Searching the whole archive¶

The search box baked into every generated page only sees the messages on that page. To search everything — every message, every session, every project — start the local server:

claude-code-log serve

Then open http://127.0.0.1:8010/search.html, or follow the “Search inside all transcripts…” link on the index page or on any project's combined transcript.

Why it needs a server¶

Archive search reads the SQLite cache, and a page opened from file:// cannot reach a database. The generated HTML stays canonical and keeps working without the server — this only adds an origin so the page can ask for search results.

Everything else is unchanged: the same pages, the same styling, the same in-page search on file://.

Two searches, on purpose¶

	Where	What it searches
In-page search (`/`)	Any session page	The messages on that page — substring, optional regex
Index page search	`index.html`	Session titles and first-message previews — a conversation finder
Archive search	`search.html`	Every message body in every project — full text

They are deliberately separate. On a real archive the index page's box covers about 0.17% of the text — it is there to answer "which conversation was that?", not "where did I ever discuss this?".

Query syntax¶

Archive search is token-based (SQLite FTS5), not substring:

Query	Meaning
`pydantic`	words starting with `pydantic`
`tokeni`	words starting with `tokeni` — finds `tokenizer`
`"cache"`	the whole word `cache`, never `cached`
`cache invalidation`	both words, anywhere in the message
`"cache invalidation"`	that exact phrase

A word you type is matched as a prefix, so a half-typed one still finds what you meant; quote it to match it whole. Punctuation is handled for you — vis-timeline searches for the two words together, and no input can produce a syntax error.

Note

Words of one or two characters are matched whole even unquoted. A two-letter prefix matches a third of the archive, which is both slow (hundreds of milliseconds, on every keystroke on the way to a longer word) and useless as a result list.

Note

Because it is token-based, a mid-word substring like ydant finds nothing — matching starts at the beginning of a word. Use the in-page search on a session page for substring or regex matching.

Filters¶

Project restricts the search to one project. Following a link from a project's combined transcript pre-selects that project; you can always widen it back to all projects.

Search in picks which parts of a message are searched. Every group has its own toggle:

Toggle	What it covers	On by default
Text	What you and Claude wrote to each other	✅
Thinking	Claude's extended-thinking blocks	✅
Tool input	Tool names and the arguments they were called with	✅
Tool results	What tools returned — file dumps, tracebacks	❌
Attachments	Pasted and attached content	✅
Meta	Session summaries, generated titles, system messages	✅

Tool results are off by default because they are roughly 69% of a typical archive's text and mostly file dumps and tracebacks, so they drown out prose unless you actually want them. Turning the others off is the way to go the other direction: Text alone finds what was said about something, without every command that mentioned it.

A reset button appears as soon as the selection differs from the server's default, and switching everything off says so rather than reporting no matches.

The query and every filter live in the URL, so results are shareable and the Back button works. The fields parameter takes the same spec as --search-fields below, so ?fields=text and ?fields=+tool_result both set the toggles.

Note

A group that --index-fields left out of the index can't be searched whatever the toggle says, so its checkbox is disabled and the page notes it as not indexed.

Jumping to a result¶

Result links land on the exact message, scroll it into view, and highlight the term — including when the match is inside a collapsed tool-result block, which the page opens for you.

A few message types (progress markers, generated titles, queued commands) have no card of their own; those results link to the session instead.

Configuring what gets searched¶

By default everything is indexed but tool results are excluded from searching, so you can turn them on per-search without a rebuild.

--search-fields sets where the page's toggles start; you can still change them per search afterwards.

# Change the default search scope
claude-code-log serve --search-fields text,thinking
claude-code-log serve --search-fields +tool_result     # add to the default
claude-code-log serve --search-fields -thinking        # remove from it
claude-code-log serve --search-fields all
claude-code-log serve --search-fields none             # every toggle off

# Or via the environment
CLAUDE_CODE_LOG_SEARCH_FIELDS=text,thinking claude-code-log serve

The field groups are text, thinking, tool_input, tool_result, attachment and meta. A spec is either an absolute list, all/none, or +/- deltas against the default — mixing an absolute name with a delta is an error rather than a guess.

To shrink the index itself — at the cost of not being able to search those groups at all — narrow --index-fields (or CLAUDE_CODE_LOG_INDEX_FIELDS). Dropping tool_result took a real 253 MB index down to 94 MB. Changing this forces a rebuild.

The index¶

The first serve builds a full-text index inside the existing cache database. On a 532,000-message archive that takes about 50 seconds, once, with a progress bar; afterwards each start only re-reads transcripts that changed, which is milliseconds.

It adds roughly 20% to the cache file (263 MB on a 1.29 GB cache).

claude-code-log serve --reindex     # rebuild from scratch
claude-code-log serve --no-index    # start without touching the index

The index is derived data: deleting the cache rebuilds both.

Staying current between server runs¶

Once an index exists, ordinary claude-code-log runs keep it up to date — you don't have to restart the server to search new conversations.

That costs about 25% extra on a conversion in which every transcript changed (3.3 s → 4.1 s over 7,516 messages); a normal run touches a handful of files and the difference isn't noticeable. Nothing is added for people who have never built an index.

To opt out and let the next serve catch up instead:

CLAUDE_CODE_LOG_SEARCH_AUTO_INDEX=0 claude-code-log

Other options¶

claude-code-log serve --port 9000
claude-code-log serve --open-browser
claude-code-log serve --no-convert     # skip the startup conversion
claude-code-log serve --projects-dir /path/to/projects

The server binds to 127.0.0.1 only, and rejects requests that don't address it as loopback — a page on the open web can otherwise point a hostname at your machine and read the archive through your browser.