session-boot

Session Boot Sequence (run on every new session)

Steps 1-3d: Parse session-boot.sh output. The session-boot.sh SessionStart hook reads pipeline-state.md, context-brief.md, error-patterns.md, pipeline-config.json, counts custom agents, and checks the CLAUDE_AGENT_TEAMS env var. It outputs structured JSON with: pipeline_active, phase, feature, stale_context, warn_agents[], branching_strategy, agent_teams_enabled, agent_teams_env, custom_agent_count, ci_watch_enabled, darwin_enabled, dashboard_mode, project_name, sentinel_enabled, deps_agent_enabled, token_budget_warning_threshold (number or null; absent key treated as null -- no budget gate for Medium, Large-only estimate with hard pause), stop_reason (present when a previous terminal pipeline has a recorded stop reason, or inferred as "session_crashed" when a stale pipeline is detected -- see session_crashed inference rule below). Eva parses this JSON to populate session state. Derive agent_teams_available from agent_teams_enabled && agent_teams_env.

Eva must parse state_dir from the boot JSON and use it as the value for docs/pipeline for the duration of the session. All reads and writes to session-specific files (pipeline-state.md, context-brief.md, investigation-ledger.md, last-qa-report.md) must use this resolved path. error-patterns.md always stays at docs/pipeline/error-patterns.md regardless of state_dir. If state_dir is absent or empty, fall back to docs/pipeline/.

session_crashed inference rule: When session-boot detects a stale pipeline (stale_context: true OR pipeline_active: true with a non-idle phase), AND the PIPELINE_STATUS JSON in pipeline-state.md has no stop_reason key (or its value is null), session-boot includes "stop_reason": "session_crashed" in its JSON output. Eva uses this inferred stop reason when announcing the stale pipeline to the user and when capturing T3 telemetry retroactively (if applicable). The session_crashed value is inferred -- it is never written to pipeline-state.md by Eva directly (Eva cannot write during a crash). Eva may write session_crashed to pipeline-state.md after boot only when she is explicitly recovering a stale pipeline and choosing to record the inferred stop reason for posterity.

1. Brain health check -- call atelier_stats. Two gates:
   • Gate 1: Is the tool available? (If not → brain not configured, skip)
   • Gate 2: Does it return brain_enabled: true? (If not → brain disabled by user)
   • Both pass → set brain_available: true in pipeline state
   • Either fails → set brain_available: false, log reason, proceed baseline

4b. Telemetry hydration (if brain_available: true) -- call atelier_hydrate with the project sessions path. Derive the path from CLAUDE_PROJECT_DIR using the same convention session-hydrate.sh used: ~/.claude/projects/-{CLAUDE_PROJECT_DIR with / replaced by -} (leading / becomes -). Example: CLAUDE_PROJECT_DIR=/Users/alice/projects/myapp → ~/.claude/projects/-Users-alice-projects-myapp. Non-blocking: atelier_hydrate returns {status: "queued"} immediately. Do not await the result. Move directly to step 5 without waiting for hydration to complete. When brain_available: false: skip this step entirely.

1. Brain context retrieval (if brain_available: true) -- call agent_search with query derived from current feature area. Inject results alongside context-brief.md. 5b. Telemetry trend query (OPTIONAL -- skip unless a pipeline is about to start or the user explicitly asks about pipeline trends) -- if brain_available: true, call agent_search with query "telemetry pipeline summary", filter: { telemetry_tier: 3 }, limit 10. Filter results client-side: keep only records where source_phase == 'telemetry'.