ohoshit.
A DOOR INTO THE ARCHIVEEST. AFTER WE LOST THAT ONE CONVERSATION

ohoshit

/ oh o shit / an extremely useful exclamation.

You definitely
talked about this.

Find the conversation. Follow the evidence. Pick up the thread.

Open the trapdoor

Agent history, across harnesses and machines.
Comfortably at home in your terminal.

A grown-up tool.
A perfectly stupid hat.

01 / THE BROWSER

REAL APP. SYNTHETIC CONVERSATIONS.
The real ohoshit terminal browser showing synthetic Codex, Claude, and OpenClaw conversations on one demonstration host.
One place to find where you left off.↑↓ browse Enter open / search
THREE HARNESSES. ONE HISTORY BROWSER.
  • Codex
  • Claude
  • OpenClaw
02 / FOLLOW THE THREAD

Past you
left a trail.

Somewhere between the first question and the seventeenth tool call, you figured something out. There should be a way back.

[01]

Start with the request.

Conversations open as request chapters. See what was asked and the response outline, then step into the part you need.

A real ohoshit terminal view grouping a synthetic conversation into requests and response outlines.
Chapters, with the original words inside.Enter
[02]

See how it happened.

Open replies, unfold tool activity, inspect checkpoints, and follow parent and subagent links. The evidence keeps its source.

A real ohoshit terminal view showing the replies and folded tool activity inside one synthetic request chapter.
Replies, tools, and links to the source.Esc

Find your own words.

Search conversation prose, or narrow the hunt to request messages. Filter by host, workspace, harness, date, and status.

Keep the family tree.

Follow verified parent, child, and prior-window relationships. Uncertain history stays visibly uncertain.

Reopen the real thing.

Get native resume or fork commands for supported sessions, with the right identity, host, and workspace.

03 / WHEN THE TRAIL GETS WEIRD

Send a wizard.
Keep the receipts.

Configure a model runner, then investigate a question across your history. Come back to saved findings, source citations, and the search conversation itself.

Model investigations are opt-in. Findings open the original conversation so you can check the evidence yourself.

aask the archive i revisit investigations
A QUESTION FOR THE ARCHIVE
“Why did we abandon that approach, and what finally worked?”

The kind of question a filename has never answered particularly well.

FOLLOW THE CITATION
A FEW SENSIBLE SPELLS

Care for the archive.

01Original logs stay on their source machines. Browsing reads retained history without rewriting it.

02Identity comes before convenience. Exact session IDs and verified native keys keep old windows distinct from current conversations.

03Memory should be portable. Saved investigations keep their question, progress, findings, and links to the source conversation.

04 / STATUS REPORT FROM THE WORKSHOP

Built.
Being unboxed.

EARLY
PRIVATE
PREVIEW

The terminal browser works locally today, with configurable SSH hosts when your archive spans more machines. A free, open-source release is planned.

The source is private today. Public installation instructions and downloads will arrive with the release.

Explore the real screenshots

Working today

  • Codex, Claude, and OpenClaw history
  • Request chapters and expandable event timelines
  • Host and workspace shelves, search, and source links
  • Native continuation and saved investigations

Still on the workbench

Public packaging, broader adapters, and a batch summary interface. The xylophone has made itself comfortable.

For preview testers Real commands, after an authorized source checkout

From the source checkout, open the local browser, search request messages, or browse your configured SSH hosts:

bin/ohoshit
bin/ohoshit --query "why SQLite?" --search-in human
bin/ohoshit --hosts all

Investigations require an explicitly configured model runner. The bundled Codex runner uses your ChatGPT subscription; the custom-runner protocol supports other harnesses. There is no public checkout or install link yet.

Can it reopen every old conversation?

It offers native continuation where the source identity and harness support it. Retained historical OpenClaw windows remain browsable; they don’t get a command that silently opens a newer conversation.

Does an investigation send history to a model?

Yes. Once you configure and explicitly start an investigation, its model runner can read relevant history to answer the question. The bundled Codex runner uses your configured ChatGPT sign-in; other harnesses can connect through the custom-runner protocol. Ordinary browsing is read-only, and original transcripts remain on their source hosts.

Do I need a fleet of machines?

The browser works locally by default. Add configured SSH hosts when your archive spans more machines. No five-computer minimum. A single sufficiently chaotic laptop is a fine beginning.