Activity

Activity shows memory use inside local agent sessions. It projects provider-specific logs into one small hierarchy:

agent activity → user request → memory search → memory chunks made available to the agent

This is intentionally not a general transcript model or chat-history importer. Assistant prose, unrelated tools, and ChatGPT conversations.json exports are out of scope.

What it renders

Three columns, newest session first:

Column	Shows
Sidebar	The existing global sidebar, with Activity selected.
Content	Agent activity for bound workspaces. Each row shows a DSH or Codex host badge, title, and time.
Detail	The user request, the exact memory query written by the agent, search duration, and the selected memory chunks rendered inline as Markdown from their historical snapshots.

The host badge is part of session identity: session ids only need to be unique within a host.

Each memory search offers Retrieval Process when a run identity is available. It opens that run's summary and full candidate table in the workspace, temporarily hiding the session list while keeping the global sidebar. Back returns to the same session and search. This drill-down belongs to one activation; a request can contain several searches. The detail also reuses Report Inaccurate and Review Evidence to capture and review an evaluation case for that run. The shared diagnostics view also supports a cross-session retrieval-run list.

What a memory search means

The user request and memory query are different data:

• User request is the human message recorded by the agent harness.
• Memory query is the exact text the agent sent to memory.activate. The daemon trims it but does not perform LLM query rewriting, intent extraction, conversation expansion, or hidden filter inference.

The raw query goes through exact id/path/title matching, BM25 full-text search, semantic vector search, reciprocal-rank fusion, and cross-encoder reranking. The final assembly removes overlapping chunks and applies relevance, per-file, fragment-count, and token-budget limits. Activity shows only the chunks made available to the agent. Model scores and excluded candidates appear in the Retrieval Process detail when requested, keeping the session's reading flow focused on requests and selected chunks.

The delivery state is a context delta, not a memory edit:

• add: the chunk was sent to the agent because it was not in the previous activation state;
• replace: a newer version replaced the version the agent already had;
• reuse: the unchanged chunk was already available to the agent, so its text could be omitted from the wire response.

The UI translates these values to plain-language delivery labels and never presents add as creating or modifying stored memory.

Session sources

DSH

DSH sessions live at ~/.dsh/sessions/<encoded-workspace>/<session>/session.jsonl.zstd. The file is an append-only, multi-frame zstd archive: DSH writes each JSONL record as an independent frame. A reader must consume all complete frames; decoding only the first frame returns the session header and makes every task count zero.

The projection consumes:

• session and session/title for identity, workspace, time, and title;
• user/message only when source.kind == "user" for human tasks;
• tool/call for the current mcp__clumsies__memory activation shape, {"op":{"activate":{"query":"...","state":"..."}}}, plus the legacy mcp__clumsies__activate / top-level query shape;
• tool/result, paired to the call by callId, for the structured activation result or error.

Codex Desktop/App

Codex rollouts are discovered under ~/.codex/sessions and ~/.codex/archived_sessions. session_meta.payload.cwd binds a rollout to a workspace; active and archived copies are de-duplicated by session id.

The projection consumes structured human response_item records, legacy user_message events, and completed MCP calls. A current activation is recorded atomically as:

event_msg.payload.type = "item_completed"
payload.item = {
  type: "McpToolCall",
  id: "...",
  server: "clumsies",
  tool: "memory",
  arguments: { op: { activate: { query: "...", state?: "..." } } },
  result?: { structuredContent: { run_id?, fragments, ... }, isError? },
  error?: { message: "..." }
}

Legacy mcp_tool_call_end events and the earlier dedicated tool names memory/activate and activate, including JSON-encoded arguments, are accepted as compatibility input. Other MCP servers and other memory operations are ignored.

Projection and retrieval-run identity

Each genuine human message starts a task. Subsequent clumsies activations belong to that task until the next human message. DSH call/result records are paired by callId; a Codex McpToolCall completion already contains both sides.

For new logs, structuredContent.run_id is the authoritative link to retrieval_runs; selected candidates from that exact run supply stable chunk identity, heading, result order, and preview text. The query text is display data, not an identity key. The same lookup supplies optional total_us for a terminal run; it stays absent while the run is running or unavailable. This is the run's recorded retrieval duration, not model response time or a sum of all stage timings. Chunk counts include reuse and do not mean newly sent content.

Retrieval history stores only a bounded preview in each candidate row, but it also retains the complete resource body used by that run. Each visible truncated or empty chunk preview loads its complete text for an inline Markdown preview using run_id + unit_key and the candidate's frozen locator. It never reads the current Memory document, which may have changed since the activity occurred. Description-only units have no body byte range and therefore fall back honestly to their stored preview. Missing or cleared run records also keep the tool result's recorded preview; an unavailable snapshot does not mean no search took place.

Older logs may not contain run_id. They remain visible with any fragments or error embedded in their tool result. The daemon uses (project_id, query) only when it identifies exactly one retrieval run; it never chooses the newest among duplicates. Ambiguous logs keep run_id and run status absent.

Loading and pagination

Activity restores the last concrete project selected for this server, organization, and account before the first request. If it is unavailable, it uses the current Memory project or the first accessible project. All Projects is an explicit choice for the current session, not the launch default. No projects means no initial Activity request.

list_recalls returns summaries only, in pages of 20 (maximum 100). Discovery reads bounded log headers/previews on the blocking pool, not complete transcripts or retrieval records. Its opaque cursor continues an immutable ordering without rescanning directories. Explicit refresh discovers a new snapshot.

get_recall_session takes a summary's opaque session_token and returns one page of tasks. Only a selected session is parsed in full; later pages reuse its raw snapshot and enrich only that page's retrievals. The old 500-task/100-activation truncation is removed. Complete memory chunks are read only when a visible preview is truncated or empty.

The daemon retains eight list snapshots and four selected session bodies in memory; it does not persist another transcript archive. Evicted handles or a restarted daemon require Refresh Activity. Binding changes invalidate access to an old snapshot. Corrupt individual log headers are skipped; selected-file failures appear in the detail region with retry.

First loads use one native progress indicator in the waiting region. Refreshes retain content and selection; continuation progress stays at the list/task footer. Old responses cannot replace another project's list or another session's detail. Apple's Loading and Progress indicators guidance informs this choice; the shared view no longer combines a spinner with decorative skeleton rows.

Scope

• Sources are DSH and Codex Desktop/App rollouts for bound workspaces.
• The model is a memory-activity projection, not a reusable normalization of the full transcript and not a ChatGPT data-export parser.
• Session browsing does not change Memory, Issues, or session files. Reporting an inaccurate run and reviewing evidence update retrieval evaluation records and retain the source run for that evaluation.

Implementation map

Concern	Path
Shared projection and DSH reader	crates/daemon/src/recall.rs
Summary snapshots and task pagination	crates/daemon/src/recall/paging.rs
Codex rollout reader	crates/daemon/src/recall/codex.rs
XPC dispatch	crates/daemon/src/state.rs (list_recalls, get_recall_session, get_recall_fragment)
XPC client and models	apps/macos/Sources/Services/Daemon/DaemonXPCClient.swift, apps/macos/Sources/Libraries/Models/DaemonModels.swift
Sidebar section	apps/macos/Sources/Libraries/Models/MemoryModels.swift (WorkspaceSection.sessions)
Activity UI, host badge, and chunk detail	apps/macos/Sources/Features/Activity/ActivityView.swift, ActivityModel.swift
Workspace wiring	apps/macos/Sources/Features/Workspace/WorkspaceView.swift