What changed—and
why it matters.

Language-aware architecture guidance

AnaxiGraph became a more complete architecture sidekick: parser-backed JavaScript and TypeScript analysis joined Python, while the Living Architecture Charter and fresh-eyes reviews connected system understanding to practical guidance.

• Added Tree-sitter analysis for JavaScript, JSX, TypeScript, and TSX, including explicit capabilities, diagnostics, source spans, visibility, module forms, and framework facts.
• Resolved relative imports, aliases, TypeScript project configuration, workspace packages, and package entrypoints without executing target code.
• Introduced the Living Architecture Charter, actor-neutral guidance, continuous reassessment, and resumable fresh-eyes architecture review.
• Converged runtime ownership around one supervised service and one writer, and reduced the default agent surface to ten bounded MCP tools.
• Expanded release evidence to 634 tests, 91.87% measured coverage, 26 browser contracts, and a 3,000-file JavaScript/TypeScript benchmark.

Durable semantic and pattern intelligence

Semantic mapping moved from an experiment to a resumable workflow. Coding agents could fund the reasoning with their own context and tokens, write validated dossiers back to the AnaxiIndex, and turn those dossiers into evidence-backed pattern and architecture advice.

• Added a durable host executor with lease recovery, restart-safe progress, parallel work, and caller-controlled provider, model, and reasoning settings.
• Introduced agent-reviewed semantic taxonomy and fingerprint reuse so unchanged modules do not need to be understood again.
• Shipped a versioned pattern catalog, sparse candidate selection, review and calibration workflows, and pattern intelligence in the CLI and dashboard.
• Added bounded graph operations and goal-specific architecture decisions for coding-agent workflows.
• Refactored agent, semantic, and evidence responsibilities while raising the suite to 466 tests and 90.7% measured coverage.

A practical local engineering companion

The initial observatory became something a developer could actually keep beside a repository: one-command startup, explicit agent onboarding, scalable temporal storage, ranked findings, and reproducible releases.

• Added one-command local startup, guided Docker onboarding, clean shutdown, browser launch, and idempotent Codex and Claude MCP connection.
• Made history imports incremental, adaptive, cancellable, and reconstructable from canonical temporal facts instead of repeatedly materializing every file.
• Ranked findings by risk, churn, and blast radius, with a focused top-20 attention view and a separate complete ledger.
• Made relationship resolution honest by distinguishing resolved, ambiguous, unresolved, and external edges.
• Added a shared Codex/Claude plugin plus reproducible packages, checksums, SBOMs, attestations, and multi-architecture containers.

The first public architecture observatory

The first PyPI release established AnaxiGraph's foundation: read-only repository scanning, a versioned SQLite graph, temporal architecture replay, finding lifecycle, a local dashboard, MCP access, and the first agent-funded semantic workflow.

• Scanned repositories into structural modules and relationships while keeping the target repository read-only.
• Stored snapshots in the AnaxiIndex and replayed how the architecture changed through Git history.
• Exposed architecture, graph, history, findings, and work-envelope views in a zero-runtime-dependency dashboard.
• Added MCP access so a coding agent could query the same system model visible to a human in the dashboard.
• Separated deterministic structural facts from provider-neutral semantic claims and their provenance.