Changelog

Unreleased (2026-09-25)

Catalog sync/list/route for OpenCode and OMP with exact model IDs, public Zen/Go sources, OpenRouter reference tariffs, manually audited Artificial Analysis v1.5/account evidence, provenance, fixed score bands and fail-closed billing checks.
Opt-in external goal loop with protected evaluator, isolated detached git worktrees, current quality-gate verification, budgets, evidence, stop/resume and no automatic merge or publish.
Native OMP package installation with seven agents and skills; Hermes remains skills-only. Claude Code/OpenCode/OMP goal-loop runners use constrained role tools.
Full 2026-09-25 Zen/Go catalog CSV with public-source hashes, missing-score/price coverage and account-eligibility exclusions; CLI and integration regressions.

Autonomous evaluator and quality gate require Linux bubblewrap and run without network access; agent CLIs retain network access for their configured providers. Neither the worktree nor this sandbox guarantees provider-only egress.
Detached attempts keep .git read-only to untrusted processes, revalidate the worktree before staging, and gate/score separate clean copies of each candidate commit. Unreported paid usage blocks automatic resume.
OMP discovery and dispatch validate each advertised thinking level and preserve slash-bearing model IDs; routing requires a matching exact harness or explicit proxy approval and an explicit budget for Zen. OpenRouter long-context overrides affect comparative costs, and archived CSV staleness can be replayed at a fixed UTC time.

Claude Code plugin packaging (.claude-plugin/plugin.json, marketplace.json); workflows moved to skills/<name>/SKILL.md.
dev-router: tiers 0 to 3 with security triggers that raise the tier, so small requests skip heavy phases.
lean-code: reuse ladder, dependency justification, diff budget per tier.
wf-bugfix and wf-refactor.
Agents: explorer, planner, implementer, reviewer, security-auditor, ui-critic, verifier, each with only the tools its role needs.
scripts/quality-gate: deterministic checks per tier, scoped to the diff; missing tools reported as unverified; new dependencies listed.
Hooks: Stop gate, shell command guard, formatter after edits.
opencode adapter and scripts/install.py for Claude Code, opencode and Hermes Agent.
taste-skill (design-taste-frontend, redesign-existing-projects) vendored for wf-frontend.
qa/lint_skills.py, qa/test_gate.py, qa/check_upstream.py and CI.

The five workflows were cut to about 100 lines each; phases state their minimum tier; grep-count checks replaced by the gate and real tools.
Security references moved to OWASP Top 10:2025, ASVS 5.0 and OWASP LLM Top 10 (2025).
Machine-specific paths (~/.hermes/..., vision_analyze) and references to skills that do not exist publicly removed.

The keyword-count quality harness (qa/wf_quality_harness.py, test_quality.py, test_tests.py, snapshots). It scored occurrences of words like VERIFICAR and NUNCA, so its "+98.6%" measured wording, not quality; it also failed on a fresh clone because it looked for skills in ~/.hermes.
apisec-inc's api-security-review from wf-backend (name collision with the vendored OWASP skill).