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
Every agent declares a harness-neutral tier: (light / standard / deep) and effort: instead of a fixed model: sonnet. The tier follows one question: is a wrong answer from this agent loud or quiet?
agents/model-profiles.json maps tiers to models per harness. Pick a cost profile at install: --profile hemat | seimbang (default) | maksimal.
--models <file> sets your own model ID per tier — required for OpenCode/ZCode, and an override for Claude Code/Codex when a model is renamed.
code-reviewer now runs at standard (Sonnet): on a seeded-bug benchmark every tier caught 27/27 bugs (single-file and context-dependent multi-file) at about half Opus's cost. For large PRs the code-review skill tells the caller to run the reviewer one tier up. code-critic and security-review stay deep.
Codex
New target eef-install --target codex-agents (--global for ~/.codex/agents): 46 custom-agent TOML files, read-only sandbox for review/audit agents.
Installer
eef-install --target claude-hooks installs the fs-guard and subagent-telemetry hooks and merges them into settings.json without touching other keys.
eef-install doctor reports what is installed, what is stale, and which model profile is active.
Skills
Orchestrator skills (research-ops, click-path-audit, spec-mining, gan-harness) escalate one failing delegate one tier up instead of the whole fan-out.
config-hygiene documents the advisor tool and ships SubagentStop telemetry plus a report for re-tiering agents from evidence.
git-and-release-ops now loads for quick git questions (it was not firing).
Lighter sessions
Agent descriptions no longer restate their skill's description: per-session context drops from ~16.7k to ~12.2k estimated tokens. CI caps it at 13.5k.