Implements detection and recovery for when agents fail to reach consensus:
- Orchestrator exits with code 2 on consensus failure (distinct from error=1)
- Records failed run context (proposals, agent states, conflicts) to Dragonfly
- Provides fallback options: rerun same, rerun with GAMMA, escalate tier, accept partial
- Adds UI alert with action buttons for user-driven recovery
- Adds failure details modal and downloadable failure report
- Only marks pipeline complete when consensus achieved or user accepts fallback
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Vault token issuance per pipeline with 2-hour TTL
- Automatic token renewal loop every 30 minutes
- Error budget tracking with threshold-based revocation
- Observability-driven token revocation for policy violations
- Diagnostic pipeline spawning on error threshold breach
- Structured handoff reports for error recovery
- Agent lifecycle status API
- New API endpoints: /api/pipeline/token, /api/pipeline/errors,
/api/observability/handoff, /api/observability/diagnostic
Orchestrator now reports errors to parent pipeline's observability
system via PIPELINE_ID environment variable.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
