v0.2.14

Agentao 0.2.14

Maintenance follow-up to 0.2.13 GA. The headline is a tool-call
resilience layer for local / open-source LLM backends whose
function-call output drifts from the OpenAI tool-call schema, so the
runtime can keep calling tools instead of crashing on minor
formatting violations. The same release also carries per-session
isolation polish, replay-schema drift gating in CI, the GitHub Actions
Node 24 prep, and disk-bounded log rotation.

No breaking changes. The full Added / Changed / Fixed / Packaging
breakdown is in the CHANGELOG.md [0.2.14] entry.

Highlights

• Tool-call repair / outbound sanitize subsystem (agentao/runtime/):
  three cooperating modules sit between the LLM and the tool dispatcher
  so models like GLM, DeepSeek, Kimi and local Ollama still land in a
  runnable shape.

  • arg_repair.py: conservative JSON repair for malformed function
    arguments — double-encoded JSON, fenced JSON, lenient Python
    literals, trailing commas, bracket imbalance. No punctuation
    guessing.
  • name_repair.py: fuzzy matching that maps near-miss tool names
    (CamelCase / suffix variants) onto a registered tool when the score
    is unambiguous.
  • sanitize.py: outbound scrubbing — replaces lone UTF-16 surrogates
    and re-emits canonical compact JSON for repaired arguments before
    assistant / tool messages reach strict provider APIs.

  Wired into chat_loop, tool_planning, and tool_runner; repair is
  invisible to the model itself (only logged), so prompt-cache
  behaviour is preserved. Coverage:
  tests/test_tool_argument_repair.py, tests/test_tool_name_repair.py,
  tests/test_outbound_sanitize.py, helper tests/support/tool_calls.py.
  Documented in developer-guide §5.1 ("Tool-call normalization").

• Per-instance background-task store: the background-agent registry
  is now per-Agentao instance rather than process-global, so
  concurrent ACP sessions and multi-tenant embeddings no longer leak
  handles across each other. Adds path-containment guards and prompt
  diagnostics surfacing.

• Replay JSON Schema export + CI drift gate: agentao/replay/
  ships an exported JSON Schema (tests/test_replay_schema.py) and
  fails fast in CI when agentao/replay/events.py evolves without the
  schema being regenerated.

• ToolRunner decomposition: the monolithic tool_runner was
  split into focused tool_planning, tool_runner (executor), and
  tool_result_formatter modules under agentao/runtime/. Public
  Agentao.chat() contract preserved.

• Bounded log rotation: agentao.log now uses
  RotatingFileHandler(maxBytes=10_000_000, backupCount=5) instead of
  a plain FileHandler, capping disk footprint at ~60 MB. The home-dir
  fallback (~/.agentao/agentao.log) gets the same rotation. Long-
  running sessions that previously grew the log into the hundreds of
  megabytes now self-cap.

Packaging / CI

• actions/upload-artifact v4 → v7, actions/download-artifact v4 →
  v8, actions/setup-python v5 → v6 — clears the GitHub Node 24
  default-runtime cutover (2026-06-02). (upload-artifact has no v8
  line yet; v7 is the current major.) setup-uv had already moved
  v6 → v7 in 0.2.14.dev0.
• Version pins refreshed from 0.2.13 to 0.2.14 across docs/ACP.md
  and the developer-guide install / version-check examples.

Release Summary

• Version: 0.2.14
• Git tag: v0.2.14
• GitHub release type: regular release, not pre-release
• Publish workflow: .github/workflows/publish.yml

Install

pip install -U agentao

Maintainer Checklist

1. Ensure agentao/__init__.py reports 0.2.14.

2. Run the smoke path:
   uv run python -m pytest tests/ && uv build && uv run twine check dist/*.

3. Push the tag:

   git push origin v0.2.14

4. Create the GitHub release for v0.2.14.

5. Leave Set as a pre-release unchecked so
   .github/workflows/publish.yml publishes to PyPI (the workflow's
   tag-vs-package version consistency check will validate alignment
   before upload).

Full Changelog

See CHANGELOG.md for the 0.2.14 entry.