Changelog

All notable changes to Tractum (formerly claudeway). Format based on Keep a Changelog; versions adhere to Semantic Versioning. Dates are ISO 8601 (UTC).

Unreleased

No changes yet.

0.5.0 — 2026-10-17

Feature release — .curt linkage, self-verifying MCP receipts, and the claudeway → tractum rename.

Added

• .curt linkage (tractum.curt_link): a quorum receipt is the DECISION; the .curt transcript(s) the consensus was computed from are the EVIDENCE. build_curt_transcript renders a ConsensusResult as a conforming .curt v1 document (per-message agent provenance via the agent-context extension), link_receipt embeds transcript refs ({id, sha256}) INSIDE the signed payload — so the aggregate signature covers the linkage — emit_curt_decision emits receipt + evidence as one .curt artifact (the consensus-receipt extension), and verify_curt_link verifies the whole chain through the existing crypto paths. Unlinked payloads stay byte-identical to the 0.4.0 wire format.
• MCP reach_consensus always signs. Without a configured operator key the server now signs each receipt with a fresh ephemeral keypair instead of returning an unsigned receipt that verify_consensus would reject. The ephemeral public key rides in the response (public_key; metadata.operator_key: "ephemeral" marks the mode).
• TRACTUM_REQUIRE_PINNED_KEY=1 for deployments that must never ship an ephemeral-signed receipt: without a configured TRACTUM_SIGNING_KEY the call fails loudly with a clear error instead of signing with a throwaway key.
• Dry-run mode. TRACTUM_MCP_DRY_RUN=1 or a per-call "dry_run": true runs deterministic offline stub agents — per-agent keypairs, real statement signatures, zero API calls. The receipt self-identifies (swarm mcp-consensus-dry-run, stub answers, metadata dry_run) so it can never pass as a live decision.
• MCP verify_consensus accepts expected_agent_pubkeys and pins each agent's statement to a known key via verify_receipt_detailed (0.4.0's docs promised this; the tool now actually honors it).

Fixed

• MCP docs matched to the code (docs/mcp-server.md): specialists take instructions (not perspective); removed the phantom "sign": true parameter; removed the nonexistent CLAUDEWAY_MCP_TOKEN bearer-auth section (the HTTP transport has no built-in auth — now documented as such); documented the real env vars.
• .curt message-id collision: message ids used only the first 12 chars of the content-addressed conversation id, so two transcripts whose ids shared leading digest chars (~1-in-256) collided when merged into a decision artifact and emit_curt_decision silently dropped evidence. Ids now embed the full conversation id (regression test added).

Changed

• Renamed: claudeway → tractum. The Python package, module, and CLI are now tractum.
• Imports: from claudeway import ... → from tractum import ...; module directory claudeway/ → tractum/; console script claudeway-mcp → tractum-mcp.
• Env vars: CLAUDEWAY_SIGNING_KEY → TRACTUM_SIGNING_KEY, CLAUDEWAY_TEST_RELAY → TRACTUM_TEST_RELAY, CLAUDEWAY_TEST_LANGGRAPH → TRACTUM_TEST_LANGGRAPH, CLAUDEWAY_RUN_LIVE → TRACTUM_RUN_LIVE, and CLAUDEWAY_DEMO_RELAY → TRACTUM_DEMO_RELAY.
• PyPI: the package publishes as tractum (the claudeway name never shipped to PyPI; 0.2.0–0.4.0 were GitHub-only).
• Wire-format type tags are UNCHANGED for backward compatibility: receipts still carry claudeway.consensus.v1 / claudeway.consensus.v2 / claudeway.agent-statement.v1 payload types and the claudeway.receipt.v1 envelope type — all shipped publicly since v0.2.0/0.3.2, and the verifier accepts them forever. The W3C VC context URL is likewise unchanged. Nostr d-tags and the client tag ARE renamed (claudeway-consensus → tractum-consensus, claudeway-transparency → tractum-transparency, client claudeway → tractum): these are self-asserted metadata on self-signed events, not verification inputs, so old events stay valid and old relays keep working. Existing receipts and .curt linkages verify without migration.
• Newly written .curt transcripts use generator: "tractum" and tractum--prefixed content-addressed conversation ids; existing ids are opaque and remain valid (verification is hash-based).
• License unchanged: BUSL-1.1, Licensor Jordan Newell.

Migration (from claudeway 0.4.x)

pip uninstall claudeway   # if installed from the repo under the old name
pip install tractum       # PyPI at launch; until then install from the repo

# old                          # new
from claudeway import Swarm    from tractum import Swarm

Old	New
from claudeway import ...	from tractum import ...
claudeway-mcp	tractum-mcp
CLAUDEWAY_SIGNING_KEY	TRACTUM_SIGNING_KEY
CLAUDEWAY_TEST_RELAY	TRACTUM_TEST_RELAY
CLAUDEWAY_TEST_LANGGRAPH	TRACTUM_TEST_LANGGRAPH
CLAUDEWAY_RUN_LIVE	TRACTUM_RUN_LIVE
CLAUDEWAY_DEMO_RELAY	TRACTUM_DEMO_RELAY
claudeway.consensus.* / claudeway.agent-statement.v1 / claudeway.receipt.v1 type tags	unchanged — backward-compat wire format

0.4.0 — 2026-10-01

Feature release — per-agent signed statements and k-of-n quorum receipts. The receipt stops being one operator's claim about what the agents said and becomes a countable multi-sig: every agent signs its own stance, and anyone can verify the signatures individually.

Added

• Per-agent signed statements. Every agent in a Swarm now gets a signing keypair (pinned via AgentConfig.signing_key, otherwise an ephemeral key is provisioned per swarm) and signs a claudeway.agent-statement.v1 statement over its stance — answer, confidence, model, role, task, round. Statements ride on AgentResponse.statement and survive dict round-trips through Task.result.
• v2 receipt payload (claudeway.consensus.v2). When statements are present, ConsensusReceipt.from_result emits a payload carrying the statements plus a quorum threshold k (default: unanimous). The operator's aggregate signature covers the agent signatures too. v1 receipts are produced unchanged when no statements exist — the verifier handles both forever.
• verify_receipt_detailed(receipt, expected_operator_pubkey="", expected_agent_pubkeys={}) — full verification report: aggregate signature, per-statement signature/hash checks, statement↔response cross-checks (response-mismatch:answer/confidence), cross-task replay detection (task-id-mismatch), operator and per-agent key pinning (key-mismatch), and the quorum count. Returns a ReceiptVerification with per-statement reasons.
• Statement helpers — SignatureBackend.sign_statement / verify_statement (implemented once on the ABC, inherited by Ed25519 and ML-DSA-65 backends; mixed-algorithm statements verify per statement), plus module-level sign_statement / verify_statement / statement_hash.
• MCP verify_consensus upgrade — the tool now accepts optional expected_operator_pubkey / expected_agent_pubkeys and returns the full statement-by-statement verdict, not just a boolean.

Security

• New adversarial tests for the statement attack surface: operator fabrication of statements under fresh keys (caught by per-agent pinning), statement tampering inside a signed receipt (aggregate hash guard), cross-task statement replay (task-id-mismatch), statement/response desync (response-mismatch), and stolen-agent-key isolation.
• Honest threat-model note: a signature proves self-consistency, not key custody. Without per-agent pinning, a malicious operator can mint self-consistent statements under fresh keys — pinning is what makes fabrication detectable. See docs/THREAT-MODEL.md.

Compatibility

• v1 receipts verify exactly as before; verify_receipt semantics are unchanged. Responses without statements keep their v1 dict shape byte-for-byte, so existing payload hashes and transparency-log anchors are unaffected.

0.3.2 — 2026-07-26

Maintenance release — CI/docs only, no package behavior change.

Changed

• Publish workflow uses OIDC trusted publishing (no API tokens). PyPI trusted-publisher configuration corrected; the GitHub Actions workflow now mints a short-lived OIDC token at publish time. Removes the need for any long-lived PYPI_API_TOKEN secret on the repo.
• README ships with a canonical Open Graph image and banner. Social previews (Twitter/X, LinkedIn, Slack, Discord) now render the Claudeway card instead of a blank tile.

0.3.1 — 2026-07-26

Hotfix release. Unblocks pip install claudeway[benchmark] on Windows.

Fixed

• [benchmark] extra no longer pulls litellm. The litellm>=1.18 dep transitively requires a Rust build toolchain that fails on Windows. The killer demo (examples/killer_demo.py) now imports litellm lazily with a try/except — token-cost tracking is degraded on Windows when litellm is absent (it was already unreliable on Windows + litellm 1.18+ per the inline comment), but the demo itself runs. Full token tracking on Linux/macOS is restored by pip install litellm separately. The Docker compose appliance (which is Linux) installs litellm in its Dockerfile.

0.3.0 — 2026-07-25

The "everyone has agents, nobody has agreement" release. Claudeway ships as the coordination layer that LangGraph, CrewAI, Microsoft Agent Framework, and Buzzrooms all lack, with post-quantum-ready signatures and a reproducible benchmark showing Claudeway beats both single Claude and CrewAI on hard questions.

Added

• LangGraph adapter (claudeway.adapters.langgraph): drop Claudeway consensus into a StateGraph as a single node. Two entry points — make_consensus_node(swarm) and build_consensus_graph(swarm). Lazy import; install with pip install claudeway[langgraph].
• Microsoft Agent Framework adapter (claudeway.adapters.maf): Claudeway as a MAF workflow + agent. MAF is Microsoft's unified successor to AutoGen + Semantic Kernel. Install with pip install claudeway[maf].
• CrewAI adapter (claudeway.adapters.crewai): Claudeway consensus as a @tool and as a Flow. Inverts the killer demo — a CrewAI crew can call Claudeway for agreement. Install with pip install claudeway[crewai].
• Post-quantum signature backend (claudeway.signing_pq.MLDSABackend): ML-DSA-65 / FIPS 204 signatures. Same receipt, same canonical payload hash — swap in at the SignatureBackend ABC. Pure-Python (dilithium-py), installs everywhere with no native build tools. Install with pip install claudeway[pq].
• Consensus transparency log (claudeway.transparency): RFC 6962-style append-only log of consensus results, anchored to Nostr on a cadence. Tamper-evident beyond the per-receipt signature.
• Streaming consensus events: OnEvent callback hook surfaces AgentCompleted and ConsensusResolved events as they happen. Observable from any adapter (MAF, LangGraph, custom host).
• Adversarial security test suite (tests/test_adversarial.py): 20 tests across 7 attack classes — signature forgery, receipt tampering, transport malleability, key reuse, replay, downgrade, and timestamp manipulation. Documented in docs/THREAT-MODEL.md.
• Threat model (docs/THREAT-MODEL.md): explicit attacker assumptions, what the signatures defend against, what they don't.
• Docs site (jordannewell.github.io/claudeway): mkdocs-material + mkdocstrings API reference. Auto-builds on push to main via the docs.yml workflow.
• Community policies: SECURITY.md (private vuln disclosure + SLA), CONTRIBUTING.md (branch model, commit rules, test commands), CODE_OF_CONDUCT.md (Contributor Covenant 2.1), issue templates.
• Killer benchmark (examples/killer_demo.py): same hard question → single Claude vs CrewAI vs Claudeway, blind LLM-judge scoring. Claudeway scores 18/20 (range 17–19) vs CrewAI 13/20 vs single Claude 11/20. Mann-Whitney U for non-parametric significance.

Fixed

• Stable signing key in adapters: the LangGraph/CrewAI/MAF adapters now reuse a stable Ed25519 key across runs instead of regenerating per call, so signatures verify across runs and receipts are diffable.
• Tolerant structured-output parser: <answer> now carries the full response, parser falls back to "everything after <answer>" when small models truncate before </answer>. Latent ConsensusResult.to_dict() crash on dict-form responses fixed.
• Substantive <answer> contract: prompt explicitly requires the full multi-paragraph analysis inside <answer>, not a one-word verdict.
• Agent.think no longer passes tools=None as a kwarg (Anthropic SDK rejects it; previously suppressed when no tools were registered).
• asyncio.run_coroutine_threadsafe self-deadlock in adapter async→sync bridges: replaced with a fresh daemon worker thread + own loop. Works whether the caller is sync, async, or already in a thread.
• Dropped dead secp256k1 fallback in transports._nostr_sign. The path called a lib.schnorr_sign API that never existed in coincurve; coincurve ships prebuilt wheels on all targets, so the CFFI fallback was dead complexity.
• [nostr] extra mismatch in pyproject.toml: was pynostr, code imports coincurve. Fixed.

Changed

• README rewritten around the killer demo and LangGraph adapter.

0.2.0 — 2026-07-25

Added

• Verifiable multi-agent consensus for Claude — initial public release.
• Real consensus — WeightedVote (cheap, default) and Debate (revision round when agents disagree, cost-guarded early-exit).
• Real decomposition — Coordinator parses a JSON plan and runs specialists in dependency-respecting parallel order.
• Concurrent execution — N-agent swarms issue all Claude calls in parallel via asyncio.gather.
• Ed25519 signed receipts — ConsensusReceipt, Ed25519Backend, swappable SignatureBackend ABC (post-quantum-ready).
• Three transports — JSON receipt, W3C Verifiable Credential, Nostr NIP-78 event (kind: 30078).
• MCP server (claudeway-mcp): exposes reach_consensus and verify_consensus as MCP tools for any MCP-capable agent.
• Buzz adapter (Nostr transport): to_nostr_event produces spec-correct events. Verified under nak (reference Nostr CLI); end-to-end tested against nak serve. Demo at examples/buzz_consensus_demo.py.
• Lean extras model — base install is 4 dependencies. [mcp], [nostr], [runner], [persist], [dev] are opt-in.
• CI — matrix (Ubuntu + Windows) × Python 3.11/3.12/3.13, ruff, pytest, wheel build + contents verification.
• LICENSE (MIT, Jordan Newell 2026), README + QUICKSTART rewritten around the coordination wedge.

0.1.0 — 2026-07-24

Added

• Initial pre-release platform implementation (Next.js dashboard, FastAPI runner scaffolding, port management). Superseded by the 0.2.0 rewrite around the SDK + consensus wedge. The dashboard and multi-tenant runner are deferred — see docs/DEPRECATION.md.