Releases: B67687/Agent-Stack
Releases · B67687/Agent-Stack
Release list
v1.3 — LSP Expansion + TS7 Native LSP + OMO Model Routing
v1.3 — 2026-07-20
LSP Infrastructure
- 19 LSPs installed (up from 15): added kotlin-ls, texlab, jdtls (Java), csharp-ls
- TypeScript 7 native LSP: resolve-ts-lsp proxy detects TS version, dispatches to Go binary directly, fixes
initializednotification bug - zls + jdtls + lua-ls fixes: same
initializedpatch for zls, Java 21 wrapper for jdtls, working directory fix for lua-ls - LSP config file: created
lsp.json— daemon reads from separate config (notopencode.jsonc) - LSP daemon config discovery documented:
LSP_TOOLS_MCP_USER_CONFIG+LSP_TOOLS_MCP_PROJECT_CONFIGenv vars
OMO Model Routing
- Diversified models across 18 agents + 9 categories
- Flash + MiMo V2.5 handle high-volume agents (sisyphus, hephaestus)
- MiniMax M3, Qwen3.7 Plus, DeepSeek V4 Pro for specialized agents
- 3 tiers: Go high-rate (31k/5hr) → Go medium (3-4k/5hr) → Zen free
- Model concurrency settings expanded to cover all Go-tier models
Config & Documentation
- ADR 013: TypeScript 7 Native LSP Migration
- ADR 014: LSP Infrastructure Expansion
- check-updates.sh: 72h cooldown policy (throttle + release-age gate)
- Config audit: removed dead lsp section from opencode.jsonc, split oversized agent prompts (build: 4298→350 chars, architect: 3431→200 chars) into 3 .mdc rule files
- AGENTS.md updated to reflect new model routing
- Atlas removed from disabled_hooks (can now auto-start)
Testing
- 32 regression tests (up from 27): 4 new tests (LSP structure, wrapper scripts, model concurrency, disabled providers)
- JSONC validation with proper parser (catches trailing commas, brace mismatches)
- Pre-commit-verify.sh enhanced with JSONC + lsp.json validation
- Fixed stale test values (DCP turns 3→5, model allowlist expanded to 7 families)
v1.0.0 — Agent-Stack
Agent-Stack v1.0.0 — The authoritative OpenCode + OMO configuration with documentation.
What's Inside
- opencode.jsonc — DeepSeek-only provider, granular bash permissions, 9 agent definitions, verification gates, SearXNG MCP, context-mode tooling
- oh-my-openagent.jsonc — team-mode, ralph loop, 17 agent definitions with three-tier fallback hierarchy, dynamic context pruning, verification prompt_append
- 7 auto-rules (.opencode/rules/) — path-glob activated for Rust, Python, TS/React, Config, Git, Go, Agent-behavior
- 8 scripts — verify.sh, regression-test.sh (29 tests), health-check.sh, pre-commit-verify.sh, redact-config.sh, cost-report.sh, model-verifier.sh, post-agent-log.sh
- 12 ADRs — all architectural decisions documented
- 7 docs — README, ARCHITECTURE, AGENTS, WORKFLOW, CONFIG_MAP, CHANGELOG, TROUBLESHOOTING
Key Features
- Single-command verification:
./scripts/verify.sh - Attribution-free git: surgical dist edits, hook defense-in-depth
- Three-tier model fallback: deepseek-v4-flash → mimo-v2.5-free → zen-free
- Self-hosted SearXNG search (no API costs)
- Dynamic context pruning (10-turn protection)
- Pre-commit validation with JSON5 parse, sensitive file check, config existence