Skip to content

Release notes — v0.2.1

Tag: v0.2.1
Package version: 0.2.1 (pyproject.toml, agent_llm_wiki_matrix.__version__, alwm version)

v0.2.1 strengthens the comparative campaign and longitudinal evaluation workflow.

Default just ci stays deterministic and offline; live backends and browser binaries remain opt-in (see docs/workflows/live-verification.md).

Compared to v0.2.0

v0.2.0 shipped six-axis fingerprints on benchmark and campaign manifests and longitudinal reporting improvements. v0.2.1 layers operator-facing docs (walkthrough, wiki, tracking, ADR), campaign_semantic_summary artifacts (rollups may show zero semantic cells when scoring stays deterministic), runtime / retry / judge-phase observability in manifests and Markdown, explicit MCP fixture scope (alwm browser run-mcp), and richer browser evidence + browser_realism.v1—without changing offline CI requirements.

Highlights

  • Campaign documentation — Consolidation and walkthrough improvements across workflows, wiki index, tracking checklist, and ADR for campaign orchestration.
  • MCP browser runner — Truthful partial classification in audits and architecture docs; alwm browser run-mcp supports fixture-backed runs (--scenario-id / --fixture). A full remote MCP protocol path is not claimed as complete.
  • Campaign semantic summary — Artifacts and reporting for semantic / repeat-judge rollups on campaign outputs (see docs/workflows/benchmark-campaigns.md).
  • Runtime observability — Benchmark and campaign execution surface timing, retry, and judge-phase summaries in manifests and Markdown where applicable (see docs/workflows/benchmarking.md).
  • Browser evidence — Richer fixture-backed BrowserEvidence: DOM excerpts, screenshot metadata, and extensions; browser_realism.v1 rubric dimensions for grounding, hallucination resistance, and source fidelity (see docs/architecture/browser.md).

Known boundaries

  • MCP remains fixture-backed / partial rather than a full remote protocol implementation.
  • Playwright remains optional, not part of default CI.
  • Browser realism is improved but still intentionally bounded by deterministic fixture mode in default pipelines.

Documentation map

Doc Purpose
CHANGELOG.md Versioned change list (section [0.2.1])
docs/workflows/campaign-walkthrough.md Campaign walkthrough (committed examples)
docs/workflows/longitudinal-reporting.md Longitudinal analysis
docs/architecture/browser.md Browser runners, MCP scope, evidence schema
docs/release-readiness.md Capability scope and opt-in paths

Upgrade / tagging

After pulling main at this release point:

git tag -a v0.2.1 -m "Release v0.2.1"
git push origin v0.2.1

Ensure pyproject.toml and src/agent_llm_wiki_matrix/__init__.py match 0.2.1 before publishing the tag.