Every release, as published.

What's Changed

• chore(ci): pin distributed action/workflow examples to v0.2.0 by @lemduc in https://github.com/lemduc/arcade-agent/pull/28
• fix(metrics): make TurboMQ the sum of cluster factors (not the mean) by @lemduc in https://github.com/lemduc/arcade-agent/pull/26
• feat: polyglot multilang parse (roadmap #18 MVP) by @tuannx in https://github.com/lemduc/arcade-agent/pull/19
• feat(ci): enrich architecture drift report with actionable details by @tuannx in https://github.com/lemduc/arcade-agent/pull/30
• Exclude JVM test sources consistently by @tuannx in https://github.com/lemduc/arcade-agent/pull/32
• fix(deps): cap mcp below 2.0 — 2.0.0 removed mcp.server.fastmcp by @lemduc in https://github.com/lemduc/arcade-agent/pull/35
• chore: add MIT LICENSE by @lemduc in https://github.com/lemduc/arcade-agent/pull/36
• fix(reporting): keep comparison labels consistent in Mermaid graphs by @tuannx in https://github.com/lemduc/arcade-agent/pull/22
• docs: fix stale setup instructions in CLAUDE.md by @lemduc in https://github.com/lemduc/arcade-agent/pull/37
• fix(reporting): make comparison names and Mermaid node ids collision-safe by @lemduc in https://github.com/lemduc/arcade-agent/pull/38
• feat: architectural changelog (roadmap #10) + compare() split/merge correctness fix by @lemduc in https://github.com/lemduc/arcade-agent/pull/33
• feat(parser): add Rust support (roadmap #16b) — reland of #18 + #29 by @lemduc in https://github.com/lemduc/arcade-agent/pull/39
• fix(reporting): route every Mermaid emitter through one sanitizer by @lemduc in https://github.com/lemduc/arcade-agent/pull/40

Full Changelog: https://github.com/lemduc/arcade-agent/compare/v0.2.0...v0.3.0

Highlights

• Kotlin parser (#17) — tree-sitter-based structural parsing for .kt/.kts: classes, objects, companion objects, nested types, top-level functions, import aliases, and inheritance edges, with crash-safe iterative traversal and exact-FQN regression fixtures. Requires the [languages] extra.
• analyze — one-call async pipeline (#16) — runs ingest → parse → recover → smells → metrics in a single MCP call, off the event loop via a background thread. Returns per-stage session IDs, preserved even on partial failure (PartialAnalysisError), so agents never redo completed stages. This brings the MCP server to 18 tools.
• Cold-analysis performance — Python reference extraction rewritten around a static annotation-field table (#14, ~2.2× faster cold parse) and component-dependency lookups now use an on-demand membership index (#15, hot path 413ms → 1.9ms median).

Also in this release

• Atomic parse-cache writes (temp file + os.replace) — safe under concurrent analyses (#16).
• ToolDef.is_async so adapters can dispatch coroutine tools correctly (#16).
• max_tokens budgeting now applies to the composite analyze summary (#16).
• README: badge row, tool-count reconciliation, and docs link (#23, #24).

Docs

Full documentation now lives at https://arcade-agent.dev — getting started, all 18 MCP tools with parameters/returns/examples from real captured runs, and CI drift-baseline setup.

Install: pip install "arcade-agent[languages,mcp]"

Full changelog: https://github.com/lemduc/arcade-agent/compare/v0.1.1...v0.2.0

Completes Phase 3 and Phase 4 of the roadmap: arcade-agent now ships the full agent context-tool cluster, plus a packaged CI action for architectural drift detection.

This is also the first release whose tag makes lemduc/arcade-agent/actions/analyze@v0.1.1 resolvable.

Agent context tools (#12)

Four composable tools for token-efficient codebase understanding, all exposed over MCP:

• api_surface — extract public interfaces only, without implementation bodies or private members. Public is derived structurally (naming convention + external-dependent signal), never read from a field.
• diff_impact — blast-radius analysis for a set of changed files: affected entities, enclosing components, the transitive reverse-dependency closure, and potentially broken public contracts.
• context_for_task — given a natural-language task, return the minimal ranked set of files to read, each with a role (direct match / dependency / dependent / component sibling) and a reason.
• dependency_cone — upstream and/or downstream reachability from an entity or file, with depth control and per-direction node caps.

Backed by a new cycle-safe traversal helper (algorithms/traversal.py) that diff_impact was refactored onto.

CI analysis action (#11)

actions/analyze composite action and reusable workflow, with architecture drift detection against a committed baseline.

Architecture scoring (#5)

Strengthened the balanced architecture score.

Fixes

• _paths_match no longer falls back to bare-basename equality, so unrelated files sharing a name (src/auth/models.py vs src/billing/models.py) are no longer conflated. This previously skewed diff_impact's blast radius and let dependency_cone attribute one file's dependents to another.
• dependency_cone now resolves file targets in order (exact FQN, exact path, then path-suffix) and reports ambiguous: true with candidate_files when a target matches several distinct files, instead of silently unioning them.
• Pinned gh-action-pypi-publish to v1.14.0 (#10).

Full changelog: https://github.com/lemduc/arcade-agent/compare/v0.1.0...v0.1.1

Initial release of arcade-agent — a framework-agnostic tool library for software architecture analysis.