McRitchie Studio

Agent Context

Root cause: bin/release (release.rb) does its real work as subprocesses (git, gh, bin/qa-server, heroku run conductor snippets) inside ONE Claude Bash tool call, and opens/closes its own agent-activity span (open_role_span REMOTE 'sweep -> deploy RC to QA', line ~1066/1404). The PostToolUse capture hook (bin/atomic-capture-hook) only fires on Claude tool calls, so the REMOTE span brackets a period with zero capturable actions -> 'No raw actions attributed'. Also the session's parent WORKFLOW span gets auto-closed when bin/release writes the marker. Fix (operator chose Real rows / self-report): make bin/release's step() best-effort POST an AgentAction to /api/v1/agent_actions attributed to the currently-open activity (agent_activity_id from the local marker read_open_activity), mirroring atomic-capture-hook auth (POST /api/v1/auth {secret} to mint the agent token). Each step (sweep! xN, per-app QA deploy, boot-wait, record qa_url, qa_green! flip) becomes a real row. MUST be best-effort/non-fatal like the existing narration + record_deploy_intent helpers: any error (down endpoint, no token, PG outage) is swallowed and the release continues. Member-flip token usage still rides each member's assembled TaskEvent via qa_green! usage_by_slug (unchanged). Applies to ship's Avi span too. Deploy-side subprocess actions have no LLM token cost, so cost stays 0 on those rows by design.