v0.2.12

Agentao 0.2.12

0.2.12 ships the Headless Runtime v1 contract for ACPManager —
the non-interactive embedding target described in
docs/features/headless-runtime.md —
together with the Week 2 diagnostics surface, the Week 3
InteractionPolicy model, and the Week 4 lifecycle / recovery story.
A second, independent track lands in the same release: the
/crystallize evidence + feedback loop, which grounds skill-draft
generation in actual tool activity and adds a durable user-feedback
rewrite path.

Two breaking changes ride along: ACPManager.get_status() converges on
the typed list[ServerStatus] shape, and .agentao/acp.json no longer
accepts the bare-string form of nonInteractivePolicy. Migration
guidance is in the changelog entry and in developer-guide appendix E.7.

Highlights

• Public headless entry points: prompt_once, send_prompt, with
  single-active-turn concurrency pinned to AcpErrorCode.SERVER_BUSY.
  send_prompt_nonblocking is classified internal / unstable and is
  not part of the embedding contract.
• Typed status snapshot (ServerStatus): v1 core fields —
  server / state / pid / has_active_turn — extended additively
  by Week 2 diagnostics (active_session_id, last_error,
  last_error_at, inbox_pending, interaction_pending,
  config_warnings).
• ACPManager.readiness(name) / .is_ready(name): typed 4-valued
  gating signal ("ready" | "busy" | "failed" | "not_ready") — prefer
  over string-matching on state.
• InteractionPolicy + interaction_policy= override on
  send_prompt / prompt_once: per-call override wins over the server
  default (nonInteractivePolicy); None falls back to the default.
• Lifecycle & recovery: deterministic release order on every
  failure path, classify_process_death pure classifier, sticky
  is_fatal(name) / restart_count(name) surfaces,
  AcpServerConfig.max_recoverable_restarts (default 3).
• Daemon-style regression suite pinned in
  tests/test_headless_runtime.py::TestDaemonRegression.
• /crystallize evidence + feedback loop: SkillEvidence and
  SkillFeedbackEntry dataclasses extend SkillDraft with structured
  tool-activity grounding (tool calls, results, key files, workflow
  steps) and a durable feedback_history. collect_crystallize_evidence
  / render_crystallize_context produce the evidence block;
  feedback_prompt + FEEDBACK_SYSTEM_PROMPT drive user-feedback
  rewrites. suggest_prompt() and refine_prompt() gained an optional
  evidence_text= so all three prompts share the same grounding. Draft
  JSON is forward- and backward-compatible (legacy payloads load with
  empty evidence/history). Covered by
  tests/test_skill_crystallize_enhancement.py (15 tests).

See the CHANGELOG.md [0.2.12] entry for the
full Added / Changed breakdown including the two breaking changes.

Breaking Changes

1. ACPManager.get_status() returns list[ServerStatus] (not
   list[dict]). The legacy "name" dict key becomes
   ServerStatus.server. No get_status_typed() side channel and no
   permanent dict alias — this is a one-shot convergence.
2. nonInteractivePolicy bare-string config form is removed.
   .agentao/acp.json must use the structured object form —
   "nonInteractivePolicy": {"mode": "reject_all" | "accept_all"}.
   Bare strings raise AcpConfigError at config-load time.

Migration: see
developer-guide appendix E.7
(zh mirror exists).