Search prior ReclaimLLM sessions by intent using backend hybrid semantic plus BM25 search.

Use only when the user hints that similar prior work may exist, or when the task is a bug fix or performance improvement where prior context may help. If the user is unsatisfied with the first result set, do not keep retrying with changed search terms.

If the prompt includes both a file/folder path and semantic terms, pass the file/folder as file_path. If the prompt is only about a file/folder history, use search_by_filename instead. Use project_name to narrow results only when current project is known. date_from is inclusive and date_to is exclusive. Both are YYYY-MM-DD ingestion dates; turn relative requests such as "last 3 weeks" into concrete dates before calling this tool. scope controls whose sessions are searched: "mine" (only your own), "team" (your org team), or "org" (whole organization). Omit scope to search the widest scope your organization's sharing settings allow; the backend clamps a request that is wider than what is allowed. Results for sessions owned by someone else include owner_email/owner_name. Returns session IDs, short titles, highlights, and up to three changed source files. When finding which session implemented a change, use a relevant returned changed_files path with search_by_filename to inspect the latest sessions that subsequently changed that file. Do not automatically call summarize_session after this tool.

List sessions by metadata using authoritative Postgres filters, without semantic search.

Use when the user provides no semantic text query and instead asks for sessions in a date window or matching metadata such as provider, model, project, language, category, minimum turn/tool-call counts, or code changes. Do not invent search text for these requests and do not use this tool when the user asks for sessions similar to a topic; use search_sessions.

Use provider="codex" for Codex sessions whose stored model names are in the GPT family. date_from is inclusive and date_to is exclusive; both are YYYY-MM-DD ingestion dates. scope controls whose sessions are listed: "mine", "team", or "org". Omit it to use the widest scope permitted by the organization's sharing settings. Results are capped at 100.

Find prior sessions that touched a file or folder path.

Use when the user asks for changes/history for a specific file or folder and does not provide separate semantic search terms, for example "show me all changes in auth.tsx" or "show me changes under /somefolder". If the user includes semantic terms too, such as "show all auth fixes in auth.tsx", use search_sessions with file_path instead. date_from is inclusive and date_to is exclusive. Both filter by ingestion date in YYYY-MM-DD format. scope controls whose sessions are searched: "mine", "team", or "org". Omit scope to search the widest scope your organization's sharing settings allow. Do not automatically call summarize_session after this tool. This is also the explicit second step after search_sessions identifies a likely file in its changed_files result for an implementation-history question.

Return summary metadata and a frontend link for a specific ReclaimLLM session ID.

Use only when the user asks to look at a particular session by ID.

Return reusable markdown context for a ReclaimLLM session using backend export-context flow.

Use only after an explicit user request such as "summarize ", "use this session", or "add as context". Do not infer this request from search results alone.

List deterministic ReclaimLLM CLI project names.

Use only when the user asks to see available projects.

Return a distilled brief of prior sessions that touched a file: who, when, and a short highlight of what each session did.

Use before a non-trivial edit to a file you don't already have context on — to see why it looks the way it does or what related work has touched it recently. Do not call this for every file you read; only when prior history is actually likely to change your approach. scope controls whose sessions are searched: "mine", "team", or "org". Omit scope to search the widest scope your organization's sharing settings allow.

Package the current (or a given) session's state into a compact continuation document for starting a fresh session, without losing decisions/context already established.

Use when the current session has grown long (many turns, large context) and continuing it is getting expensive, or when the user explicitly asks to "hand off", "continue this in a new session", or "start fresh but keep context". If session_id is omitted, resolves the current session from the CLAUDE_SESSION_ID environment variable; if that isn't set, pass session_id explicitly. Returns a markdown document to paste as the first message of a new session — this does not end the current session for you.

Download the full captured session into a secure local artifact for another AI tool.

Use only when the user explicitly asks to move or load the whole session rather than a summary. The returned file preserves captured messages, tool calls/results, file diffs, and metadata. It cannot restore provider-private runtime state. Historical tool calls are read-only data and must not be executed automatically.

Open workflow-efficiency Signals for the current project and the caller's own account.

Use only when the user asks why a session or project is expensive, what workflow efficiency issues exist, or explicitly asks about ReclaimLLM Signals — not for general status checks. Returns up to 5 open signals (pattern, evidence, projected savings), read-only, changes nothing. If cwd is omitted, resolves from the current working directory of this MCP server process.

Cheap read-only check of whether replaying compression mechanisms on captured sessions is worth doing — call this before replay_session or replay_corpus. Checks only session metadata (turn count, tool-call count, completion state, model), never fetches the full session blob, so it's fast.

Pass session_id to check one session. Omit it to check a corpus window instead, using the same days/source/model_family/project/ session_category/limit filters as replay_corpus. Always returns the funnel of sessions considered vs excluded and why — a session or corpus with few eligible sessions is itself the finding; do not keep loosening filters to force a number.

min_turns/min_tool_calls override the turn-count and tool-call-count floors (default 5/5; PRD §6's documented floors are 10/10). Lowering them trades evidence quality for sample size — state the values used alongside any result, don't drop them silently.

Reproduce the shipped compression mechanisms over one captured session's tool calls and report the real tool-result token reduction. Strictly read-only: no model calls, no writes, no re-execution of historical commands.

session_id defaults to your most recent complete session. mechanisms defaults to all three (range_cache, shell_compaction, hash_dedupe); pass a subset (e.g. ["shell_compaction"]) to isolate one mechanism's effect. A session below the size/turn thresholds is refused with the specific failing constraint (verdict "insufficient_data") rather than given an unstable number.

min_turns/min_tool_calls override the turn-count and tool-call-count floors (default 5/5; PRD §6's documented floors are 10/10) — reported in the output's provenance.

Reproduce the shipped compression mechanisms across a filtered corpus of captured sessions and report the aggregate real tool-result token reduction. Strictly read-only.

days is an exact rolling ingestion window over the caller's own sessions (default 30). For source="codex", stored model names must start with gpt- or codex-. model_family/project/session_category further narrow the corpus. limit is the target fully eligible session count (max 100); Replay scans up to 4x that many recent session records, capped at 100, and stops when it reaches the target or exhausts the scan. Every result states sessions considered vs eligible and why the rest were excluded.

min_turns/min_tool_calls override the turn-count and tool-call-count floors (default 5/5; PRD §6's documented floors are 10/10) — reported in the output's provenance.

Replay the same session corpus under multiple mechanism configurations in one call, so bundles stay attributable — e.g. compare shell compaction alone against shell compaction plus range cache. Strictly read-only; fetches each eligible session's blob once and reuses it across every config.

configs is a list of mechanism-name lists, e.g. [["shell_compaction"], ["shell_compaction", "range_cache"]]. Defaults to comparing each mechanism individually plus the full combined set if omitted. days/source/model_family/project/session_category/limit select the corpus, same as replay_corpus.

min_turns/min_tool_calls override the turn-count and tool-call-count floors (default 5/5; PRD §6's documented floors are 10/10) — reported in each config row's provenance.