Skip to content

Releases: clearclown/tirami

Forge v0.3.0 — Launch-ready

09 Apr 16:04

Choose a tag to compare

Codename: Launch-ready. First release verified end-to-end on real hardware with a real GGUF model.

  • 426 tests passing (Rust workspace)
  • 27 pytest (Python SDK)
  • 16 SPEC-AUDIT (forge-economics)
  • 686 tests (forge-mesh production runtime)
  • 95/95 verify-impl.sh conformance assertions GREEN
  • 43 match / 0 drift theory ↔ implementation audit

What's in v0.3.0

See CHANGELOG.md for the full Phase 1-12 history.

Highlights since v0.1.0:

  • L2/L3/L4 Rust workspace — forge-bank (finance), forge-mind (self-improvement), forge-agora (marketplace)
  • Signed reputation gossip with outlier-resistant weighted-median consensus
  • Collusion detection (tight cluster + volume spike + Tarjan SCC round-robin)
  • Persistent state (BankServices / Marketplace / ForgeMindAgent survive restarts)
  • Prometheus /metrics (11 series including collusion scores)
  • Bitcoin OP_RETURN anchoring of the trade Merkle root
  • Real token-by-token streaming via SSE (~2-4 ms inter-chunk on Apple Silicon Metal)
  • OpenAI tools / function calling (drop-in compat for Cursor / Claude Code / ChatGPT desktop)
  • zkML / BitVM / federated training scaffolds (Phase 13 will plug in real backends)
  • One-command demo: bash scripts/demo-e2e.sh

Quickstart

git clone https://github.com/clearclown/forge && cd forge
bash scripts/demo-e2e.sh

See README.md and docs/compatibility.md.

Empirical validation (2026-04-09, Apple Silicon Metal)

✓ node PID 26860, model loaded after 1×2s
✓ 3 real chat completions via SmolLM2-135M on Apple Silicon Metal
✓ balance: contributed=41 CU, reputation=0.5 (DEFAULT_REPUTATION)
✓ PortfolioManager.tick() → action=lend
✓ RiskModel VaR 99%: 692 CU (DEFAULT_RATE=0.02, LGD=0.50, σ=2.33)
✓ Marketplace.find() returned 1 matches
✓ ForgeMindAgent initialized with EchoMetaOptimizer
✓ improve(1) → decision=Revert (correct)
✓ trust_penalty=0.0 (below MIN_TRADES_FOR_ANALYSIS)
✓ forge_trade_count_total 3
✓ merkle_root: 094f69461a6b339c75b7455b90c2c146943261c41f9023173fb142a91864ffb1
✓ Bitcoin OP_RETURN payload: 6a284652474501000000094f69461a6b339c75b7455b90c2c146943261c41f9023173fb142a91864ffb1

All Phase 1-12 endpoints verified with live data.

Known limitations

  • crates.io publish is deferred to Phase 13. The forge-core and forge-cli crate names are squatted on crates.io by unrelated parties. Install via: cargo install --git https://github.com/clearclown/forge forge-cli
  • PyPI: install forge-sdk==0.3.0 and forge-cu-mcp==0.3.0 (upload pending at release time — check PyPI if not yet available)
  • Single-operator deployments only — no production SLAs, no external security audit. Run for research, curiosity, or small-scale use. Full production is Phase 13+ scope.

License

MIT — see LICENSE.

forge-sdk 0.3.0

09 Apr 16:04

Choose a tag to compare

forge-sdk v0.3.0 — Python SDK covering all Phase 1-12 endpoints. 38 methods total, 27 pytest tests.

Phase 8 L2/L3/L4 additions (20 new methods):

  • Bank (8): bank_portfolio, bank_tick, bank_set_strategy, bank_set_risk, bank_list_futures, bank_create_future, bank_risk_assessment, bank_optimize
  • Agora (7): agora_register, agora_list_agents, agora_reputation, agora_find, agora_stats, agora_snapshot, agora_restore
  • Mind (5): mind_init, mind_state, mind_improve, mind_budget, mind_stats

Install: pip install forge-sdk==0.3.0

Companion MCP server: pip install forge-cu-mcp==0.3.0

forge-cu-mcp 0.3.0

09 Apr 16:04

Choose a tag to compare

forge-cu-mcp v0.3.0 — MCP server exposing Forge to Claude Code / Cursor / ChatGPT desktop. 36 tools total, 20 new Phase 8 L2/L3/L4 tools.

Install: pip install forge-cu-mcp==0.3.0

Companion SDK: pip install forge-sdk==0.3.0

Forge v0.1.0

02 Apr 14:54

Choose a tag to compare

Forge v0.1.0 Pre-release
Pre-release

Initial public MVP release of Forge.

Current reference implementation:

  • Encrypted seed/worker inference over Iroh QUIC
  • Loopback-first HTTP API with optional bearer protection
  • Local CU-native ledger, persisted snapshots, and settlement export
  • Capability handshake and topology planning groundwork

Not included in this release:

  • Real partial-layer loading by LayerRange
  • Forward-driven split inference runtime
  • Topology-driven multi-stage execution
  • Graceful fallback and reconnection behavior for split inference

This release should be treated as a public prerelease baseline for stabilization work.