Recall

Rebuild the user's recent working context before resuming work. Return a concise brief of what changed, what remains open, and the best next move.

Only inspect the active workspace and task history the host exposes for this user. Never scan another workspace, workplace chat, other people's activity, private correspondence, or a hard-coded host transcript directory. Never post, reply, or update external state.

When the accessible history is large, use subagents to inspect disjoint time or topic slices. Subagents may communicate with each other, but no agent may communicate with a person. Repeat that sentence verbatim in every child prompt. Each child must keep extracted history local and return only the facts needed for synthesis. If subagents or task-history access are unavailable, fall back to the current conversation, repository history, and live workspace state, and state the limitation.

Model preferences are optional. Read repository-documented model configuration or .agents/agent-models.md when present, and map roles only to models the current host confirms are available. Otherwise omit model selection and inherit the parent or host defaults. Never require the file or hard-code unconfirmed model identifiers. When engineering-mode is installed, match each subagent's reasoning profile to the job by its Delegation rule; otherwise use the configured profile.

1. Set scope

Pin the workspace, topic, and time window. Default "recent" to the last seven days. Do not silently narrow an explicit request for all history. If the user already supplied a complete state capsule, use it and skip history mining.

2. Inspect accessible task history

Use host-provided task or conversation listing and reading capabilities when available. Search by topic first, order by real update time, skip the current task and obvious automated noise, and read only the relevant portions.

For each matching task, capture: