v0.17.0: Scoring Architecture Overhaul

Major Changes

Scoring Architecture (SDK → Server-side rules)

• SDK now emits factual flags (http_4xx, has_tool_calls, heartbeat, provider_error, etc.)
• New scoring_rules.py engine classifies records: interaction / heartbeat / system_error / infra_error / tool_intermediate
• Rules fetched from server, cached locally, with built-in defaults fallback
• Zero SDK upgrades needed when scoring rules change

Tool Chain Detection

• Proxy extracts tool_call names/inputs from streaming responses (OpenAI + Anthropic)
• Tool-only responses now stored in vault (no more empty output)
• aggregate_interactions() merges tool chains into single logical interactions

Heartbeat Detection

• Auto-detects heartbeat messages, flags them
• Excluded from scoring (reliability, latency, interaction count)
• Shown separately in atlast stats

Provider Error Detection

• Billing, quota, auth errors classified as system_error (not agent's fault)
• Does not affect agent Trust Score

Stats

• atlast stats now shows interaction breakdown with excluded counts
• 859 tests passed, +46 new tests, 0 regressions

v0.16.0

Chain linking, batch scheduler, show command, latency fix

v0.15.0 — Auto-setup proxy on init

atlast init now auto-starts proxy and configures LLM routing. Zero user intervention.

v0.14.0

Scanner removed. Use atlast proxy.

v0.13.0 — Automatic Scanner Service

Major

• atlast init now auto-starts a persistent scanner service — detects all OpenClaw agents, installs launchd (macOS) / systemd (Linux) / cron (fallback)
• API errors are now recorded — overloaded, auth failures, etc. show up as evidence
• Dedup protection — rescanning won't create duplicate records

Doctor

• Shows scanner service status (running/stopped)
• --fix auto-starts scanner if not running

No manual steps needed. Install → init → done.

v0.12.1 — Scanner Fix

Fix scanner model/token extraction + tool calls not marked incomplete

v0.12.0 — User-Friendly CLI + Doctor Command

What's New

🩺 atlast doctor — One-command environment check

• atlast doctor --fix checks Python, storage, identity, PyNaCl, server, disk
• Auto-fixes everything it can

🎨 User-Friendly CLI

• No more scary DID hashes or 'unverified' warnings
• Server failures show 'offline mode' not error stacktrace
• All commands use friendly short IDs

🐛 Bug Fixes

• Fix crash on repeated atlast init
• Python 3.9 compatibility
• Chain integrity false-positive display

v0.11.1 — Stress Test Bug Fixes

Bug Fixes (from deep user simulation testing)

• Fix: bip39_english.txt + dashboard_assets/ now included in PyPI package
• Fix: atlast discover crash (endpoints dict/list compat)
• Fix: atlast push now shows attestation ID instead of blank
• Fix: All ruff lint errors resolved (W293 + E702)
• Fix: VAULT_DIR import restored for test compatibility

Improvements

• Better error messages for atlast proxy/run suggesting wrap() alternative
• New docs: Stress Test Report + Issue Log

Full Test Results

9 frameworks × 105 tasks = 4,595 records. 100% chain integrity. See report for details.

v0.11.0 — Stress Test Validated

Changes

• Fixed batch.py sort order + chain hash in LangChain/CrewAI adapters
• Fixed 8 CLI crashes (backup-key, discover, dashboard, push output)
• Fixed all ruff lint errors (225 → 0)
• Stress tested with 9 agent frameworks, 4,595 records, 100% chain integrity
• See docs/STRESS-TEST-V3-FINAL-REPORT.md for full results

v0.10.0 — Query Engine, Dashboard, Vault V2

What's New

Query & Audit Engine

• atlast search, atlast audit, atlast trace, atlast timeline — full CLI query suite
• SQLite-powered local index, --json output for agent consumption

Local Web Dashboard

• atlast dashboard → problem-oriented UI at localhost:3827
• Auto-detects: error spikes, confidence drops, latency anomalies
• Activity tab with agent/error filters, full-text search, evidence chain trace

Vault V2 (Proxy Storage)

• Smart deduplication: stores only new content per conversation turn
• 13x storage reduction for long conversations
• Government-grade audit: SHA-256 hash verification of complete request/response

Critical Fixes

• EAS fail-closed: production never generates fake UIDs
• Auth enforced: unauthenticated batch uploads rejected
• Anchor coordinator: prevents concurrent anchor races

Packages

• PyPI: atlast-ecp==0.10.0 ✅
• npm: atlast-ecp-ts@0.3.0 ✅

Also

• Per-agent rate limiting
• MCP query tools for AI agent integration
• Demo data generator: atlast demo --days 60
• docs updated: sepolia → base-mainnet

Full Changelog: v0.9.0...v0.10.0