Platform

Pi / OpenClaw (Pi Agent) — but applies to all platforms using CONTEXT_MODE_PROJECT_DIR to opt into a global shared DB.

Feature description

When opting into a global knowledge base via CONTEXT_MODE_PROJECT_DIR=$HOME (or any single shared path), ctx_search has no way to scope results to the current project. The only user-facing filter parameters today are source (label partial-match), sort, contentType, and limit.

The problem: auto-captured session memory — indexed under fixed source labels like session-events, project/<file>, user/<file>, memory/<file> — cannot be namespaced by project because those labels are generated automatically and contain no project identifier. Using source: "myproject/" only works for content explicitly indexed by the user with a matching prefix; it silently misses all auto-memory.

The internal plumbing already supports per-project filtering at every layer:

• SessionDB.searchEvents(query, limit, projectDir, source) — projectDir param exists and filters correctly
• searchAutoMemory(queries, limit, projectDir, configDir, adapter) — projectDir param exists
• ContentStore chunks carry session_id / event_id FK attribution (written via index({ attribution })) which links back to a session that has a project_dir

The gap is that ctx_search never surfaces projectDir as a user-facing parameter, so it is inaccessible from MCP tool calls.

Use case

I want a single global knowledge base so I can search across all my projects in one ctx_search call when I need cross-project recall. But I also want to narrow results to just the current project during normal work — without having to maintain per-project DB configs or restart sessions.

Concrete workflow:

# default: search only current project
ctx_search(queries: ["apk install"])
# → only returns results from /home/jet/src/temp

# explicit global search
ctx_search(queries: ["apk install"], project: "global")
# → returns results from all projects

Today the only workaround is either:

1. Stay on per-project DBs (lose cross-project recall entirely), or
2. Use the global DB and get undifferentiated results from all projects mixed together

Related: #663 fixed cross-project auto-memory leakage in per-project DB mode. This feature is the complementary ask for global DB mode — intentional cross-project sharing with opt-in per-project scoping.

Proposed solution

Add an optional project parameter to ctx_search:

project?: "current" | "global" | string
// "current" (default when global DB is active) → filter to current session's projectDir
// "global"                                     → no filter, search all projects (today's behaviour)
// "/path/to/project"                           → filter to that specific project

When CONTEXT_MODE_PROJECT_DIR is not set (default per-project DB mode), the parameter is a no-op — the DB is already isolated by hash, so every search is implicitly project-scoped.

Implementation sketch:

1. ContentStore chunk attribution — the attribution FK (session_id, event_id) already written on indexed chunks can join to session_events.project_dir. A WHERE project_dir = ? filter on the FTS5 join would scope FTS5 results without schema changes beyond adding the join.

2. searchAllSources in src/search/unified.ts — already receives projectDir in SearchAllSourcesOpts and passes it to sessionDB.searchEvents and searchAutoMemory. The missing piece is forwarding it to the ContentStore search path (Source 1 in searchAllSources).

3. ctx_search tool handler in src/server.ts — add project to the Zod schema, resolve "current" to getProjectDir(), "global" to undefined, and thread the resolved value through to searchAllSources.

Backwards compat: when project is omitted the behaviour is identical to today.