AnaxiGraph 0.4.0

AnaxiGraph 0.4.0

AnaxiGraph 0.4.0 completes roadmap 4.2 by making its shared architecture model substantially more
useful to both humans and coding agents and by replacing shallow JavaScript-family analysis with
honest parser-backed evidence.

Highlights

• Adds Tree-sitter-backed JavaScript, JSX, TypeScript, and TSX analysis with explicit capability
  declarations, parser recovery diagnostics, source spans, visibility, modern module forms, and
  selected framework-level facts.
• Resolves relative imports, repository aliases, TypeScript baseUrl/paths/project references,
  workspace packages, package exports/imports, and entry points from indexed configuration evidence
  without executing target repository tooling.
• Persists IR v2 reference forms and resolution provenance while keeping old IR v1 snapshots
  readable and conservatively preserving ambiguous, unresolved, external, and computed edges.
• Delivers the shared Living Architecture Charter, actor-neutral guidance workflow, fresh-eyes
  counterfactual review, continuous post-change reassessment, and stable responsibility-map
  vocabulary completed in roadmap 4.2.
• Keeps normal deployment to one supervised service and one repository-scoped writer, with a
  bounded ten-tool MCP profile and five task-centered dashboard journeys.

Safety and compatibility

• AnaxiGraph remains read-only toward the repositories it analyzes.
• TypeScript facts are syntax evidence, not type-checker or runtime claims.
• Dynamic imports remain explicit unknowns; AnaxiGraph does not invent a target or recommend unsafe
  deletion when relationship evidence is incomplete.
• Existing indexes remain readable. A new scan records parser and IR v2 capability truth for current
  JavaScript-family files.

Validation

• 634 Python tests pass at 91.87% coverage.
• All 26 containerized browser contracts pass.
• The hardened sidecar, first-user workflow, deterministic self-analysis, release archives, clean
  wheel install, and parser-enabled container smoke pass.
• A retained 3,000-source-file JS/TS benchmark completes its cold scan in 4.825 seconds and analyzes
  exactly one source file per one-file incremental history frame on the recorded Linux runner.

Public verification

• Main CI, the
  multi-architecture container workflow,
  and the protected release workflow
  all completed successfully.
• PyPI 0.4.0 and the
  GitHub release publish matching wheel
  and source-distribution digests. The GitHub bundle also retains the dual-client plugin, SPDX SBOM,
  dependency-license inventory, checksums, release contract, and container digest.
• ghcr.io/hcekne/anaxigraph:0.4.0 supports linux/amd64 and linux/arm64 and resolves to
  sha256:3f7f3f2ed5f4bbe0b58996fe75d8fa14b7d032fe6d6d72c67133643ed1b45c3b.
  Pulling that exact digest reports AnaxiGraph 0.4.0, and its GitHub provenance verifies against
  the tagged container workflow.
• A no-cache install from public PyPI installs the exact pinned Tree-sitter dependencies and parses
  TSX with builtin-typescript-tree-sitter; release checksums and attestations verify independently.

Full comparison: v0.3.0...v0.4.0

AnaxiGraph 0.3.0

AnaxiGraph 0.3.0 turns the architecture index into a durable, autonomous semantic mapping system and adds evidence-backed coding-pattern intelligence.

Highlights

Semantic mapping that finishes

• Adds a durable host executor that keeps semantic work alive outside one coding-agent session.
• Recovers expired leases and stranded jobs, resumes an existing queue, and runs independent work in parallel.
• Keeps provider, model, and reasoning choices under caller control; no model name is built into module logic.
• Reuses unchanged semantic dossiers across executor, model, and package-version changes.

Architecture and pattern intelligence

• Builds and agent-reviews a semantic taxonomy without requiring a human approval step.
• Adds bounded graph operations and goal-specific architecture decisions for coding agents.
• Ships an extensible 128-pattern catalog spanning functions, modules, boundaries, workflows, reliability, data, concurrency, and subsystem architecture.
• Adds multi-level pattern evidence, autonomous pattern reviews, calibration, CLI queries, and dashboard exploration.

Findings people and agents can understand

• Rewrites built-in architecture findings to say what was observed, why it matters, what to do next, and when no change may be needed.
• Adds a versioned plain_language response shared by REST, MCP, and the dashboard while preserving exact machine evidence for tools and gates.
• Upgrades old stored finding copy when it is read, so an existing index becomes understandable without a destructive rebuild.
• Refactors the architecture rule evaluator below the repository's 500-line file ceiling and removes its accepted complexity debt.

Validation

• 466 Python tests and 15 browser contracts passed.
• 90.7% total coverage and 97.5% changed-code coverage.
• Verified on Python 3.11 and 3.12, Ubuntu and macOS package installs, hardened sidecar smoke tests, first-user performance, and multi-architecture container builds.

Full comparison: v0.2.0...v0.3.0

AnaxiGraph 0.2.0

AnaxiGraph 0.2.0 turns the architecture observatory into a practical local engineering companion: one command starts the dashboard and AnaxiMCP, and a connected coding agent can build durable semantic understanding with its own model context and tokens.

Highlights

• One-command local startup with external state, adaptive background history, clean shutdown, and optional browser opening.
• Explicit, idempotent Codex and Claude MCP connection with dry-run, project/user scopes, backups, and anaxigraph doctor.
• A shared Codex and Claude plugin for scope, impact, findings, and the complete agent-funded semantic dossier loop.
• Delta-driven Git history and canonical temporal facts that avoid re-reading and re-materializing unchanged files.
• Ranked top-20 attention queue plus a lossless, filterable findings ledger with evidence and lifecycle state.
• Honest relationship provenance for resolved, ambiguous, unresolved, and external references.
• Hardened generated Docker sidecars and a simplified four-step onboarding path.
• Reproducible wheel, source distribution, plugin bundle, SBOM, checksums, attestations, and versioned multi-architecture container release contracts.

Start locally

uvx anaxigraph up . --open --semantic agent --connect codex

For Claude Code, replace codex with claude. The durable Docker sidecar remains available through anaxigraph init . --start.

Safety boundary

The REST and MCP endpoints remain unauthenticated in 0.2.0. Keep them on loopback or behind an SSH tunnel; authenticated shared/team deployment is planned for a later phase.

Publication record

The protected workflow built and attested the release bundle successfully. Its initial OIDC exchange was rejected because the corresponding trusted publisher had not yet been registered in the existing PyPI project. With explicit maintainer authorization, the exact downloaded and checksum-verified workflow artifacts were uploaded through the emergency project-token path. PyPI's recorded SHA-256 digests match the attested GitHub assets, and a clean public uvx anaxigraph up journey passed. Register the documented release.yml / pypi publisher before the next version so routine releases use short-lived OIDC credentials.

Trusted publisher verification

The missing PyPI publisher was subsequently registered. Protected workflow run 32412357679 successfully exchanged the exact hcekne/anaxigraph / release.yml / pypi GitHub identity for a short-lived PyPI token, masked it, and discarded it without building or uploading an artifact. The temporary diagnostic branch allowance was removed; routine publication remains restricted to reviewed v* tags.