oracle

Truth-verification harness with structured research, budget-aware rate-limit gating, multi-workspace MCP onboarding, transcript-corpus-derived behavioural guards, and (new in 0.6.0) a CodeRabbit-style session-summary checkpoint loop. Verification: SessionStart hook injects a protocol; PreToolUse hook on Bash intercepts unpinned installs across 13 package managers; tiered PreToolUse + PostToolUse gating on the bundled firecrawl-mcp@3.2.1 surface (silent under 80% used, soft remind 80-95%, ask 95-100% or single-call >=15% or rolling-hour over 5000 cr, deny over 100%). Eight slash commands: /oracle:setup, /oracle:verify, /oracle:research [--intensity], /oracle:vet, /oracle:budget, /oracle:mcp-fleet, /oracle:narrator, /oracle:bugfix (new in 0.10.0: diagnose-then-fix workflow -- ground in git history before touching code, root-cause fix with Boy Scout cleanup, regression tests, hotfix-by-default shipping). Five Opus-4.7 subagents. Seven auto-trigger knowledge skills. Session-summary checkpoint loop (UserPromptSubmit + Stop + SessionEnd): cumulative agent-active milliseconds accrue on Stop only so user-AFK time is structurally excluded; UserPromptSubmit fires a CodeRabbit-style review every 30 minutes of active work or every 50 turns (whichever lands first), with deterministic tool-call counts plus an LLM narrative + three self-review questions generated by the narrator model selected via /oracle:narrator (defaults to claude-sonnet-4-6; aliases sonnet/opus/haiku; `off` disables the LLM tier). safe-edit-guard PreToolUse hook on Edit/Write/MultiEdit/NotebookEdit reminds when target path has not been Read within a 30-minute freshness window; track-reads PostToolUse hook records read state per session. Targets ~1,128 read-before-edit errors plus ~300 speculative-path failures plus 676 long-session interrupts observed across the local 1,186-transcript corpus. Requires FIRECRAWL_API_KEY.