Session search: find any coding-agent conversation by name or by what was said in it, and resume it

[samithaj]

Problem

Once you have run more than a handful of coding-agent sessions there is no way to find one. The project rail can only show what fits in a sidebar — it caps discovered sessions per project — so conversations that exist on disk are simply unreachable. On this machine that is 73 project directories and several thousand Claude transcripts.

Two questions go unanswered: "which conversation was the one about X?" and, more painfully, "where is that thing I pasted?"

Proposal

A popup (Cmd+Shift+K, plus a magnifier in the rail header) over every known session — live tabs, stored session handles, and transcripts discovered on disk — deduped by (agent, session_id). Enter resumes: if that session's pane is still open in the right directory it activates that tab, otherwise it opens a new one with the resume command prefilled but not executed.

Results are staged, and the staging is the design:

• Names match synchronously against data already in memory and render on the same frame as the keystroke.
• Content matches inside transcript text and appends below, under its own header, with a one-line snippet and a progress readout. Name results never reorder.

Content search reads transcripts off the main thread into an in-memory digest keyed on (path, len, mtime) — the len component makes growth append-only, so a transcript that grew re-reads only [old_len, EOF). Matching is literal (memchr::memmem), never fuzzy, so an identifier like POA-2236 stays one token. What counts as content is decided structurally rather than by regex: a "type":"user" envelope whose blocks carry a tool_use_id is a tool result, not something the user typed — tool output and pasted file bodies are excluded, and the footer says so rather than letting a half-answer look complete.

Three constraints worth recording for anyone implementing this

1. The mixer must stay synchronous. Registering an AsyncDataSource withholds all results — including the instant name matches — for up to INITIAL_RESULTS_TIMEOUT (500 ms) while the palette still shows the previous query's rows, and that is also why Enter is a no-op while loading (search_bar.rs). Feeding a sync source from a model that owns its own spawn avoids both.
2. Section order is a tier, and the direction is counterintuitive: the mixer sorts ascending on (priority_tier, score, source_order) and the search bar then reverses for top-down, so a higher tier renders higher. This inverts silently with no compile error.
3. SelectionUpdate::Top does not skip static separators. A section header at index 0 gets selected when the popup opens, and the user's first Enter does nothing at all.

Implementation (fork, reusing the existing command-palette view rather than a new modal): https://github.com/samithaj/warp/tree/samithaj/session-search — built on #14720.

Scope note

Content search covers Claude Code only, because Codex keeps no per-cwd transcript store — its sessions are date-keyed rollouts. Codex sessions still appear in name search via recorded handles.

[oz-for-oss]

@samithaj

Here's what I found while triaging this issue:

This looks like a feature gap rather than a bug. Current conversation search appears to cover known conversation names/metadata, but not a full indexed search across all on-disk coding-agent transcripts with a resume action.

Maintainer details

I concluded that feature request for a unified coding-agent session search/resume popup that can search both conversation names and transcript content across live, persisted, and on-disk sessions..

Triage summary

This is an enhancement request for a broader coding-agent session discovery surface: search by conversation/session name immediately, then append content matches from transcript text, and resume the matching session by focusing an open pane or pre-filling a local resume command.

Evidence from the current codebase

• The current conversation list view is backed by ConversationListViewModel, which caches AgentConversationEntryIds from AgentConversationsModel and filters them with match_indices_case_insensitive against item.display.title only.
• The inline conversation menu uses ConversationMenuDataSource and query_conversation_entries, which also matches only entry.display.title; the zero-state and search result paths are bounded by DEFAULT_RESULT_COUNT / MAX_SEARCH_RESULTS.
• AgentConversationsModel merges active ambient tasks, loaded conversations, cloud metadata, and local conversation metadata, but the loaded set is intentionally bounded (INITIAL_TASK_AMOUNT, MAX_PERSONAL_TASKS, MAX_TEAM_TASKS, and MAX_HISTORICAL_CONVERSATIONS) and is not a complete transcript-store scan.
• Claude transcript helpers know how to read/write/re-hydrate Claude transcript envelopes for resume paths, but there is no indexed content-search data source that scans local Claude project JSONL files and contributes staged results to the command palette or rail.
• The reporter's notes about SearchMixer/SearchBar align with the current implementation: mixed sync/async search sources buffer initial results up to INITIAL_RESULTS_TIMEOUT, sort by (priority_tier, score, source_order), then reverse for top-down rendering; SelectionUpdate::Top currently selects the computed top index without explicitly skipping static separators.

Duplicate check

I scanned the open issue list. #14720 is closely related because it asks for durable/resumable CLI-agent conversation handles, but it does not cover the global search-by-name/content popup requested here. Other matches (#8664, #11836, #6882, #9155, #6185) are adjacent history, organization, or session-search requests, but do not duplicate this specific session-search-and-resume proposal.

Routing notes

This should route to agent conversation/session restoration plus search/command-palette owners. Because the request introduces a new discoverability surface and rail affordance, it likely needs product/design review or mocks before implementation.

This is my automated analysis and may be incorrect. A maintainer will verify the details.

Powered by Oz