v0.2.16

Agentao 0.2.16

Maintenance release that closes the embedded-harness M2 / M3
milestones. Agentao(...) is now a pure-injection construction
surface — nothing in the constructor implicitly reads os.environ,
Path.home(), Path.cwd(), or <wd>/.agentao/*.json unless the
caller routes through agentao.embedding.build_from_environment().
CLI and ACP both go through the factory, so end-user behavior is
unchanged. Embedded hosts get a deterministic, side-effect-free
construction surface plus an await agent.arun(...) async path,
opt-in replay / sandbox / bg_store, and a one-cycle
soft-deprecation for Agentao() constructed without
working_directory= (TypeError in 0.3.0).

No breaking changes for CLI / ACP users. The full Added / Changed /
Deprecated breakdown is in the CHANGELOG.md [0.2.16] entry.

Highlights

• agentao.capabilities — FileSystem / ShellExecutor protocols
  with LocalFileSystem / LocalShellExecutor defaults that are
  byte-equivalent to today's behavior. File, search, and shell tools
  route through the capability layer, so embedded hosts can swap in
  Docker exec, virtual filesystems, audit-log shells, or remote
  runners without monkey-patching subprocess / pathlib.
• agentao.embedding.build_from_environment(...) factory.
  Single chokepoint for every implicit .env / .agentao/permissions.json
  / .agentao/mcp.json / .agentao/replay.json / .agentao/sandbox.json
  / cwd / home read that used to live in Agentao.__init__. CLI
  (agentao/cli/...) and ACP (session_new, session_load) both
  call through it; subsystem fallbacks are now dead code from their
  perspective.
• Pure-injection construction. Agentao.__init__ accepts
  explicit llm_client, logger, memory_manager, skill_manager,
  project_instructions, mcp_manager, filesystem, and shell
  injections. When skill_manager or project_instructions is
  injected, the auto-discovery / <wd>/AGENTAO.md disk-read paths
  are skipped entirely.
• Async public surface — Agentao.arun(...). Bridges the sync
  chat pipeline through loop.run_in_executor, so async hosts can
  await agent.arun(...) without rolling their own thread bridge.
  Cancellation, replay, and max_iterations behave identically
  across chat() and arun(). Internal runtime stays sync (no
  AsyncTransport).
• Opt-in replay / sandbox / bg_store. Three new keyword-
  only kwargs on Agentao.__init__. Each defaults to None =
  fully disabled; embedded hosts that don't ask for the feature pay
  zero cost. Notable: when bg_store=None, the run_in_background
  field is schema-level removed from sub-agent tool definitions
  (not expose-then-error), so the LLM cannot be tempted to call a
  disabled feature and ACP / OpenAI catalogs do not advertise it.
  CLI continues to construct all three through the factory.
• Subsystem constructors no longer fall back to environment.
  LLMClient(api_key=, base_url=, model=), MemoryManager(project_root=, global_root=),
  PermissionEngine(project_root=), and load_mcp_config(project_root=)
  raise TypeError if their required arguments are missing — no
  more silent reads of os.environ, Path.cwd(), or Path.home().
  Direct construction in embedded-host or test code may need a
  one-line update; CLI / ACP behavior is unchanged because the
  factory has been passing these explicitly since the M2 cut.
• Sub-agent credential split bug fixed. Sub-agent construction
  in agentao/agents/tools.py no longer re-reads provider env vars
  ({PROVIDER}_API_KEY / _BASE_URL). Children inherit the
  parent's already-resolved LLM config, so a mid-run env mutation
  cannot create a credential split between parent and sub-agent.

Deprecation (one-cycle warning before 0.3.0)

• Agentao() constructed without working_directory= emits a
  DeprecationWarning pointing callers at the factory or the
  explicit working_directory=Path(...) form. The argument becomes
  a required keyword in 0.3.0. Internal docs, examples, and tests
  have already been migrated; pytest -W error::DeprecationWarning tests/
  is clean on this release.

Tests

• tests/test_agent_subsystems_optional.py (new) — pins the
  replay_config=None / sandbox_policy=None / bg_store=None
  contract: bare construction, schema removal of run_in_background,
  short-circuited /agent CLI subcommands, factory parity.
• tests/test_no_subsystem_fallback_reads.py (new) — trip-wires
  Path.cwd() and Path.home() while constructing LLMClient /
  MemoryManager / PermissionEngine / load_mcp_config, so any
  re-introduced fallback surfaces as a loud failure rather than
  silent steering into the user's home directory.
• Existing test suite (tests/test_acp_*, tests/test_bg_task_store.py,
  tests/test_mcp_config.py, tests/test_per_session_cwd.py,
  tests/test_permissions.py, etc.) updated to pass explicit
  project_root= / user_root= / capability arguments.

Release Summary

• Version: 0.2.16
• Git tag: v0.2.16
• 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 0.2.16 (drop the .dev0
   suffix from 0.2.16.dev0).

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 v0.2.16

4. Create the GitHub release for v0.2.16.

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).

6. Close embedded-harness GitHub issue #15 (Issue 9, opt-in
   Replay / Sandbox / BackgroundTaskStore) once the
   release is on PyPI. Issues #7-#13 (capability protocols,
   factory, fallback deletion, arun, soft deprecation) were
   already closed when c1f802f and b3b403f landed and do
   not need to be re-closed. Issues #14 (Issue 8, hard break)
   and #16-#18 (M5: MemoryStore / MCPRegistry /
   docs/EMBEDDING.md) stay open and are scheduled for 0.3.0
   and follow-ups.

Full Changelog

See CHANGELOG.md for the 0.2.16 entry.