agent-recap: local search across Codex and Claude Code transcripts

[rogermas05]

I built agent-recap because finding an old answer across Codex and Claude Code sessions meant grepping raw JSONL.

It searches selected Codex and Claude Code JSONL transcripts together in a browser tab. Open a result to read nearby messages and the answer. No account, installation, uploads, analytics or browser storage.

Try the synthetic Codex example:

1. Open the transcript explorer. Sample sessions load automatically.
2. Replace the search with webhook.
3. Open Read nearby messages. The answer describes a sliding window and a Retry-After header, even though it doesn’t repeat “webhook”.

You can also select your own JSONL files, up to 50 MB total. Keep private transcripts out of this discussion.

For a large rollout file, the Node 20+ CLI exports readable text to a new file:

npm exec --yes --package https://github.com/rogermas05/agent-recap/releases/download/v0.3.0/agent-recap-0.3.0.tgz -- agent-recap export ./rollout.jsonl --agent codex --out ./readable.txt

It leaves the source unchanged and refuses existing output. Export omits data/base64 string values before parsing, with an 8 MiB limit on the remaining record. Other large fields can still cause a whole record to be skipped, with counts reported.

This is user/assistant text in file order. It omits tool results, reasoning and live state. It does not reconstruct branches, restore deleted files or make an old task resumable. The browser previews up to two messages either side of a match, with long text capped and labeled.

MIT source, limits and tests. If you try the sample, did the nearby answer give you what you needed, or would your real workflow require something else, such as tool output?

[ryanroundhouse]

I had an AI assistant try the public synthetic browser example. Searching webhook and opening Read nearby messages showed the sliding-window / Retry-After answer. webhook Retry-After also found the same session, even though the words occur in different messages. That answers the recall question in your sample.

One small first-use improvement: after a no-match search, Clear all → Use sample sessions loaded the two samples but kept the old query, so the page still showed zero results. A Reset sample search action, or a hint that the existing filter is still applied, would make that state easier to understand.

For reviewing a coding-agent change, I would want optional tool evidence before relying on the answer: the relevant command and exit status, with a clear “not present” state if the transcript lacks them. I would keep that separate from the nearby-text view and avoid exposing every tool result by default. Your existing warning that tools and runtime state are not shown is useful.

Only synthetic samples were checked here, with no private transcript import or CLI installation. I'm Ryan, founder of Zozo HQ; the check and this feedback were prepared with AI assistance.

[rogermas05]

Thanks Ryan, this is really useful.

The sample reset bug is fixed and live. Use sample sessions now restores the sample search, and a search with no matches says it is still filtering every session.

Tool evidence makes sense for reviewing a change. I'm adding it as an opt-in view next to the nearby messages: the command, its exit status, and a clear "not in transcript" state when the log doesn't have them. Tracking it here: https://github.com/rogermas05/agent-recap/issues/1

[ryanroundhouse]

Thanks for the quick fix. My AI assistant rechecked the synthetic demo: the no-match hint is clear, and Use sample sessions resets the search correctly. The opt-in evidence view in #1 matches what I meant.

[rogermas05]

Shipped. Each search result now has a collapsed Tool evidence panel next to the nearby messages. It lists the commands that ran between that turn's prompt and the next one, with the exit status where the log records it. When there are no tool records it says "none in this turn" instead of implying anything ran. Tool output is never kept or shown.

It reads Claude Code tool_use/tool_result and Codex function_call/custom_tool_call/local_shell_call with their outputs. The Codex sample now includes one test command, so searching webhook shows it.

Thanks again for the idea.

[ryanroundhouse]

Thanks for shipping the tool-evidence view and following up!