You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Combined Audit Skill (/audit) — Single entry point to run any combination of eval, health, and doc audits. Asks scoping questions one at a time, spawns up to 5 agents in parallel, produces all intake docs in one directory for a single /pipeline run
Unified Audit Flow (flows/audit-flow.md) — Merged-plan model: one planner reads all intake docs and creates one plan with phases tagged [HYGIENIST], [IMPLEMENTER], [FORTIFIER], [DOC-ENGINEER]. Tags route each phase to the correct implementer/reviewer pair
Verification Stage — Lightweight final gate replacing expensive re-evaluation. One reviewer agent verifies specific file:line findings from intake docs instead of re-running 3-5 evaluator/auditor agents
Signal Validation — Intake skills now validate completion signals (EVAL_HIRE_COMPLETE, AUDIT_COMPLETE, DOC_AUDIT_COMPLETE) before writing intake docs. Truncated agent output is detected and reported
VERIFIED/UNVERIFIED Signals — New pipeline signals persisted to feedback.md for state recovery across interruptions
Per-Pillar Threshold Overrides — Users can set custom thresholds or exclude specific pillars from the remediation gate
Cross-Evaluator Calibration — Normalizes scores across evaluator lenses before planning; divergences ≥3 points flagged as signal
Changed
Agent Reuse via SendMessage — Planner, Plan Reviewer, Implementer, and Reviewer agents are spawned once and continued via SendMessage for subsequent iterations. Preserves context instead of re-reading codebase from scratch each iteration
Strict Agent Spawning Rules — One agent at a time, no duplicates, no per-phase planners, no background agents, no parallel agents in the pipeline orchestrator
Intake Questions — Added "known pain points" (universal) and "deployment target" (health). Merged scope+constraints into single questions. Dropped redundant "context" question from eval
Token Budget — Changed from hard 50k target to flexible guideline. Planner sizes phases to the work; single-phase plans OK for small scopes