They are a request to a cloud metadata service for the machine’s own credentials, something set to run again on its own after the session, such as a crontab, a launch agent, a systemd unit or a scheduled task, and an agent loosening its own or another agent’s guardrails: one started with its approvals off, a broad allow rule added, a deny rule taken out, or the ranwhat hook removed. The hook asks before the high and critical ones, a call that would switch the hook off among them, once the ranwhat it runs is upgraded.
  • New: ranwhat ignore stops listing an action you expected. Each action now has an id, the 16 characters at the end of its line. ignore ID leaves that action out of later reports, and ignore all leaves out every action listed now, so later runs list just what is new. Entries go in ~/.ranwhat/ignore, a plain-text file, and --no-ignore lists everything again. A secret is never left out, and a change an agent makes to ranwhat itself, such as to that file or the hook, stays listed until you acknowledge it with ignore ID on a terminal.
  • New: ranwhat push sends this machine’s findings to your Plus account, so every linked machine’s findings can be read side by side at account.ranwhat.com. It sends rule ids, severities, action ids, agents, times, counts, the days read and ranwhat’s version, never a command, a path or a secret, and it asks before it sends. --dry-run prints the exact JSON and sends nothing.
  • For Claude Code, a flagged action says when part of it, such as a URL, appeared earlier in the same session: in the output of a web fetch, an MCP tool or a file it read, or in your own message. That is where the text was seen, not proof that it caused the call, and the output itself is never quoted.
  • A command run through env, nohup, timeout, sudo -E or a similar wrapper is judged as the bare command would be, and the files Codex edits are judged like any other agent’s writes.
  • The new rules and agents can report calls and secrets that 0.7.0 did not, so a pipeline that runs check --fail-on may start failing after the upgrade. An export PATH line an agent appends to a shell startup file is now reported as medium, so --fail-on medium fails on it. The GitHub Action’s new ignore-file input reads a committed .ranwhatignore as it was at the pull request’s base, or before the push, and what that file leaves out is not counted. Each action line in the text report now ends with its id, so a script that reads the text rather than --json sees changed lines.

0.7.0

Callers, revoke, bench and thirteen watch rules

  • New: ranwhat callers asks the provider’s own log who used a key: AWS CloudTrail by access key ID, and a GitHub organisation’s audit log by token hash. --mine CIDR marks your own addresses, and callers outside them are flagged. It is read-only.
  • New: ranwhat revoke lists the Claude Code settings that let an agent act without asking: bypassPermissions, enableAllProjectMcpServers, and allow rules for any shell command, curl, rm, git push or a credential file. --apply takes them back and keeps a copy of each file.
  • New: ranwhat bench plants fake secrets and risky commands in a temporary directory, runs the detection check uses over them, and scores it against an answer key. What ranwhat is known not to catch is listed apart from the score. It sends nothing, and exits 3 when an expectation is not met.
  • watch has four new rules, for thirteen in all, up from nine. They are a script fetched from the network and run in the same command, files made writable by every user, container volumes or every container removed, and a command hidden in base64, hex or printf escapes and then run. That last one is decoded without being run and judged by the other rules, so an encoded rm -rf ~ is reported as that deletion too.
  • The new rules can report calls that 0.6.0 did not, so a pipeline that runs check --fail-on may start failing after the upgrade.
  • clean’s review links the provider’s rotation page for each leaked key, and open N opens it.

0.6.0

Twenty coding agents, accounts and a Claude Code hook

  • check, watch and clean read Cursor, Windsurf, Cline, Roo Code, OpenCode, Continue, Aider and Goose too, twenty coding agents in all, each from the history it keeps on disk. Windsurf is read from the plaintext transcripts its Cascade hook writes.
  • New: ranwhat login links a machine to an optional account at account.ranwhat.com. It prints a code to type at ranwhat.com/device, signed in, and the machine gets a token of its own for update. ranwhat whoami says whose a token is, and ranwhat logout revokes the machine’s token and deletes it. Nothing that runs locally needs an account.
  • New: ranwhat reach lists the MCP servers your agents are configured with and what each runs or connects to, flags a secret written inline or a package fetched at no fixed version, and lists the credential files no Claude Code deny rule covers. It writes nothing.
  • New: ranwhat hook install adds an opt-in Claude Code hook, so a call watch would rate high or critical waits for your yes before it runs. It judges on your machine and sends nothing.
  • check, watch and clean take --fail-on SEVERITY, which exits 3 when anything at or above it is found, and --sarif PATH for code scanning. The repository’s action.yml runs them in GitHub Actions.
  • keep in clean’s review is remembered, so a value you keep is not reported again.
  • The bundled catalogue rates 204 scopes across twelve providers, up from 104, taking in the 100 the Plus feed added after 0.5.0.
  • On a terminal, scan, live and update --status may add one dim line about the Plus feed, never with --json. RANWHAT_NO_HINTS=1 turns it off.
  • watch flags aws s3 rb as a cloud delete.

0.5.0

Twelve coding agents

  • check, watch and clean read twelve coding agents, each from where it keeps its history: Claude Code, Codex, Gemini CLI, GitHub Copilot CLI, Qwen Code, Grok Build, Droid, Kimi Code, Kimi CLI, Pi, Muse Code and OpenClaw.
  • New: ranwhat sources lists every agent and where it looked. --source ID reads one agent, and --path ID=PATH reads one from somewhere else.
  • clean knows each value by a salted fingerprint, so masking it once masks every copy, including one another program wrote with different escaping.
  • Agents’ databases, such as Codex’s thread index and OpenClaw’s, are searched for secrets read only, and the report says how to remove one in the agent itself.
  • On Windows, suggested commands are quoted to paste into cmd or PowerShell.

0.4.0

Sharper watch and clean

  • watch spots a local file sent off the machine: piped or posted to curl, wget or nc, including through head, gzip, jq or sed.
  • watch reads what the shell really runs inside quotes, $'...' and a wrapped shell, and ranks deleting your home directory critical however it is spelled, \rm included.
  • watch masks the secrets in the commands it quotes, and check masks in its watch section every secret its clean section found.
  • Reads the transcripts that subagents write.
  • Fewer false alarms: placeholders, settings constants and AWS ARNs read as what they are, not as secrets.
  • clean and watch stay linear in time on hostile input, such as a megabyte of crafted text or one enormous shell command.

0.3.0

One command for both, and a faster clean

  • New: ranwhat check runs watch and clean in one read-only pass. It changes nothing; masking stays a choice you make in clean.
  • clean went from 89 seconds to 2.1 on a 39 MB transcript, with the same findings. It no longer searches pasted screenshots, which cannot hold a credential.
  • Output fits your terminal’s width, and NO_COLOR is honoured.

0.2.1

Commands you can run

  • Run through uvx, the overview suggests uvx ranwhat demo, because a bare ranwhat is not on your PATH there.
  • The capability catalogue covers HubSpot, Discord, GitLab, Atlassian, Microsoft Graph, Sentry and Shopify.

0.2.0

A catalogue that can be refreshed

  • ranwhat update refreshes the capability catalogue from ranwhat’s feed, for Plus subscribers once Plus opens. It sends the subscription token and nothing else.
  • Without a feed every command still runs on the catalogue bundled with the release, and a damaged cache falls back to it.

0.1.1

A first run that answers

  • A bare ranwhat prints what each command does and where to start, instead of a usage error.

0.1.0

First release

  • watch reads what Claude Code and OpenClaw ran and surfaces the irreversible actions worth knowing about.
  • clean finds credentials left in Claude Code’s transcripts and masks them when you ask.
  • scan and live score what an agent’s credentials allow, and demo shows a scan on an example.