Yagami 0.8.3

Changed

• Include the MCP ownership marker in the published PyPI description, and validate the README configured in package metadata before registry publication.

Application and security improvements

This is the first complete GitHub release of the accumulated 0.8 improvements. It includes the following changes in addition to the ownership metadata fix above.

The signed 0.8.0 and 0.8.1 candidates were withdrawn before publication while dependency and container security checks were completed; their immutable tags are retained. This release includes all intended 0.8 improvements.

Security

• Apply signed Debian security updates to the pinned container runtime, including the Perl and PCRE2 fixes reported by the release scan. Refresh that stage on every CI/release build, and enforce the high/critical container scan before creating immutable release tags.

• Update Mako to 1.4.3 in the reproducible Python and container locks to fix the newly disclosed Windows TemplateLookup path traversal advisory.

• Bound calculator input complexity and intermediate arithmetic work to reject resource-exhaustion attempts before expensive operations.

• Enforce a whole-operation HTTPS fetch deadline, validate every redirect, reject compressed responses before decoding, and omit sensitive URLs from error output.

• Refresh Python, JavaScript, and build dependencies to address reported advisories, including the newly disclosed source-map-js and selector-parser issues.

• Inspect caller-supplied tool schemas during policy preview so malicious tool

  descriptions cannot bypass context-risk quarantine.

• Preserve medical sensitivity inherited from prior messages and record

  governed tool outcomes using content-free names, status, safe error codes,

  and byte counts rather than arguments or results.

• Classify colloquial personal-health disclosures as medical data and preserve the established sensitivity across text follow-ups using content-free decision evidence.

• Route high-confidence first-person health disclosures directly to local medical handling, including when an image-generation phrase is also present, without exposing them to a semantic classifier or cloud image backend.

• Refresh transitive PostCSS, nanoid, and undici packages to versions without the current published npm advisories.

• Update the reproducible Python and container locks to cryptography 50.0.0, which resolves CVE-2026-69247.

• Use a safe consumer-health prompt by default; clinician assumptions now require an explicit trusted clinical purpose.

• Delete session-scoped memory vectors and reversible privacy tokens when a conversation is deleted.

Added

• Add 75 UI tests and 40 cross-browser checks covering recovery, races, dialogs, keyboard access, overflow, mobile layouts, and automated accessibility scans.

• Add native governed tool calling for Ollama and explicit, chat-format-gated

  tool support for llama.cpp, including a sensitive local-tool flagship demo.

• Add a schema-validated containment corpus covering identifiers, clinical

  history, secrets, retrieved-context contamination, tool policy, prompt

  injection, and benign cloud controls.

• Add a zero-configuration loopback-only Compose demo, comparison and

  configuration guides, an architecture diagram, social preview metadata, and

  maintainer-ready contributor and launch drafts.

Changed

• Redesign the control surface with consistent typography, responsive history and session-insight panels, accessible dialogs, starter actions, and clearer connection states.

• Upgrade paired React/React DOM and Vitest/coverage packages together; align packaging metadata and workflow action versions. JavaScript development requires supported Node 22.22.2+, 24.15.0+, or 26+.

• Make provider SDKs, PDF ingestion, and OS keyring support optional extras;

  improve yagami doctor with optional-capability status and actionable next

  commands.

• Prefer an installed Ollama model automatically in yagami demo; when none is

  available, present a clearly labeled policy-only fallback that never echoes

  user content as though it were an AI answer.

Fixed

• Preserve drafts through reconnects and failed operations, prevent stale responses from overwriting current sessions, and restore feedback state when persistence fails.

• Clear prior-session privacy and budget data during transitions and support response actions by keyboard and touch.

• Reject malformed chat image fields without terminating the WebSocket, allowing the next valid turn on the same connection.

• Add memory-saver, balanced, and performance profiles for explicit Ollama model retention and optional background preloading, with content-free readiness health and truthful UI phases for policy checks, model loading, and generation.

• Keep default Ollama embedding chunks below the configured all-minilm context without reducing the prior per-message retention ceiling.

• Hide abandoned empty chat sessions and allow completed streams to finalize cleanly when a client disconnects immediately after the done event.