SmartPerfetto v1.11.0

[1.11.0] - 2026-09-16

Added

• Evidence-backed final delivery: conclusion claims are checked against retained
  execution captures and one bounded, no-tool semantic review of the full answer.
  Web, CLI, HTML reports and snapshots show each claim's verification result,
  including failed and unchecked claims, while keeping the conclusion body intact.
• CLI session markers (✓, ~, !, ✗) and deliveryVerdict in JSON/NDJSON
  complete events. CLI turns persist their evidence bundle so exported reports
  keep claim verification.
• AGENT_MAX_RUN_TIMEOUT_MS / OPENAI_MAX_RUN_TIMEOUT_MS: the OpenAI runtime
  extends its initial deadline while tools return data or the provider is still
  producing output, and reserves one no-tool call to deliver a limited
  partial / timeout conclusion from returned data.

Changed

• Startup analysis collects scheduling, blocking, CPU frequency and main-thread
  state evidence against explicit investigation requirements, and final reports
  keep exact values or mark rounded ones.
• All five runtimes share one conclusion declaration protocol, and MCP tool
  descriptions are loaded from strategy templates.
• Missing-credential errors, smp doctor and runtime health include setup
  guidance. The CLI Provider store stays separate from a source Web backend
  unless both use the same SMARTPERFETTO_BACKEND_DATA_DIR.

Removed

• The Claude Agent SDK runtime no longer falls back to a local Claude Code login,
  and ANTHROPIC_BASE_URL alone no longer counts as configured, including for
  gateways that previously worked without a key. Analysis now stops early with
  setup guidance instead. To migrate, add a provider under
  AI Assistant Settings → Providers, or set ANTHROPIC_API_KEY /
  ANTHROPIC_AUTH_TOKEN (plus ANTHROPIC_BASE_URL for a compatible gateway),
  Bedrock or Vertex configuration in backend/.env, ~/.smartperfetto/env or
  the Docker env.

Fixed

• Delivery diagnostics name semantic review failures (answer text that differs
  from its declared claim, undeclared assertions) instead of reporting an
  unclassified check failure.
• CLI source supplements rebuild the turn evidence fingerprint, so later report
  export no longer reports valid evidence as mismatched.
• Qoder provider connection tests accept a local qodercli login, matching the
  runtime.

Known issues

• With real providers, long startup and scrolling reports are frequently
  delivered as partial: the semantic review rejects rounded values written
  without an approximation marker and assertions without a matching claim
  declaration. The answer is still delivered with the failing checks named;
  treat it as unverified.