@tencent-rtc/trtc-agent-skills 0.1.0

package/skills/trtc-topic/tests/test_stop_require_apply.py

@@ -0,0 +1,99 @@
+"""Tests for stop_require_apply_evidence.py — Stop hook.
+
+Stop hook runs when the AI's turn ends. We use it to refuse to end on
+states that mean "code was written but not validated":
+
+    code_written  → exit 2 (block) — apply.py never ran
+    apply_failed  → exit 2 (block) — apply rejected the code, AI must regenerate
+
+Other states pass:
+    not_started, slice_read, apply_passed, all_done, queue-not-initialised,
+    session-missing.
+"""
+from __future__ import annotations
+
+import subprocess
+import sys
+from pathlib import Path
+
+GUARDRAILS_DIR = Path(__file__).resolve().parents[1] / "guardrails"
+STOP_SCRIPT = GUARDRAILS_DIR / "stop_require_apply_evidence.py"
+
+STATE_MACHINE_DIR = Path(__file__).resolve().parents[1] / "scripts" / "lib"
+sys.path.insert(0, str(STATE_MACHINE_DIR))
+import state_machine  # noqa: E402
+sys.path.pop(0)
+
+
+def _run(session_path: Path) -> subprocess.CompletedProcess:
+    return subprocess.run(
+        [sys.executable, str(STOP_SCRIPT)],
+        text=True,
+        capture_output=True,
+        env={"TRTC_SESSION_PATH": str(session_path), "PATH": "/usr/bin:/bin"},
+    )
+
+
+class TestStopHook:
+    def test_allows_when_session_missing(self, tmp_path):
+        result = _run(tmp_path / "missing.yaml")
+        assert result.returncode == 0
+
+    def test_allows_when_queue_not_initialised(self, session_factory):
+        path = session_factory()
+        result = _run(path)
+        assert result.returncode == 0
+
+    def test_allows_in_not_started(self, session_factory):
+        path = session_factory()
+        state_machine.init_queue(path)
+        result = _run(path)
+        assert result.returncode == 0
+
+    def test_allows_in_slice_read(self, session_factory):
+        path = session_factory()
+        state_machine.init_queue(path)
+        state_machine.advance(path, "mark_slice_read")
+        result = _run(path)
+        assert result.returncode == 0
+
+    def test_blocks_in_code_written(self, session_factory):
+        path = session_factory()
+        state_machine.init_queue(path)
+        state_machine.advance(path, "mark_slice_read")
+        state_machine.advance(path, "mark_code_written")
+        result = _run(path)
+        assert result.returncode == 2
+        # Helpful error mentions apply.py and slice id
+        assert "apply" in result.stderr.lower()
+        assert "login-auth" in result.stderr
+
+    def test_blocks_in_apply_failed(self, session_factory):
+        path = session_factory()
+        state_machine.init_queue(path)
+        state_machine.advance(path, "mark_slice_read")
+        state_machine.advance(path, "mark_code_written")
+        state_machine.advance(path, "mark_apply_failed")
+        result = _run(path)
+        assert result.returncode == 2
+        assert "apply_failed" in result.stderr or "regenerate" in result.stderr.lower()
+
+    def test_allows_in_apply_passed(self, session_factory):
+        path = session_factory()
+        state_machine.init_queue(path)
+        state_machine.advance(path, "mark_slice_read")
+        state_machine.advance(path, "mark_code_written")
+        state_machine.advance(path, "mark_apply_passed")
+        result = _run(path)
+        assert result.returncode == 0
+
+    def test_allows_in_all_done(self, session_factory):
+        path = session_factory(confirmed_plan=["conference/login-auth"])
+        state_machine.init_queue(path)
+        state_machine.advance(path, "mark_slice_read")
+        state_machine.advance(path, "mark_code_written")
+        state_machine.advance(path, "mark_apply_passed")
+        state_machine.advance(path, "mark_user_confirmed")
+        # Now in all_done
+        result = _run(path)
+        assert result.returncode == 0

package/skills/trtc-topic/tests/test_topic_skill_invariants.py

@@ -0,0 +1,130 @@
+"""Topic skill SKILL.md invariants — guards the post-redesign shape.
+
+After the topic-skill redesign (2026-05-15), `topic/SKILL.md` was reduced
+from 512 lines to ~390. Four structural invariants survived from that
+work and need to stay true:
+
+1. The "Apply Evidence Block" section was deleted — it instructed AI to
+   run its own grep/ls/tsc, duplicating apply.py and giving AI an obvious
+   bypass path.
+2. The State Machine Guide was split out to
+   `topic/scripts/STATE-MACHINE-GUIDE.md` and topic/SKILL.md just
+   references it.
+3. The "Calling apply" section is compact — the full request schema lives
+   in `apply/SKILL.md` Phase 0; topic only references it.
+4. topic/SKILL.md size stays within budget so we don't silently re-inline
+   the things we split out.
+
+Originally this file also enforced the §3.5 cross-skill handoff
+convention ("must use Skill tool, never Read"). That convention was
+walked back on 2026-05-18: the runtime never supported it (sub-skills
+under `trtc/X/` aren't top-level discoverable, so `Skill(skill='trtc-X')`
+returns "Unknown skill"). Plain Read handoff is the only working option
+today. The §3.5-specific tests were removed; this file kept its other
+four invariants. See `docs/topic-skill-long-term-design.md` §10 for the
+walk-back rationale.
+"""
+from __future__ import annotations
+
+from pathlib import Path
+
+import pytest
+
+ROOT = Path(__file__).resolve().parents[3]  # repo root
+TOPIC_SKILL = ROOT / "skills" / "trtc-topic" / "SKILL.md"
+STATE_MACHINE_GUIDE = (
+    ROOT / "skills" / "trtc-topic" / "scripts" / "STATE-MACHINE-GUIDE.md"
+)
+EXECUTION_UNITS = ROOT / "skills" / "trtc-topic" / "references" / "execution-units.yaml"
+
+
+def test_apply_evidence_block_remains_deleted():
+    """Regression: the deleted 'Apply Evidence Block' must not return.
+
+    Was a multi-paragraph block instructing AI to run its own grep/ls/tsc
+    in addition to apply.py. Two-track verification was demo-test-2 root
+    cause: AI satisfied one track (the SKILL.md grep) by comment-stuffing
+    while apply.py also passed because it ran the same kind of substring
+    grep. Single-source verification is enforced by deletion.
+    """
+    text = TOPIC_SKILL.read_text()
+    assert "Apply Evidence Block (MANDATORY" not in text, (
+        "topic/SKILL.md still contains 'Apply Evidence Block (MANDATORY' — "
+        "this was deleted because it instructed AI to run its own grep, "
+        "which duplicates apply.py and AI learned to bypass."
+    )
+    assert "P1 Imports:" not in text, (
+        "topic/SKILL.md still has P1 Imports instruction for AI-run grep"
+    )
+
+
+def test_state_machine_guide_exists_and_is_referenced():
+    """The state machine guide must exist and topic/SKILL.md must point to it."""
+    assert STATE_MACHINE_GUIDE.exists(), (
+        f"State Machine Guide missing at {STATE_MACHINE_GUIDE}. "
+        "topic/SKILL.md was reduced under the assumption this guide is split out."
+    )
+    topic_text = TOPIC_SKILL.read_text()
+    assert "STATE-MACHINE-GUIDE.md" in topic_text, (
+        "topic/SKILL.md must reference scripts/STATE-MACHINE-GUIDE.md, "
+        "otherwise AI has no pointer to the operator's manual."
+    )
+
+
+def test_delivery_unit_grouping_is_declarative():
+    """Automatic unit grouping must stay visible outside state_machine.py."""
+    assert EXECUTION_UNITS.exists(), "delivery-unit grouping config is missing"
+    topic_text = TOPIC_SKILL.read_text()
+    guide_text = STATE_MACHINE_GUIDE.read_text()
+    assert "references/execution-units.yaml" in topic_text
+    assert "references/execution-units.yaml" in guide_text
+
+    state_machine_text = (
+        ROOT / "skills" / "trtc-topic" / "scripts" / "lib" / "state_machine.py"
+    ).read_text()
+    assert "_DEFAULT_UNIT_GROUPS" not in state_machine_text
+    assert "会议基础链路" not in state_machine_text
+
+
+def test_topic_skill_size_reduced():
+    """Sanity check: topic/SKILL.md should be substantially smaller than the
+    original 512-line version. If it grows back, something has been re-inlined."""
+    text = TOPIC_SKILL.read_text()
+    line_count = len(text.splitlines())
+    # The topic skill has grown with ui_mode, runtime verification, and
+    # delivery-unit execution. Keep a budget so the old anti-patterns don't
+    # quietly return, but do not enforce the obsolete 2026-05 compact size.
+    assert line_count < 760, (
+        f"topic/SKILL.md is {line_count} lines — above the current budget. "
+        f"Check if Apply Evidence Block, state machine manual, or Calling apply "
+        f"contract was re-inlined."
+    )
+
+
+def test_calling_apply_section_is_compact():
+    """The 'Calling apply' section must not contain the removed full request schema.
+
+    The full I/O contract lives in apply/SKILL.md Phase 0; topic just references it.
+    """
+    text = TOPIC_SKILL.read_text()
+    # Locate the "### Calling apply" section bounds.
+    start_marker = "### Calling apply"
+    if start_marker not in text:
+        pytest.fail("'### Calling apply' section missing from topic/SKILL.md")
+    section_start = text.index(start_marker)
+    # Section ends at the next '### ' heading.
+    after = text[section_start + len(start_marker):]
+    next_section_at = after.find("\n### ")
+    section_text = after[:next_section_at] if next_section_at >= 0 else after
+
+    # Mode selection table: should be in apply/SKILL.md only, not in topic.
+    bad_markers = [
+        "**Mode selection rules:**",
+        "**Request construction**",
+        "| Situation | mode |",
+    ]
+    found = [m for m in bad_markers if m in section_text]
+    assert not found, (
+        f"topic/SKILL.md 'Calling apply' section re-inlined apply request schema "
+        f"({found}). Keep the contract in apply/SKILL.md only; topic just references it."
+    )