intervention-detector

Intervention detection identifies patterns in confidence trends, recurring feedback, or stalled deliverables to trigger escalation decisions. Use confidence-trend data and iteration counts with configurable thresholds to automatically signal when agent swaps, specialist injections, or scope adjustments are required.

Intervention detection monitors three core patterns: confidence plateaus where scores stagnate despite iterations, recurring feedback indicating unresolved issues, and deliverables stalls signaling progress blockage. Each pattern threshold is configurable to match your workflow requirements.

Run detect-intervention.sh with iteration-count and confidence-trend parameters to produce JSON-formatted intervention decisions. The Bash-based tooling processes your data through configurable thresholds and outputs actionable orchestration commands for downstream escalation workflows.

Intervention detection requires jq for JSON processing and bc for numerical threshold calculations. Both are standard Unix utilities that enable the Bash scripts to parse confidence metrics and trigger intervention gates based on your configured decision logic.

Yes. Intervention detection works with an Intervention Orchestrator to unite detection and orchestration, managing escalation gates and intervention workflows automatically. Detection outputs feed downstream orchestration to coordinate agent swaps, specialist injections, or scope simplifications.

Confidence plateaus indicate stagnant score progression across iterations despite continued processing, while recurring feedback signals repeated unresolved issues or comments. Detecting both patterns helps distinguish between convergence failures and systemic blockers requiring different intervention types.