Search docs: sessions, prompts and archived tool output

The gateway saves each turn of your Claude Code and Codex sessions as it passes. You search them in the console. Your agent searches the same store with context-mode-search, and reads archived tool output back byte for byte, except secrets and e-mail addresses, which are masked before storage.

What is saved

Each request carries the conversation. The gateway appends the new turns to your account's store, grouped by workspace (a project and a worktree) and by session. Claude Code and Codex turns go to the same store. Client preamble, such as Codex's environment block, is stripped first.

Sessions from before you connected the gateway are not in the store.

Search in the console

Open Search in the sidebar at console.context-mode.com. Type at least 2 characters.

• Search runs over every workspace in your account that has saved turns. The line under the box says how many workspaces it searched and how many turns matched.
• Results arrive as each workspace answers. A workspace that fails is counted, and searching again retries it.
• Each hit shows your message and the reply, with the project, the date and your words marked. Several hits in one exchange show once.
• Each workspace shows its top 50 matching turns and the full count. Add a word to narrow it.
• Click a hit to open the reader: the exchange and four turns either side. Show earlier and Show later add six turns at a time. Esc closes it.

To open a session again in your terminal, see Session resume.

The agent's search tool

The gateway adds context-mode-search to the tools of Claude Code and Codex requests and answers the call itself. Nothing runs on your machine. The tool has two modes.

• Query. With q, it returns the top matches from your archive, ranked by BM25, each with its docid and a snippet.
• Read. With docid, it returns the archived bytes as stored, one page at a time.

A read answers with total_bytes, offset, returned_bytes and next_offset. The agent passes next_offset back until it is null. An id with nothing behind it answers found: false.

When Context Saving shortens a result, the label it leaves names the call to make:

context-mode-search {"q":"full output","docid":"context-mode-toolcap-9f2c1b-184320","offset":0}

How large output is stored

The archive is a SQLite database with a full-text index, one per account. A large output is split into pages of 512,000 characters under one docid. A read crosses pages, so total_bytes and next_offset cover the whole output. The test suite stores 5 MB and reads it back identical. On the deployed gateway, Context Saving labels paged back to the stored output byte for byte.

Masking. Before archived text is stored, the gateway replaces known secret formats with [redacted-secret] and every e-mail address with [redacted-email]. So a git log reads back with its author addresses masked. This keeps secrets out of the archive.

Who can read it

• Each account has its own store. The console needs a sign-in, and a workspace from another account answers 403.
• Tool output the gateway archived stays readable with Memory off. Archived conversation needs Memory on; with it off, those ids answer as not found.

Compared with local history

Limits

• The console searches the first 4,000 characters of each tool result. The agent can read the full archived output.
• Full-text search covers the first page of a large output. Later pages are reached by docid and offset.

Compared with other tools: see the landscape.