Context Health Check

Context Health Check

Use when a DevFlow task may be drifting, repeating failures, spreading its diff, missing evidence, or approaching a checkpoint/compact boundary.

Procedure

1. Inspect git status, git diff --stat, .planning/devflow/STATE.md, the latest verification record, and active OpenSpec change files. If .planning/devflow/implementation-readiness/<change>/ exists, inspect its Requirement, Evidence, current Receipt, semantic bindings, and override state without rewriting any document.
2. Run the health check:

python3 scripts/context_health_check.py --repo <repo> --write-report --json

3. Interpret the decision:
   • continue: proceed with the next planned step.
   • reconcile: stop implementation and align against git diff, files, tests, and OpenSpec.
   • checkpoint_compact: run the existing checkpoint-compact flow before expanding work.
   • checkpoint_new_thread: create a checkpoint and use the report's minimal next context in a new thread.

Goal and Delegation Gate

Context health repairs drift; intake or planning owns the primary Goal Suitability Gate. If goal.status is missing, stale, conflicting, or weak, use define-goal and require evidence, boundaries, non-goals, stop conditions, and the Goal Quality Gate.

Evaluate subAgent usefulness at planning, execution, context-health, and review boundaries. Recommend it for repeated investigation pressure or command failures, independent diff spread, or a bounded review or delegation need. Resolve every disposition: pending with record_context_health_disposition.py.

Before changing a goal or recording any delegation disposition, read references/goal-and-delegation.md. Record accepted only when explicit user authorization or an approved delegated workflow exists and the Agent Task Contract validates; otherwise record declined or blocked. A read-only explorer or reviewer must keep its stated scope; any edit or expansion is a Human Gate for the main agent.

Historical Recovery and Privacy

For work that predates DevFlow events, read references/session-recovery.md before importing session history. Missing usage or attribution stays unknown. Store only sanitized metadata, never prompts, file or command bodies, or raw tool payloads.

Readiness drift is not context drift to auto-repair. Report the stable issue codes and exact next action; never select a provider, refresh evidence, record an override, or promote a Requirement from this diagnostic route.