ccr_search before starting work, so the agent doesn’t redo or contradict something, and ccr_log_event after a real outcome. Tool names keep the ccr_ prefix. For what makes a good event, see Writing Context events.
Event shape
ccr_search
Semantic search, so phrase queries the way the event would describe the work (“login test flaky retry”), not as questions.
Returns
{ matches: [{ score, record }] }, best first. score is cosine similarity; scores around 0.35 are loose matches.
ccr_recent
The newest events, for vague requests like “continue” or “what changed today” that don’t give search much to work with.limit 1–100, default 20. Returns { records: [Event] }.
ccr_log_event
Each write counts toward the monthly Context quota.
origin.surface values: slaunt_dm, slaunt_web, external_llm, claude_desktop, claude_code_cli, chatgpt_desktop, chatgpt_web, codex_desktop, codex_cli, gemini, openai_client, mcp_client, manual, api.
ccr_issue_log
Records a reproducible bug as a structured issue attached to an existing event (the deploy or change it relates to). It requires aworkspace_people_list call earlier in the session, so the agent has looked at who owns the area; it never mentions or assigns people itself.
Event detail tools
For drilling into one event found by search. All takeevent_id.