Skip to content

v0.1.0 — first SDK release

Choose a tag to compare

View all tags
@jbellsolutions jbellsolutions released this 26 Apr 10:01
· 24 commits to main since this release
v0.1.0
b98e304

Changelog

All notable changes to forge are documented here. The format follows
Keep a Changelog and forge adheres to
Semantic Versioning starting with v0.1.0.

[0.1.0] — 2026-04-26

First public release. Full 8-layer harness, MCP server, Claude Code skill +
slash command, GitHub-published.

Added

  • L0 kernel: AgentLoop, HookBus lifecycle (SessionStart /
    PreToolUse / PostToolUse / SessionEnd), Verdict (READY / WARNING /
    BLOCKED), AgentDef, provider-as-profile YAML loader.
  • L1 memory: ReasoningBank (RETRIEVE→JUDGE→DISTILL→CONSOLIDATE→ROUTE)
    with pluggable embedders (hash / OpenAI / Voyage / ONNX MiniLM); GitJournal
    for resumable sessions; ObsidianVault with wiki-link parsing, backlinks
    graph, and frontmatter; cross-project genome at ~/.forge/genome.json;
    ClaudeDir filesystem contract.
  • L2 tools: ToolRegistry with three-tier classification
    (mcp / computer_browser / cli), per-agent allow/deny enforcement; built-in
    ShellTool, FSReadTool / FSWriteTool (sandbox-anchored),
    HttpFetchTool, EchoTool, Obsidian tools, CLISubprocessTool family
    (ClaudeCodeTool, CodexCLITool, GeminiCLITool); real MCPClientPool
    over the official mcp SDK; ComposioAdapter over the Composio SDK;
    in-process MCP adapter for tests.
  • L3 healing: ErrorType taxonomy + regex classifier; CircuitBreaker
    with CLOSED / OPEN / HALF_OPEN transitions; CircuitRegistry;
    attach_healing(hooks) to wire breakers into the hook bus.
  • L4 swarm: Topology × Consensus enums; Spawner for solo /
    parallel-council / hierarchy; RoleCouncilSpawner injecting per-member
    system prompts (optimist / skeptic / pragmatist).
  • L5 skills: SkillStore with versioned SKILL.md + per-version
    runs.jsonl; EvalGate (MIN_SAMPLES=50, CONFIDENCE_MARGIN=0.05);
    autosynth proposer; SkillSearchIndex over body + outcomes.
  • L7 observability: TraceStore writing full-fidelity JSONL per session;
    Telemetry with per-profile pricing tables; OTelExporter (no-op when
    opentelemetry is not installed); read-only dashboard CLI.
  • Recursion: TraceAnalyzer, rule-based propose(), LLM-driven
    propose_with_llm() honoring the AutoAgent regularizer ("would this still
    help if this specific task vanished?"), # === FIXED ADAPTER BOUNDARY ===
    sentinel, ResultsLedger (TSV); recurse_once() orchestrator.
  • CLI: forge console script with subcommands doctor, run, recurse,
    recurse-loop, dashboard, skill, vault, heartbeat, mcp.
  • MCP server: 12 tools (forge_council, forge_recurse, forge_vault_*,
    forge_memory_*, forge_skill_*, forge_doctor, forge_dashboard).
  • Claude Code integration: slash command ~/.claude/commands/forge.md +
    skill ~/.claude/skills/forge/SKILL.md, both with auto-trigger triggers.
  • Schedulers: crontab, launchd, GitHub Actions templates under
    forge/scheduler/.
  • Examples: mock operator vertical, real-MCP operator_real (basic +
    tool-using council), recursion_demo for the full self-mod cycle.
  • Tests: 70 passing across kernel, providers, tools, healing, swarm,
    memory, skills, observability, recursion, CLI.

Notes

  • macOS Sonoma+ sets com.apple.provenance xattr (UF_HIDDEN) on files which
    causes Python 3.14 to skip .pth editable-install markers. Workaround:
    .venv/bin/forge is a bash wrapper exporting PYTHONPATH; the editable
    install still works for direct python -m forge.cli invocations.
  • The Anthropic + OpenRouter + Composio integrations are LIVE-VERIFIED:
    council emits real verdicts (~$0.10/run), recursion proposer emits
    regularizer-honoring diffs against real traces, Composio sees 1,048 apps,
    OpenRouter routes to DeepSeek-chat correctly.
Assets 6
Loading