Tags: openexec/openexec
Tags
build: stamp version, commit, and build date from the git tag The version constant lied after every release (a 0.13.1 binary reported v0.12.1). Version is now a var stamped via -ldflags from git describe; the in-source default only appears in untagged local builds. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
fix: canonicalize both sides when relativizing legacy stored paths Legacy migration kept absolute paths on macOS because the evaluated repo root crosses the /var -> /private/var symlink while stored paths do not. Caught by TestRepositoryIdentityAndLegacyMigration on a Mac host after the branch passed only in a Linux sandbox. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
release: v0.9.0 feat(providers): named API endpoints with switching - ProviderConfig + Providers map + ActiveProvider on ExecutionConfig. - ActiveAPI() resolves the active endpoint with read-time fallback to legacy api_provider/api_base_url/api_key/api_model fields, so existing configs keep working unchanged. - openexec config provider list/use/add/remove subcommands. - init flow: AgenticsNZ preset (token-only) + custom OpenAI-compatible flow; entries are saved under a user-chosen name so multiple endpoints (e.g. agentics-work, agentics-personal, vllm-local) can coexist. feat(agent): OpenAIProvider accepts custom Name + Models + ModelInfo, and looks up <NAME>_API_KEY from the environment when APIKey is empty. build: add -s -w -trimpath for smaller, reproducible binaries.
release: v0.8.2 Skills System: - SKILL.md loader with multi-source discovery (builtin/user/imported/project) - Keyword-based skill selection injected into prompt briefing - CLI: openexec skills list/info/search/import/create/enable/disable - Claude Code skill import from ~/.claude/skills/ OpenAI-Compatible API Provider: - Configure any OpenAI-format API (Kimi, Mistral, Ollama, etc.) - Full agentic loop via HTTP with tool execution (read/write/shell/patch) - Config: api_provider, api_base_url, api_key, api_model - Interactive setup in openexec init Coordinator-Pattern Multi-Agent: - Frontier model decomposes tasks into independent subtasks - Workers execute in parallel with dependency-aware wave scheduling - Git diff capture and sequential patch merging - Config: coordinator_model, worker_model, worker_api_provider
release: v0.8.0 Major milestone: all subsystems wired into execution pipeline. New features: - Deterministic routing (always-on): classifies mode, toolset, repo zones, sensitivity, and frontier model requirements for every task - BitNet routing (opt-in): local 1-bit LLM with auto-download model manager - Quality Gates V2: auto-detects project type, runs lint/test/format gates - Caching: knowledge cache + tool result cache for repeat runs - Predictive loading: pre-fetches files based on task description - Memory system: extracts patterns from completed stages, injects in future runs - Checkpointing: deterministic crash recovery after each stage - Multi-agent parallel: splits agentic stages across parallel workers - Context pruner: intelligent file selection to reduce token usage - Knowledge CLI: openexec knowledge index/symbols/status commands - Runner preflight: actionable error when CLI tool is missing Bug fixes: - Complete SQLite migration to modernc.org/sqlite (CGO_ENABLED=0 builds work) - Fix SanitizeJSON corrupting paths with // and arrays with objects - Fix parser channel buffer overflow - Fix blueprint callback wiring in Loop.runBlueprint() - Fix MCP shell commands in danger-full-access mode Documentation: - README rewritten to match implementation (removed misleading claims) - Architecture docs updated with accurate status markers - Deleted outdated/misleading docs, moved aspirational ones to docs/future/ - Removed 960 lines of aspirational documentation Dead code cleanup: - Removed standalone loop tests and dead code - Updated DCP integration tests for suggest-only mode
PreviousNext